/* If-conversion support.
- Copyright (C) 2000, 2001, 2002, 2003, 2004 Free Software Foundation, Inc.
+ Copyright (C) 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007
+ Free Software Foundation, Inc.
This file is part of GCC.
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, 59 Temple Place - Suite 330, Boston, MA
- 02111-1307, USA. */
+ Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA
+ 02110-1301, USA. */
#include "config.h"
#include "system.h"
#include "tm_p.h"
#include "cfgloop.h"
#include "target.h"
+#include "timevar.h"
+#include "tree-pass.h"
+#include "vec.h"
+#include "vecprim.h"
#ifndef HAVE_conditional_execution
#define MAX_CONDITIONAL_EXECUTE (BRANCH_COST + 1)
#endif
-#define NULL_EDGE ((struct edge_def *)NULL)
-#define NULL_BLOCK ((struct basic_block_def *)NULL)
+#define NULL_BLOCK ((basic_block) NULL)
/* # of IF-THEN or IF-THEN-ELSE blocks we looked at */
static int num_possible_if_blocks;
/* Forward references. */
static int count_bb_insns (basic_block);
-static int total_bb_rtx_cost (basic_block);
+static bool cheap_bb_rtx_cost_p (basic_block, int);
static rtx first_active_insn (basic_block);
static rtx last_active_insn (basic_block, int);
static basic_block block_fallthru (basic_block);
basic_block, int);
static void noce_emit_move_insn (rtx, rtx);
static rtx block_has_only_trap (basic_block);
-static void mark_loop_exit_edges (void);
\f
-/* Sets EDGE_LOOP_EXIT flag for all loop exits. */
-static void
-mark_loop_exit_edges (void)
-{
- struct loops loops;
- basic_block bb;
- edge e;
-
- flow_loops_find (&loops, LOOP_TREE);
- free_dominance_info (CDI_DOMINATORS);
-
- if (loops.num > 1)
- {
- FOR_EACH_BB (bb)
- {
- for (e = bb->succ; e; e = e->succ_next)
- {
- if (find_common_loop (bb->loop_father, e->dest->loop_father)
- != bb->loop_father)
- e->flags |= EDGE_LOOP_EXIT;
- else
- e->flags &= ~EDGE_LOOP_EXIT;
- }
- }
- }
-
- flow_loops_free (&loops);
-}
-
/* Count the number of non-jump active insns in BB. */
static int
return count;
}
-/* Count the total insn_rtx_cost of non-jump active insns in BB.
- This function returns -1, if the cost of any instruction could
- not be estimated. */
+/* Determine whether the total insn_rtx_cost on non-jump insns in
+ basic block BB is less than MAX_COST. This function returns
+ false if the cost of any instruction could not be estimated. */
-static int
-total_bb_rtx_cost (basic_block bb)
+static bool
+cheap_bb_rtx_cost_p (basic_block bb, int max_cost)
{
int count = 0;
rtx insn = BB_HEAD (bb);
{
int cost = insn_rtx_cost (PATTERN (insn));
if (cost == 0)
- return -1;
+ return false;
+
+ /* If this instruction is the load or set of a "stack" register,
+ such as a floating point register on x87, then the cost of
+ speculatively executing this insn may need to include
+ the additional cost of popping its result off of the
+ register stack. Unfortunately, correctly recognizing and
+ accounting for this additional overhead is tricky, so for
+ now we simply prohibit such speculative execution. */
+#ifdef STACK_REGS
+ {
+ rtx set = single_set (insn);
+ if (set && STACK_REG_P (SET_DEST (set)))
+ return false;
+ }
+#endif
+
count += cost;
+ if (count >= max_cost)
+ return false;
}
else if (CALL_P (insn))
- return -1;
-
+ return false;
+
if (insn == BB_END (bb))
break;
insn = NEXT_INSN (insn);
}
- return count;
+ return true;
}
/* Return the first non-jump active insn in the basic block. */
block_fallthru (basic_block bb)
{
edge e;
+ edge_iterator ei;
- for (e = bb->succ;
- e != NULL_EDGE && (e->flags & EDGE_FALLTHRU) == 0;
- e = e->succ_next)
- ;
+ FOR_EACH_EDGE (e, ei, bb->succs)
+ if (e->flags & EDGE_FALLTHRU)
+ break;
return (e) ? e->dest : NULL_BLOCK;
}
if (NOTE_P (insn))
goto insn_done;
- if (!NONJUMP_INSN_P (insn) && !CALL_P (insn))
- abort ();
+ gcc_assert(NONJUMP_INSN_P (insn) || CALL_P (insn));
/* Remove USE insns that get in the way. */
if (reload_completed && GET_CODE (PATTERN (insn)) == USE)
{
rtx start, end;
rtx t, f;
+ enum rtx_code f_code;
bb = block_fallthru (bb);
start = first_active_insn (bb);
if (! t)
goto fail;
- f = gen_rtx_fmt_ee (reverse_condition (GET_CODE (t)),
- GET_MODE (t),
- XEXP (t, 0),
- XEXP (t, 1));
+ f_code = reversed_comparison_code (t, BB_END (bb));
+ if (f_code == UNKNOWN)
+ goto fail;
+ f = gen_rtx_fmt_ee (f_code, GET_MODE (t), XEXP (t, 0), XEXP (t, 1));
if (ce_info->and_and_p)
{
t = gen_rtx_AND (GET_MODE (t), true_expr, t);
struct noce_if_info
{
+ /* A basic block that ends in a simple conditional jump. */
basic_block test_bb;
+
+ /* The jump that ends TEST_BB. */
+ rtx jump;
+
+ /* The jump condition. */
+ rtx cond;
+
+ /* New insns should be inserted before this one. */
+ rtx cond_earliest;
+
+ /* Insns in the THEN and ELSE block. There is always just this
+ one insns in those blocks. The insns are single_set insns.
+ If there was no ELSE block, INSN_B is the last insn before
+ COND_EARLIEST, or NULL_RTX. In the former case, the insn
+ operands are still valid, as if INSN_B was moved down below
+ the jump. */
rtx insn_a, insn_b;
- rtx x, a, b;
- rtx jump, cond, cond_earliest;
- /* True if "b" was originally evaluated unconditionally. */
- bool b_unconditional;
+
+ /* The SET_SRC of INSN_A and INSN_B. */
+ rtx a, b;
+
+ /* The SET_DEST of INSN_A. */
+ rtx x;
};
static rtx noce_emit_store_flag (struct noce_if_info *, rtx, int, int);
static void
noce_emit_move_insn (rtx x, rtx y)
{
- enum machine_mode outmode, inmode;
+ enum machine_mode outmode;
rtx outer, inner;
int bitpos;
if (GET_CODE (x) != STRICT_LOW_PART)
{
- emit_move_insn (x, y);
+ rtx seq, insn, target;
+ optab ot;
+
+ start_sequence ();
+ /* Check that the SET_SRC is reasonable before calling emit_move_insn,
+ otherwise construct a suitable SET pattern ourselves. */
+ insn = (OBJECT_P (y) || CONSTANT_P (y) || GET_CODE (y) == SUBREG)
+ ? emit_move_insn (x, y)
+ : emit_insn (gen_rtx_SET (VOIDmode, x, y));
+ seq = get_insns ();
+ end_sequence ();
+
+ if (recog_memoized (insn) <= 0)
+ {
+ if (GET_CODE (x) == ZERO_EXTRACT)
+ {
+ rtx op = XEXP (x, 0);
+ unsigned HOST_WIDE_INT size = INTVAL (XEXP (x, 1));
+ unsigned HOST_WIDE_INT start = INTVAL (XEXP (x, 2));
+
+ /* store_bit_field expects START to be relative to
+ BYTES_BIG_ENDIAN and adjusts this value for machines with
+ BITS_BIG_ENDIAN != BYTES_BIG_ENDIAN. In order to be able to
+ invoke store_bit_field again it is necessary to have the START
+ value from the first call. */
+ if (BITS_BIG_ENDIAN != BYTES_BIG_ENDIAN)
+ {
+ if (MEM_P (op))
+ start = BITS_PER_UNIT - start - size;
+ else
+ {
+ gcc_assert (REG_P (op));
+ start = BITS_PER_WORD - start - size;
+ }
+ }
+
+ gcc_assert (start < (MEM_P (op) ? BITS_PER_UNIT : BITS_PER_WORD));
+ store_bit_field (op, size, start, GET_MODE (x), y);
+ return;
+ }
+
+ switch (GET_RTX_CLASS (GET_CODE (y)))
+ {
+ case RTX_UNARY:
+ ot = code_to_optab[GET_CODE (y)];
+ if (ot)
+ {
+ start_sequence ();
+ target = expand_unop (GET_MODE (y), ot, XEXP (y, 0), x, 0);
+ if (target != NULL_RTX)
+ {
+ if (target != x)
+ emit_move_insn (x, target);
+ seq = get_insns ();
+ }
+ end_sequence ();
+ }
+ break;
+
+ case RTX_BIN_ARITH:
+ case RTX_COMM_ARITH:
+ ot = code_to_optab[GET_CODE (y)];
+ if (ot)
+ {
+ start_sequence ();
+ target = expand_binop (GET_MODE (y), ot,
+ XEXP (y, 0), XEXP (y, 1),
+ x, 0, OPTAB_DIRECT);
+ if (target != NULL_RTX)
+ {
+ if (target != x)
+ emit_move_insn (x, target);
+ seq = get_insns ();
+ }
+ end_sequence ();
+ }
+ break;
+
+ default:
+ break;
+ }
+ }
+
+ emit_insn (seq);
return;
}
outer = XEXP (x, 0);
inner = XEXP (outer, 0);
outmode = GET_MODE (outer);
- inmode = GET_MODE (inner);
bitpos = SUBREG_BYTE (outer) * BITS_PER_UNIT;
store_bit_field (inner, GET_MODE_BITSIZE (outmode), bitpos, outmode, y);
}
rtx a = if_info->a;
rtx b = if_info->b;
rtx x = if_info->x;
+ rtx orig_a, orig_b;
rtx insn_a, insn_b;
rtx tmp, target;
int is_mem = 0;
+ int insn_cost;
enum rtx_code code;
/* A conditional move from two memory sources is equivalent to a
insn_a = if_info->insn_a;
insn_b = if_info->insn_b;
+ /* Total insn_rtx_cost should be smaller than branch cost. Exit
+ if insn_rtx_cost can't be estimated. */
+ if (insn_a)
+ {
+ insn_cost = insn_rtx_cost (PATTERN (insn_a));
+ if (insn_cost == 0 || insn_cost > COSTS_N_INSNS (BRANCH_COST))
+ return FALSE;
+ }
+ else
+ insn_cost = 0;
+
+ if (insn_b)
+ {
+ insn_cost += insn_rtx_cost (PATTERN (insn_b));
+ if (insn_cost == 0 || insn_cost > COSTS_N_INSNS (BRANCH_COST))
+ return FALSE;
+ }
+
/* Possibly rearrange operands to make things come out more natural. */
if (reversed_comparison_code (if_info->cond, if_info->jump) != UNKNOWN)
{
start_sequence ();
+ orig_a = a;
+ orig_b = b;
+
/* If either operand is complex, load it into a register first.
The best way to do this is to copy the original insn. In this
way we preserve any clobbers etc that the insn may have had.
}
if (! general_operand (b, GET_MODE (b)))
{
- rtx set;
+ rtx set, last;
if (no_new_pseudos)
goto end_seq_and_fail;
if (is_mem)
{
tmp = gen_reg_rtx (GET_MODE (b));
- tmp = emit_insn (gen_rtx_SET (VOIDmode,
- tmp,
- b));
+ tmp = gen_rtx_SET (VOIDmode, tmp, b);
}
else if (! insn_b)
goto end_seq_and_fail;
tmp = copy_rtx (insn_b);
set = single_set (tmp);
SET_DEST (set) = b;
- tmp = emit_insn (PATTERN (tmp));
+ tmp = PATTERN (tmp);
}
+
+ /* If insn to set up A clobbers any registers B depends on, try to
+ swap insn that sets up A with the one that sets up B. If even
+ that doesn't help, punt. */
+ last = get_last_insn ();
+ if (last && modified_in_p (orig_b, last))
+ {
+ tmp = emit_insn_before (tmp, get_insns ());
+ if (modified_in_p (orig_a, tmp))
+ goto end_seq_and_fail;
+ }
+ else
+ tmp = emit_insn (tmp);
+
if (recog_memoized (tmp) < 0)
goto end_seq_and_fail;
}
rtx prev_insn;
/* First, look to see if we put a constant in a register. */
- prev_insn = PREV_INSN (if_info->cond_earliest);
+ prev_insn = prev_nonnote_insn (if_info->cond_earliest);
if (prev_insn
&& INSN_P (prev_insn)
&& GET_CODE (PATTERN (prev_insn)) == SET)
}
cond = canonicalize_condition (if_info->jump, cond, reverse,
- earliest, target, false);
+ earliest, target, false, true);
if (! cond || ! reg_mentioned_p (target, cond))
return NULL;
if (no_new_pseudos)
return FALSE;
- /* Recognize A and B as constituting an ABS or NABS. */
+ /* Recognize A and B as constituting an ABS or NABS. The canonical
+ form is a branch around the negation, taken when the object is the
+ first operand of a comparison against 0 that evaluates to true. */
a = if_info->a;
b = if_info->b;
if (GET_CODE (a) == NEG && rtx_equal_p (XEXP (a, 0), b))
if (rtx_equal_p (XEXP (cond, 0), b))
c = XEXP (cond, 1);
else if (rtx_equal_p (XEXP (cond, 1), b))
- c = XEXP (cond, 0);
+ {
+ c = XEXP (cond, 0);
+ negate = !negate;
+ }
else
return FALSE;
- /* Verify that C is zero. Search backward through the block for
- a REG_EQUAL note if necessary. */
+ /* Verify that C is zero. Search one step backward for a
+ REG_EQUAL note or a simple source if necessary. */
if (REG_P (c))
{
- rtx insn, note = NULL;
- for (insn = earliest;
- insn != BB_HEAD (if_info->test_bb);
- insn = PREV_INSN (insn))
- if (INSN_P (insn)
- && ((note = find_reg_note (insn, REG_EQUAL, c))
- || (note = find_reg_note (insn, REG_EQUIV, c))))
- break;
- if (! note)
+ rtx set, insn = prev_nonnote_insn (earliest);
+ if (insn
+ && (set = single_set (insn))
+ && rtx_equal_p (SET_DEST (set), c))
+ {
+ rtx note = find_reg_equal_equiv_note (insn);
+ if (note)
+ c = XEXP (note, 0);
+ else
+ c = SET_SRC (set);
+ }
+ else
return FALSE;
- c = XEXP (note, 0);
}
if (MEM_P (c)
&& GET_CODE (XEXP (c, 0)) == SYMBOL_REF
rtx cond, t, m, c, seq;
enum machine_mode mode;
enum rtx_code code;
+ bool b_unconditional;
if (no_new_pseudos)
return FALSE;
return FALSE;
/* This is only profitable if T is cheap, or T is unconditionally
- executed/evaluated in the original insn sequence. */
+ executed/evaluated in the original insn sequence. The latter
+ happens if INSN_B was taken from TEST_BB, or if there was no
+ INSN_B which can happen for e.g. conditional stores to memory. */
+ b_unconditional = (if_info->insn_b == NULL_RTX
+ || BLOCK_FOR_INSN (if_info->insn_b) == if_info->test_bb);
if (rtx_cost (t, SET) >= COSTS_N_INSNS (2)
- && (!if_info->b_unconditional
+ && (!b_unconditional
|| t != if_info->b))
return FALSE;
}
+/* Optimize away "if (x & C) x |= C" and similar bit manipulation
+ transformations. */
+
+static int
+noce_try_bitop (struct noce_if_info *if_info)
+{
+ rtx cond, x, a, result, seq;
+ enum machine_mode mode;
+ enum rtx_code code;
+ int bitnum;
+
+ x = if_info->x;
+ cond = if_info->cond;
+ code = GET_CODE (cond);
+
+ /* Check for no else condition. */
+ if (! rtx_equal_p (x, if_info->b))
+ return FALSE;
+
+ /* Check for a suitable condition. */
+ if (code != NE && code != EQ)
+ return FALSE;
+ if (XEXP (cond, 1) != const0_rtx)
+ return FALSE;
+ cond = XEXP (cond, 0);
+
+ /* ??? We could also handle AND here. */
+ if (GET_CODE (cond) == ZERO_EXTRACT)
+ {
+ if (XEXP (cond, 1) != const1_rtx
+ || GET_CODE (XEXP (cond, 2)) != CONST_INT
+ || ! rtx_equal_p (x, XEXP (cond, 0)))
+ return FALSE;
+ bitnum = INTVAL (XEXP (cond, 2));
+ mode = GET_MODE (x);
+ if (BITS_BIG_ENDIAN)
+ bitnum = GET_MODE_BITSIZE (mode) - 1 - bitnum;
+ if (bitnum < 0 || bitnum >= HOST_BITS_PER_WIDE_INT)
+ return FALSE;
+ }
+ else
+ return FALSE;
+
+ a = if_info->a;
+ if (GET_CODE (a) == IOR || GET_CODE (a) == XOR)
+ {
+ /* Check for "if (X & C) x = x op C". */
+ if (! rtx_equal_p (x, XEXP (a, 0))
+ || GET_CODE (XEXP (a, 1)) != CONST_INT
+ || (INTVAL (XEXP (a, 1)) & GET_MODE_MASK (mode))
+ != (unsigned HOST_WIDE_INT) 1 << bitnum)
+ return FALSE;
+
+ /* if ((x & C) == 0) x |= C; is transformed to x |= C. */
+ /* if ((x & C) != 0) x |= C; is transformed to nothing. */
+ if (GET_CODE (a) == IOR)
+ result = (code == NE) ? a : NULL_RTX;
+ else if (code == NE)
+ {
+ /* if ((x & C) == 0) x ^= C; is transformed to x |= C. */
+ result = gen_int_mode ((HOST_WIDE_INT) 1 << bitnum, mode);
+ result = simplify_gen_binary (IOR, mode, x, result);
+ }
+ else
+ {
+ /* if ((x & C) != 0) x ^= C; is transformed to x &= ~C. */
+ result = gen_int_mode (~((HOST_WIDE_INT) 1 << bitnum), mode);
+ result = simplify_gen_binary (AND, mode, x, result);
+ }
+ }
+ else if (GET_CODE (a) == AND)
+ {
+ /* Check for "if (X & C) x &= ~C". */
+ if (! rtx_equal_p (x, XEXP (a, 0))
+ || GET_CODE (XEXP (a, 1)) != CONST_INT
+ || (INTVAL (XEXP (a, 1)) & GET_MODE_MASK (mode))
+ != (~((HOST_WIDE_INT) 1 << bitnum) & GET_MODE_MASK (mode)))
+ return FALSE;
+
+ /* if ((x & C) == 0) x &= ~C; is transformed to nothing. */
+ /* if ((x & C) != 0) x &= ~C; is transformed to x &= ~C. */
+ result = (code == EQ) ? a : NULL_RTX;
+ }
+ else
+ return FALSE;
+
+ if (result)
+ {
+ start_sequence ();
+ noce_emit_move_insn (x, result);
+ seq = end_ifcvt_sequence (if_info);
+ if (!seq)
+ return FALSE;
+
+ emit_insn_before_setloc (seq, if_info->jump,
+ INSN_LOCATOR (if_info->insn_a));
+ }
+ return TRUE;
+}
+
+
/* Similar to get_condition, only the resulting condition must be
valid at JUMP, instead of at EARLIEST. */
static rtx
noce_get_condition (rtx jump, rtx *earliest)
{
- rtx cond, set, tmp, insn;
+ rtx cond, set, tmp;
bool reverse;
if (! any_condjump_p (jump))
/* Otherwise, fall back on canonicalize_condition to do the dirty
work of manipulating MODE_CC values and COMPARE rtx codes. */
+ return canonicalize_condition (jump, cond, reverse, earliest,
+ NULL_RTX, false, true);
+}
- tmp = canonicalize_condition (jump, cond, reverse, earliest, NULL_RTX,
- false);
- if (!tmp)
- return NULL_RTX;
+/* Initialize for a simple IF-THEN or IF-THEN-ELSE block. We will not
+ be using conditional execution. Set some fields of IF_INFO based
+ on CE_INFO: test_bb, cond, jump, cond_earliest. Return TRUE if
+ things look OK. */
- /* We are going to insert code before JUMP, not before EARLIEST.
- We must therefore be certain that the given condition is valid
- at JUMP by virtue of not having been modified since. */
- for (insn = *earliest; insn != jump; insn = NEXT_INSN (insn))
- if (INSN_P (insn) && modified_in_p (tmp, insn))
- break;
- if (insn == jump)
- return tmp;
-
- /* The condition was modified. See if we can get a partial result
- that doesn't follow all the reversals. Perhaps combine can fold
- them together later. */
- tmp = XEXP (tmp, 0);
- if (!REG_P (tmp) || GET_MODE_CLASS (GET_MODE (tmp)) != MODE_INT)
- return NULL_RTX;
- tmp = canonicalize_condition (jump, cond, reverse, earliest, tmp,
- false);
- if (!tmp)
- return NULL_RTX;
+static int
+noce_init_if_info (struct ce_if_block *ce_info, struct noce_if_info *if_info)
+{
+ basic_block test_bb = ce_info->test_bb;
+ rtx cond, jump;
- /* For sanity's sake, re-validate the new result. */
- for (insn = *earliest; insn != jump; insn = NEXT_INSN (insn))
- if (INSN_P (insn) && modified_in_p (tmp, insn))
- return NULL_RTX;
+ /* If test is comprised of && or || elements, don't handle it unless
+ it is the special case of && elements without an ELSE block. */
+ if (ce_info->num_multiple_test_blocks)
+ {
+ if (ce_info->else_bb || !ce_info->and_and_p)
+ return FALSE;
+
+ ce_info->test_bb = test_bb = ce_info->last_test_bb;
+ ce_info->num_multiple_test_blocks = 0;
+ ce_info->num_and_and_blocks = 0;
+ ce_info->num_or_or_blocks = 0;
+ }
+
+ /* If this is not a standard conditional jump, we can't parse it. */
+ jump = BB_END (test_bb);
+ cond = noce_get_condition (jump, &if_info->cond_earliest);
+ if (!cond)
+ return FALSE;
+
+ /* If the conditional jump is more than just a conditional
+ jump, then we can not do if-conversion on this block. */
+ if (! onlyjump_p (jump))
+ return FALSE;
+
+ /* We must be comparing objects whose modes imply the size. */
+ if (GET_MODE (XEXP (cond, 0)) == BLKmode)
+ return FALSE;
- return tmp;
+ if_info->test_bb = test_bb;
+ if_info->cond = cond;
+ if_info->jump = jump;
+
+ return TRUE;
}
/* Return true if OP is ok for if-then-else processing. */
return ! may_trap_p (op);
}
+/* Return true if a write into MEM may trap or fault. */
+
+static bool
+noce_mem_write_may_trap_or_fault_p (rtx mem)
+{
+ rtx addr;
+
+ if (MEM_READONLY_P (mem))
+ return true;
+
+ if (may_trap_or_fault_p (mem))
+ return true;
+
+ addr = XEXP (mem, 0);
+
+ /* Call target hook to avoid the effects of -fpic etc.... */
+ addr = targetm.delegitimize_address (addr);
+
+ while (addr)
+ switch (GET_CODE (addr))
+ {
+ case CONST:
+ case PRE_DEC:
+ case PRE_INC:
+ case POST_DEC:
+ case POST_INC:
+ case POST_MODIFY:
+ addr = XEXP (addr, 0);
+ break;
+ case LO_SUM:
+ case PRE_MODIFY:
+ addr = XEXP (addr, 1);
+ break;
+ case PLUS:
+ if (GET_CODE (XEXP (addr, 1)) == CONST_INT)
+ addr = XEXP (addr, 0);
+ else
+ return false;
+ break;
+ case LABEL_REF:
+ return true;
+ case SYMBOL_REF:
+ if (SYMBOL_REF_DECL (addr)
+ && decl_readonly_section (SYMBOL_REF_DECL (addr), 0))
+ return true;
+ return false;
+ default:
+ return false;
+ }
+
+ return false;
+}
+
/* Given a simple IF-THEN or IF-THEN-ELSE block, attempt to convert it
without using conditional execution. Return TRUE if we were
successful at converting the block. */
basic_block test_bb = ce_info->test_bb; /* test block */
basic_block then_bb = ce_info->then_bb; /* THEN */
basic_block else_bb = ce_info->else_bb; /* ELSE or NULL */
+ basic_block join_bb;
struct noce_if_info if_info;
rtx insn_a, insn_b;
rtx set_a, set_b;
??? For future expansion, look for multiple X in such patterns. */
- /* If test is comprised of && or || elements, don't handle it unless it is
- the special case of && elements without an ELSE block. */
- if (ce_info->num_multiple_test_blocks)
- {
- if (else_bb || ! ce_info->and_and_p)
- return FALSE;
-
- ce_info->test_bb = test_bb = ce_info->last_test_bb;
- ce_info->num_multiple_test_blocks = 0;
- ce_info->num_and_and_blocks = 0;
- ce_info->num_or_or_blocks = 0;
- }
-
- /* If this is not a standard conditional jump, we can't parse it. */
- jump = BB_END (test_bb);
- cond = noce_get_condition (jump, &if_info.cond_earliest);
- if (! cond)
+ if (!noce_init_if_info (ce_info, &if_info))
return FALSE;
- /* If the conditional jump is more than just a conditional
- jump, then we can not do if-conversion on this block. */
- if (! onlyjump_p (jump))
- return FALSE;
-
- /* We must be comparing objects whose modes imply the size. */
- if (GET_MODE (XEXP (cond, 0)) == BLKmode)
- return FALSE;
+ cond = if_info.cond;
+ jump = if_info.jump;
/* Look for one of the potential sets. */
insn_a = first_active_insn (then_bb);
{
if (no_new_pseudos || GET_MODE (x) == BLKmode)
return FALSE;
+
+ if (GET_MODE (x) == ZERO_EXTRACT
+ && (GET_CODE (XEXP (x, 1)) != CONST_INT
+ || GET_CODE (XEXP (x, 2)) != CONST_INT))
+ return FALSE;
+
x = gen_reg_rtx (GET_MODE (GET_CODE (x) == STRICT_LOW_PART
? XEXP (x, 0) : x));
}
return FALSE;
/* Set up the info block for our subroutines. */
- if_info.test_bb = test_bb;
- if_info.cond = cond;
- if_info.jump = jump;
if_info.insn_a = insn_a;
if_info.insn_b = insn_b;
if_info.x = x;
if_info.a = a;
if_info.b = b;
- if_info.b_unconditional = else_bb == 0;
/* Try optimizations in some approximation of a useful order. */
/* ??? Should first look to see if X is live incoming at all. If it
}
/* Disallow the "if (...) x = a;" form (with an implicit "else x = x;")
- for most optimizations if writing to x may trap, i.e. it's a memory
- other than a static var or a stack slot. */
- if (! set_b
- && MEM_P (orig_x)
- && ! MEM_NOTRAP_P (orig_x)
- && rtx_addr_can_trap_p (XEXP (orig_x, 0)))
- {
- if (HAVE_conditional_move)
- {
- if (noce_try_cmove (&if_info))
- goto success;
- if (! HAVE_conditional_execution
- && noce_try_cmove_arith (&if_info))
- goto success;
- }
- return FALSE;
- }
+ for optimizations if writing to x may trap or fault, i.e. it's a memory
+ other than a static var or a stack slot, is misaligned on strict
+ aligned machines or is read-only.
+ If x is a read-only memory, then the program is valid only if we
+ avoid the store into it. If there are stores on both the THEN and
+ ELSE arms, then we can go ahead with the conversion; either the
+ program is broken, or the condition is always false such that the
+ other memory is selected. */
+ if (!set_b && MEM_P (orig_x) && noce_mem_write_may_trap_or_fault_p (orig_x))
+ return FALSE;
if (noce_try_move (&if_info))
goto success;
if (noce_try_store_flag (&if_info))
goto success;
+ if (noce_try_bitop (&if_info))
+ goto success;
if (noce_try_minmax (&if_info))
goto success;
if (noce_try_abs (&if_info))
return FALSE;
success:
- /* The original sets may now be killed. */
- delete_insn (insn_a);
-
- /* Several special cases here: First, we may have reused insn_b above,
- in which case insn_b is now NULL. Second, we want to delete insn_b
- if it came from the ELSE block, because follows the now correct
- write that appears in the TEST block. However, if we got insn_b from
- the TEST block, it may in fact be loading data needed for the comparison.
- We'll let life_analysis remove the insn if it's really dead. */
- if (insn_b && else_bb)
- delete_insn (insn_b);
-
- /* The new insns will have been inserted immediately before the jump. We
- should be able to remove the jump with impunity, but the condition itself
- may have been modified by gcse to be shared across basic blocks. */
- delete_insn (jump);
/* If we used a temporary, fix it up now. */
if (orig_x != x)
{
+ rtx seq;
+
start_sequence ();
noce_emit_move_insn (orig_x, x);
- insn_b = get_insns ();
+ seq = get_insns ();
set_used_flags (orig_x);
- unshare_all_rtl_in_chain (insn_b);
+ unshare_all_rtl_in_chain (seq);
end_sequence ();
- emit_insn_after_setloc (insn_b, BB_END (test_bb), INSN_LOCATOR (insn_a));
+ emit_insn_before_setloc (seq, BB_END (test_bb), INSN_LOCATOR (insn_a));
}
- /* Merge the blocks! */
- merge_if_block (ce_info);
+ /* The original THEN and ELSE blocks may now be removed. The test block
+ must now jump to the join block. If the test block and the join block
+ can be merged, do so. */
+
+ join_bb = single_succ (then_bb);
+ if (else_bb)
+ {
+ delete_basic_block (else_bb);
+ num_true_changes++;
+ }
+ else
+ remove_edge (find_edge (test_bb, join_bb));
+
+ remove_edge (find_edge (then_bb, join_bb));
+ redirect_edge_and_branch_force (single_succ_edge (test_bb), join_bb);
+ delete_basic_block (then_bb);
+ num_true_changes++;
+
+ if (can_merge_blocks_p (test_bb, join_bb))
+ {
+ merge_blocks (test_bb, join_bb);
+ num_true_changes++;
+ }
+
+ num_updated_if_blocks++;
+ return TRUE;
+}
+
+/* Check whether a block is suitable for conditional move conversion.
+ Every insn must be a simple set of a register to a constant or a
+ register. For each assignment, store the value in the array VALS,
+ indexed by register number, then store the register number in
+ REGS. COND is the condition we will test. */
+
+static int
+check_cond_move_block (basic_block bb, rtx *vals, VEC (int, heap) *regs, rtx cond)
+{
+ rtx insn;
+
+ /* We can only handle simple jumps at the end of the basic block.
+ It is almost impossible to update the CFG otherwise. */
+ insn = BB_END (bb);
+ if (JUMP_P (insn) && !onlyjump_p (insn))
+ return FALSE;
+
+ FOR_BB_INSNS (bb, insn)
+ {
+ rtx set, dest, src;
+
+ if (!INSN_P (insn) || JUMP_P (insn))
+ continue;
+ set = single_set (insn);
+ if (!set)
+ return FALSE;
+
+ dest = SET_DEST (set);
+ src = SET_SRC (set);
+ if (!REG_P (dest)
+ || (SMALL_REGISTER_CLASSES && HARD_REGISTER_P (dest)))
+ return FALSE;
+
+ if (!CONSTANT_P (src) && !register_operand (src, VOIDmode))
+ return FALSE;
+
+ if (side_effects_p (src) || side_effects_p (dest))
+ return FALSE;
+
+ if (may_trap_p (src) || may_trap_p (dest))
+ return FALSE;
+
+ /* Don't try to handle this if the source register was
+ modified earlier in the block. */
+ if ((REG_P (src)
+ && vals[REGNO (src)] != NULL)
+ || (GET_CODE (src) == SUBREG && REG_P (SUBREG_REG (src))
+ && vals[REGNO (SUBREG_REG (src))] != NULL))
+ return FALSE;
+
+ /* Don't try to handle this if the destination register was
+ modified earlier in the block. */
+ if (vals[REGNO (dest)] != NULL)
+ return FALSE;
+
+ /* Don't try to handle this if the condition uses the
+ destination register. */
+ if (reg_overlap_mentioned_p (dest, cond))
+ return FALSE;
+
+ /* Don't try to handle this if the source register is modified
+ later in the block. */
+ if (!CONSTANT_P (src)
+ && modified_between_p (src, insn, NEXT_INSN (BB_END (bb))))
+ return FALSE;
+
+ vals[REGNO (dest)] = src;
+
+ VEC_safe_push (int, heap, regs, REGNO (dest));
+ }
return TRUE;
}
+
+/* Given a basic block BB suitable for conditional move conversion,
+ a condition COND, and arrays THEN_VALS and ELSE_VALS containing the
+ register values depending on COND, emit the insns in the block as
+ conditional moves. If ELSE_BLOCK is true, THEN_BB was already
+ processed. The caller has started a sequence for the conversion.
+ Return true if successful, false if something goes wrong. */
+
+static bool
+cond_move_convert_if_block (struct noce_if_info *if_infop,
+ basic_block bb, rtx cond,
+ rtx *then_vals, rtx *else_vals,
+ bool else_block_p)
+{
+ enum rtx_code code;
+ rtx insn, cond_arg0, cond_arg1;
+
+ code = GET_CODE (cond);
+ cond_arg0 = XEXP (cond, 0);
+ cond_arg1 = XEXP (cond, 1);
+
+ FOR_BB_INSNS (bb, insn)
+ {
+ rtx set, target, dest, t, e;
+ unsigned int regno;
+
+ if (!INSN_P (insn) || JUMP_P (insn))
+ continue;
+ set = single_set (insn);
+ gcc_assert (set && REG_P (SET_DEST (set)));
+
+ dest = SET_DEST (set);
+ regno = REGNO (dest);
+
+ t = then_vals[regno];
+ e = else_vals[regno];
+
+ if (else_block_p)
+ {
+ /* If this register was set in the then block, we already
+ handled this case there. */
+ if (t)
+ continue;
+ t = dest;
+ gcc_assert (e);
+ }
+ else
+ {
+ gcc_assert (t);
+ if (!e)
+ e = dest;
+ }
+
+ target = noce_emit_cmove (if_infop, dest, code, cond_arg0, cond_arg1,
+ t, e);
+ if (!target)
+ return false;
+
+ if (target != dest)
+ noce_emit_move_insn (dest, target);
+ }
+
+ return true;
+}
+
+/* Given a simple IF-THEN or IF-THEN-ELSE block, attempt to convert it
+ using only conditional moves. Return TRUE if we were successful at
+ converting the block. */
+
+static int
+cond_move_process_if_block (struct ce_if_block *ce_info)
+{
+ basic_block test_bb = ce_info->test_bb;
+ basic_block then_bb = ce_info->then_bb;
+ basic_block else_bb = ce_info->else_bb;
+ basic_block join_bb;
+ struct noce_if_info if_info;
+ rtx jump, cond, seq, loc_insn;
+ int max_reg, size, c, reg;
+ rtx *then_vals;
+ rtx *else_vals;
+ VEC (int, heap) *then_regs = NULL;
+ VEC (int, heap) *else_regs = NULL;
+ unsigned int i;
+
+ if (!HAVE_conditional_move || no_new_pseudos)
+ return FALSE;
+
+ memset (&if_info, 0, sizeof if_info);
+
+ if (!noce_init_if_info (ce_info, &if_info))
+ return FALSE;
+
+ cond = if_info.cond;
+ jump = if_info.jump;
+
+ /* Build a mapping for each block to the value used for each
+ register. */
+ max_reg = max_reg_num ();
+ size = (max_reg + 1) * sizeof (rtx);
+ then_vals = (rtx *) alloca (size);
+ else_vals = (rtx *) alloca (size);
+ memset (then_vals, 0, size);
+ memset (else_vals, 0, size);
+
+ /* Make sure the blocks are suitable. */
+ if (!check_cond_move_block (then_bb, then_vals, then_regs, cond)
+ || (else_bb && !check_cond_move_block (else_bb, else_vals, else_regs, cond)))
+ return FALSE;
+
+ /* Make sure the blocks can be used together. If the same register
+ is set in both blocks, and is not set to a constant in both
+ cases, then both blocks must set it to the same register. We
+ have already verified that if it is set to a register, that the
+ source register does not change after the assignment. Also count
+ the number of registers set in only one of the blocks. */
+ c = 0;
+ for (i = 0; VEC_iterate (int, then_regs, i, reg); i++)
+ {
+ if (!then_vals[reg] && !else_vals[reg])
+ continue;
+
+ if (!else_vals[reg])
+ ++c;
+ else
+ {
+ if (!CONSTANT_P (then_vals[reg])
+ && !CONSTANT_P (else_vals[reg])
+ && !rtx_equal_p (then_vals[reg], else_vals[reg]))
+ return FALSE;
+ }
+ }
+
+ /* Finish off c for MAX_CONDITIONAL_EXECUTE. */
+ for (i = 0; VEC_iterate (int, else_regs, i, reg); ++i)
+ if (!then_vals[reg])
+ ++c;
+
+ /* Make sure it is reasonable to convert this block. What matters
+ is the number of assignments currently made in only one of the
+ branches, since if we convert we are going to always execute
+ them. */
+ if (c > MAX_CONDITIONAL_EXECUTE)
+ return FALSE;
+
+ /* Try to emit the conditional moves. First do the then block,
+ then do anything left in the else blocks. */
+ start_sequence ();
+ if (!cond_move_convert_if_block (&if_info, then_bb, cond,
+ then_vals, else_vals, false)
+ || (else_bb
+ && !cond_move_convert_if_block (&if_info, else_bb, cond,
+ then_vals, else_vals, true)))
+ {
+ end_sequence ();
+ return FALSE;
+ }
+ seq = end_ifcvt_sequence (&if_info);
+ if (!seq)
+ return FALSE;
+
+ loc_insn = first_active_insn (then_bb);
+ if (!loc_insn)
+ {
+ loc_insn = first_active_insn (else_bb);
+ gcc_assert (loc_insn);
+ }
+ emit_insn_before_setloc (seq, jump, INSN_LOCATOR (loc_insn));
+
+ join_bb = single_succ (then_bb);
+ if (else_bb)
+ {
+ delete_basic_block (else_bb);
+ num_true_changes++;
+ }
+ else
+ remove_edge (find_edge (test_bb, join_bb));
+
+ remove_edge (find_edge (then_bb, join_bb));
+ redirect_edge_and_branch_force (single_succ_edge (test_bb), join_bb);
+ delete_basic_block (then_bb);
+ num_true_changes++;
+
+ if (can_merge_blocks_p (test_bb, join_bb))
+ {
+ merge_blocks (test_bb, join_bb);
+ num_true_changes++;
+ }
+
+ num_updated_if_blocks++;
+
+ VEC_free (int, heap, then_regs);
+ VEC_free (int, heap, else_regs);
+
+ return TRUE;
+}
+
\f
/* Attempt to convert an IF-THEN or IF-THEN-ELSE block into
straight line code. Return true if successful. */
static int
process_if_block (struct ce_if_block * ce_info)
{
- if (! reload_completed
- && noce_process_if_block (ce_info))
- return TRUE;
+ /* Only perform the noce transformations before register allocation.
+ They could be made to run later, but this would require a lot of
+ work, and it doesn't seem to be worth it. */
+ if (! reload_completed)
+ {
+ if (noce_process_if_block (ce_info))
+ return TRUE;
+
+ if (HAVE_conditional_move
+ && cond_move_process_if_block (ce_info))
+ return TRUE;
+ }
if (HAVE_conditional_execution && reload_completed)
{
if (then_bb)
{
- if (combo_bb->global_live_at_end)
- COPY_REG_SET (combo_bb->global_live_at_end,
- then_bb->global_live_at_end);
merge_blocks (combo_bb, then_bb);
num_true_changes++;
}
/* The outgoing edge for the current COMBO block should already
be correct. Verify this. */
- if (combo_bb->succ == NULL_EDGE)
- {
- if (find_reg_note (last, REG_NORETURN, NULL))
- ;
- else if (NONJUMP_INSN_P (last)
- && GET_CODE (PATTERN (last)) == TRAP_IF
- && TRAP_CONDITION (PATTERN (last)) == const_true_rtx)
- ;
- else
- abort ();
- }
+ if (EDGE_COUNT (combo_bb->succs) == 0)
+ gcc_assert (find_reg_note (last, REG_NORETURN, NULL)
+ || (NONJUMP_INSN_P (last)
+ && GET_CODE (PATTERN (last)) == TRAP_IF
+ && (TRAP_CONDITION (PATTERN (last))
+ == const_true_rtx)));
+ else
/* There should still be something at the end of the THEN or ELSE
blocks taking us to our final destination. */
- else if (JUMP_P (last))
- ;
- else if (combo_bb->succ->dest == EXIT_BLOCK_PTR
- && CALL_P (last)
- && SIBLING_CALL_P (last))
- ;
- else if ((combo_bb->succ->flags & EDGE_EH)
- && can_throw_internal (last))
- ;
- else
- abort ();
+ gcc_assert (JUMP_P (last)
+ || (EDGE_SUCC (combo_bb, 0)->dest == EXIT_BLOCK_PTR
+ && CALL_P (last)
+ && SIBLING_CALL_P (last))
+ || ((EDGE_SUCC (combo_bb, 0)->flags & EDGE_EH)
+ && can_throw_internal (last)));
}
/* The JOIN block may have had quite a number of other predecessors too.
have only one remaining edge from our if-then-else diamond. If there
is more than one remaining edge, it must come from elsewhere. There
may be zero incoming edges if the THEN block didn't actually join
- back up (as with a call to abort). */
- else if ((join_bb->pred == NULL
- || join_bb->pred->pred_next == NULL)
+ back up (as with a call to a non-return function). */
+ else if (EDGE_COUNT (join_bb->preds) < 2
&& join_bb != EXIT_BLOCK_PTR)
{
/* We can merge the JOIN. */
- if (combo_bb->global_live_at_end)
- COPY_REG_SET (combo_bb->global_live_at_end,
- join_bb->global_live_at_end);
-
merge_blocks (combo_bb, join_bb);
num_true_changes++;
}
/* The outgoing edge for the current COMBO block should already
be correct. Verify this. */
- if (combo_bb->succ->succ_next != NULL_EDGE
- || combo_bb->succ->dest != join_bb)
- abort ();
+ gcc_assert (single_succ_p (combo_bb)
+ && single_succ (combo_bb) == join_bb);
/* Remove the jump and cruft from the end of the COMBO block. */
if (join_bb != EXIT_BLOCK_PTR)
- tidy_fallthru_edge (combo_bb->succ);
+ tidy_fallthru_edge (single_succ_edge (combo_bb));
}
num_updated_if_blocks++;
edge else_edge;
/* The kind of block we're looking for has exactly two successors. */
- if ((then_edge = test_bb->succ) == NULL_EDGE
- || (else_edge = then_edge->succ_next) == NULL_EDGE
- || else_edge->succ_next != NULL_EDGE)
+ if (EDGE_COUNT (test_bb->succs) != 2)
return NULL;
+ then_edge = EDGE_SUCC (test_bb, 0);
+ else_edge = EDGE_SUCC (test_bb, 1);
+
/* Neither edge should be abnormal. */
if ((then_edge->flags & EDGE_COMPLEX)
|| (else_edge->flags & EDGE_COMPLEX))
rtx insn;
rtx end;
int n_insns = 0;
+ edge_iterator ei;
if (!cur_bb || !target_bb)
return -1;
/* If no edges, obviously it doesn't jump or fallthru. */
- if (cur_bb->succ == NULL_EDGE)
+ if (EDGE_COUNT (cur_bb->succs) == 0)
return FALSE;
- for (cur_edge = cur_bb->succ;
- cur_edge != NULL_EDGE;
- cur_edge = cur_edge->succ_next)
+ FOR_EACH_EDGE (cur_edge, ei, cur_bb->succs)
{
if (cur_edge->flags & EDGE_COMPLEX)
/* Anything complex isn't what we want. */
basic_block then_bb = ce_info->then_bb;
basic_block else_bb = ce_info->else_bb;
basic_block join_bb = NULL_BLOCK;
- edge then_succ = then_bb->succ;
- edge else_succ = else_bb->succ;
- int then_predecessors;
- int else_predecessors;
edge cur_edge;
basic_block next;
+ edge_iterator ei;
ce_info->last_test_bb = test_bb;
were && tests (which jump to the else block) or || tests (which jump to
the then block). */
if (HAVE_conditional_execution && reload_completed
- && test_bb->pred != NULL_EDGE
- && test_bb->pred->pred_next == NULL_EDGE
- && test_bb->pred->flags == EDGE_FALLTHRU)
+ && single_pred_p (test_bb)
+ && single_pred_edge (test_bb)->flags == EDGE_FALLTHRU)
{
- basic_block bb = test_bb->pred->src;
+ basic_block bb = single_pred (test_bb);
basic_block target_bb;
int max_insns = MAX_CONDITIONAL_EXECUTE;
int n_insns;
total_insns += n_insns;
blocks++;
- if (bb->pred == NULL_EDGE || bb->pred->pred_next != NULL_EDGE)
+ if (!single_pred_p (bb))
break;
- bb = bb->pred->src;
+ bb = single_pred (bb);
n_insns = block_jumps_and_fallthru_p (bb, target_bb);
}
while (n_insns >= 0 && (total_insns + n_insns) <= max_insns);
}
}
- /* Count the number of edges the THEN and ELSE blocks have. */
- then_predecessors = 0;
- for (cur_edge = then_bb->pred;
- cur_edge != NULL_EDGE;
- cur_edge = cur_edge->pred_next)
+ /* The THEN block of an IF-THEN combo must have exactly one predecessor,
+ other than any || blocks which jump to the THEN block. */
+ if ((EDGE_COUNT (then_bb->preds) - ce_info->num_or_or_blocks) != 1)
+ return FALSE;
+
+ /* The edges of the THEN and ELSE blocks cannot have complex edges. */
+ FOR_EACH_EDGE (cur_edge, ei, then_bb->preds)
{
- then_predecessors++;
if (cur_edge->flags & EDGE_COMPLEX)
return FALSE;
}
- else_predecessors = 0;
- for (cur_edge = else_bb->pred;
- cur_edge != NULL_EDGE;
- cur_edge = cur_edge->pred_next)
+ FOR_EACH_EDGE (cur_edge, ei, else_bb->preds)
{
- else_predecessors++;
if (cur_edge->flags & EDGE_COMPLEX)
return FALSE;
}
- /* The THEN block of an IF-THEN combo must have exactly one predecessor,
- other than any || blocks which jump to the THEN block. */
- if ((then_predecessors - ce_info->num_or_or_blocks) != 1)
- return FALSE;
-
/* The THEN block of an IF-THEN combo must have zero or one successors. */
- if (then_succ != NULL_EDGE
- && (then_succ->succ_next != NULL_EDGE
- || (then_succ->flags & EDGE_COMPLEX)
+ if (EDGE_COUNT (then_bb->succs) > 0
+ && (!single_succ_p (then_bb)
+ || (single_succ_edge (then_bb)->flags & EDGE_COMPLEX)
|| (flow2_completed && tablejump_p (BB_END (then_bb), NULL, NULL))))
return FALSE;
Check for the last insn of the THEN block being an indirect jump, which
is listed as not having any successors, but confuses the rest of the CE
code processing. ??? we should fix this in the future. */
- if (then_succ == NULL)
+ if (EDGE_COUNT (then_bb->succs) == 0)
{
- if (else_bb->pred->pred_next == NULL_EDGE)
+ if (single_pred_p (else_bb))
{
rtx last_insn = BB_END (then_bb);
/* If the THEN block's successor is the other edge out of the TEST block,
then we have an IF-THEN combo without an ELSE. */
- else if (then_succ->dest == else_bb)
+ else if (single_succ (then_bb) == else_bb)
{
join_bb = else_bb;
else_bb = NULL_BLOCK;
/* If the THEN and ELSE block meet in a subsequent block, and the ELSE
has exactly one predecessor and one successor, and the outgoing edge
is not complex, then we have an IF-THEN-ELSE combo. */
- else if (else_succ != NULL_EDGE
- && then_succ->dest == else_succ->dest
- && else_bb->pred->pred_next == NULL_EDGE
- && else_succ->succ_next == NULL_EDGE
- && ! (else_succ->flags & EDGE_COMPLEX)
+ else if (single_succ_p (else_bb)
+ && single_succ (then_bb) == single_succ (else_bb)
+ && single_pred_p (else_bb)
+ && ! (single_succ_edge (else_bb)->flags & EDGE_COMPLEX)
&& ! (flow2_completed && tablejump_p (BB_END (else_bb), NULL, NULL)))
- join_bb = else_succ->dest;
+ join_bb = single_succ (else_bb);
/* Otherwise it is not an IF-THEN or IF-THEN-ELSE combination. */
else
we checked the FALLTHRU flag, those are already adjacent to the last IF
block. */
/* ??? As an enhancement, move the ELSE block. Have to deal with
- BLOCK notes, if by no other means than aborting the merge if they
+ BLOCK notes, if by no other means than backing out the merge if they
exist. Sticky enough I don't want to think about it now. */
next = then_bb;
if (else_bb && (next = next->next_bb) != else_bb)
if (seq == NULL)
return FALSE;
- num_true_changes++;
-
/* Emit the new insns before cond_earliest. */
emit_insn_before_setloc (seq, cond_earliest, INSN_LOCATOR (trap));
/* Delete the trap block if possible. */
remove_edge (trap_bb == then_bb ? then_edge : else_edge);
- if (trap_bb->pred == NULL)
- delete_basic_block (trap_bb);
-
- /* If the non-trap block and the test are now adjacent, merge them.
- Otherwise we must insert a direct branch. */
- if (test_bb->next_bb == other_bb)
+ if (EDGE_COUNT (trap_bb->preds) == 0)
{
- struct ce_if_block new_ce_info;
- delete_insn (jump);
- memset (&new_ce_info, '\0', sizeof (new_ce_info));
- new_ce_info.test_bb = test_bb;
- new_ce_info.then_bb = NULL;
- new_ce_info.else_bb = NULL;
- new_ce_info.join_bb = other_bb;
- merge_if_block (&new_ce_info);
+ delete_basic_block (trap_bb);
+ num_true_changes++;
}
+
+ /* Wire together the blocks again. */
+ if (current_ir_type () == IR_RTL_CFGLAYOUT)
+ single_succ_edge (test_bb)->flags |= EDGE_FALLTHRU;
else
{
rtx lab, newjump;
LABEL_NUSES (lab) += 1;
JUMP_LABEL (newjump) = lab;
emit_barrier_after (newjump);
+ }
+ delete_insn (jump);
- delete_insn (jump);
+ if (can_merge_blocks_p (test_bb, other_bb))
+ {
+ merge_blocks (test_bb, other_bb);
+ num_true_changes++;
}
+ num_updated_if_blocks++;
return TRUE;
}
return NULL_RTX;
/* The block must have no successors. */
- if (bb->succ)
+ if (EDGE_COUNT (bb->succs) > 0)
return NULL_RTX;
/* The only instruction in the THEN block must be the trap. */
{
basic_block then_bb = then_edge->dest;
basic_block else_bb = else_edge->dest, new_bb;
- edge then_succ = then_bb->succ;
- int then_bb_index, bb_cost;
+ int then_bb_index;
/* If we are partitioning hot/cold basic blocks, we don't want to
mess up unconditional or indirect jumps that cross between hot
- and cold sections. */
-
- if (flag_reorder_blocks_and_partition
- && ((BB_END (then_bb)
- && find_reg_note (BB_END (then_bb), REG_CROSSING_JUMP, NULL_RTX))
- || (BB_END (else_bb)
- && find_reg_note (BB_END (else_bb), REG_CROSSING_JUMP,
- NULL_RTX))))
+ and cold sections.
+
+ Basic block partitioning may result in some jumps that appear to
+ be optimizable (or blocks that appear to be mergeable), but which really
+ must be left untouched (they are required to make it safely across
+ partition boundaries). See the comments at the top of
+ bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
+
+ if ((BB_END (then_bb)
+ && find_reg_note (BB_END (then_bb), REG_CROSSING_JUMP, NULL_RTX))
+ || (BB_END (test_bb)
+ && find_reg_note (BB_END (test_bb), REG_CROSSING_JUMP, NULL_RTX))
+ || (BB_END (else_bb)
+ && find_reg_note (BB_END (else_bb), REG_CROSSING_JUMP,
+ NULL_RTX)))
return FALSE;
/* THEN has one successor. */
- if (!then_succ || then_succ->succ_next != NULL)
+ if (!single_succ_p (then_bb))
return FALSE;
/* THEN does not fall through, but is not strange either. */
- if (then_succ->flags & (EDGE_COMPLEX | EDGE_FALLTHRU))
+ if (single_succ_edge (then_bb)->flags & (EDGE_COMPLEX | EDGE_FALLTHRU))
return FALSE;
/* THEN has one predecessor. */
- if (then_bb->pred->pred_next != NULL)
+ if (!single_pred_p (then_bb))
return FALSE;
/* THEN must do something. */
test_bb->index, then_bb->index);
/* THEN is small. */
- bb_cost = total_bb_rtx_cost (then_bb);
- if (bb_cost < 0 || bb_cost >= COSTS_N_INSNS (BRANCH_COST))
+ if (! cheap_bb_rtx_cost_p (then_bb, COSTS_N_INSNS (BRANCH_COST)))
return FALSE;
/* Registers set are dead, or are predicable. */
if (! dead_or_predicable (test_bb, then_bb, else_bb,
- then_bb->succ->dest, 1))
+ single_succ (then_bb), 1))
return FALSE;
/* Conversion went ok, including moving the insns and fixing up the
jump. Adjust the CFG to match. */
- bitmap_operation (test_bb->global_live_at_end,
- else_bb->global_live_at_start,
- then_bb->global_live_at_end, BITMAP_IOR);
+ bitmap_ior (test_bb->il.rtl->global_live_at_end,
+ else_bb->il.rtl->global_live_at_start,
+ then_bb->il.rtl->global_live_at_end);
+
+
+ /* We can avoid creating a new basic block if then_bb is immediately
+ followed by else_bb, i.e. deleting then_bb allows test_bb to fall
+ thru to else_bb. */
+
+ if (then_bb->next_bb == else_bb
+ && then_bb->prev_bb == test_bb
+ && else_bb != EXIT_BLOCK_PTR)
+ {
+ redirect_edge_succ (FALLTHRU_EDGE (test_bb), else_bb);
+ new_bb = 0;
+ }
+ else
+ new_bb = redirect_edge_and_branch_force (FALLTHRU_EDGE (test_bb),
+ else_bb);
- new_bb = redirect_edge_and_branch_force (FALLTHRU_EDGE (test_bb), else_bb);
then_bb_index = then_bb->index;
delete_basic_block (then_bb);
if (new_bb)
{
new_bb->index = then_bb_index;
- BASIC_BLOCK (then_bb_index) = new_bb;
+ SET_BASIC_BLOCK (then_bb_index, new_bb);
+ /* Since the fallthru edge was redirected from test_bb to new_bb,
+ we need to ensure that new_bb is in the same partition as
+ test bb (you can not fall through across section boundaries). */
+ BB_COPY_PARTITION (new_bb, test_bb);
}
/* We've possibly created jump to next insn, cleanup_cfg will solve that
later. */
{
basic_block then_bb = then_edge->dest;
basic_block else_bb = else_edge->dest;
- edge else_succ = else_bb->succ;
- int bb_cost;
+ edge else_succ;
rtx note;
/* If we are partitioning hot/cold basic blocks, we don't want to
mess up unconditional or indirect jumps that cross between hot
- and cold sections. */
-
- if (flag_reorder_blocks_and_partition
- && ((BB_END (then_bb)
- && find_reg_note (BB_END (then_bb), REG_CROSSING_JUMP, NULL_RTX))
- || (BB_END (else_bb)
- && find_reg_note (BB_END (else_bb), REG_CROSSING_JUMP,
- NULL_RTX))))
+ and cold sections.
+
+ Basic block partitioning may result in some jumps that appear to
+ be optimizable (or blocks that appear to be mergeable), but which really
+ must be left untouched (they are required to make it safely across
+ partition boundaries). See the comments at the top of
+ bb-reorder.c:partition_hot_cold_basic_blocks for complete details. */
+
+ if ((BB_END (then_bb)
+ && find_reg_note (BB_END (then_bb), REG_CROSSING_JUMP, NULL_RTX))
+ || (BB_END (test_bb)
+ && find_reg_note (BB_END (test_bb), REG_CROSSING_JUMP, NULL_RTX))
+ || (BB_END (else_bb)
+ && find_reg_note (BB_END (else_bb), REG_CROSSING_JUMP,
+ NULL_RTX)))
return FALSE;
/* ELSE has one successor. */
- if (!else_succ || else_succ->succ_next != NULL)
+ if (!single_succ_p (else_bb))
return FALSE;
+ else
+ else_succ = single_succ_edge (else_bb);
/* ELSE outgoing edge is not complex. */
if (else_succ->flags & EDGE_COMPLEX)
return FALSE;
/* ELSE has one predecessor. */
- if (else_bb->pred->pred_next != NULL)
+ if (!single_pred_p (else_bb))
return FALSE;
/* THEN is not EXIT. */
- if (then_bb->index < 0)
+ if (then_bb->index < NUM_FIXED_BLOCKS)
return FALSE;
/* ELSE is predicted or SUCC(ELSE) postdominates THEN. */
note = find_reg_note (BB_END (test_bb), REG_BR_PROB, NULL_RTX);
if (note && INTVAL (XEXP (note, 0)) >= REG_BR_PROB_BASE / 2)
;
- else if (else_succ->dest->index < 0
+ else if (else_succ->dest->index < NUM_FIXED_BLOCKS
|| dominated_by_p (CDI_POST_DOMINATORS, then_bb,
else_succ->dest))
;
test_bb->index, else_bb->index);
/* ELSE is small. */
- bb_cost = total_bb_rtx_cost (else_bb);
- if (bb_cost < 0 || bb_cost >= COSTS_N_INSNS (BRANCH_COST))
+ if (! cheap_bb_rtx_cost_p (else_bb, COSTS_N_INSNS (BRANCH_COST)))
return FALSE;
/* Registers set are dead, or are predicable. */
/* Conversion went ok, including moving the insns and fixing up the
jump. Adjust the CFG to match. */
- bitmap_operation (test_bb->global_live_at_end,
- then_bb->global_live_at_start,
- else_bb->global_live_at_end, BITMAP_IOR);
+ bitmap_ior (test_bb->il.rtl->global_live_at_end,
+ then_bb->il.rtl->global_live_at_start,
+ else_bb->il.rtl->global_live_at_end);
delete_basic_block (else_bb);
head = BB_HEAD (merge_bb);
end = BB_END (merge_bb);
+ /* If merge_bb ends with a tablejump, predicating/moving insn's
+ into test_bb and then deleting merge_bb will result in the jumptable
+ that follows merge_bb being removed along with merge_bb and then we
+ get an unresolved reference to the jumptable. */
+ if (tablejump_p (end, NULL, NULL))
+ return FALSE;
+
if (LABEL_P (head))
head = NEXT_INSN (head);
if (NOTE_P (head))
that any registers modified are dead at the branch site. */
rtx insn, cond, prev;
- regset_head merge_set_head, tmp_head, test_live_head, test_set_head;
regset merge_set, tmp, test_live, test_set;
struct propagate_block_info *pbi;
- int i, fail = 0;
+ unsigned i, fail = 0;
+ bitmap_iterator bi;
/* Check for no calls or trapping operations. */
for (insn = head; ; insn = NEXT_INSN (insn))
TEST_SET = set of registers set between EARLIEST and the
end of the block. */
- tmp = INITIALIZE_REG_SET (tmp_head);
- merge_set = INITIALIZE_REG_SET (merge_set_head);
- test_live = INITIALIZE_REG_SET (test_live_head);
- test_set = INITIALIZE_REG_SET (test_set_head);
+ tmp = ALLOC_REG_SET (®_obstack);
+ merge_set = ALLOC_REG_SET (®_obstack);
+ test_live = ALLOC_REG_SET (®_obstack);
+ test_set = ALLOC_REG_SET (®_obstack);
/* ??? bb->local_set is only valid during calculate_global_regs_live,
so we must recompute usage for MERGE_BB. Not so bad, I suppose,
since we've already asserted that MERGE_BB is small. */
+ /* If we allocated new pseudos (e.g. in the conditional move
+ expander called from noce_emit_cmove), we must resize the
+ array first. */
+ if (max_regno < max_reg_num ())
+ {
+ max_regno = max_reg_num ();
+ allocate_reg_info (max_regno, FALSE, FALSE);
+ }
propagate_block (merge_bb, tmp, merge_set, merge_set, 0);
/* For small register class machines, don't lengthen lifetimes of
hard registers before reload. */
if (SMALL_REGISTER_CLASSES && ! reload_completed)
{
- EXECUTE_IF_SET_IN_BITMAP
- (merge_set, 0, i,
- {
- if (i < FIRST_PSEUDO_REGISTER
- && ! fixed_regs[i]
- && ! global_regs[i])
+ EXECUTE_IF_SET_IN_BITMAP (merge_set, 0, i, bi)
+ {
+ if (i < FIRST_PSEUDO_REGISTER
+ && ! fixed_regs[i]
+ && ! global_regs[i])
fail = 1;
- });
+ }
}
/* For TEST, we're interested in a range of insns, not a whole block.
Moreover, we're interested in the insns live from OTHER_BB. */
- COPY_REG_SET (test_live, other_bb->global_live_at_start);
+ COPY_REG_SET (test_live, other_bb->il.rtl->global_live_at_start);
pbi = init_propagate_block_info (test_bb, test_live, test_set, test_set,
0);
/* We can perform the transformation if
MERGE_SET & (TEST_SET | TEST_LIVE)
and
- TEST_SET & merge_bb->global_live_at_start
+ TEST_SET & merge_bb->il.rtl->global_live_at_start
are empty. */
- bitmap_operation (tmp, test_set, test_live, BITMAP_IOR);
- bitmap_operation (tmp, tmp, merge_set, BITMAP_AND);
- EXECUTE_IF_SET_IN_BITMAP(tmp, 0, i, fail = 1);
-
- bitmap_operation (tmp, test_set, merge_bb->global_live_at_start,
- BITMAP_AND);
- EXECUTE_IF_SET_IN_BITMAP(tmp, 0, i, fail = 1);
+ if (bitmap_intersect_p (test_set, merge_set)
+ || bitmap_intersect_p (test_live, merge_set)
+ || bitmap_intersect_p (test_set,
+ merge_bb->il.rtl->global_live_at_start))
+ fail = 1;
FREE_REG_SET (tmp);
FREE_REG_SET (merge_set);
if (other_bb != new_dest)
{
- if (old_dest)
- LABEL_NUSES (old_dest) -= 1;
- if (new_label)
- LABEL_NUSES (new_label) += 1;
- JUMP_LABEL (jump) = new_label;
- if (reversep)
- invert_br_probabilities (jump);
+ redirect_jump_2 (jump, old_dest, new_label, 0, reversep);
redirect_edge_succ (BRANCH_EDGE (test_bb), new_dest);
if (reversep)
/* Move the insns out of MERGE_BB to before the branch. */
if (head != NULL)
{
+ rtx insn;
+
if (end == BB_END (merge_bb))
BB_END (merge_bb) = PREV_INSN (head);
if (squeeze_notes (&head, &end))
return TRUE;
+ /* PR 21767: When moving insns above a conditional branch, REG_EQUAL
+ notes might become invalid. */
+ insn = head;
+ do
+ {
+ rtx note, set;
+
+ if (! INSN_P (insn))
+ continue;
+ note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
+ if (! note)
+ continue;
+ set = single_set (insn);
+ if (!set || !function_invariant_p (SET_SRC (set)))
+ remove_note (insn, note);
+ } while (insn != end && (insn = NEXT_INSN (insn)));
+
reorder_insns (head, end, PREV_INSN (earliest));
}
\f
/* Main entry point for all if-conversion. */
-void
+static void
if_convert (int x_life_data_ok)
{
basic_block bb;
num_true_changes = 0;
life_data_ok = (x_life_data_ok != 0);
- if ((! targetm.cannot_modify_jumps_p ())
- && (!flag_reorder_blocks_and_partition || !no_new_pseudos))
- mark_loop_exit_edges ();
+ loop_optimizer_init (AVOID_CFG_MODIFICATIONS);
+ if (current_loops)
+ {
+ mark_loop_exit_edges ();
+ loop_optimizer_finalize ();
+ }
+ free_dominance_info (CDI_DOMINATORS);
/* Compute postdominators if we think we'll use them. */
if (HAVE_conditional_execution || life_data_ok)
verify_flow_info ();
#endif
}
+\f
+static bool
+gate_handle_if_conversion (void)
+{
+ return (optimize > 0);
+}
+
+/* If-conversion and CFG cleanup. */
+static unsigned int
+rest_of_handle_if_conversion (void)
+{
+ if (flag_if_conversion)
+ {
+ if (dump_file)
+ dump_flow_info (dump_file, dump_flags);
+ cleanup_cfg (CLEANUP_EXPENSIVE);
+ reg_scan (get_insns (), max_reg_num ());
+ if_convert (0);
+ }
+
+ timevar_push (TV_JUMP);
+ cleanup_cfg (CLEANUP_EXPENSIVE);
+ reg_scan (get_insns (), max_reg_num ());
+ timevar_pop (TV_JUMP);
+ return 0;
+}
+
+struct tree_opt_pass pass_rtl_ifcvt =
+{
+ "ce1", /* name */
+ gate_handle_if_conversion, /* gate */
+ rest_of_handle_if_conversion, /* execute */
+ NULL, /* sub */
+ NULL, /* next */
+ 0, /* static_pass_number */
+ TV_IFCVT, /* tv_id */
+ 0, /* properties_required */
+ 0, /* properties_provided */
+ 0, /* properties_destroyed */
+ 0, /* todo_flags_start */
+ TODO_dump_func, /* todo_flags_finish */
+ 'C' /* letter */
+};
+
+static bool
+gate_handle_if_after_combine (void)
+{
+ return (optimize > 0 && flag_if_conversion);
+}
+
+
+/* Rerun if-conversion, as combine may have simplified things enough
+ to now meet sequence length restrictions. */
+static unsigned int
+rest_of_handle_if_after_combine (void)
+{
+ no_new_pseudos = 0;
+ if_convert (1);
+ no_new_pseudos = 1;
+ return 0;
+}
+
+struct tree_opt_pass pass_if_after_combine =
+{
+ "ce2", /* name */
+ gate_handle_if_after_combine, /* gate */
+ rest_of_handle_if_after_combine, /* execute */
+ NULL, /* sub */
+ NULL, /* next */
+ 0, /* static_pass_number */
+ TV_IFCVT, /* tv_id */
+ 0, /* properties_required */
+ 0, /* properties_provided */
+ 0, /* properties_destroyed */
+ 0, /* todo_flags_start */
+ TODO_dump_func |
+ TODO_ggc_collect, /* todo_flags_finish */
+ 'C' /* letter */
+};
+
+
+static bool
+gate_handle_if_after_reload (void)
+{
+ return (optimize > 0);
+}
+
+static unsigned int
+rest_of_handle_if_after_reload (void)
+{
+ /* Last attempt to optimize CFG, as scheduling, peepholing and insn
+ splitting possibly introduced more crossjumping opportunities. */
+ cleanup_cfg (CLEANUP_EXPENSIVE
+ | CLEANUP_UPDATE_LIFE
+ | (flag_crossjumping ? CLEANUP_CROSSJUMP : 0));
+ if (flag_if_conversion2)
+ if_convert (1);
+ return 0;
+}
+
+
+struct tree_opt_pass pass_if_after_reload =
+{
+ "ce3", /* name */
+ gate_handle_if_after_reload, /* gate */
+ rest_of_handle_if_after_reload, /* execute */
+ NULL, /* sub */
+ NULL, /* next */
+ 0, /* static_pass_number */
+ TV_IFCVT2, /* tv_id */
+ 0, /* properties_required */
+ 0, /* properties_provided */
+ 0, /* properties_destroyed */
+ 0, /* todo_flags_start */
+ TODO_dump_func |
+ TODO_ggc_collect, /* todo_flags_finish */
+ 'E' /* letter */
+};
OSDN Git Service
