/* Instruction scheduling pass. This file computes dependencies between
instructions.
Copyright (C) 1992, 1993, 1994, 1995, 1996, 1997, 1998,
- 1999, 2000, 2001, 2002, 2003 Free Software Foundation, Inc.
+ 1999, 2000, 2001, 2002, 2003, 2004 Free Software Foundation, Inc.
Contributed by Michael Tiemann (tiemann@cygnus.com) Enhanced by,
and currently maintained by, Jim Wilson (wilson@cygnus.com)
#include "cselib.h"
#include "df.h"
-extern char *reg_known_equiv_p;
-extern rtx *reg_known_value;
static regset_head reg_pending_sets_head;
static regset_head reg_pending_clobbers_head;
has enough entries to represent a dependency on any other insn in
the insn chain. All bitmap for true dependencies cache is
allocated then the rest two ones are also allocated. */
-static sbitmap *true_dependency_cache;
-static sbitmap *anti_dependency_cache;
-static sbitmap *output_dependency_cache;
+static bitmap_head *true_dependency_cache;
+static bitmap_head *anti_dependency_cache;
+static bitmap_head *output_dependency_cache;
+int cache_size;
/* To speed up checking consistency of formed forward insn
dependencies we use the following cache. Another possible solution
could be switching off checking duplication of insns in forward
dependencies. */
#ifdef ENABLE_CHECKING
-static sbitmap *forward_dependency_cache;
+static bitmap_head *forward_dependency_cache;
#endif
static int deps_may_trap_p (rtx);
{
rtx addr = XEXP (mem, 0);
- if (REG_P (addr)
- && REGNO (addr) >= FIRST_PSEUDO_REGISTER
- && reg_known_value[REGNO (addr)])
- addr = reg_known_value[REGNO (addr)];
+ if (REG_P (addr) && REGNO (addr) >= FIRST_PSEUDO_REGISTER)
+ {
+ rtx t = get_reg_known_value (REGNO (addr));
+ if (t)
+ addr = t;
+ }
return rtx_addr_can_trap_p (addr);
}
\f
get_condition (rtx insn)
{
rtx pat = PATTERN (insn);
- rtx cond;
+ rtx src;
if (pat == 0)
return 0;
+
if (GET_CODE (pat) == COND_EXEC)
return COND_EXEC_TEST (pat);
- if (GET_CODE (insn) != JUMP_INSN)
- return 0;
- if (GET_CODE (pat) != SET || SET_SRC (pat) != pc_rtx)
- return 0;
- if (GET_CODE (SET_DEST (pat)) != IF_THEN_ELSE)
- return 0;
- pat = SET_DEST (pat);
- cond = XEXP (pat, 0);
- if (GET_CODE (XEXP (cond, 1)) == LABEL_REF
- && XEXP (cond, 2) == pc_rtx)
- return cond;
- else if (GET_CODE (XEXP (cond, 2)) == LABEL_REF
- && XEXP (cond, 1) == pc_rtx)
- return gen_rtx_fmt_ee (reverse_condition (GET_CODE (cond)), GET_MODE (cond),
- XEXP (cond, 0), XEXP (cond, 1));
- else
+
+ if (!any_condjump_p (insn) || !onlyjump_p (insn))
return 0;
+
+ src = SET_SRC (pc_set (insn));
+#if 0
+ /* The previous code here was completely invalid and could never extract
+ the condition from a jump. This code does the correct thing, but that
+ triggers latent bugs later in the scheduler on ports with conditional
+ execution. So this is disabled for now. */
+ if (XEXP (src, 2) == pc_rtx)
+ return XEXP (src, 0);
+ else if (XEXP (src, 1) == pc_rtx)
+ {
+ rtx cond = XEXP (src, 0);
+ enum rtx_code revcode = reversed_comparison_code (cond, insn);
+
+ if (revcode == UNKNOWN)
+ return 0;
+ return gen_rtx_fmt_ee (revcode, GET_MODE (cond), XEXP (cond, 0),
+ XEXP (cond, 1));
+ }
+#endif
+
+ return 0;
}
/* Return nonzero if conditions COND1 and COND2 can never be both true. */
static int
conditions_mutex_p (rtx cond1, rtx cond2)
{
- if (GET_RTX_CLASS (GET_CODE (cond1)) == '<'
- && GET_RTX_CLASS (GET_CODE (cond2)) == '<'
- && GET_CODE (cond1) == reverse_condition (GET_CODE (cond2))
+ if (COMPARISON_P (cond1)
+ && COMPARISON_P (cond2)
+ && GET_CODE (cond1) == reversed_comparison_code (cond2, NULL)
&& XEXP (cond1, 0) == XEXP (cond2, 0)
&& XEXP (cond1, 1) == XEXP (cond2, 1))
return 1;
/* We can get a dependency on deleted insns due to optimizations in
the register allocation and reloading or due to splitting. Any
such dependency is useless and can be ignored. */
- if (GET_CODE (elem) == NOTE)
+ if (NOTE_P (elem))
return 0;
/* flow.c doesn't handle conditional lifetimes entirely correctly;
/* ??? add_dependence is the wrong place to be eliding dependencies,
as that forgets that the condition expressions themselves may
be dependent. */
- if (GET_CODE (insn) != CALL_INSN && GET_CODE (elem) != CALL_INSN)
+ if (!CALL_P (insn) && !CALL_P (elem))
{
cond1 = get_condition (insn);
cond2 = get_condition (elem);
No need for interblock dependences with calls, since
calls are not moved between blocks. Note: the edge where
elem is a CALL is still required. */
- if (GET_CODE (insn) == CALL_INSN
+ if (CALL_P (insn)
&& (INSN_BB (elem) != INSN_BB (insn)))
return 0;
#endif
{
enum reg_note present_dep_type = 0;
- if (anti_dependency_cache == NULL || output_dependency_cache == NULL)
- abort ();
- if (TEST_BIT (true_dependency_cache[INSN_LUID (insn)], INSN_LUID (elem)))
+ gcc_assert (anti_dependency_cache);
+ gcc_assert (output_dependency_cache);
+ if (bitmap_bit_p (&true_dependency_cache[INSN_LUID (insn)],
+ INSN_LUID (elem)))
/* Do nothing (present_set_type is already 0). */
;
- else if (TEST_BIT (anti_dependency_cache[INSN_LUID (insn)],
+ else if (bitmap_bit_p (&anti_dependency_cache[INSN_LUID (insn)],
INSN_LUID (elem)))
present_dep_type = REG_DEP_ANTI;
- else if (TEST_BIT (output_dependency_cache[INSN_LUID (insn)],
+ else if (bitmap_bit_p (&output_dependency_cache[INSN_LUID (insn)],
INSN_LUID (elem)))
present_dep_type = REG_DEP_OUTPUT;
else
may be changed. */
if (true_dependency_cache != NULL)
{
- if (REG_NOTE_KIND (link) == REG_DEP_ANTI)
- RESET_BIT (anti_dependency_cache[INSN_LUID (insn)],
- INSN_LUID (elem));
- else if (REG_NOTE_KIND (link) == REG_DEP_OUTPUT
- && output_dependency_cache)
- RESET_BIT (output_dependency_cache[INSN_LUID (insn)],
- INSN_LUID (elem));
- else
- abort ();
+ enum reg_note kind = REG_NOTE_KIND (link);
+ switch (kind)
+ {
+ case REG_DEP_ANTI:
+ bitmap_clear_bit (&anti_dependency_cache[INSN_LUID (insn)],
+ INSN_LUID (elem));
+ break;
+ case REG_DEP_OUTPUT:
+ gcc_assert (output_dependency_cache);
+ bitmap_clear_bit (&output_dependency_cache[INSN_LUID (insn)],
+ INSN_LUID (elem));
+ break;
+ default:
+ gcc_unreachable ();
+ }
}
#endif
if (true_dependency_cache != NULL)
{
if ((int) REG_NOTE_KIND (link) == 0)
- SET_BIT (true_dependency_cache[INSN_LUID (insn)],
- INSN_LUID (elem));
+ bitmap_set_bit (&true_dependency_cache[INSN_LUID (insn)],
+ INSN_LUID (elem));
else if (REG_NOTE_KIND (link) == REG_DEP_ANTI)
- SET_BIT (anti_dependency_cache[INSN_LUID (insn)],
- INSN_LUID (elem));
+ bitmap_set_bit (&anti_dependency_cache[INSN_LUID (insn)],
+ INSN_LUID (elem));
else if (REG_NOTE_KIND (link) == REG_DEP_OUTPUT)
- SET_BIT (output_dependency_cache[INSN_LUID (insn)],
- INSN_LUID (elem));
+ bitmap_set_bit (&output_dependency_cache[INSN_LUID (insn)],
+ INSN_LUID (elem));
}
#endif
return 0;
if (true_dependency_cache != NULL)
{
if ((int) dep_type == 0)
- SET_BIT (true_dependency_cache[INSN_LUID (insn)], INSN_LUID (elem));
+ bitmap_set_bit (&true_dependency_cache[INSN_LUID (insn)], INSN_LUID (elem));
else if (dep_type == REG_DEP_ANTI)
- SET_BIT (anti_dependency_cache[INSN_LUID (insn)], INSN_LUID (elem));
+ bitmap_set_bit (&anti_dependency_cache[INSN_LUID (insn)], INSN_LUID (elem));
else if (dep_type == REG_DEP_OUTPUT)
- SET_BIT (output_dependency_cache[INSN_LUID (insn)], INSN_LUID (elem));
+ bitmap_set_bit (&output_dependency_cache[INSN_LUID (insn)], INSN_LUID (elem));
}
#endif
return 1;
mem = shallow_copy_rtx (mem);
XEXP (mem, 0) = cselib_subst_to_values (XEXP (mem, 0));
}
- link = alloc_EXPR_LIST (VOIDmode, mem, *mem_list);
+ link = alloc_EXPR_LIST (VOIDmode, canon_rtx (mem), *mem_list);
*mem_list = link;
deps->pending_lists_length++;
dest = XEXP (dest, 0);
}
- if (GET_CODE (dest) == REG)
+ if (REG_P (dest))
{
regno = REGNO (dest);
If so, mark all of them just like the first. */
if (regno < FIRST_PSEUDO_REGISTER)
{
- int i = HARD_REGNO_NREGS (regno, GET_MODE (dest));
+ int i = hard_regno_nregs[regno][GET_MODE (dest)];
if (code == SET)
{
while (--i >= 0)
purpose already. */
else if (regno >= deps->max_reg)
{
- if (GET_CODE (PATTERN (insn)) != USE
- && GET_CODE (PATTERN (insn)) != CLOBBER)
- abort ();
+ gcc_assert (GET_CODE (PATTERN (insn)) == USE
+ || GET_CODE (PATTERN (insn)) == CLOBBER);
}
else
{
/* Pseudos that are REG_EQUIV to something may be replaced
by that during reloading. We need only add dependencies for
the address in the REG_EQUIV note. */
- if (!reload_completed
- && reg_known_equiv_p[regno]
- && GET_CODE (reg_known_value[regno]) == MEM)
- sched_analyze_2 (deps, XEXP (reg_known_value[regno], 0), insn);
+ if (!reload_completed && get_reg_known_equiv_p (regno))
+ {
+ rtx t = get_reg_known_value (regno);
+ if (MEM_P (t))
+ sched_analyze_2 (deps, XEXP (t, 0), insn);
+ }
/* Don't let it cross a call after scheduling if it doesn't
already cross one. */
add_dependence_list (insn, deps->last_function_call, REG_DEP_ANTI);
}
}
- else if (GET_CODE (dest) == MEM)
+ else if (MEM_P (dest))
{
/* Writing memory. */
rtx t = dest;
cselib_lookup (XEXP (t, 0), Pmode, 1);
XEXP (t, 0) = cselib_subst_to_values (XEXP (t, 0));
}
+ t = canon_rtx (t);
if (deps->pending_lists_length > MAX_PENDING_LIST_LENGTH)
{
case CC0:
/* User of CC0 depends on immediately preceding insn. */
set_sched_group_p (insn);
+ /* Don't move CC0 setter to another block (it can set up the
+ same flag for previous CC0 users which is safe). */
+ CANT_MOVE (prev_nonnote_insn (insn)) = 1;
return;
#endif
int regno = REGNO (x);
if (regno < FIRST_PSEUDO_REGISTER)
{
- int i = HARD_REGNO_NREGS (regno, GET_MODE (x));
+ int i = hard_regno_nregs[regno][GET_MODE (x)];
while (--i >= 0)
SET_REGNO_REG_SET (reg_pending_uses, regno + i);
}
purpose already. */
else if (regno >= deps->max_reg)
{
- if (GET_CODE (PATTERN (insn)) != USE
- && GET_CODE (PATTERN (insn)) != CLOBBER)
- abort ();
+ gcc_assert (GET_CODE (PATTERN (insn)) == USE
+ || GET_CODE (PATTERN (insn)) == CLOBBER);
}
else
{
/* Pseudos that are REG_EQUIV to something may be replaced
by that during reloading. We need only add dependencies for
the address in the REG_EQUIV note. */
- if (!reload_completed
- && reg_known_equiv_p[regno]
- && GET_CODE (reg_known_value[regno]) == MEM)
- sched_analyze_2 (deps, XEXP (reg_known_value[regno], 0), insn);
+ if (!reload_completed && get_reg_known_equiv_p (regno))
+ {
+ rtx t = get_reg_known_value (regno);
+ if (MEM_P (t))
+ sched_analyze_2 (deps, XEXP (t, 0), insn);
+ }
/* If the register does not already cross any calls, then add this
insn to the sched_before_next_call list so that it will still
cselib_lookup (XEXP (t, 0), Pmode, 1);
XEXP (t, 0) = cselib_subst_to_values (XEXP (t, 0));
}
+ t = canon_rtx (t);
pending = deps->pending_read_insns;
pending_mem = deps->pending_read_mems;
while (pending)
}
for (u = deps->last_pending_memory_flush; u; u = XEXP (u, 1))
- if (GET_CODE (XEXP (u, 0)) != JUMP_INSN
+ if (!JUMP_P (XEXP (u, 0))
|| deps_may_trap_p (x))
add_dependence (insn, XEXP (u, 0), REG_DEP_ANTI);
sched_analyze_2 (deps, x, insn);
/* Mark registers CLOBBERED or used by called function. */
- if (GET_CODE (insn) == CALL_INSN)
+ if (CALL_P (insn))
{
for (link = CALL_INSN_FUNCTION_USAGE (insn); link; link = XEXP (link, 1))
{
reg_pending_barrier = MOVE_BARRIER;
}
- if (GET_CODE (insn) == JUMP_INSN)
+ if (JUMP_P (insn))
{
rtx next;
next = next_nonnote_insn (insn);
- if (next && GET_CODE (next) == BARRIER)
+ if (next && BARRIER_P (next))
reg_pending_barrier = TRUE_BARRIER;
else
{
tmp = SET_DEST (set);
if (GET_CODE (tmp) == SUBREG)
tmp = SUBREG_REG (tmp);
- if (GET_CODE (tmp) == REG)
+ if (REG_P (tmp))
dest_regno = REGNO (tmp);
else
goto end_call_group;
tmp = SET_SRC (set);
if (GET_CODE (tmp) == SUBREG)
tmp = SUBREG_REG (tmp);
- if (GET_CODE (tmp) == REG)
+ if ((GET_CODE (tmp) == PLUS
+ || GET_CODE (tmp) == MINUS)
+ && REG_P (XEXP (tmp, 0))
+ && REGNO (XEXP (tmp, 0)) == STACK_POINTER_REGNUM
+ && dest_regno == STACK_POINTER_REGNUM)
+ src_regno = STACK_POINTER_REGNUM;
+ else if (REG_P (tmp))
src_regno = REGNO (tmp);
else
goto end_call_group;
if (src_regno < FIRST_PSEUDO_REGISTER
|| dest_regno < FIRST_PSEUDO_REGISTER)
{
- set_sched_group_p (insn);
+ /* If we are inside a post-call group right at the start of the
+ scheduling region, we must not add a dependency. */
+ if (deps->in_post_call_group_p == post_call_initial)
+ {
+ SCHED_GROUP_P (insn) = 1;
+ deps->in_post_call_group_p = post_call;
+ }
+ else
+ set_sched_group_p (insn);
CANT_MOVE (insn) = 1;
}
else
{
end_call_group:
- deps->in_post_call_group_p = false;
+ deps->in_post_call_group_p = not_post_call;
}
}
}
rtx loop_notes = 0;
if (current_sched_info->use_cselib)
- cselib_init ();
+ cselib_init (true);
+ /* Before reload, if the previous block ended in a call, show that
+ we are inside a post-call group, so as to keep the lifetimes of
+ hard registers correct. */
+ if (! reload_completed && !LABEL_P (head))
+ {
+ insn = prev_nonnote_insn (head);
+ if (insn && CALL_P (insn))
+ deps->in_post_call_group_p = post_call_initial;
+ }
for (insn = head;; insn = NEXT_INSN (insn))
{
rtx link, end_seq, r0, set;
- if (GET_CODE (insn) == INSN || GET_CODE (insn) == JUMP_INSN)
+ if (NONJUMP_INSN_P (insn) || JUMP_P (insn))
{
/* Clear out the stale LOG_LINKS from flow. */
free_INSN_LIST_list (&LOG_LINKS (insn));
/* Make each JUMP_INSN a scheduling barrier for memory
references. */
- if (GET_CODE (insn) == JUMP_INSN)
+ if (JUMP_P (insn))
{
/* Keep the list a reasonable size. */
if (deps->pending_flush_length++ > MAX_PENDING_LIST_LENGTH)
sched_analyze_insn (deps, PATTERN (insn), insn, loop_notes);
loop_notes = 0;
}
- else if (GET_CODE (insn) == CALL_INSN)
+ else if (CALL_P (insn))
{
int i;
/* Before reload, begin a post-call group, so as to keep the
lifetimes of hard registers correct. */
if (! reload_completed)
- deps->in_post_call_group_p = true;
+ deps->in_post_call_group_p = post_call;
}
/* See comments on reemit_notes as to why we do this.
??? Actually, the reemit_notes just say what is done, not why. */
- if (GET_CODE (insn) == NOTE
+ if (NOTE_P (insn)
&& (NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG
|| NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_END
|| NOTE_LINE_NUMBER (insn) == NOTE_INSN_EH_REGION_BEG
|| NOTE_LINE_NUMBER (insn) == NOTE_INSN_EH_REGION_END)
rtx_region = GEN_INT (NOTE_EH_HANDLER (insn));
else
- rtx_region = GEN_INT (0);
+ rtx_region = const0_rtx;
loop_notes = alloc_EXPR_LIST (REG_SAVE_NOTE,
rtx_region,
the outermost libcall sequence. */
&& deps->libcall_block_tail_insn == 0
/* The sequence must start with a clobber of a register. */
- && GET_CODE (insn) == INSN
+ && NONJUMP_INSN_P (insn)
&& GET_CODE (PATTERN (insn)) == CLOBBER
- && (r0 = XEXP (PATTERN (insn), 0), GET_CODE (r0) == REG)
- && GET_CODE (XEXP (PATTERN (insn), 0)) == REG
+ && (r0 = XEXP (PATTERN (insn), 0), REG_P (r0))
+ && REG_P (XEXP (PATTERN (insn), 0))
/* The CLOBBER must also have a REG_LIBCALL note attached. */
&& (link = find_reg_note (insn, REG_LIBCALL, NULL_RTX)) != 0
&& (end_seq = XEXP (link, 0)) != 0
return;
}
}
- abort ();
+ gcc_unreachable ();
}
\f
However, if we have enabled checking we might as well go
ahead and verify that add_dependence worked properly. */
- if (GET_CODE (from) == NOTE
- || INSN_DELETED_P (from)
- || (forward_dependency_cache != NULL
- && TEST_BIT (forward_dependency_cache[INSN_LUID (from)],
- INSN_LUID (to)))
- || (forward_dependency_cache == NULL
- && find_insn_list (to, INSN_DEPEND (from))))
- abort ();
+ gcc_assert (!NOTE_P (from));
+ gcc_assert (!INSN_DELETED_P (from));
+ if (forward_dependency_cache)
+ gcc_assert (!bitmap_bit_p (&forward_dependency_cache[INSN_LUID (from)],
+ INSN_LUID (to)));
+ else
+ gcc_assert (!find_insn_list (to, INSN_DEPEND (from)));
+
+ /* ??? If bitmap_bit_p is a predicate, what is this supposed to do? */
if (forward_dependency_cache != NULL)
- SET_BIT (forward_dependency_cache[INSN_LUID (from)],
- INSN_LUID (to));
+ bitmap_bit_p (&forward_dependency_cache[INSN_LUID (from)],
+ INSN_LUID (to));
#endif
new_link = alloc_INSN_LIST (to, INSN_DEPEND (from));
deps->last_pending_memory_flush = 0;
deps->last_function_call = 0;
deps->sched_before_next_call = 0;
- deps->in_post_call_group_p = false;
+ deps->in_post_call_group_p = not_post_call;
deps->libcall_block_tail_insn = 0;
}
free_INSN_LIST_list (&deps->last_pending_memory_flush);
/* Without the EXECUTE_IF_SET, this loop is executed max_reg * nr_regions
- times. For a test case with 42000 regs and 8000 small basic blocks,
+ times. For a testcase with 42000 regs and 8000 small basic blocks,
this loop accounted for nearly 60% (84 sec) of the total -O2 runtime. */
EXECUTE_IF_SET_IN_REG_SET (&deps->reg_last_in_use, 0, i,
{
what we consider "very high". */
if (luid / n_basic_blocks > 100 * 5)
{
- true_dependency_cache = sbitmap_vector_alloc (luid, luid);
- sbitmap_vector_zero (true_dependency_cache, luid);
- anti_dependency_cache = sbitmap_vector_alloc (luid, luid);
- sbitmap_vector_zero (anti_dependency_cache, luid);
- output_dependency_cache = sbitmap_vector_alloc (luid, luid);
- sbitmap_vector_zero (output_dependency_cache, luid);
+ int i;
+ true_dependency_cache = xmalloc (luid * sizeof (bitmap_head));
+ anti_dependency_cache = xmalloc (luid * sizeof (bitmap_head));
+ output_dependency_cache = xmalloc (luid * sizeof (bitmap_head));
#ifdef ENABLE_CHECKING
- forward_dependency_cache = sbitmap_vector_alloc (luid, luid);
- sbitmap_vector_zero (forward_dependency_cache, luid);
+ forward_dependency_cache = xmalloc (luid * sizeof (bitmap_head));
#endif
+ for (i = 0; i < luid; i++)
+ {
+ bitmap_initialize (&true_dependency_cache[i], 0);
+ bitmap_initialize (&anti_dependency_cache[i], 0);
+ bitmap_initialize (&output_dependency_cache[i], 0);
+#ifdef ENABLE_CHECKING
+ bitmap_initialize (&forward_dependency_cache[i], 0);
+#endif
+ }
+ cache_size = luid;
}
}
{
if (true_dependency_cache)
{
- sbitmap_vector_free (true_dependency_cache);
+ int i;
+
+ for (i = 0; i < cache_size; i++)
+ {
+ bitmap_clear (&true_dependency_cache[i]);
+ bitmap_clear (&anti_dependency_cache[i]);
+ bitmap_clear (&output_dependency_cache[i]);
+#ifdef ENABLE_CHECKING
+ bitmap_clear (&forward_dependency_cache[i]);
+#endif
+ }
+ free (true_dependency_cache);
true_dependency_cache = NULL;
- sbitmap_vector_free (anti_dependency_cache);
+ free (anti_dependency_cache);
anti_dependency_cache = NULL;
- sbitmap_vector_free (output_dependency_cache);
+ free (output_dependency_cache);
output_dependency_cache = NULL;
#ifdef ENABLE_CHECKING
- sbitmap_vector_free (forward_dependency_cache);
+ free (forward_dependency_cache);
forward_dependency_cache = NULL;
#endif
}