static int
get_uncond_jump_length (void)
{
rtx label, jump;
int length;
label = emit_label_before (gen_label_rtx (), get_insns ());
jump = emit_jump_insn (gen_jump (label));
length = get_attr_length (jump);
delete_insn (jump);
delete_insn (label);
return length;
}
static unsigned int
postreload_load (void)
{
basic_block bb;
init_alias_analysis ();
FOR_EACH_BB (bb)
{
rtx insn;
htab_load = htab_create (10, load_htab_hash, load_htab_eq, NULL);
FOR_BB_INSNS (bb, insn)
{
rtx set;
struct load **load;
/* Set reg_kill, invalidate entries if there is an
aliasing store or if the registers making up the address
change. */
htab_traverse_noresize
(htab_load, find_reg_kill_and_mem_invalidate, insn);
set = single_set (insn);
if (interesting_second_load (set, &load, insn))
{
rtx move;
move = gen_move_insn (SET_DEST (set), (*load)->reg);
/* Make sure we can generate a move. */
extract_insn (move);
if (! constrain_operands (1))
continue;
move = emit_insn_before (move, (*load)->reg_kill);
delete_insn (insn);
if (dump_file)
{
fputs ("Replaced this load:\n ", dump_file);
print_inline_rtx (dump_file, insn, 2);
fputs ("\n with this move:\n ", dump_file);
print_inline_rtx (dump_file, move, 2);
fputs ("\n\n", dump_file);
}
}
else if (interesting_load (set))
alloc_load (set);
else if (CALL_P (insn))
htab_empty (htab_load);
}
htab_empty (htab_load);
}
void
reemit_insn_block_notes (void)
{
tree cur_block = DECL_INITIAL (cfun->decl);
rtx insn, note;
insn = get_insns ();
if (!active_insn_p (insn))
insn = next_active_insn (insn);
for (; insn; insn = next_active_insn (insn))
{
tree this_block;
/* Avoid putting scope notes between jump table and its label. */
if (JUMP_P (insn)
&& (GET_CODE (PATTERN (insn)) == ADDR_VEC
|| GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC))
continue;
this_block = insn_scope (insn);
/* For sequences compute scope resulting from merging all scopes
of instructions nested inside. */
if (GET_CODE (PATTERN (insn)) == SEQUENCE)
{
int i;
rtx body = PATTERN (insn);
this_block = NULL;
for (i = 0; i < XVECLEN (body, 0); i++)
this_block = choose_inner_scope (this_block,
insn_scope (XVECEXP (body, 0, i)));
}
if (! this_block)
continue;
if (this_block != cur_block)
{
change_scope (insn, cur_block, this_block);
cur_block = this_block;
}
}
/* change_scope emits before the insn, not after. */
note = emit_note (NOTE_INSN_DELETED);
change_scope (note, cur_block, DECL_INITIAL (cfun->decl));
delete_insn (note);
reorder_blocks ();
}
/* Some old code expects exactly one BARRIER as the NEXT_INSN of a
non-fallthru insn. This is not generally true, as multiple barriers
may have crept in, or the BARRIER may be separated from the last
real insn by one or more NOTEs.
This simple pass moves barriers and removes duplicates so that the
old code is happy.
*/
unsigned int
cleanup_barriers (void)
{
rtx insn, next, prev;
for (insn = get_insns (); insn; insn = next)
{
next = NEXT_INSN (insn);
if (BARRIER_P (insn))
{
prev = prev_nonnote_insn (insn);
if (BARRIER_P (prev))
delete_insn (insn);
else if (prev != PREV_INSN (insn))
reorder_insns (insn, insn, prev);
}
}
return 0;
}
void
find_comparison_dom_walker::before_dom_children (basic_block bb)
{
struct comparison *last_cmp;
rtx_insn *insn, *next, *last_clobber;
bool last_cmp_valid;
bool need_purge = false;
bitmap killed;
killed = BITMAP_ALLOC (NULL);
/* The last comparison that was made. Will be reset to NULL
once the flags are clobbered. */
last_cmp = NULL;
/* True iff the last comparison has not been clobbered, nor
have its inputs. Used to eliminate duplicate compares. */
last_cmp_valid = false;
/* The last insn that clobbered the flags, if that insn is of
a form that may be valid for eliminating a following compare.
To be reset to NULL once the flags are set otherwise. */
last_clobber = NULL;
/* Propagate the last live comparison throughout the extended basic block. */
if (single_pred_p (bb))
{
last_cmp = (struct comparison *) single_pred (bb)->aux;
if (last_cmp)
last_cmp_valid = last_cmp->inputs_valid;
}
for (insn = BB_HEAD (bb); insn; insn = next)
{
rtx src;
next = (insn == BB_END (bb) ? NULL : NEXT_INSN (insn));
if (!NONDEBUG_INSN_P (insn))
continue;
/* Compute the set of registers modified by this instruction. */
bitmap_clear (killed);
df_simulate_find_defs (insn, killed);
src = conforming_compare (insn);
if (src)
{
rtx eh_note = NULL;
if (cfun->can_throw_non_call_exceptions)
eh_note = find_reg_note (insn, REG_EH_REGION, NULL);
if (last_cmp_valid && can_eliminate_compare (src, eh_note, last_cmp))
{
if (eh_note)
need_purge = true;
delete_insn (insn);
continue;
}
last_cmp = XCNEW (struct comparison);
last_cmp->insn = insn;
last_cmp->prev_clobber = last_clobber;
last_cmp->in_a = XEXP (src, 0);
last_cmp->in_b = XEXP (src, 1);
last_cmp->eh_note = eh_note;
last_cmp->orig_mode = GET_MODE (src);
all_compares.safe_push (last_cmp);
/* It's unusual, but be prepared for comparison patterns that
also clobber an input, or perhaps a scratch. */
last_clobber = NULL;
last_cmp_valid = true;
}
/* Notice if this instruction kills the flags register. */
else if (bitmap_bit_p (killed, targetm.flags_regnum))
{
/* See if this insn could be the "clobber" that eliminates
a future comparison. */
last_clobber = (arithmetic_flags_clobber_p (insn) ? insn : NULL);
/* In either case, the previous compare is no longer valid. */
last_cmp = NULL;
last_cmp_valid = false;
}
/* Notice if this instruction uses the flags register. */
else if (last_cmp)
find_flags_uses_in_insn (last_cmp, insn);
/* Notice if any of the inputs to the comparison have changed. */
if (last_cmp_valid
&& (bitmap_bit_p (killed, REGNO (last_cmp->in_a))
|| (REG_P (last_cmp->in_b)
&& bitmap_bit_p (killed, REGNO (last_cmp->in_b)))))
last_cmp_valid = false;
}
static void
combine_stack_adjustments_for_block (basic_block bb)
{
HOST_WIDE_INT last_sp_adjust = 0;
rtx last_sp_set = NULL_RTX;
rtx last2_sp_set = NULL_RTX;
struct csa_reflist *reflist = NULL;
rtx insn, next, set;
struct record_stack_refs_data data;
bool end_of_block = false;
for (insn = BB_HEAD (bb); !end_of_block ; insn = next)
{
end_of_block = insn == BB_END (bb);
next = NEXT_INSN (insn);
if (! INSN_P (insn))
continue;
set = single_set_for_csa (insn);
if (set)
{
rtx dest = SET_DEST (set);
rtx src = SET_SRC (set);
/* Find constant additions to the stack pointer. */
if (dest == stack_pointer_rtx
&& GET_CODE (src) == PLUS
&& XEXP (src, 0) == stack_pointer_rtx
&& CONST_INT_P (XEXP (src, 1)))
{
HOST_WIDE_INT this_adjust = INTVAL (XEXP (src, 1));
/* If we've not seen an adjustment previously, record
it now and continue. */
if (! last_sp_set)
{
last_sp_set = insn;
last_sp_adjust = this_adjust;
continue;
}
/* If not all recorded refs can be adjusted, or the
adjustment is now too large for a constant addition,
we cannot merge the two stack adjustments.
Also we need to be careful to not move stack pointer
such that we create stack accesses outside the allocated
area. We can combine an allocation into the first insn,
or a deallocation into the second insn. We can not
combine an allocation followed by a deallocation.
The only somewhat frequent occurrence of the later is when
a function allocates a stack frame but does not use it.
For this case, we would need to analyze rtl stream to be
sure that allocated area is really unused. This means not
only checking the memory references, but also all registers
or global memory references possibly containing a stack
frame address.
Perhaps the best way to address this problem is to teach
gcc not to allocate stack for objects never used. */
/* Combine an allocation into the first instruction. */
if (STACK_GROWS_DOWNWARD ? this_adjust <= 0 : this_adjust >= 0)
{
if (try_apply_stack_adjustment (last_sp_set, reflist,
last_sp_adjust + this_adjust,
this_adjust))
{
/* It worked! */
maybe_move_args_size_note (last_sp_set, insn, false);
delete_insn (insn);
last_sp_adjust += this_adjust;
continue;
}
}
/* Otherwise we have a deallocation. Do not combine with
a previous allocation. Combine into the second insn. */
else if (STACK_GROWS_DOWNWARD
? last_sp_adjust >= 0 : last_sp_adjust <= 0)
{
if (try_apply_stack_adjustment (insn, reflist,
last_sp_adjust + this_adjust,
-last_sp_adjust))
{
/* It worked! */
maybe_move_args_size_note (insn, last_sp_set, true);
delete_insn (last_sp_set);
last_sp_set = insn;
last_sp_adjust += this_adjust;
free_csa_reflist (reflist);
reflist = NULL;
continue;
}
}
/* Combination failed. Restart processing from here. If
deallocation+allocation conspired to cancel, we can
delete the old deallocation insn. */
if (last_sp_set)
{
if (last_sp_adjust == 0)
{
maybe_move_args_size_note (insn, last_sp_set, true);
delete_insn (last_sp_set);
}
else
last2_sp_set = last_sp_set;
}
free_csa_reflist (reflist);
reflist = NULL;
last_sp_set = insn;
last_sp_adjust = this_adjust;
continue;
}
/* Find a store with pre-(dec|inc)rement or pre-modify of exactly
the previous adjustment and turn it into a simple store. This
is equivalent to anticipating the stack adjustment so this must
be an allocation. */
if (MEM_P (dest)
&& ((STACK_GROWS_DOWNWARD
? (GET_CODE (XEXP (dest, 0)) == PRE_DEC
&& last_sp_adjust
== (HOST_WIDE_INT) GET_MODE_SIZE (GET_MODE (dest)))
: (GET_CODE (XEXP (dest, 0)) == PRE_INC
&& last_sp_adjust
== -(HOST_WIDE_INT) GET_MODE_SIZE (GET_MODE (dest))))
|| ((STACK_GROWS_DOWNWARD
? last_sp_adjust >= 0 : last_sp_adjust <= 0)
&& GET_CODE (XEXP (dest, 0)) == PRE_MODIFY
&& GET_CODE (XEXP (XEXP (dest, 0), 1)) == PLUS
&& XEXP (XEXP (XEXP (dest, 0), 1), 0)
== stack_pointer_rtx
&& GET_CODE (XEXP (XEXP (XEXP (dest, 0), 1), 1))
== CONST_INT
&& INTVAL (XEXP (XEXP (XEXP (dest, 0), 1), 1))
== -last_sp_adjust))
&& XEXP (XEXP (dest, 0), 0) == stack_pointer_rtx
&& !reg_mentioned_p (stack_pointer_rtx, src)
&& memory_address_p (GET_MODE (dest), stack_pointer_rtx)
&& try_apply_stack_adjustment (insn, reflist, 0,
-last_sp_adjust))
{
if (last2_sp_set)
maybe_move_args_size_note (last2_sp_set, last_sp_set, false);
else
maybe_move_args_size_note (insn, last_sp_set, true);
delete_insn (last_sp_set);
free_csa_reflist (reflist);
reflist = NULL;
last_sp_set = NULL_RTX;
last_sp_adjust = 0;
continue;
}
}
data.insn = insn;
data.reflist = reflist;
if (!CALL_P (insn) && last_sp_set
&& !for_each_rtx (&PATTERN (insn), record_stack_refs, &data))
{
reflist = data.reflist;
continue;
}
reflist = data.reflist;
/* Otherwise, we were not able to process the instruction.
Do not continue collecting data across such a one. */
if (last_sp_set
&& (CALL_P (insn)
|| reg_mentioned_p (stack_pointer_rtx, PATTERN (insn))))
{
if (last_sp_set && last_sp_adjust == 0)
{
force_move_args_size_note (bb, last2_sp_set, last_sp_set);
delete_insn (last_sp_set);
}
free_csa_reflist (reflist);
reflist = NULL;
last2_sp_set = NULL_RTX;
last_sp_set = NULL_RTX;
last_sp_adjust = 0;
}
}
if (last_sp_set && last_sp_adjust == 0)
{
force_move_args_size_note (bb, last2_sp_set, last_sp_set);
delete_insn (last_sp_set);
}
if (reflist)
free_csa_reflist (reflist);
}
static bool
attempt_change (rtx new_addr, rtx inc_reg)
{
/* There are four cases: For the two cases that involve an add
instruction, we are going to have to delete the add and insert a
mov. We are going to assume that the mov is free. This is
fairly early in the backend and there are a lot of opportunities
for removing that move later. In particular, there is the case
where the move may be dead, this is what dead code elimination
passes are for. The two cases where we have an inc insn will be
handled mov free. */
basic_block bb = BLOCK_FOR_INSN (mem_insn.insn);
rtx mov_insn = NULL;
int regno;
rtx mem = *mem_insn.mem_loc;
enum machine_mode mode = GET_MODE (mem);
rtx new_mem;
int old_cost = 0;
int new_cost = 0;
bool speed = optimize_bb_for_speed_p (bb);
PUT_MODE (mem_tmp, mode);
XEXP (mem_tmp, 0) = new_addr;
old_cost = (set_src_cost (mem, speed)
+ set_rtx_cost (PATTERN (inc_insn.insn), speed));
new_cost = set_src_cost (mem_tmp, speed);
/* The first item of business is to see if this is profitable. */
if (old_cost < new_cost)
{
if (dump_file)
fprintf (dump_file, "cost failure old=%d new=%d\n", old_cost, new_cost);
return false;
}
/* Jump through a lot of hoops to keep the attributes up to date. We
do not want to call one of the change address variants that take
an offset even though we know the offset in many cases. These
assume you are changing where the address is pointing by the
offset. */
new_mem = replace_equiv_address_nv (mem, new_addr);
if (! validate_change (mem_insn.insn, mem_insn.mem_loc, new_mem, 0))
{
if (dump_file)
fprintf (dump_file, "validation failure\n");
return false;
}
/* From here to the end of the function we are committed to the
change, i.e. nothing fails. Generate any necessary movs, move
any regnotes, and fix up the reg_next_{use,inc_use,def}. */
switch (inc_insn.form)
{
case FORM_PRE_ADD:
/* Replace the addition with a move. Do it at the location of
the addition since the operand of the addition may change
before the memory reference. */
mov_insn = insert_move_insn_before (inc_insn.insn,
inc_insn.reg_res, inc_insn.reg0);
move_dead_notes (mov_insn, inc_insn.insn, inc_insn.reg0);
regno = REGNO (inc_insn.reg_res);
reg_next_def[regno] = mov_insn;
reg_next_use[regno] = NULL;
regno = REGNO (inc_insn.reg0);
reg_next_use[regno] = mov_insn;
df_recompute_luids (bb);
break;
case FORM_POST_INC:
regno = REGNO (inc_insn.reg_res);
if (reg_next_use[regno] == reg_next_inc_use[regno])
reg_next_inc_use[regno] = NULL;
/* Fallthru. */
case FORM_PRE_INC:
regno = REGNO (inc_insn.reg_res);
reg_next_def[regno] = mem_insn.insn;
reg_next_use[regno] = NULL;
break;
case FORM_POST_ADD:
mov_insn = insert_move_insn_before (mem_insn.insn,
inc_insn.reg_res, inc_insn.reg0);
move_dead_notes (mov_insn, inc_insn.insn, inc_insn.reg0);
/* Do not move anything to the mov insn because the instruction
pointer for the main iteration has not yet hit that. It is
still pointing to the mem insn. */
regno = REGNO (inc_insn.reg_res);
reg_next_def[regno] = mem_insn.insn;
reg_next_use[regno] = NULL;
regno = REGNO (inc_insn.reg0);
reg_next_use[regno] = mem_insn.insn;
if ((reg_next_use[regno] == reg_next_inc_use[regno])
|| (reg_next_inc_use[regno] == inc_insn.insn))
reg_next_inc_use[regno] = NULL;
df_recompute_luids (bb);
break;
case FORM_last:
default:
gcc_unreachable ();
}
if (!inc_insn.reg1_is_const)
{
regno = REGNO (inc_insn.reg1);
reg_next_use[regno] = mem_insn.insn;
if ((reg_next_use[regno] == reg_next_inc_use[regno])
|| (reg_next_inc_use[regno] == inc_insn.insn))
reg_next_inc_use[regno] = NULL;
}
delete_insn (inc_insn.insn);
if (dump_file && mov_insn)
{
fprintf (dump_file, "inserting mov ");
dump_insn_slim (dump_file, mov_insn);
}
/* Record that this insn has an implicit side effect. */
add_reg_note (mem_insn.insn, REG_INC, inc_reg);
if (dump_file)
{
fprintf (dump_file, "****success ");
dump_insn_slim (dump_file, mem_insn.insn);
}
return true;
}
void
find_comparison_dom_walker::before_dom_children (basic_block bb)
{
struct comparison *last_cmp;
rtx insn, next, last_clobber;
bool last_cmp_valid;
bool need_purge = false;
bitmap killed;
killed = BITMAP_ALLOC (NULL);
/* The last comparison that was made. Will be reset to NULL
once the flags are clobbered. */
last_cmp = NULL;
/* True iff the last comparison has not been clobbered, nor
have its inputs. Used to eliminate duplicate compares. */
last_cmp_valid = false;
/* The last insn that clobbered the flags, if that insn is of
a form that may be valid for eliminating a following compare.
To be reset to NULL once the flags are set otherwise. */
last_clobber = NULL;
/* Propagate the last live comparison throughout the extended basic block. */
if (single_pred_p (bb))
{
last_cmp = (struct comparison *) single_pred (bb)->aux;
if (last_cmp)
last_cmp_valid = last_cmp->inputs_valid;
}
for (insn = BB_HEAD (bb); insn; insn = next)
{
rtx src;
next = (insn == BB_END (bb) ? NULL_RTX : NEXT_INSN (insn));
if (!NONDEBUG_INSN_P (insn))
continue;
/* Compute the set of registers modified by this instruction. */
bitmap_clear (killed);
df_simulate_find_defs (insn, killed);
src = conforming_compare (insn);
if (src)
{
enum machine_mode src_mode = GET_MODE (src);
rtx eh_note = NULL;
if (flag_non_call_exceptions)
eh_note = find_reg_note (insn, REG_EH_REGION, NULL);
if (!last_cmp_valid)
goto dont_delete;
/* Take care that it's in the same EH region. */
if (flag_non_call_exceptions
&& !rtx_equal_p (eh_note, last_cmp->eh_note))
goto dont_delete;
/* Make sure the compare is redundant with the previous. */
if (!rtx_equal_p (last_cmp->in_a, XEXP (src, 0))
|| !rtx_equal_p (last_cmp->in_b, XEXP (src, 1)))
goto dont_delete;
/* New mode must be compatible with the previous compare mode. */
{
enum machine_mode new_mode
= targetm.cc_modes_compatible (last_cmp->orig_mode, src_mode);
if (new_mode == VOIDmode)
goto dont_delete;
if (new_mode != last_cmp->orig_mode)
{
rtx x, flags = gen_rtx_REG (src_mode, targetm.flags_regnum);
/* Generate new comparison for substitution. */
x = gen_rtx_COMPARE (new_mode, XEXP (src, 0), XEXP (src, 1));
x = gen_rtx_SET (VOIDmode, flags, x);
if (!validate_change (last_cmp->insn,
&PATTERN (last_cmp->insn), x, false))
goto dont_delete;
last_cmp->orig_mode = new_mode;
}
}
/* All tests and substitutions succeeded! */
if (eh_note)
need_purge = true;
delete_insn (insn);
continue;
dont_delete:
last_cmp = XCNEW (struct comparison);
last_cmp->insn = insn;
last_cmp->prev_clobber = last_clobber;
last_cmp->in_a = XEXP (src, 0);
last_cmp->in_b = XEXP (src, 1);
last_cmp->eh_note = eh_note;
last_cmp->orig_mode = src_mode;
all_compares.safe_push (last_cmp);
/* It's unusual, but be prepared for comparison patterns that
also clobber an input, or perhaps a scratch. */
last_clobber = NULL;
last_cmp_valid = true;
}
/* Notice if this instruction kills the flags register. */
else if (bitmap_bit_p (killed, targetm.flags_regnum))
{
/* See if this insn could be the "clobber" that eliminates
a future comparison. */
last_clobber = (arithmetic_flags_clobber_p (insn) ? insn : NULL);
/* In either case, the previous compare is no longer valid. */
last_cmp = NULL;
last_cmp_valid = false;
continue;
}
/* Notice if this instruction uses the flags register. */
else if (last_cmp)
find_flags_uses_in_insn (last_cmp, insn);
/* Notice if any of the inputs to the comparison have changed. */
if (last_cmp_valid
&& (bitmap_bit_p (killed, REGNO (last_cmp->in_a))
|| (REG_P (last_cmp->in_b)
&& bitmap_bit_p (killed, REGNO (last_cmp->in_b)))))
last_cmp_valid = false;
}
static basic_block
expand_gimple_tailcall (basic_block bb, tree stmt, bool *can_fallthru)
{
rtx last2, last;
edge e;
edge_iterator ei;
int probability;
gcov_type count;
last2 = last = get_last_insn ();
expand_expr_stmt (stmt);
for (last = NEXT_INSN (last); last; last = NEXT_INSN (last))
if (CALL_P (last) && SIBLING_CALL_P (last))
goto found;
maybe_dump_rtl_for_tree_stmt (stmt, last2);
*can_fallthru = true;
return NULL;
found:
/* ??? Wouldn't it be better to just reset any pending stack adjust?
Any instructions emitted here are about to be deleted. */
do_pending_stack_adjust ();
/* Remove any non-eh, non-abnormal edges that don't go to exit. */
/* ??? I.e. the fallthrough edge. HOWEVER! If there were to be
EH or abnormal edges, we shouldn't have created a tail call in
the first place. So it seems to me we should just be removing
all edges here, or redirecting the existing fallthru edge to
the exit block. */
probability = 0;
count = 0;
for (ei = ei_start (bb->succs); (e = ei_safe_edge (ei)); )
{
if (!(e->flags & (EDGE_ABNORMAL | EDGE_EH)))
{
if (e->dest != EXIT_BLOCK_PTR)
{
e->dest->count -= e->count;
e->dest->frequency -= EDGE_FREQUENCY (e);
if (e->dest->count < 0)
e->dest->count = 0;
if (e->dest->frequency < 0)
e->dest->frequency = 0;
}
count += e->count;
probability += e->probability;
remove_edge (e);
}
else
ei_next (&ei);
}
/* This is somewhat ugly: the call_expr expander often emits instructions
after the sibcall (to perform the function return). These confuse the
find_sub_basic_blocks code, so we need to get rid of these. */
last = NEXT_INSN (last);
gcc_assert (BARRIER_P (last));
*can_fallthru = false;
while (NEXT_INSN (last))
{
/* For instance an sqrt builtin expander expands if with
sibcall in the then and label for `else`. */
if (LABEL_P (NEXT_INSN (last)))
{
*can_fallthru = true;
break;
}
delete_insn (NEXT_INSN (last));
}
e = make_edge (bb, EXIT_BLOCK_PTR, EDGE_ABNORMAL | EDGE_SIBCALL);
e->probability += probability;
e->count += count;
BB_END (bb) = last;
update_bb_for_insn (bb);
if (NEXT_INSN (last))
{
bb = create_basic_block (NEXT_INSN (last), get_last_insn (), bb);
last = BB_END (bb);
if (BARRIER_P (last))
BB_END (bb) = PREV_INSN (last);
}
maybe_dump_rtl_for_tree_stmt (stmt, last2);
return bb;
}
static rtx
duplicate_insn_chain (rtx from, rtx to)
{
rtx insn, last;
/* Avoid updating of boundaries of previous basic block. The
note will get removed from insn stream in fixup. */
last = emit_note (NOTE_INSN_DELETED);
/* Create copy at the end of INSN chain. The chain will
be reordered later. */
for (insn = from; insn != NEXT_INSN (to); insn = NEXT_INSN (insn))
{
switch (GET_CODE (insn))
{
case INSN:
case CALL_INSN:
case JUMP_INSN:
/* Avoid copying of dispatch tables. We never duplicate
tablejumps, so this can hit only in case the table got
moved far from original jump. */
if (GET_CODE (PATTERN (insn)) == ADDR_VEC
|| GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC)
break;
emit_copy_of_insn_after (insn, get_last_insn ());
break;
case CODE_LABEL:
break;
case BARRIER:
emit_barrier ();
break;
case NOTE:
switch (NOTE_LINE_NUMBER (insn))
{
/* In case prologue is empty and function contain label
in first BB, we may want to copy the block. */
case NOTE_INSN_PROLOGUE_END:
case NOTE_INSN_LOOP_VTOP:
case NOTE_INSN_LOOP_CONT:
case NOTE_INSN_LOOP_BEG:
case NOTE_INSN_LOOP_END:
/* Strip down the loop notes - we don't really want to keep
them consistent in loop copies. */
case NOTE_INSN_DELETED:
case NOTE_INSN_DELETED_LABEL:
/* No problem to strip these. */
case NOTE_INSN_EPILOGUE_BEG:
case NOTE_INSN_FUNCTION_END:
/* Debug code expect these notes to exist just once.
Keep them in the master copy.
??? It probably makes more sense to duplicate them for each
epilogue copy. */
case NOTE_INSN_FUNCTION_BEG:
/* There is always just single entry to function. */
case NOTE_INSN_BASIC_BLOCK:
break;
/* There is no purpose to duplicate prologue. */
case NOTE_INSN_BLOCK_BEG:
case NOTE_INSN_BLOCK_END:
/* The BLOCK_BEG/BLOCK_END notes should be eliminated when BB
reordering is in the progress. */
case NOTE_INSN_EH_REGION_BEG:
case NOTE_INSN_EH_REGION_END:
/* Should never exist at BB duplication time. */
abort ();
break;
case NOTE_INSN_REPEATED_LINE_NUMBER:
emit_note_copy (insn);
break;
default:
if (NOTE_LINE_NUMBER (insn) < 0)
abort ();
/* It is possible that no_line_number is set and the note
won't be emitted. */
emit_note_copy (insn);
}
break;
default:
abort ();
}
}
insn = NEXT_INSN (last);
delete_insn (last);
return insn;
}
void
insn_locators_initialize (void)
{
tree block = NULL;
tree last_block = NULL;
rtx insn, next;
int loc = 0;
int line_number = 0, last_line_number = 0;
char *file_name = NULL, *last_file_name = NULL;
prologue_locator = epilogue_locator = 0;
VARRAY_INT_INIT (block_locators_locs, 32, "block_locators_locs");
VARRAY_TREE_INIT (block_locators_blocks, 32, "block_locators_blocks");
VARRAY_INT_INIT (line_locators_locs, 32, "line_locators_locs");
VARRAY_INT_INIT (line_locators_lines, 32, "line_locators_lines");
VARRAY_INT_INIT (file_locators_locs, 32, "file_locators_locs");
VARRAY_CHAR_PTR_INIT (file_locators_files, 32, "file_locators_files");
for (insn = get_insns (); insn; insn = next)
{
next = NEXT_INSN (insn);
if ((active_insn_p (insn)
&& GET_CODE (PATTERN (insn)) != ADDR_VEC
&& GET_CODE (PATTERN (insn)) != ADDR_DIFF_VEC)
|| !NEXT_INSN (insn)
|| (!prologue_locator && file_name))
{
if (last_block != block)
{
loc++;
VARRAY_PUSH_INT (block_locators_locs, loc);
VARRAY_PUSH_TREE (block_locators_blocks, block);
last_block = block;
}
if (last_line_number != line_number)
{
loc++;
VARRAY_PUSH_INT (line_locators_locs, loc);
VARRAY_PUSH_INT (line_locators_lines, line_number);
last_line_number = line_number;
}
if (last_file_name != file_name)
{
loc++;
VARRAY_PUSH_INT (file_locators_locs, loc);
VARRAY_PUSH_CHAR_PTR (file_locators_files, file_name);
last_file_name = file_name;
}
}
if (!prologue_locator && file_name)
prologue_locator = loc;
if (!NEXT_INSN (insn))
epilogue_locator = loc;
if (active_insn_p (insn))
INSN_LOCATOR (insn) = loc;
else if (GET_CODE (insn) == NOTE)
{
switch (NOTE_LINE_NUMBER (insn))
{
case NOTE_INSN_BLOCK_BEG:
block = NOTE_BLOCK (insn);
delete_insn (insn);
break;
case NOTE_INSN_BLOCK_END:
block = BLOCK_SUPERCONTEXT (block);
if (block && TREE_CODE (block) == FUNCTION_DECL)
block = 0;
delete_insn (insn);
break;
default:
if (NOTE_LINE_NUMBER (insn) > 0)
{
line_number = NOTE_LINE_NUMBER (insn);
file_name = (char *)NOTE_SOURCE_FILE (insn);
}
break;
}
}
}
/* Tag the blocks with a depth number so that change_scope can find
the common parent easily. */
set_block_levels (DECL_INITIAL (cfun->decl), 0);
}
static void
doloop_modify (struct loop *loop, struct niter_desc *desc,
rtx doloop_seq, rtx condition, rtx count)
{
rtx counter_reg;
rtx tmp, noloop = NULL_RTX;
rtx sequence;
rtx jump_insn;
rtx jump_label;
int nonneg = 0, irr;
bool increment_count;
basic_block loop_end = desc->out_edge->src;
enum machine_mode mode;
jump_insn = BB_END (loop_end);
if (dump_file)
{
fprintf (dump_file, "Doloop: Inserting doloop pattern (");
if (desc->const_iter)
fprintf (dump_file, HOST_WIDEST_INT_PRINT_DEC, desc->niter);
else
fputs ("runtime", dump_file);
fputs (" iterations).\n", dump_file);
}
/* Discard original jump to continue loop. The original compare
result may still be live, so it cannot be discarded explicitly. */
delete_insn (jump_insn);
counter_reg = XEXP (condition, 0);
if (GET_CODE (counter_reg) == PLUS)
counter_reg = XEXP (counter_reg, 0);
mode = GET_MODE (counter_reg);
increment_count = false;
switch (GET_CODE (condition))
{
case NE:
/* Currently only NE tests against zero and one are supported. */
if (XEXP (condition, 1) == const1_rtx)
{
increment_count = true;
noloop = const1_rtx;
}
else if (XEXP (condition, 1) == const0_rtx)
noloop = const0_rtx;
else
abort ();
break;
case GE:
/* Currently only GE tests against zero are supported. */
if (XEXP (condition, 1) != const0_rtx)
abort ();
noloop = constm1_rtx;
/* The iteration count does not need incrementing for a GE test. */
increment_count = false;
/* Determine if the iteration counter will be non-negative.
Note that the maximum value loaded is iterations_max - 1. */
if (desc->niter_max
<= ((unsigned HOST_WIDEST_INT) 1
<< (GET_MODE_BITSIZE (mode) - 1)))
nonneg = 1;
break;
/* Abort if an invalid doloop pattern has been generated. */
default:
abort ();
}
if (increment_count)
count = simplify_gen_binary (PLUS, mode, count, const1_rtx);
/* Insert initialization of the count register into the loop header. */
start_sequence ();
tmp = force_operand (count, counter_reg);
convert_move (counter_reg, tmp, 1);
sequence = get_insns ();
end_sequence ();
emit_insn_after (sequence, BB_END (loop_preheader_edge (loop)->src));
if (desc->noloop_assumptions)
{
rtx ass = copy_rtx (desc->noloop_assumptions);
basic_block preheader = loop_preheader_edge (loop)->src;
basic_block set_zero
= loop_split_edge_with (loop_preheader_edge (loop), NULL_RTX);
basic_block new_preheader
= loop_split_edge_with (loop_preheader_edge (loop), NULL_RTX);
basic_block bb;
edge te;
gcov_type cnt;
/* Expand the condition testing the assumptions and if it does not pass,
reset the count register to 0. */
add_test (XEXP (ass, 0), preheader, set_zero);
EDGE_SUCC (preheader, 0)->flags &= ~EDGE_FALLTHRU;
cnt = EDGE_SUCC (preheader, 0)->count;
EDGE_SUCC (preheader, 0)->probability = 0;
EDGE_SUCC (preheader, 0)->count = 0;
irr = EDGE_SUCC (preheader, 0)->flags & EDGE_IRREDUCIBLE_LOOP;
te = make_edge (preheader, new_preheader, EDGE_FALLTHRU | irr);
te->probability = REG_BR_PROB_BASE;
te->count = cnt;
set_immediate_dominator (CDI_DOMINATORS, new_preheader, preheader);
set_zero->count = 0;
set_zero->frequency = 0;
for (ass = XEXP (ass, 1); ass; ass = XEXP (ass, 1))
{
bb = loop_split_edge_with (te, NULL_RTX);
te = EDGE_SUCC (bb, 0);
add_test (XEXP (ass, 0), bb, set_zero);
make_edge (bb, set_zero, irr);
}
start_sequence ();
convert_move (counter_reg, noloop, 0);
sequence = get_insns ();
end_sequence ();
emit_insn_after (sequence, BB_END (set_zero));
}
/* Some targets (eg, C4x) need to initialize special looping
registers. */
#ifdef HAVE_doloop_begin
{
rtx init;
unsigned level = get_loop_level (loop) + 1;
init = gen_doloop_begin (counter_reg,
desc->const_iter ? desc->niter_expr : const0_rtx,
desc->niter_max,
GEN_INT (level));
if (init)
{
start_sequence ();
emit_insn (init);
sequence = get_insns ();
end_sequence ();
emit_insn_after (sequence, BB_END (loop_preheader_edge (loop)->src));
}
}
#endif
/* Insert the new low-overhead looping insn. */
emit_jump_insn_after (doloop_seq, BB_END (loop_end));
jump_insn = BB_END (loop_end);
jump_label = block_label (desc->in_edge->dest);
JUMP_LABEL (jump_insn) = jump_label;
LABEL_NUSES (jump_label)++;
/* Ensure the right fallthru edge is marked, for case we have reversed
the condition. */
desc->in_edge->flags &= ~EDGE_FALLTHRU;
desc->out_edge->flags |= EDGE_FALLTHRU;
/* Add a REG_NONNEG note if the actual or estimated maximum number
of iterations is non-negative. */
if (nonneg)
{
REG_NOTES (jump_insn)
= gen_rtx_EXPR_LIST (REG_NONNEG, NULL_RTX, REG_NOTES (jump_insn));
}
}
/* Do transforms 3) and 4) on INSN if applicable. */
static bool
mod_subtract_transform (rtx insn)
{
rtx set, set_src, set_dest, op1, op2, value, histogram;
enum rtx_code code;
enum machine_mode mode;
gcov_type wrong_values, counts[2], count, all;
edge e;
int i, prob1, prob2;
set = single_set (insn);
if (!set)
return false;
set_src = SET_SRC (set);
set_dest = SET_DEST (set);
code = GET_CODE (set_src);
mode = GET_MODE (set_dest);
if (code != UMOD)
return false;
op1 = XEXP (set_src, 0);
op2 = XEXP (set_src, 1);
for (histogram = REG_NOTES (insn);
histogram;
histogram = XEXP (histogram, 1))
if (REG_NOTE_KIND (histogram) == REG_VALUE_PROFILE
&& XEXP (XEXP (histogram, 0), 0) == GEN_INT (HIST_TYPE_INTERVAL))
break;
if (!histogram)
return false;
histogram = XEXP (XEXP (histogram, 0), 1);
value = XEXP (histogram, 0);
histogram = XEXP (histogram, 1);
all = 0;
for (i = 0; i < 2; i++)
{
counts[i] = INTVAL (XEXP (histogram, 0));
all += counts[i];
histogram = XEXP (histogram, 1);
}
wrong_values = INTVAL (XEXP (histogram, 0));
histogram = XEXP (histogram, 1);
wrong_values += INTVAL (XEXP (histogram, 0));
all += wrong_values;
/* We require that we use just subtractions in at least 50% of all
evaluations. */
count = 0;
for (i = 0; i < 2; i++)
{
count += counts[i];
if (count * 2 >= all)
break;
}
if (i == 2)
return false;
if (dump_file)
fprintf (dump_file, "Mod subtract transformation on insn %d\n",
INSN_UID (insn));
/* Compute probability of taking the optimal path(s). */
prob1 = (counts[0] * REG_BR_PROB_BASE + all / 2) / all;
prob2 = (counts[1] * REG_BR_PROB_BASE + all / 2) / all;
e = split_block (BLOCK_FOR_INSN (insn), PREV_INSN (insn));
delete_insn (insn);
insert_insn_on_edge (
gen_mod_subtract (mode, code, set_dest,
op1, op2, i, prob1, prob2), e);
return true;
}
/* Do transform 2) on INSN if applicable. */
static bool
mod_pow2_value_transform (rtx insn)
{
rtx set, set_src, set_dest, op1, op2, value, histogram;
enum rtx_code code;
enum machine_mode mode;
gcov_type wrong_values, count;
edge e;
int i, all, prob;
set = single_set (insn);
if (!set)
return false;
set_src = SET_SRC (set);
set_dest = SET_DEST (set);
code = GET_CODE (set_src);
mode = GET_MODE (set_dest);
if (code != UMOD)
return false;
op1 = XEXP (set_src, 0);
op2 = XEXP (set_src, 1);
for (histogram = REG_NOTES (insn);
histogram;
histogram = XEXP (histogram, 1))
if (REG_NOTE_KIND (histogram) == REG_VALUE_PROFILE
&& XEXP (XEXP (histogram, 0), 0) == GEN_INT (HIST_TYPE_POW2))
break;
if (!histogram)
return false;
histogram = XEXP (XEXP (histogram, 0), 1);
value = XEXP (histogram, 0);
histogram = XEXP (histogram, 1);
wrong_values =INTVAL (XEXP (histogram, 0));
histogram = XEXP (histogram, 1);
count = 0;
for (i = 0; i < GET_MODE_BITSIZE (mode); i++)
{
count += INTVAL (XEXP (histogram, 0));
histogram = XEXP (histogram, 1);
}
if (!rtx_equal_p (op2, value))
return false;
/* We require that we hit a power of two at least half of all evaluations. */
if (count < wrong_values)
return false;
if (dump_file)
fprintf (dump_file, "Mod power of 2 transformation on insn %d\n",
INSN_UID (insn));
/* Compute probability of taking the optimal path. */
all = count + wrong_values;
prob = (count * REG_BR_PROB_BASE + all / 2) / all;
e = split_block (BLOCK_FOR_INSN (insn), PREV_INSN (insn));
delete_insn (insn);
insert_insn_on_edge (
gen_mod_pow2 (mode, code, set_dest, op1, op2, prob), e);
return true;
}
/* Do transform 1) on INSN if applicable. */
static bool
divmod_fixed_value_transform (rtx insn)
{
rtx set, set_src, set_dest, op1, op2, value, histogram;
enum rtx_code code;
enum machine_mode mode;
gcov_type val, count, all;
edge e;
int prob;
set = single_set (insn);
if (!set)
return false;
set_src = SET_SRC (set);
set_dest = SET_DEST (set);
code = GET_CODE (set_src);
mode = GET_MODE (set_dest);
if (code != DIV && code != MOD && code != UDIV && code != UMOD)
return false;
op1 = XEXP (set_src, false);
op2 = XEXP (set_src, 1);
for (histogram = REG_NOTES (insn);
histogram;
histogram = XEXP (histogram, 1))
if (REG_NOTE_KIND (histogram) == REG_VALUE_PROFILE
&& XEXP (XEXP (histogram, 0), 0) == GEN_INT (HIST_TYPE_SINGLE_VALUE))
break;
if (!histogram)
return false;
histogram = XEXP (XEXP (histogram, 0), 1);
value = XEXP (histogram, 0);
histogram = XEXP (histogram, 1);
val = INTVAL (XEXP (histogram, 0));
histogram = XEXP (histogram, 1);
count = INTVAL (XEXP (histogram, 0));
histogram = XEXP (histogram, 1);
all = INTVAL (XEXP (histogram, 0));
/* We require that count be at least half of all; this means
that for the transformation to fire the value must be constant
at least 50% of time (and 75% gives the guarantee of usage). */
if (!rtx_equal_p (op2, value) || 2 * count < all)
return false;
if (dump_file)
fprintf (dump_file, "Div/mod by constant transformation on insn %d\n",
INSN_UID (insn));
/* Compute probability of taking the optimal path. */
prob = (count * REG_BR_PROB_BASE + all / 2) / all;
e = split_block (BLOCK_FOR_INSN (insn), PREV_INSN (insn));
delete_insn (insn);
insert_insn_on_edge (
gen_divmod_fixed_value (mode, code, set_dest,
op1, op2, val, prob), e);
return true;
}
