static bool
nds32_expand_setmem_loop_v3m (rtx dstmem, rtx size, rtx value)
{
rtx base_reg = copy_to_mode_reg (Pmode, XEXP (dstmem, 0));
rtx need_align_bytes = gen_reg_rtx (SImode);
rtx last_2_bit = gen_reg_rtx (SImode);
rtx byte_loop_base = gen_reg_rtx (SImode);
rtx byte_loop_size = gen_reg_rtx (SImode);
rtx remain_size = gen_reg_rtx (SImode);
rtx new_base_reg;
rtx value4byte, value4doubleword;
rtx byte_mode_size;
rtx last_byte_loop_label = gen_label_rtx ();
size = force_reg (SImode, size);
value4doubleword = nds32_gen_dup_8_byte_to_double_word_value (value);
value4byte = simplify_gen_subreg (QImode, value4doubleword, DImode,
subreg_lowpart_offset (QImode, DImode));
emit_move_insn (byte_loop_size, size);
emit_move_insn (byte_loop_base, base_reg);
/* Jump to last byte loop if size is less than 16. */
emit_cmp_and_jump_insns (size, gen_int_mode (16, SImode), LE, NULL,
SImode, 1, last_byte_loop_label);
/* Make sure align to 4 byte first since v3m can't unalign access. */
emit_insn (gen_andsi3 (last_2_bit,
base_reg,
gen_int_mode (0x3, SImode)));
emit_insn (gen_subsi3 (need_align_bytes,
gen_int_mode (4, SImode),
last_2_bit));
/* Align to 4 byte. */
new_base_reg = emit_setmem_byte_loop (base_reg,
need_align_bytes,
value4byte,
true);
/* Calculate remain size. */
emit_insn (gen_subsi3 (remain_size, size, need_align_bytes));
/* Set memory word by word. */
byte_mode_size = emit_setmem_doubleword_loop (new_base_reg,
remain_size,
value4doubleword);
emit_move_insn (byte_loop_base, new_base_reg);
emit_move_insn (byte_loop_size, byte_mode_size);
emit_label (last_byte_loop_label);
/* And set memory for remain bytes. */
emit_setmem_byte_loop (byte_loop_base, byte_loop_size, value4byte, false);
return true;
}
bool
nds32_expand_strlen (rtx result, rtx str,
rtx target_char, rtx align ATTRIBUTE_UNUSED)
{
rtx base_reg, backup_base_reg;
rtx ffb_result;
rtx target_char_ptr, length;
rtx loop_label, tmp;
if (optimize_size || optimize < 3)
return false;
gcc_assert (MEM_P (str));
gcc_assert (CONST_INT_P (target_char) || REG_P (target_char));
base_reg = copy_to_mode_reg (SImode, XEXP (str, 0));
loop_label = gen_label_rtx ();
ffb_result = gen_reg_rtx (Pmode);
tmp = gen_reg_rtx (SImode);
backup_base_reg = gen_reg_rtx (SImode);
/* Emit loop version of strlen.
move $backup_base, $base
.Lloop:
lmw.bim $tmp, [$base], $tmp, 0
ffb $ffb_result, $tmp, $target_char ! is there $target_char?
beqz $ffb_result, .Lloop
add $last_char_ptr, $base, $ffb_result
sub $length, $last_char_ptr, $backup_base */
/* move $backup_base, $base */
emit_move_insn (backup_base_reg, base_reg);
/* .Lloop: */
emit_label (loop_label);
/* lmw.bim $tmp, [$base], $tmp, 0 */
emit_insn (gen_unaligned_load_update_base_w (base_reg, tmp, base_reg));
/* ffb $ffb_result, $tmp, $target_char ! is there $target_char? */
emit_insn (gen_unspec_ffb (ffb_result, tmp, target_char));
/* beqz $ffb_result, .Lloop */
emit_cmp_and_jump_insns (ffb_result, const0_rtx, EQ, NULL,
SImode, 1, loop_label);
/* add $target_char_ptr, $base, $ffb_result */
target_char_ptr = expand_binop (Pmode, add_optab, base_reg,
ffb_result, NULL_RTX, 0, OPTAB_WIDEN);
/* sub $length, $target_char_ptr, $backup_base */
length = expand_binop (Pmode, sub_optab, target_char_ptr,
backup_base_reg, NULL_RTX, 0, OPTAB_WIDEN);
emit_move_insn (result, length);
return true;
}
static rtx
emit_setmem_doubleword_loop (rtx itr, rtx size, rtx value)
{
rtx word_mode_label = gen_label_rtx ();
rtx word_mode_end_label = gen_label_rtx ();
rtx byte_mode_size = gen_reg_rtx (SImode);
rtx byte_mode_size_tmp = gen_reg_rtx (SImode);
rtx word_mode_end = gen_reg_rtx (SImode);
rtx size_for_word = gen_reg_rtx (SImode);
/* and $size_for_word, $size, #~0x7 */
size_for_word = expand_binop (SImode, and_optab, size,
gen_int_mode (~0x7, SImode),
NULL_RTX, 0, OPTAB_WIDEN);
emit_move_insn (byte_mode_size, size);
/* beqz $size_for_word, .Lbyte_mode_entry */
emit_cmp_and_jump_insns (size_for_word, const0_rtx, EQ, NULL,
SImode, 1, word_mode_end_label);
/* add $word_mode_end, $dst, $size_for_word */
word_mode_end = expand_binop (Pmode, add_optab, itr, size_for_word,
NULL_RTX, 0, OPTAB_WIDEN);
/* andi $byte_mode_size, $size, 0x7 */
byte_mode_size_tmp = expand_binop (SImode, and_optab, size, GEN_INT (0x7),
NULL_RTX, 0, OPTAB_WIDEN);
emit_move_insn (byte_mode_size, byte_mode_size_tmp);
/* .Lword_mode: */
emit_label (word_mode_label);
/* ! word-mode set loop
smw.bim $value4word, [$dst_itr], $value4word, 0
bne $word_mode_end, $dst_itr, .Lword_mode */
emit_insn (gen_unaligned_store_update_base_dw (itr,
itr,
value));
emit_cmp_and_jump_insns (word_mode_end, itr, NE, NULL,
Pmode, 1, word_mode_label);
emit_label (word_mode_end_label);
return byte_mode_size;
}
static rtx
emit_setmem_byte_loop (rtx itr, rtx size, rtx value, bool need_end)
{
rtx end = gen_reg_rtx (Pmode);
rtx byte_mode_label = gen_label_rtx ();
rtx end_label = gen_label_rtx ();
value = force_reg (QImode, value);
if (need_end)
end = expand_binop (Pmode, add_optab, itr, size,
NULL_RTX, 0, OPTAB_WIDEN);
/* beqz $byte_mode_size, .Lend
add $byte_mode_end, $dst_itr, $byte_mode_size */
emit_cmp_and_jump_insns (size, const0_rtx, EQ, NULL,
SImode, 1, end_label);
if (!need_end)
end = expand_binop (Pmode, add_optab, itr, size,
NULL_RTX, 0, OPTAB_WIDEN);
/* .Lbyte_mode: */
emit_label (byte_mode_label);
/* ! byte-mode set loop
sbi.bi $value, [$dst_itr] ,1
bne $byte_mode_end, $dst_itr, .Lbyte_mode */
nds32_emit_post_inc_load_store (value, itr, QImode, false);
emit_cmp_and_jump_insns (end, itr, NE, NULL,
Pmode, 1, byte_mode_label);
/* .Lend: */
emit_label (end_label);
if (need_end)
return end;
else
return NULL_RTX;
}
void
ubsan_expand_si_overflow_mul_check (gimple stmt)
{
rtx res, op0, op1;
tree lhs, fn, arg0, arg1;
rtx_code_label *done_label, *do_error;
rtx target = NULL_RTX;
lhs = gimple_call_lhs (stmt);
arg0 = gimple_call_arg (stmt, 0);
arg1 = gimple_call_arg (stmt, 1);
done_label = gen_label_rtx ();
do_error = gen_label_rtx ();
do_pending_stack_adjust ();
op0 = expand_normal (arg0);
op1 = expand_normal (arg1);
machine_mode mode = TYPE_MODE (TREE_TYPE (arg0));
if (lhs)
target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE);
enum insn_code icode = optab_handler (mulv4_optab, mode);
if (icode != CODE_FOR_nothing)
{
struct expand_operand ops[4];
rtx_insn *last = get_last_insn ();
res = gen_reg_rtx (mode);
create_output_operand (&ops[0], res, mode);
create_input_operand (&ops[1], op0, mode);
create_input_operand (&ops[2], op1, mode);
create_fixed_operand (&ops[3], do_error);
if (maybe_expand_insn (icode, 4, ops))
{
last = get_last_insn ();
if (profile_status_for_fn (cfun) != PROFILE_ABSENT
&& JUMP_P (last)
&& any_condjump_p (last)
&& !find_reg_note (last, REG_BR_PROB, 0))
add_int_reg_note (last, REG_BR_PROB, PROB_VERY_UNLIKELY);
emit_jump (done_label);
}
else
{
delete_insns_since (last);
icode = CODE_FOR_nothing;
}
}
if (icode == CODE_FOR_nothing)
{
struct separate_ops ops;
machine_mode hmode
= mode_for_size (GET_MODE_PRECISION (mode) / 2, MODE_INT, 1);
ops.op0 = arg0;
ops.op1 = arg1;
ops.op2 = NULL_TREE;
ops.location = gimple_location (stmt);
if (GET_MODE_2XWIDER_MODE (mode) != VOIDmode
&& targetm.scalar_mode_supported_p (GET_MODE_2XWIDER_MODE (mode)))
{
machine_mode wmode = GET_MODE_2XWIDER_MODE (mode);
ops.code = WIDEN_MULT_EXPR;
ops.type
= build_nonstandard_integer_type (GET_MODE_PRECISION (wmode), 0);
res = expand_expr_real_2 (&ops, NULL_RTX, wmode, EXPAND_NORMAL);
rtx hipart = expand_shift (RSHIFT_EXPR, wmode, res,
GET_MODE_PRECISION (mode), NULL_RTX, 0);
hipart = gen_lowpart (mode, hipart);
res = gen_lowpart (mode, res);
rtx signbit = expand_shift (RSHIFT_EXPR, mode, res,
GET_MODE_PRECISION (mode) - 1,
NULL_RTX, 0);
/* RES is low half of the double width result, HIPART
the high half. There was overflow if
HIPART is different from RES < 0 ? -1 : 0. */
emit_cmp_and_jump_insns (signbit, hipart, EQ, NULL_RTX, mode,
false, done_label, PROB_VERY_LIKELY);
}
else if (hmode != BLKmode
&& 2 * GET_MODE_PRECISION (hmode) == GET_MODE_PRECISION (mode))
{
rtx_code_label *large_op0 = gen_label_rtx ();
rtx_code_label *small_op0_large_op1 = gen_label_rtx ();
rtx_code_label *one_small_one_large = gen_label_rtx ();
rtx_code_label *both_ops_large = gen_label_rtx ();
rtx_code_label *after_hipart_neg = gen_label_rtx ();
rtx_code_label *after_lopart_neg = gen_label_rtx ();
rtx_code_label *do_overflow = gen_label_rtx ();
rtx_code_label *hipart_different = gen_label_rtx ();
unsigned int hprec = GET_MODE_PRECISION (hmode);
rtx hipart0 = expand_shift (RSHIFT_EXPR, mode, op0, hprec,
NULL_RTX, 0);
hipart0 = gen_lowpart (hmode, hipart0);
rtx lopart0 = gen_lowpart (hmode, op0);
rtx signbit0 = expand_shift (RSHIFT_EXPR, hmode, lopart0, hprec - 1,
NULL_RTX, 0);
rtx hipart1 = expand_shift (RSHIFT_EXPR, mode, op1, hprec,
NULL_RTX, 0);
hipart1 = gen_lowpart (hmode, hipart1);
rtx lopart1 = gen_lowpart (hmode, op1);
rtx signbit1 = expand_shift (RSHIFT_EXPR, hmode, lopart1, hprec - 1,
NULL_RTX, 0);
res = gen_reg_rtx (mode);
/* True if op0 resp. op1 are known to be in the range of
halfstype. */
bool op0_small_p = false;
bool op1_small_p = false;
/* True if op0 resp. op1 are known to have all zeros or all ones
in the upper half of bits, but are not known to be
op{0,1}_small_p. */
bool op0_medium_p = false;
bool op1_medium_p = false;
/* -1 if op{0,1} is known to be negative, 0 if it is known to be
nonnegative, 1 if unknown. */
int op0_sign = 1;
int op1_sign = 1;
if (TREE_CODE (arg0) == SSA_NAME)
{
wide_int arg0_min, arg0_max;
if (get_range_info (arg0, &arg0_min, &arg0_max) == VR_RANGE)
{
unsigned int mprec0 = wi::min_precision (arg0_min, SIGNED);
unsigned int mprec1 = wi::min_precision (arg0_max, SIGNED);
if (mprec0 <= hprec && mprec1 <= hprec)
op0_small_p = true;
else if (mprec0 <= hprec + 1 && mprec1 <= hprec + 1)
op0_medium_p = true;
if (!wi::neg_p (arg0_min, TYPE_SIGN (TREE_TYPE (arg0))))
op0_sign = 0;
else if (wi::neg_p (arg0_max, TYPE_SIGN (TREE_TYPE (arg0))))
op0_sign = -1;
}
}
if (TREE_CODE (arg1) == SSA_NAME)
{
wide_int arg1_min, arg1_max;
if (get_range_info (arg1, &arg1_min, &arg1_max) == VR_RANGE)
{
unsigned int mprec0 = wi::min_precision (arg1_min, SIGNED);
unsigned int mprec1 = wi::min_precision (arg1_max, SIGNED);
if (mprec0 <= hprec && mprec1 <= hprec)
op1_small_p = true;
else if (mprec0 <= hprec + 1 && mprec1 <= hprec + 1)
op1_medium_p = true;
if (!wi::neg_p (arg1_min, TYPE_SIGN (TREE_TYPE (arg1))))
op1_sign = 0;
else if (wi::neg_p (arg1_max, TYPE_SIGN (TREE_TYPE (arg1))))
op1_sign = -1;
}
}
int smaller_sign = 1;
int larger_sign = 1;
if (op0_small_p)
{
smaller_sign = op0_sign;
larger_sign = op1_sign;
}
else if (op1_small_p)
{
smaller_sign = op1_sign;
larger_sign = op0_sign;
}
else if (op0_sign == op1_sign)
{
smaller_sign = op0_sign;
larger_sign = op0_sign;
}
if (!op0_small_p)
emit_cmp_and_jump_insns (signbit0, hipart0, NE, NULL_RTX, hmode,
false, large_op0, PROB_UNLIKELY);
if (!op1_small_p)
emit_cmp_and_jump_insns (signbit1, hipart1, NE, NULL_RTX, hmode,
false, small_op0_large_op1,
PROB_UNLIKELY);
/* If both op0 and op1 are sign extended from hmode to mode,
the multiplication will never overflow. We can do just one
hmode x hmode => mode widening multiplication. */
if (GET_CODE (lopart0) == SUBREG)
{
SUBREG_PROMOTED_VAR_P (lopart0) = 1;
SUBREG_PROMOTED_SET (lopart0, 0);
}
if (GET_CODE (lopart1) == SUBREG)
{
SUBREG_PROMOTED_VAR_P (lopart1) = 1;
SUBREG_PROMOTED_SET (lopart1, 0);
}
tree halfstype = build_nonstandard_integer_type (hprec, 0);
ops.op0 = make_tree (halfstype, lopart0);
ops.op1 = make_tree (halfstype, lopart1);
ops.code = WIDEN_MULT_EXPR;
ops.type = TREE_TYPE (arg0);
rtx thisres
= expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL);
emit_move_insn (res, thisres);
emit_jump (done_label);
emit_label (small_op0_large_op1);
/* If op0 is sign extended from hmode to mode, but op1 is not,
just swap the arguments and handle it as op1 sign extended,
op0 not. */
rtx larger = gen_reg_rtx (mode);
rtx hipart = gen_reg_rtx (hmode);
rtx lopart = gen_reg_rtx (hmode);
emit_move_insn (larger, op1);
emit_move_insn (hipart, hipart1);
emit_move_insn (lopart, lopart0);
emit_jump (one_small_one_large);
emit_label (large_op0);
if (!op1_small_p)
emit_cmp_and_jump_insns (signbit1, hipart1, NE, NULL_RTX, hmode,
false, both_ops_large, PROB_UNLIKELY);
/* If op1 is sign extended from hmode to mode, but op0 is not,
prepare larger, hipart and lopart pseudos and handle it together
with small_op0_large_op1. */
emit_move_insn (larger, op0);
emit_move_insn (hipart, hipart0);
emit_move_insn (lopart, lopart1);
emit_label (one_small_one_large);
/* lopart is the low part of the operand that is sign extended
to mode, larger is the the other operand, hipart is the
high part of larger and lopart0 and lopart1 are the low parts
of both operands.
We perform lopart0 * lopart1 and lopart * hipart widening
multiplications. */
tree halfutype = build_nonstandard_integer_type (hprec, 1);
ops.op0 = make_tree (halfutype, lopart0);
ops.op1 = make_tree (halfutype, lopart1);
rtx lo0xlo1
= expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL);
ops.op0 = make_tree (halfutype, lopart);
ops.op1 = make_tree (halfutype, hipart);
rtx loxhi = gen_reg_rtx (mode);
rtx tem = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL);
emit_move_insn (loxhi, tem);
/* if (hipart < 0) loxhi -= lopart << (bitsize / 2); */
if (larger_sign == 0)
emit_jump (after_hipart_neg);
else if (larger_sign != -1)
emit_cmp_and_jump_insns (hipart, const0_rtx, GE, NULL_RTX, hmode,
false, after_hipart_neg, PROB_EVEN);
tem = convert_modes (mode, hmode, lopart, 1);
tem = expand_shift (LSHIFT_EXPR, mode, tem, hprec, NULL_RTX, 1);
tem = expand_simple_binop (mode, MINUS, loxhi, tem, NULL_RTX,
1, OPTAB_DIRECT);
emit_move_insn (loxhi, tem);
emit_label (after_hipart_neg);
/* if (lopart < 0) loxhi -= larger; */
if (smaller_sign == 0)
emit_jump (after_lopart_neg);
else if (smaller_sign != -1)
emit_cmp_and_jump_insns (lopart, const0_rtx, GE, NULL_RTX, hmode,
false, after_lopart_neg, PROB_EVEN);
tem = expand_simple_binop (mode, MINUS, loxhi, larger, NULL_RTX,
1, OPTAB_DIRECT);
emit_move_insn (loxhi, tem);
emit_label (after_lopart_neg);
/* loxhi += (uns) lo0xlo1 >> (bitsize / 2); */
tem = expand_shift (RSHIFT_EXPR, mode, lo0xlo1, hprec, NULL_RTX, 1);
tem = expand_simple_binop (mode, PLUS, loxhi, tem, NULL_RTX,
1, OPTAB_DIRECT);
emit_move_insn (loxhi, tem);
/* if (loxhi >> (bitsize / 2)
== (hmode) loxhi >> (bitsize / 2 - 1)) */
rtx hipartloxhi = expand_shift (RSHIFT_EXPR, mode, loxhi, hprec,
NULL_RTX, 0);
hipartloxhi = gen_lowpart (hmode, hipartloxhi);
rtx lopartloxhi = gen_lowpart (hmode, loxhi);
rtx signbitloxhi = expand_shift (RSHIFT_EXPR, hmode, lopartloxhi,
hprec - 1, NULL_RTX, 0);
emit_cmp_and_jump_insns (signbitloxhi, hipartloxhi, NE, NULL_RTX,
hmode, false, do_overflow,
PROB_VERY_UNLIKELY);
/* res = (loxhi << (bitsize / 2)) | (hmode) lo0xlo1; */
rtx loxhishifted = expand_shift (LSHIFT_EXPR, mode, loxhi, hprec,
NULL_RTX, 1);
tem = convert_modes (mode, hmode, gen_lowpart (hmode, lo0xlo1), 1);
tem = expand_simple_binop (mode, IOR, loxhishifted, tem, res,
1, OPTAB_DIRECT);
if (tem != res)
emit_move_insn (res, tem);
emit_jump (done_label);
emit_label (both_ops_large);
/* If both operands are large (not sign extended from hmode),
then perform the full multiplication which will be the result
of the operation. The only cases which don't overflow are
some cases where both hipart0 and highpart1 are 0 or -1. */
ops.code = MULT_EXPR;
ops.op0 = make_tree (TREE_TYPE (arg0), op0);
ops.op1 = make_tree (TREE_TYPE (arg0), op1);
tem = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL);
emit_move_insn (res, tem);
if (!op0_medium_p)
{
tem = expand_simple_binop (hmode, PLUS, hipart0, const1_rtx,
NULL_RTX, 1, OPTAB_DIRECT);
emit_cmp_and_jump_insns (tem, const1_rtx, GTU, NULL_RTX, hmode,
true, do_error, PROB_VERY_UNLIKELY);
}
if (!op1_medium_p)
{
tem = expand_simple_binop (hmode, PLUS, hipart1, const1_rtx,
NULL_RTX, 1, OPTAB_DIRECT);
emit_cmp_and_jump_insns (tem, const1_rtx, GTU, NULL_RTX, hmode,
true, do_error, PROB_VERY_UNLIKELY);
}
/* At this point hipart{0,1} are both in [-1, 0]. If they are the
same, overflow happened if res is negative, if they are different,
overflow happened if res is positive. */
if (op0_sign != 1 && op1_sign != 1 && op0_sign != op1_sign)
emit_jump (hipart_different);
else if (op0_sign == 1 || op1_sign == 1)
emit_cmp_and_jump_insns (hipart0, hipart1, NE, NULL_RTX, hmode,
true, hipart_different, PROB_EVEN);
emit_cmp_and_jump_insns (res, const0_rtx, LT, NULL_RTX, mode, false,
do_error, PROB_VERY_UNLIKELY);
emit_jump (done_label);
emit_label (hipart_different);
emit_cmp_and_jump_insns (res, const0_rtx, GE, NULL_RTX, mode, false,
do_error, PROB_VERY_UNLIKELY);
emit_jump (done_label);
emit_label (do_overflow);
/* Overflow, do full multiplication and fallthru into do_error. */
ops.op0 = make_tree (TREE_TYPE (arg0), op0);
ops.op1 = make_tree (TREE_TYPE (arg0), op1);
tem = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL);
emit_move_insn (res, tem);
}
else
{
ops.code = MULT_EXPR;
ops.type = TREE_TYPE (arg0);
res = expand_expr_real_2 (&ops, NULL_RTX, mode, EXPAND_NORMAL);
emit_jump (done_label);
}
}
emit_label (do_error);
/* Expand the ubsan builtin call. */
push_temp_slots ();
fn = ubsan_build_overflow_builtin (MULT_EXPR, gimple_location (stmt),
TREE_TYPE (arg0), arg0, arg1);
expand_normal (fn);
pop_temp_slots ();
do_pending_stack_adjust ();
/* We're done. */
emit_label (done_label);
if (lhs)
emit_move_insn (target, res);
}
void
ubsan_expand_si_overflow_neg_check (gimple stmt)
{
rtx res, op1;
tree lhs, fn, arg1;
rtx_code_label *done_label, *do_error;
rtx target = NULL_RTX;
lhs = gimple_call_lhs (stmt);
arg1 = gimple_call_arg (stmt, 1);
done_label = gen_label_rtx ();
do_error = gen_label_rtx ();
do_pending_stack_adjust ();
op1 = expand_normal (arg1);
machine_mode mode = TYPE_MODE (TREE_TYPE (arg1));
if (lhs)
target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE);
enum insn_code icode = optab_handler (negv3_optab, mode);
if (icode != CODE_FOR_nothing)
{
struct expand_operand ops[3];
rtx_insn *last = get_last_insn ();
res = gen_reg_rtx (mode);
create_output_operand (&ops[0], res, mode);
create_input_operand (&ops[1], op1, mode);
create_fixed_operand (&ops[2], do_error);
if (maybe_expand_insn (icode, 3, ops))
{
last = get_last_insn ();
if (profile_status_for_fn (cfun) != PROFILE_ABSENT
&& JUMP_P (last)
&& any_condjump_p (last)
&& !find_reg_note (last, REG_BR_PROB, 0))
add_int_reg_note (last, REG_BR_PROB, PROB_VERY_UNLIKELY);
emit_jump (done_label);
}
else
{
delete_insns_since (last);
icode = CODE_FOR_nothing;
}
}
if (icode == CODE_FOR_nothing)
{
/* Compute the operation. On RTL level, the addition is always
unsigned. */
res = expand_unop (mode, neg_optab, op1, NULL_RTX, false);
/* Compare the operand with the most negative value. */
rtx minv = expand_normal (TYPE_MIN_VALUE (TREE_TYPE (arg1)));
emit_cmp_and_jump_insns (op1, minv, NE, NULL_RTX, mode, false,
done_label, PROB_VERY_LIKELY);
}
emit_label (do_error);
/* Expand the ubsan builtin call. */
push_temp_slots ();
fn = ubsan_build_overflow_builtin (NEGATE_EXPR, gimple_location (stmt),
TREE_TYPE (arg1), arg1, NULL_TREE);
expand_normal (fn);
pop_temp_slots ();
do_pending_stack_adjust ();
/* We're done. */
emit_label (done_label);
if (lhs)
emit_move_insn (target, res);
}
void
ubsan_expand_si_overflow_addsub_check (tree_code code, gimple stmt)
{
rtx res, op0, op1;
tree lhs, fn, arg0, arg1;
rtx_code_label *done_label, *do_error;
rtx target = NULL_RTX;
lhs = gimple_call_lhs (stmt);
arg0 = gimple_call_arg (stmt, 0);
arg1 = gimple_call_arg (stmt, 1);
done_label = gen_label_rtx ();
do_error = gen_label_rtx ();
do_pending_stack_adjust ();
op0 = expand_normal (arg0);
op1 = expand_normal (arg1);
machine_mode mode = TYPE_MODE (TREE_TYPE (arg0));
if (lhs)
target = expand_expr (lhs, NULL_RTX, VOIDmode, EXPAND_WRITE);
enum insn_code icode
= optab_handler (code == PLUS_EXPR ? addv4_optab : subv4_optab, mode);
if (icode != CODE_FOR_nothing)
{
struct expand_operand ops[4];
rtx_insn *last = get_last_insn ();
res = gen_reg_rtx (mode);
create_output_operand (&ops[0], res, mode);
create_input_operand (&ops[1], op0, mode);
create_input_operand (&ops[2], op1, mode);
create_fixed_operand (&ops[3], do_error);
if (maybe_expand_insn (icode, 4, ops))
{
last = get_last_insn ();
if (profile_status_for_fn (cfun) != PROFILE_ABSENT
&& JUMP_P (last)
&& any_condjump_p (last)
&& !find_reg_note (last, REG_BR_PROB, 0))
add_int_reg_note (last, REG_BR_PROB, PROB_VERY_UNLIKELY);
emit_jump (done_label);
}
else
{
delete_insns_since (last);
icode = CODE_FOR_nothing;
}
}
if (icode == CODE_FOR_nothing)
{
rtx_code_label *sub_check = gen_label_rtx ();
int pos_neg = 3;
/* Compute the operation. On RTL level, the addition is always
unsigned. */
res = expand_binop (mode, code == PLUS_EXPR ? add_optab : sub_optab,
op0, op1, NULL_RTX, false, OPTAB_LIB_WIDEN);
/* If we can prove one of the arguments (for MINUS_EXPR only
the second operand, as subtraction is not commutative) is always
non-negative or always negative, we can do just one comparison
and conditional jump instead of 2 at runtime, 3 present in the
emitted code. If one of the arguments is CONST_INT, all we
need is to make sure it is op1, then the first
emit_cmp_and_jump_insns will be just folded. Otherwise try
to use range info if available. */
if (code == PLUS_EXPR && CONST_INT_P (op0))
{
rtx tem = op0;
op0 = op1;
op1 = tem;
}
else if (CONST_INT_P (op1))
;
else if (code == PLUS_EXPR && TREE_CODE (arg0) == SSA_NAME)
{
wide_int arg0_min, arg0_max;
if (get_range_info (arg0, &arg0_min, &arg0_max) == VR_RANGE)
{
if (!wi::neg_p (arg0_min, TYPE_SIGN (TREE_TYPE (arg0))))
pos_neg = 1;
else if (wi::neg_p (arg0_max, TYPE_SIGN (TREE_TYPE (arg0))))
pos_neg = 2;
}
if (pos_neg != 3)
{
rtx tem = op0;
op0 = op1;
op1 = tem;
}
}
if (pos_neg == 3 && !CONST_INT_P (op1) && TREE_CODE (arg1) == SSA_NAME)
{
wide_int arg1_min, arg1_max;
if (get_range_info (arg1, &arg1_min, &arg1_max) == VR_RANGE)
{
if (!wi::neg_p (arg1_min, TYPE_SIGN (TREE_TYPE (arg1))))
pos_neg = 1;
else if (wi::neg_p (arg1_max, TYPE_SIGN (TREE_TYPE (arg1))))
pos_neg = 2;
}
}
/* If the op1 is negative, we have to use a different check. */
if (pos_neg == 3)
emit_cmp_and_jump_insns (op1, const0_rtx, LT, NULL_RTX, mode,
false, sub_check, PROB_EVEN);
/* Compare the result of the operation with one of the operands. */
if (pos_neg & 1)
emit_cmp_and_jump_insns (res, op0, code == PLUS_EXPR ? GE : LE,
NULL_RTX, mode, false, done_label,
PROB_VERY_LIKELY);
/* If we get here, we have to print the error. */
if (pos_neg == 3)
{
emit_jump (do_error);
emit_label (sub_check);
}
/* We have k = a + b for b < 0 here. k <= a must hold. */
if (pos_neg & 2)
emit_cmp_and_jump_insns (res, op0, code == PLUS_EXPR ? LE : GE,
NULL_RTX, mode, false, done_label,
PROB_VERY_LIKELY);
}
emit_label (do_error);
/* Expand the ubsan builtin call. */
push_temp_slots ();
fn = ubsan_build_overflow_builtin (code, gimple_location (stmt),
TREE_TYPE (arg0), arg0, arg1);
expand_normal (fn);
pop_temp_slots ();
do_pending_stack_adjust ();
/* We're done. */
emit_label (done_label);
if (lhs)
emit_move_insn (target, res);
}
void
do_compare_rtx_and_jump (rtx op0, rtx op1, enum rtx_code code, int unsignedp,
enum machine_mode mode, rtx size, rtx if_false_label,
rtx if_true_label, int prob)
{
rtx tem;
rtx dummy_label = NULL_RTX;
rtx last;
/* Reverse the comparison if that is safe and we want to jump if it is
false. Also convert to the reverse comparison if the target can
implement it. */
if ((! if_true_label
|| ! can_compare_p (code, mode, ccp_jump))
&& (! FLOAT_MODE_P (mode)
|| code == ORDERED || code == UNORDERED
|| (! HONOR_NANS (mode) && (code == LTGT || code == UNEQ))
|| (! HONOR_SNANS (mode) && (code == EQ || code == NE))))
{
enum rtx_code rcode;
if (FLOAT_MODE_P (mode))
rcode = reverse_condition_maybe_unordered (code);
else
rcode = reverse_condition (code);
/* Canonicalize to UNORDERED for the libcall. */
if (can_compare_p (rcode, mode, ccp_jump)
|| (code == ORDERED && ! can_compare_p (ORDERED, mode, ccp_jump)))
{
tem = if_true_label;
if_true_label = if_false_label;
if_false_label = tem;
code = rcode;
prob = inv (prob);
}
}
/* If one operand is constant, make it the second one. Only do this
if the other operand is not constant as well. */
if (swap_commutative_operands_p (op0, op1))
{
tem = op0;
op0 = op1;
op1 = tem;
code = swap_condition (code);
}
do_pending_stack_adjust ();
code = unsignedp ? unsigned_condition (code) : code;
if (0 != (tem = simplify_relational_operation (code, mode, VOIDmode,
op0, op1)))
{
if (CONSTANT_P (tem))
{
rtx label = (tem == const0_rtx || tem == CONST0_RTX (mode))
? if_false_label : if_true_label;
if (label)
emit_jump (label);
return;
}
code = GET_CODE (tem);
mode = GET_MODE (tem);
op0 = XEXP (tem, 0);
op1 = XEXP (tem, 1);
unsignedp = (code == GTU || code == LTU || code == GEU || code == LEU);
}
if (! if_true_label)
dummy_label = if_true_label = gen_label_rtx ();
if (GET_MODE_CLASS (mode) == MODE_INT
&& ! can_compare_p (code, mode, ccp_jump))
{
switch (code)
{
case LTU:
do_jump_by_parts_greater_rtx (mode, 1, op1, op0,
if_false_label, if_true_label, prob);
break;
case LEU:
do_jump_by_parts_greater_rtx (mode, 1, op0, op1,
if_true_label, if_false_label,
inv (prob));
break;
case GTU:
do_jump_by_parts_greater_rtx (mode, 1, op0, op1,
if_false_label, if_true_label, prob);
break;
case GEU:
do_jump_by_parts_greater_rtx (mode, 1, op1, op0,
if_true_label, if_false_label,
inv (prob));
break;
case LT:
do_jump_by_parts_greater_rtx (mode, 0, op1, op0,
if_false_label, if_true_label, prob);
break;
case LE:
do_jump_by_parts_greater_rtx (mode, 0, op0, op1,
if_true_label, if_false_label,
inv (prob));
break;
case GT:
do_jump_by_parts_greater_rtx (mode, 0, op0, op1,
if_false_label, if_true_label, prob);
break;
case GE:
do_jump_by_parts_greater_rtx (mode, 0, op1, op0,
if_true_label, if_false_label,
inv (prob));
break;
case EQ:
do_jump_by_parts_equality_rtx (mode, op0, op1, if_false_label,
if_true_label, prob);
break;
case NE:
do_jump_by_parts_equality_rtx (mode, op0, op1, if_true_label,
if_false_label, inv (prob));
break;
default:
gcc_unreachable ();
}
}
else
{
if (GET_MODE_CLASS (mode) == MODE_FLOAT
&& ! can_compare_p (code, mode, ccp_jump)
&& can_compare_p (swap_condition (code), mode, ccp_jump))
{
rtx tmp;
code = swap_condition (code);
tmp = op0;
op0 = op1;
op1 = tmp;
}
else if (GET_MODE_CLASS (mode) == MODE_FLOAT
&& ! can_compare_p (code, mode, ccp_jump)
/* Never split ORDERED and UNORDERED. These must be implemented. */
&& (code != ORDERED && code != UNORDERED)
/* Split a floating-point comparison if we can jump on other
conditions... */
&& (have_insn_for (COMPARE, mode)
/* ... or if there is no libcall for it. */
|| code_to_optab[code] == NULL))
{
enum rtx_code first_code;
bool and_them = split_comparison (code, mode, &first_code, &code);
/* If there are no NaNs, the first comparison should always fall
through. */
if (!HONOR_NANS (mode))
gcc_assert (first_code == (and_them ? ORDERED : UNORDERED));
else
{
if (and_them)
{
rtx dest_label;
/* If we only jump if true, just bypass the second jump. */
if (! if_false_label)
{
if (! dummy_label)
dummy_label = gen_label_rtx ();
dest_label = dummy_label;
}
else
dest_label = if_false_label;
do_compare_rtx_and_jump (op0, op1, first_code, unsignedp, mode,
size, dest_label, NULL_RTX, prob);
}
else
do_compare_rtx_and_jump (op0, op1, first_code, unsignedp, mode,
size, NULL_RTX, if_true_label, prob);
}
}
last = get_last_insn ();
emit_cmp_and_jump_insns (op0, op1, code, size, mode, unsignedp,
if_true_label);
if (prob != -1 && profile_status != PROFILE_ABSENT)
{
for (last = NEXT_INSN (last);
last && NEXT_INSN (last);
last = NEXT_INSN (last))
if (JUMP_P (last))
break;
if (!last
|| !JUMP_P (last)
|| NEXT_INSN (last)
|| !any_condjump_p (last))
{
if (dump_file)
fprintf (dump_file, "Failed to add probability note\n");
}
else
{
gcc_assert (!find_reg_note (last, REG_BR_PROB, 0));
add_reg_note (last, REG_BR_PROB, GEN_INT (prob));
}
}
}
if (if_false_label)
emit_jump (if_false_label);
if (dummy_label)
emit_label (dummy_label);
}
void
do_compare_rtx_and_jump (rtx op0, rtx op1, enum rtx_code code, int unsignedp,
enum machine_mode mode, rtx size, rtx if_false_label,
rtx if_true_label)
{
rtx tem;
int dummy_true_label = 0;
/* Reverse the comparison if that is safe and we want to jump if it is
false. */
if (! if_true_label && ! FLOAT_MODE_P (mode))
{
if_true_label = if_false_label;
if_false_label = 0;
code = reverse_condition (code);
}
/* If one operand is constant, make it the second one. Only do this
if the other operand is not constant as well. */
if (swap_commutative_operands_p (op0, op1))
{
tem = op0;
op0 = op1;
op1 = tem;
code = swap_condition (code);
}
do_pending_stack_adjust ();
code = unsignedp ? unsigned_condition (code) : code;
if (0 != (tem = simplify_relational_operation (code, mode, VOIDmode,
op0, op1)))
{
if (CONSTANT_P (tem))
{
rtx label = (tem == const0_rtx || tem == CONST0_RTX (mode))
? if_false_label : if_true_label;
if (label)
emit_jump (label);
return;
}
code = GET_CODE (tem);
mode = GET_MODE (tem);
op0 = XEXP (tem, 0);
op1 = XEXP (tem, 1);
unsignedp = (code == GTU || code == LTU || code == GEU || code == LEU);
}
if (! if_true_label)
{
dummy_true_label = 1;
if_true_label = gen_label_rtx ();
}
if (GET_MODE_CLASS (mode) == MODE_INT
&& ! can_compare_p (code, mode, ccp_jump))
{
switch (code)
{
case LTU:
do_jump_by_parts_greater_rtx (mode, 1, op1, op0,
if_false_label, if_true_label);
break;
case LEU:
do_jump_by_parts_greater_rtx (mode, 1, op0, op1,
if_true_label, if_false_label);
break;
case GTU:
do_jump_by_parts_greater_rtx (mode, 1, op0, op1,
if_false_label, if_true_label);
break;
case GEU:
do_jump_by_parts_greater_rtx (mode, 1, op1, op0,
if_true_label, if_false_label);
break;
case LT:
do_jump_by_parts_greater_rtx (mode, 0, op1, op0,
if_false_label, if_true_label);
break;
case LE:
do_jump_by_parts_greater_rtx (mode, 0, op0, op1,
if_true_label, if_false_label);
break;
case GT:
do_jump_by_parts_greater_rtx (mode, 0, op0, op1,
if_false_label, if_true_label);
break;
case GE:
do_jump_by_parts_greater_rtx (mode, 0, op1, op0,
if_true_label, if_false_label);
break;
case EQ:
do_jump_by_parts_equality_rtx (mode, op0, op1, if_false_label,
if_true_label);
break;
case NE:
do_jump_by_parts_equality_rtx (mode, op0, op1, if_true_label,
if_false_label);
break;
default:
gcc_unreachable ();
}
}
else
emit_cmp_and_jump_insns (op0, op1, code, size, mode, unsignedp,
if_true_label);
if (if_false_label)
emit_jump (if_false_label);
if (dummy_true_label)
emit_label (if_true_label);
}
void
do_compare_rtx_and_jump (rtx op0, rtx op1, enum rtx_code code, int unsignedp,
enum machine_mode mode, rtx size, rtx if_false_label,
rtx if_true_label)
{
enum rtx_code ucode;
rtx tem;
int dummy_true_label = 0;
/* Reverse the comparison if that is safe and we want to jump if it is
false. */
if (! if_true_label && ! FLOAT_MODE_P (mode))
{
if_true_label = if_false_label;
if_false_label = 0;
code = reverse_condition (code);
}
/* If one operand is constant, make it the second one. Only do this
if the other operand is not constant as well. */
if (swap_commutative_operands_p (op0, op1))
{
tem = op0;
op0 = op1;
op1 = tem;
code = swap_condition (code);
}
if (flag_force_mem)
{
op0 = force_not_mem (op0);
op1 = force_not_mem (op1);
}
do_pending_stack_adjust ();
ucode = unsignedp ? unsigned_condition (code) : code;
if ((tem = simplify_relational_operation (ucode, mode, op0, op1)) != 0)
{
if (tem == const_true_rtx)
{
if (if_true_label)
emit_jump (if_true_label);
}
else
{
if (if_false_label)
emit_jump (if_false_label);
}
return;
}
#if 0
/* There's no need to do this now that combine.c can eliminate lots of
sign extensions. This can be less efficient in certain cases on other
machines. */
/* If this is a signed equality comparison, we can do it as an
unsigned comparison since zero-extension is cheaper than sign
extension and comparisons with zero are done as unsigned. This is
the case even on machines that can do fast sign extension, since
zero-extension is easier to combine with other operations than
sign-extension is. If we are comparing against a constant, we must
convert it to what it would look like unsigned. */
if ((code == EQ || code == NE) && ! unsignedp
&& GET_MODE_BITSIZE (GET_MODE (op0)) <= HOST_BITS_PER_WIDE_INT)
{
if (GET_CODE (op1) == CONST_INT
&& (INTVAL (op1) & GET_MODE_MASK (GET_MODE (op0))) != INTVAL (op1))
op1 = GEN_INT (INTVAL (op1) & GET_MODE_MASK (GET_MODE (op0)));
unsignedp = 1;
}
#endif
if (! if_true_label)
{
dummy_true_label = 1;
if_true_label = gen_label_rtx ();
}
emit_cmp_and_jump_insns (op0, op1, code, size, mode, unsignedp,
if_true_label);
if (if_false_label)
emit_jump (if_false_label);
if (dummy_true_label)
emit_label (if_true_label);
}
void
do_compare_rtx_and_jump (rtx op0, rtx op1, enum rtx_code code, int unsignedp,
machine_mode mode, rtx size,
rtx_code_label *if_false_label,
rtx_code_label *if_true_label, int prob)
{
rtx tem;
rtx_code_label *dummy_label = NULL;
/* Reverse the comparison if that is safe and we want to jump if it is
false. Also convert to the reverse comparison if the target can
implement it. */
if ((! if_true_label
|| ! can_compare_p (code, mode, ccp_jump))
&& (! FLOAT_MODE_P (mode)
|| code == ORDERED || code == UNORDERED
|| (! HONOR_NANS (mode) && (code == LTGT || code == UNEQ))
|| (! HONOR_SNANS (mode) && (code == EQ || code == NE))))
{
enum rtx_code rcode;
if (FLOAT_MODE_P (mode))
rcode = reverse_condition_maybe_unordered (code);
else
rcode = reverse_condition (code);
/* Canonicalize to UNORDERED for the libcall. */
if (can_compare_p (rcode, mode, ccp_jump)
|| (code == ORDERED && ! can_compare_p (ORDERED, mode, ccp_jump)))
{
std::swap (if_true_label, if_false_label);
code = rcode;
prob = inv (prob);
}
}
/* If one operand is constant, make it the second one. Only do this
if the other operand is not constant as well. */
if (swap_commutative_operands_p (op0, op1))
{
std::swap (op0, op1);
code = swap_condition (code);
}
do_pending_stack_adjust ();
code = unsignedp ? unsigned_condition (code) : code;
if (0 != (tem = simplify_relational_operation (code, mode, VOIDmode,
op0, op1)))
{
if (CONSTANT_P (tem))
{
rtx_code_label *label = (tem == const0_rtx
|| tem == CONST0_RTX (mode))
? if_false_label : if_true_label;
if (label)
emit_jump (label);
return;
}
code = GET_CODE (tem);
mode = GET_MODE (tem);
op0 = XEXP (tem, 0);
op1 = XEXP (tem, 1);
unsignedp = (code == GTU || code == LTU || code == GEU || code == LEU);
}
if (! if_true_label)
dummy_label = if_true_label = gen_label_rtx ();
if (GET_MODE_CLASS (mode) == MODE_INT
&& ! can_compare_p (code, mode, ccp_jump))
{
switch (code)
{
case LTU:
do_jump_by_parts_greater_rtx (mode, 1, op1, op0,
if_false_label, if_true_label, prob);
break;
case LEU:
do_jump_by_parts_greater_rtx (mode, 1, op0, op1,
if_true_label, if_false_label,
inv (prob));
break;
case GTU:
do_jump_by_parts_greater_rtx (mode, 1, op0, op1,
if_false_label, if_true_label, prob);
break;
case GEU:
do_jump_by_parts_greater_rtx (mode, 1, op1, op0,
if_true_label, if_false_label,
inv (prob));
break;
case LT:
do_jump_by_parts_greater_rtx (mode, 0, op1, op0,
if_false_label, if_true_label, prob);
break;
case LE:
do_jump_by_parts_greater_rtx (mode, 0, op0, op1,
if_true_label, if_false_label,
inv (prob));
break;
case GT:
do_jump_by_parts_greater_rtx (mode, 0, op0, op1,
if_false_label, if_true_label, prob);
break;
case GE:
do_jump_by_parts_greater_rtx (mode, 0, op1, op0,
if_true_label, if_false_label,
inv (prob));
break;
case EQ:
do_jump_by_parts_equality_rtx (mode, op0, op1, if_false_label,
if_true_label, prob);
break;
case NE:
do_jump_by_parts_equality_rtx (mode, op0, op1, if_true_label,
if_false_label, inv (prob));
break;
default:
gcc_unreachable ();
}
}
else
{
if (SCALAR_FLOAT_MODE_P (mode)
&& ! can_compare_p (code, mode, ccp_jump)
&& can_compare_p (swap_condition (code), mode, ccp_jump))
{
code = swap_condition (code);
std::swap (op0, op1);
}
else if (SCALAR_FLOAT_MODE_P (mode)
&& ! can_compare_p (code, mode, ccp_jump)
/* Never split ORDERED and UNORDERED.
These must be implemented. */
&& (code != ORDERED && code != UNORDERED)
/* Split a floating-point comparison if
we can jump on other conditions... */
&& (have_insn_for (COMPARE, mode)
/* ... or if there is no libcall for it. */
|| code_to_optab (code) == unknown_optab))
{
enum rtx_code first_code;
bool and_them = split_comparison (code, mode, &first_code, &code);
/* If there are no NaNs, the first comparison should always fall
through. */
if (!HONOR_NANS (mode))
gcc_assert (first_code == (and_them ? ORDERED : UNORDERED));
else
{
int first_prob = prob;
if (first_code == UNORDERED)
first_prob = REG_BR_PROB_BASE / 100;
else if (first_code == ORDERED)
first_prob = REG_BR_PROB_BASE - REG_BR_PROB_BASE / 100;
if (and_them)
{
rtx_code_label *dest_label;
/* If we only jump if true, just bypass the second jump. */
if (! if_false_label)
{
if (! dummy_label)
dummy_label = gen_label_rtx ();
dest_label = dummy_label;
}
else
dest_label = if_false_label;
do_compare_rtx_and_jump (op0, op1, first_code, unsignedp, mode,
size, dest_label, NULL, first_prob);
}
else
do_compare_rtx_and_jump (op0, op1, first_code, unsignedp, mode,
size, NULL, if_true_label, first_prob);
}
}
emit_cmp_and_jump_insns (op0, op1, code, size, mode, unsignedp,
if_true_label, prob);
}
if (if_false_label)
emit_jump (if_false_label);
if (dummy_label)
emit_label (dummy_label);
}
void
do_compare_rtx_and_jump (rtx op0, rtx op1, enum rtx_code code, int unsignedp,
enum machine_mode mode, rtx size, rtx if_false_label,
rtx if_true_label)
{
rtx tem;
int dummy_true_label = 0;
/* Reverse the comparison if that is safe and we want to jump if it is
false. */
if (! if_true_label && ! FLOAT_MODE_P (mode))
{
if_true_label = if_false_label;
if_false_label = 0;
code = reverse_condition (code);
}
/* If one operand is constant, make it the second one. Only do this
if the other operand is not constant as well. */
if (swap_commutative_operands_p (op0, op1))
{
tem = op0;
op0 = op1;
op1 = tem;
code = swap_condition (code);
}
do_pending_stack_adjust ();
code = unsignedp ? unsigned_condition (code) : code;
if (0 != (tem = simplify_relational_operation (code, mode, VOIDmode,
op0, op1)))
{
if (CONSTANT_P (tem))
{
rtx label = (tem == const0_rtx || tem == CONST0_RTX (mode))
? if_false_label : if_true_label;
if (label)
emit_jump (label);
return;
}
code = GET_CODE (tem);
mode = GET_MODE (tem);
op0 = XEXP (tem, 0);
op1 = XEXP (tem, 1);
unsignedp = (code == GTU || code == LTU || code == GEU || code == LEU);
}
if (! if_true_label)
{
dummy_true_label = 1;
if_true_label = gen_label_rtx ();
}
emit_cmp_and_jump_insns (op0, op1, code, size, mode, unsignedp,
if_true_label);
if (if_false_label)
emit_jump (if_false_label);
if (dummy_true_label)
emit_label (if_true_label);
}
static bool
nds32_expand_movmemsi_loop_known_size (rtx dstmem, rtx srcmem,
rtx size, rtx alignment)
{
rtx dst_base_reg, src_base_reg;
rtx dst_itr, src_itr;
rtx dstmem_m, srcmem_m, dst_itr_m, src_itr_m;
rtx dst_end;
rtx double_word_mode_loop, byte_mode_loop;
rtx tmp;
int start_regno;
bool align_to_4_bytes = (INTVAL (alignment) & 3) == 0;
unsigned HOST_WIDE_INT total_bytes = UINTVAL (size);
if (TARGET_ISA_V3M && !align_to_4_bytes)
return 0;
if (TARGET_REDUCED_REGS)
start_regno = 2;
else
start_regno = 16;
dst_itr = gen_reg_rtx (Pmode);
src_itr = gen_reg_rtx (Pmode);
dst_end = gen_reg_rtx (Pmode);
tmp = gen_reg_rtx (QImode);
double_word_mode_loop = gen_label_rtx ();
byte_mode_loop = gen_label_rtx ();
dst_base_reg = copy_to_mode_reg (Pmode, XEXP (dstmem, 0));
src_base_reg = copy_to_mode_reg (Pmode, XEXP (srcmem, 0));
if (total_bytes < 8)
{
/* Emit total_bytes less than 8 loop version of movmem.
add $dst_end, $dst, $size
move $dst_itr, $dst
.Lbyte_mode_loop:
lbi.bi $tmp, [$src_itr], #1
sbi.bi $tmp, [$dst_itr], #1
! Not readch upper bound. Loop.
bne $dst_itr, $dst_end, .Lbyte_mode_loop */
/* add $dst_end, $dst, $size */
dst_end = expand_binop (Pmode, add_optab, dst_base_reg, size,
NULL_RTX, 0, OPTAB_WIDEN);
/* move $dst_itr, $dst
move $src_itr, $src */
emit_move_insn (dst_itr, dst_base_reg);
emit_move_insn (src_itr, src_base_reg);
/* .Lbyte_mode_loop: */
emit_label (byte_mode_loop);
/* lbi.bi $tmp, [$src_itr], #1 */
nds32_emit_post_inc_load_store (tmp, src_itr, QImode, true);
/* sbi.bi $tmp, [$dst_itr], #1 */
nds32_emit_post_inc_load_store (tmp, dst_itr, QImode, false);
/* ! Not readch upper bound. Loop.
bne $dst_itr, $dst_end, .Lbyte_mode_loop */
emit_cmp_and_jump_insns (dst_itr, dst_end, NE, NULL,
SImode, 1, byte_mode_loop);
return true;
}
else if (total_bytes % 8 == 0)
{
/* Emit multiple of 8 loop version of movmem.
add $dst_end, $dst, $size
move $dst_itr, $dst
move $src_itr, $src
.Ldouble_word_mode_loop:
lmw.bim $tmp-begin, [$src_itr], $tmp-end, #0 ! $src_itr' = $src_itr
smw.bim $tmp-begin, [$dst_itr], $tmp-end, #0 ! $dst_itr' = $dst_itr
! move will delete after register allocation
move $src_itr, $src_itr'
move $dst_itr, $dst_itr'
! Not readch upper bound. Loop.
bne $double_word_end, $dst_itr, .Ldouble_word_mode_loop */
/* add $dst_end, $dst, $size */
dst_end = expand_binop (Pmode, add_optab, dst_base_reg, size,
NULL_RTX, 0, OPTAB_WIDEN);
/* move $dst_itr, $dst
move $src_itr, $src */
emit_move_insn (dst_itr, dst_base_reg);
emit_move_insn (src_itr, src_base_reg);
/* .Ldouble_word_mode_loop: */
emit_label (double_word_mode_loop);
/* lmw.bim $tmp-begin, [$src_itr], $tmp-end, #0 ! $src_itr' = $src_itr
smw.bim $tmp-begin, [$dst_itr], $tmp-end, #0 ! $dst_itr' = $dst_itr */
src_itr_m = src_itr;
dst_itr_m = dst_itr;
srcmem_m = srcmem;
dstmem_m = dstmem;
nds32_emit_mem_move_block (start_regno, 2,
&dst_itr_m, &dstmem_m,
&src_itr_m, &srcmem_m,
true);
/* move $src_itr, $src_itr'
move $dst_itr, $dst_itr' */
emit_move_insn (dst_itr, dst_itr_m);
emit_move_insn (src_itr, src_itr_m);
/* ! Not readch upper bound. Loop.
bne $double_word_end, $dst_itr, .Ldouble_word_mode_loop */
emit_cmp_and_jump_insns (dst_end, dst_itr, NE, NULL,
Pmode, 1, double_word_mode_loop);
}
else
{
/* Handle size greater than 8, and not a multiple of 8. */
return nds32_expand_movmemsi_loop_unknown_size (dstmem, srcmem,
size, alignment);
}
return true;
}
static bool
nds32_expand_movmemsi_loop_unknown_size (rtx dstmem, rtx srcmem,
rtx size,
rtx alignment)
{
/* Emit loop version of movmem.
andi $size_least_3_bit, $size, #~7
add $dst_end, $dst, $size
move $dst_itr, $dst
move $src_itr, $src
beqz $size_least_3_bit, .Lbyte_mode_entry ! Not large enough.
add $double_word_end, $dst, $size_least_3_bit
.Ldouble_word_mode_loop:
lmw.bim $tmp-begin, [$src_itr], $tmp-end, #0 ! $src_itr' = $src_itr
smw.bim $tmp-begin, [$dst_itr], $tmp-end, #0 ! $dst_itr' = $dst_itr
! move will delete after register allocation
move $src_itr, $src_itr'
move $dst_itr, $dst_itr'
! Not readch upper bound. Loop.
bne $double_word_end, $dst_itr, .Ldouble_word_mode_loop
.Lbyte_mode_entry:
beq $dst_itr, $dst_end, .Lend_label
.Lbyte_mode_loop:
lbi.bi $tmp, [$src_itr], #1
sbi.bi $tmp, [$dst_itr], #1
! Not readch upper bound. Loop.
bne $dst_itr, $dst_end, .Lbyte_mode_loop
.Lend_label:
*/
rtx dst_base_reg, src_base_reg;
rtx dst_itr, src_itr;
rtx dstmem_m, srcmem_m, dst_itr_m, src_itr_m;
rtx dst_end;
rtx size_least_3_bit;
rtx double_word_end;
rtx double_word_mode_loop, byte_mode_entry, byte_mode_loop, end_label;
rtx tmp;
rtx mask_least_3_bit;
int start_regno;
bool align_to_4_bytes = (INTVAL (alignment) & 3) == 0;
if (TARGET_ISA_V3M && !align_to_4_bytes)
return 0;
if (TARGET_REDUCED_REGS)
start_regno = 2;
else
start_regno = 16;
dst_itr = gen_reg_rtx (Pmode);
src_itr = gen_reg_rtx (Pmode);
dst_end = gen_reg_rtx (Pmode);
tmp = gen_reg_rtx (QImode);
mask_least_3_bit = GEN_INT (~7);
double_word_mode_loop = gen_label_rtx ();
byte_mode_entry = gen_label_rtx ();
byte_mode_loop = gen_label_rtx ();
end_label = gen_label_rtx ();
dst_base_reg = copy_to_mode_reg (Pmode, XEXP (dstmem, 0));
src_base_reg = copy_to_mode_reg (Pmode, XEXP (srcmem, 0));
/* andi $size_least_3_bit, $size, #~7 */
size_least_3_bit = expand_binop (SImode, and_optab, size, mask_least_3_bit,
NULL_RTX, 0, OPTAB_WIDEN);
/* add $dst_end, $dst, $size */
dst_end = expand_binop (Pmode, add_optab, dst_base_reg, size,
NULL_RTX, 0, OPTAB_WIDEN);
/* move $dst_itr, $dst
move $src_itr, $src */
emit_move_insn (dst_itr, dst_base_reg);
emit_move_insn (src_itr, src_base_reg);
/* beqz $size_least_3_bit, .Lbyte_mode_entry ! Not large enough. */
emit_cmp_and_jump_insns (size_least_3_bit, const0_rtx, EQ, NULL,
SImode, 1, byte_mode_entry);
/* add $double_word_end, $dst, $size_least_3_bit */
double_word_end = expand_binop (Pmode, add_optab,
dst_base_reg, size_least_3_bit,
NULL_RTX, 0, OPTAB_WIDEN);
/* .Ldouble_word_mode_loop: */
emit_label (double_word_mode_loop);
/* lmw.bim $tmp-begin, [$src_itr], $tmp-end, #0 ! $src_itr' = $src_itr
smw.bim $tmp-begin, [$dst_itr], $tmp-end, #0 ! $dst_itr' = $dst_itr */
src_itr_m = src_itr;
dst_itr_m = dst_itr;
srcmem_m = srcmem;
dstmem_m = dstmem;
nds32_emit_mem_move_block (start_regno, 2,
&dst_itr_m, &dstmem_m,
&src_itr_m, &srcmem_m,
true);
/* move $src_itr, $src_itr'
move $dst_itr, $dst_itr' */
emit_move_insn (dst_itr, dst_itr_m);
emit_move_insn (src_itr, src_itr_m);
/* ! Not readch upper bound. Loop.
bne $double_word_end, $dst_itr, .Ldouble_word_mode_loop */
emit_cmp_and_jump_insns (double_word_end, dst_itr, NE, NULL,
Pmode, 1, double_word_mode_loop);
/* .Lbyte_mode_entry: */
emit_label (byte_mode_entry);
/* beq $dst_itr, $dst_end, .Lend_label */
emit_cmp_and_jump_insns (dst_itr, dst_end, EQ, NULL,
Pmode, 1, end_label);
/* .Lbyte_mode_loop: */
emit_label (byte_mode_loop);
/* lbi.bi $tmp, [$src_itr], #1 */
nds32_emit_post_inc_load_store (tmp, src_itr, QImode, true);
/* sbi.bi $tmp, [$dst_itr], #1 */
nds32_emit_post_inc_load_store (tmp, dst_itr, QImode, false);
/* ! Not readch upper bound. Loop.
bne $dst_itr, $dst_end, .Lbyte_mode_loop */
emit_cmp_and_jump_insns (dst_itr, dst_end, NE, NULL,
SImode, 1, byte_mode_loop);
/* .Lend_label: */
emit_label (end_label);
return true;
}
