rtx
gen_lowpart_if_possible (machine_mode mode, rtx x)
{
rtx result = gen_lowpart_common (mode, x);
if (result)
return result;
else if (MEM_P (x))
{
/* This is the only other case we handle. */
poly_int64 offset = byte_lowpart_offset (mode, GET_MODE (x));
rtx new_rtx = adjust_address_nv (x, mode, offset);
if (! memory_address_addr_space_p (mode, XEXP (new_rtx, 0),
MEM_ADDR_SPACE (x)))
return 0;
return new_rtx;
}
else if (mode != GET_MODE (x) && GET_MODE (x) != VOIDmode
&& validate_subreg (mode, GET_MODE (x), x,
subreg_lowpart_offset (mode, GET_MODE (x))))
return gen_lowpart_SUBREG (mode, x);
else
return 0;
}
/* Transform (subreg (plus reg const)) to (plus (subreg reg) const)
when it is possible. Return X or the transformation result if the
transformation is done. */
static rtx
move_plus_up (rtx x)
{
rtx subreg_reg;
enum machine_mode x_mode, subreg_reg_mode;
if (GET_CODE (x) != SUBREG || !subreg_lowpart_p (x))
return x;
subreg_reg = SUBREG_REG (x);
x_mode = GET_MODE (x);
subreg_reg_mode = GET_MODE (subreg_reg);
if (GET_CODE (x) == SUBREG && GET_CODE (subreg_reg) == PLUS
&& GET_MODE_SIZE (x_mode) <= GET_MODE_SIZE (subreg_reg_mode)
&& CONSTANT_P (XEXP (subreg_reg, 1))
&& GET_MODE_CLASS (x_mode) == MODE_INT
&& GET_MODE_CLASS (subreg_reg_mode) == MODE_INT)
{
rtx cst = simplify_subreg (x_mode, XEXP (subreg_reg, 1), subreg_reg_mode,
subreg_lowpart_offset (x_mode,
subreg_reg_mode));
if (cst && CONSTANT_P (cst))
return gen_rtx_PLUS (x_mode, lowpart_subreg (x_mode,
XEXP (subreg_reg, 0),
subreg_reg_mode), cst);
}
return x;
}
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;
}
static bool
nds32_expand_setmem_unroll (rtx dstmem, rtx size, rtx value,
rtx align ATTRIBUTE_UNUSED,
rtx expected_align ATTRIBUTE_UNUSED,
rtx expected_size ATTRIBUTE_UNUSED)
{
unsigned maximum_regs, maximum_bytes, start_regno, regno;
rtx value4word;
rtx dst_base_reg, new_base_reg;
unsigned HOST_WIDE_INT remain_bytes, remain_words, prepare_regs, fill_per_smw;
unsigned HOST_WIDE_INT real_size;
if (TARGET_REDUCED_REGS)
{
maximum_regs = 4;
maximum_bytes = 64;
start_regno = 2;
}
else
{
maximum_regs = 8;
maximum_bytes = 128;
start_regno = 16;
}
real_size = UINTVAL (size) & GET_MODE_MASK(SImode);
if (!(CONST_INT_P (size) && real_size <= maximum_bytes))
return false;
remain_bytes = real_size;
gcc_assert (GET_MODE (value) == QImode || CONST_INT_P (value));
value4word = nds32_gen_dup_4_byte_to_word_value (value);
prepare_regs = remain_bytes / UNITS_PER_WORD;
dst_base_reg = copy_to_mode_reg (SImode, XEXP (dstmem, 0));
if (prepare_regs > maximum_regs)
prepare_regs = maximum_regs;
fill_per_smw = prepare_regs * UNITS_PER_WORD;
regno = start_regno;
switch (prepare_regs)
{
case 2:
default:
{
rtx reg0 = gen_rtx_REG (SImode, regno);
rtx reg1 = gen_rtx_REG (SImode, regno+1);
unsigned last_regno = start_regno + prepare_regs - 1;
emit_move_insn (reg0, value4word);
emit_move_insn (reg1, value4word);
rtx regd = gen_rtx_REG (DImode, regno);
regno += 2;
/* Try to utilize movd44! */
while (regno <= last_regno)
{
if ((regno + 1) <=last_regno)
{
rtx reg = gen_rtx_REG (DImode, regno);
emit_move_insn (reg, regd);
regno += 2;
}
else
{
rtx reg = gen_rtx_REG (SImode, regno);
emit_move_insn (reg, reg0);
regno += 1;
}
}
break;
}
case 1:
{
rtx reg = gen_rtx_REG (SImode, regno++);
emit_move_insn (reg, value4word);
}
break;
case 0:
break;
}
if (fill_per_smw)
for (;remain_bytes >= fill_per_smw;remain_bytes -= fill_per_smw)
{
emit_insn (nds32_expand_store_multiple (start_regno, prepare_regs,
dst_base_reg, dstmem,
true, &new_base_reg));
dst_base_reg = new_base_reg;
dstmem = gen_rtx_MEM (SImode, dst_base_reg);
}
remain_words = remain_bytes / UNITS_PER_WORD;
if (remain_words)
{
emit_insn (nds32_expand_store_multiple (start_regno, remain_words,
dst_base_reg, dstmem,
true, &new_base_reg));
dst_base_reg = new_base_reg;
dstmem = gen_rtx_MEM (SImode, dst_base_reg);
}
remain_bytes = remain_bytes - (remain_words * UNITS_PER_WORD);
if (remain_bytes)
{
value = simplify_gen_subreg (QImode, value4word, SImode,
subreg_lowpart_offset(QImode, SImode));
int offset = 0;
for (;remain_bytes;--remain_bytes, ++offset)
{
nds32_emit_load_store (value, dstmem, QImode, offset, false);
}
}
return true;
}
static bool
nds32_expand_setmem_loop (rtx dstmem, rtx size, rtx value)
{
rtx value4doubleword;
rtx value4byte;
rtx dst;
rtx byte_mode_size;
/* Emit loop version of setmem.
memset:
! prepare word
andi $tmp1, $val, 0xff ! $tmp1 <- 0x000000ab
slli $tmp2, $tmp1, 8 ! $tmp2 <- 0x0000ab00
or $tmp3, $val, $tmp2 ! $tmp3 <- 0x0000abab
slli $tmp4, $tmp3, 16 ! $tmp4 <- 0xabab0000
or $val4word, $tmp3, $tmp4 ! $value4word <- 0xabababab
and $size_for_word, $size, #-4
beqz $size_for_word, .Lword_mode_end
add $word_mode_end, $dst, $size_for_word
andi $byte_mode_size, $size, 3
.Lword_mode:
! word-mode set loop
smw.bim $value4word, [$dst], $value4word, 0
bne $word_mode_end, $dst, .Lword_mode
.Lword_mode_end:
beqz $byte_mode_size, .Lend
add $byte_mode_end, $dst, $byte_mode_size
.Lbyte_mode:
! byte-mode set loop
sbi.bi $value4word, [$dst] ,1
bne $byte_mode_end, $dst, .Lbyte_mode
.Lend: */
dst = copy_to_mode_reg (SImode, XEXP (dstmem, 0));
/* ! prepare word
andi $tmp1, $value, 0xff ! $tmp1 <- 0x000000ab
slli $tmp2, $tmp1, 8 ! $tmp2 <- 0x0000ab00
or $tmp3, $tmp1, $tmp2 ! $tmp3 <- 0x0000abab
slli $tmp4, $tmp3, 16 ! $tmp4 <- 0xabab0000
or $val4word, $tmp3, $tmp4 ! $value4word <- 0xabababab */
value4doubleword = nds32_gen_dup_8_byte_to_double_word_value (value);
/* and $size_for_word, $size, #-4
beqz $size_for_word, .Lword_mode_end
add $word_mode_end, $dst, $size_for_word
andi $byte_mode_size, $size, 3
.Lword_mode:
! word-mode set loop
smw.bim $value4word, [$dst], $value4word, 0
bne $word_mode_end, $dst, .Lword_mode
.Lword_mode_end: */
byte_mode_size = emit_setmem_doubleword_loop (dst, size, value4doubleword);
/* beqz $byte_mode_size, .Lend
add $byte_mode_end, $dst, $byte_mode_size
.Lbyte_mode:
! byte-mode set loop
sbi.bi $value, [$dst] ,1
bne $byte_mode_end, $dst, .Lbyte_mode
.Lend: */
value4byte = simplify_gen_subreg (QImode, value4doubleword, DImode,
subreg_lowpart_offset (QImode, DImode));
emit_setmem_byte_loop (dst, byte_mode_size, value4byte, false);
return true;
}
