Google Git
Sign in
gnu / gcc / d01fc6a3172f05d1618263450c1c9f13737acf57 / . / gcc / expmed.c
blob: a8039346dd2a474ee1b40fb08e131c4fbb966f67 [file] [log] [blame]
/* Medium-level subroutines: convert bit-field store and extract
and shifts, multiplies and divides to rtl instructions.
Copyright (C) 1987, 1988, 1989, 1992, 1993, 1994, 1995, 1996, 1997, 1998,
1999, 2000, 2001, 2002, 2003, 2004 Free Software Foundation, Inc.
This file is part of GCC.
GCC is free software; you can redistribute it and/or modify it under
the terms of the GNU General Public License as published by the Free
Software Foundation; either version 2, or (at your option) any later
version.
GCC is distributed in the hope that it will be useful, but WITHOUT ANY
WARRANTY; without even the implied warranty of MERCHANTABILITY or
FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
for more details.
You should have received a copy of the GNU General Public License
along with GCC; see the file COPYING. If not, write to the Free
Software Foundation, 59 Temple Place - Suite 330, Boston, MA
02111-1307, USA. */
#include "config.h"
#include "system.h"
#include "coretypes.h"
#include "tm.h"
#include "toplev.h"
#include "rtl.h"
#include "tree.h"
#include "tm_p.h"
#include "flags.h"
#include "insn-config.h"
#include "expr.h"
#include "optabs.h"
#include "real.h"
#include "recog.h"
#include "langhooks.h"
static void store_fixed_bit_field (rtx, unsigned HOST_WIDE_INT,
unsigned HOST_WIDE_INT,
unsigned HOST_WIDE_INT, rtx);
static void store_split_bit_field (rtx, unsigned HOST_WIDE_INT,
unsigned HOST_WIDE_INT, rtx);
static rtx extract_fixed_bit_field (enum machine_mode, rtx,
unsigned HOST_WIDE_INT,
unsigned HOST_WIDE_INT,
unsigned HOST_WIDE_INT, rtx, int);
static rtx mask_rtx (enum machine_mode, int, int, int);
static rtx lshift_value (enum machine_mode, rtx, int, int);
static rtx extract_split_bit_field (rtx, unsigned HOST_WIDE_INT,
unsigned HOST_WIDE_INT, int);
static void do_cmp_and_jump (rtx, rtx, enum rtx_code, enum machine_mode, rtx);
/* Nonzero means divides or modulus operations are relatively cheap for
powers of two, so don't use branches; emit the operation instead.
Usually, this will mean that the MD file will emit non-branch
sequences. */
static int sdiv_pow2_cheap, smod_pow2_cheap;
#ifndef SLOW_UNALIGNED_ACCESS
#define SLOW_UNALIGNED_ACCESS(MODE, ALIGN) STRICT_ALIGNMENT
#endif
/* For compilers that support multiple targets with different word sizes,
MAX_BITS_PER_WORD contains the biggest value of BITS_PER_WORD. An example
is the H8/300(H) compiler. */
#ifndef MAX_BITS_PER_WORD
#define MAX_BITS_PER_WORD BITS_PER_WORD
#endif
/* Reduce conditional compilation elsewhere. */
#ifndef HAVE_insv
#define HAVE_insv 0
#define CODE_FOR_insv CODE_FOR_nothing
#define gen_insv(a,b,c,d) NULL_RTX
#endif
#ifndef HAVE_extv
#define HAVE_extv 0
#define CODE_FOR_extv CODE_FOR_nothing
#define gen_extv(a,b,c,d) NULL_RTX
#endif
#ifndef HAVE_extzv
#define HAVE_extzv 0
#define CODE_FOR_extzv CODE_FOR_nothing
#define gen_extzv(a,b,c,d) NULL_RTX
#endif
/* Cost of various pieces of RTL. Note that some of these are indexed by
shift count and some by mode. */
static int add_cost, negate_cost, zero_cost;
static int shift_cost[MAX_BITS_PER_WORD];
static int shiftadd_cost[MAX_BITS_PER_WORD];
static int shiftsub_cost[MAX_BITS_PER_WORD];
static int mul_cost[NUM_MACHINE_MODES];
static int div_cost[NUM_MACHINE_MODES];
static int mul_widen_cost[NUM_MACHINE_MODES];
static int mul_highpart_cost[NUM_MACHINE_MODES];
void
init_expmed (void)
{
rtx reg, shift_insn, shiftadd_insn, shiftsub_insn;
int dummy;
int m;
enum machine_mode mode, wider_mode;
start_sequence ();
/* This is "some random pseudo register" for purposes of calling recog
to see what insns exist. */
reg = gen_rtx_REG (word_mode, 10000);
zero_cost = rtx_cost (const0_rtx, 0);
add_cost = rtx_cost (gen_rtx_PLUS (word_mode, reg, reg), SET);
shift_insn = emit_insn (gen_rtx_SET (VOIDmode, reg,
gen_rtx_ASHIFT (word_mode, reg,
const0_rtx)));
shiftadd_insn
= emit_insn (gen_rtx_SET (VOIDmode, reg,
gen_rtx_PLUS (word_mode,
gen_rtx_MULT (word_mode,
reg, const0_rtx),
reg)));
shiftsub_insn
= emit_insn (gen_rtx_SET (VOIDmode, reg,
gen_rtx_MINUS (word_mode,
gen_rtx_MULT (word_mode,
reg, const0_rtx),
reg)));
init_recog ();
shift_cost[0] = 0;
shiftadd_cost[0] = shiftsub_cost[0] = add_cost;
for (m = 1; m < MAX_BITS_PER_WORD; m++)
{
rtx c_int = GEN_INT ((HOST_WIDE_INT) 1 << m);
shift_cost[m] = shiftadd_cost[m] = shiftsub_cost[m] = 32000;
XEXP (SET_SRC (PATTERN (shift_insn)), 1) = GEN_INT (m);
if (recog (PATTERN (shift_insn), shift_insn, &dummy) >= 0)
shift_cost[m] = rtx_cost (SET_SRC (PATTERN (shift_insn)), SET);
XEXP (XEXP (SET_SRC (PATTERN (shiftadd_insn)), 0), 1) = c_int;
if (recog (PATTERN (shiftadd_insn), shiftadd_insn, &dummy) >= 0)
shiftadd_cost[m] = rtx_cost (SET_SRC (PATTERN (shiftadd_insn)), SET);
XEXP (XEXP (SET_SRC (PATTERN (shiftsub_insn)), 0), 1) = c_int;
if (recog (PATTERN (shiftsub_insn), shiftsub_insn, &dummy) >= 0)
shiftsub_cost[m] = rtx_cost (SET_SRC (PATTERN (shiftsub_insn)), SET);
}
negate_cost = rtx_cost (gen_rtx_NEG (word_mode, reg), SET);
sdiv_pow2_cheap
= (rtx_cost (gen_rtx_DIV (word_mode, reg, GEN_INT (32)), SET)
<= 2 * add_cost);
smod_pow2_cheap
= (rtx_cost (gen_rtx_MOD (word_mode, reg, GEN_INT (32)), SET)
<= 2 * add_cost);
for (mode = GET_CLASS_NARROWEST_MODE (MODE_INT);
mode != VOIDmode;
mode = GET_MODE_WIDER_MODE (mode))
{
reg = gen_rtx_REG (mode, 10000);
div_cost[(int) mode] = rtx_cost (gen_rtx_UDIV (mode, reg, reg), SET);
mul_cost[(int) mode] = rtx_cost (gen_rtx_MULT (mode, reg, reg), SET);
wider_mode = GET_MODE_WIDER_MODE (mode);
if (wider_mode != VOIDmode)
{
mul_widen_cost[(int) wider_mode]
= rtx_cost (gen_rtx_MULT (wider_mode,
gen_rtx_ZERO_EXTEND (wider_mode, reg),
gen_rtx_ZERO_EXTEND (wider_mode, reg)),
SET);
mul_highpart_cost[(int) mode]
= rtx_cost (gen_rtx_TRUNCATE
(mode,
gen_rtx_LSHIFTRT (wider_mode,
gen_rtx_MULT (wider_mode,
gen_rtx_ZERO_EXTEND
(wider_mode, reg),
gen_rtx_ZERO_EXTEND
(wider_mode, reg)),
GEN_INT (GET_MODE_BITSIZE (mode)))),
SET);
}
}
end_sequence ();
}
/* Return an rtx representing minus the value of X.
MODE is the intended mode of the result,
useful if X is a CONST_INT. */
rtx
negate_rtx (enum machine_mode mode, rtx x)
{
rtx result = simplify_unary_operation (NEG, mode, x, mode);
if (result == 0)
result = expand_unop (mode, neg_optab, x, NULL_RTX, 0);
return result;
}
/* Report on the availability of insv/extv/extzv and the desired mode
of each of their operands. Returns MAX_MACHINE_MODE if HAVE_foo
is false; else the mode of the specified operand. If OPNO is -1,
all the caller cares about is whether the insn is available. */
enum machine_mode
mode_for_extraction (enum extraction_pattern pattern, int opno)
{
const struct insn_data *data;
switch (pattern)
{
case EP_insv:
if (HAVE_insv)
{
data = &insn_data[CODE_FOR_insv];
break;
}
return MAX_MACHINE_MODE;
case EP_extv:
if (HAVE_extv)
{
data = &insn_data[CODE_FOR_extv];
break;
}
return MAX_MACHINE_MODE;
case EP_extzv:
if (HAVE_extzv)
{
data = &insn_data[CODE_FOR_extzv];
break;
}
return MAX_MACHINE_MODE;
default:
abort ();
}
if (opno == -1)
return VOIDmode;
/* Everyone who uses this function used to follow it with
if (result == VOIDmode) result = word_mode; */
if (data->operand[opno].mode == VOIDmode)
return word_mode;
return data->operand[opno].mode;
}
/* Generate code to store value from rtx VALUE
into a bit-field within structure STR_RTX
containing BITSIZE bits starting at bit BITNUM.
FIELDMODE is the machine-mode of the FIELD_DECL node for this field.
ALIGN is the alignment that STR_RTX is known to have.
TOTAL_SIZE is the size of the structure in bytes, or -1 if varying. */
/* ??? Note that there are two different ideas here for how
to determine the size to count bits within, for a register.
One is BITS_PER_WORD, and the other is the size of operand 3
of the insv pattern.
If operand 3 of the insv pattern is VOIDmode, then we will use BITS_PER_WORD
else, we use the mode of operand 3. */
rtx
store_bit_field (rtx str_rtx, unsigned HOST_WIDE_INT bitsize,
unsigned HOST_WIDE_INT bitnum, enum machine_mode fieldmode,
rtx value, HOST_WIDE_INT total_size)
{
unsigned int unit
= (GET_CODE (str_rtx) == MEM) ? BITS_PER_UNIT : BITS_PER_WORD;
unsigned HOST_WIDE_INT offset = bitnum / unit;
unsigned HOST_WIDE_INT bitpos = bitnum % unit;
rtx op0 = str_rtx;
int byte_offset;
rtx orig_value;
enum machine_mode op_mode = mode_for_extraction (EP_insv, 3);
/* Discount the part of the structure before the desired byte.
We need to know how many bytes are safe to reference after it. */
if (total_size >= 0)
total_size -= (bitpos / BIGGEST_ALIGNMENT
* (BIGGEST_ALIGNMENT / BITS_PER_UNIT));
while (GET_CODE (op0) == SUBREG)
{
/* The following line once was done only if WORDS_BIG_ENDIAN,
but I think that is a mistake. WORDS_BIG_ENDIAN is
meaningful at a much higher level; when structures are copied
between memory and regs, the higher-numbered regs
always get higher addresses. */
offset += (SUBREG_BYTE (op0) / UNITS_PER_WORD);
/* We used to adjust BITPOS here, but now we do the whole adjustment
right after the loop. */
op0 = SUBREG_REG (op0);
}
value = protect_from_queue (value, 0);
/* Use vec_extract patterns for extracting parts of vectors whenever
available. */
if (VECTOR_MODE_P (GET_MODE (op0))
&& GET_CODE (op0) != MEM
&& (vec_set_optab->handlers[(int)GET_MODE (op0)].insn_code
!= CODE_FOR_nothing)
&& fieldmode == GET_MODE_INNER (GET_MODE (op0))
&& bitsize == GET_MODE_BITSIZE (GET_MODE_INNER (GET_MODE (op0)))
&& !(bitnum % GET_MODE_BITSIZE (GET_MODE_INNER (GET_MODE (op0)))))
{
enum machine_mode outermode = GET_MODE (op0);
enum machine_mode innermode = GET_MODE_INNER (outermode);
int icode = (int) vec_set_optab->handlers[(int) outermode].insn_code;
int pos = bitnum / GET_MODE_BITSIZE (innermode);
rtx rtxpos = GEN_INT (pos);
rtx src = value;
rtx dest = op0;
rtx pat, seq;
enum machine_mode mode0 = insn_data[icode].operand[0].mode;
enum machine_mode mode1 = insn_data[icode].operand[1].mode;
enum machine_mode mode2 = insn_data[icode].operand[2].mode;
start_sequence ();
if (! (*insn_data[icode].operand[1].predicate) (src, mode1))
src = copy_to_mode_reg (mode1, src);
if (! (*insn_data[icode].operand[2].predicate) (rtxpos, mode2))
rtxpos = copy_to_mode_reg (mode1, rtxpos);
/* We could handle this, but we should always be called with a pseudo
for our targets and all insns should take them as outputs. */
if (! (*insn_data[icode].operand[0].predicate) (dest, mode0)
|| ! (*insn_data[icode].operand[1].predicate) (src, mode1)
|| ! (*insn_data[icode].operand[2].predicate) (rtxpos, mode2))
abort ();
pat = GEN_FCN (icode) (dest, src, rtxpos);
seq = get_insns ();
end_sequence ();
if (pat)
{
emit_insn (seq);
emit_insn (pat);
return dest;
}
}
if (flag_force_mem)
{
int old_generating_concat_p = generating_concat_p;
generating_concat_p = 0;
value = force_not_mem (value);
generating_concat_p = old_generating_concat_p;
}
/* If the target is a register, overwriting the entire object, or storing
a full-word or multi-word field can be done with just a SUBREG.
If the target is memory, storing any naturally aligned field can be
done with a simple store. For targets that support fast unaligned
memory, any naturally sized, unit aligned field can be done directly. */
byte_offset = (bitnum % BITS_PER_WORD) / BITS_PER_UNIT
+ (offset * UNITS_PER_WORD);
if (bitpos == 0
&& bitsize == GET_MODE_BITSIZE (fieldmode)
&& (GET_CODE (op0) != MEM
? ((GET_MODE_SIZE (fieldmode) >= UNITS_PER_WORD
|| GET_MODE_SIZE (GET_MODE (op0)) == GET_MODE_SIZE (fieldmode))
&& byte_offset % GET_MODE_SIZE (fieldmode) == 0)
: (! SLOW_UNALIGNED_ACCESS (fieldmode, MEM_ALIGN (op0))
|| (offset * BITS_PER_UNIT % bitsize == 0
&& MEM_ALIGN (op0) % GET_MODE_BITSIZE (fieldmode) == 0))))
{
if (GET_MODE (op0) != fieldmode)
{
if (GET_CODE (op0) == SUBREG)
{
if (GET_MODE (SUBREG_REG (op0)) == fieldmode
|| GET_MODE_CLASS (fieldmode) == MODE_INT
|| GET_MODE_CLASS (fieldmode) == MODE_PARTIAL_INT)
op0 = SUBREG_REG (op0);
else
/* Else we've got some float mode source being extracted into
a different float mode destination -- this combination of
subregs results in Severe Tire Damage. */
abort ();
}
if (GET_CODE (op0) == REG)
op0 = gen_rtx_SUBREG (fieldmode, op0, byte_offset);
else
op0 = adjust_address (op0, fieldmode, offset);
}
emit_move_insn (op0, value);
return value;
}
/* Make sure we are playing with integral modes. Pun with subregs
if we aren't. This must come after the entire register case above,
since that case is valid for any mode. The following cases are only
valid for integral modes. */
{
enum machine_mode imode = int_mode_for_mode (GET_MODE (op0));
if (imode != GET_MODE (op0))
{
if (GET_CODE (op0) == MEM)
op0 = adjust_address (op0, imode, 0);
else if (imode != BLKmode)
op0 = gen_lowpart (imode, op0);
else
abort ();
}
}
/* We may be accessing data outside the field, which means
we can alias adjacent data. */
if (GET_CODE (op0) == MEM)
{
op0 = shallow_copy_rtx (op0);
set_mem_alias_set (op0, 0);
set_mem_expr (op0, 0);
}
/* If OP0 is a register, BITPOS must count within a word.
But as we have it, it counts within whatever size OP0 now has.
On a bigendian machine, these are not the same, so convert. */
if (BYTES_BIG_ENDIAN
&& GET_CODE (op0) != MEM
&& unit > GET_MODE_BITSIZE (GET_MODE (op0)))
bitpos += unit - GET_MODE_BITSIZE (GET_MODE (op0));
/* Storing an lsb-aligned field in a register
can be done with a movestrict instruction. */
if (GET_CODE (op0) != MEM
&& (BYTES_BIG_ENDIAN ? bitpos + bitsize == unit : bitpos == 0)
&& bitsize == GET_MODE_BITSIZE (fieldmode)
&& (movstrict_optab->handlers[(int) fieldmode].insn_code
!= CODE_FOR_nothing))
{
int icode = movstrict_optab->handlers[(int) fieldmode].insn_code;
/* Get appropriate low part of the value being stored. */
if (GET_CODE (value) == CONST_INT || GET_CODE (value) == REG)
value = gen_lowpart (fieldmode, value);
else if (!(GET_CODE (value) == SYMBOL_REF
|| GET_CODE (value) == LABEL_REF
|| GET_CODE (value) == CONST))
value = convert_to_mode (fieldmode, value, 0);
if (! (*insn_data[icode].operand[1].predicate) (value, fieldmode))
value = copy_to_mode_reg (fieldmode, value);
if (GET_CODE (op0) == SUBREG)
{
if (GET_MODE (SUBREG_REG (op0)) == fieldmode
|| GET_MODE_CLASS (fieldmode) == MODE_INT
|| GET_MODE_CLASS (fieldmode) == MODE_PARTIAL_INT)
op0 = SUBREG_REG (op0);
else
/* Else we've got some float mode source being extracted into
a different float mode destination -- this combination of
subregs results in Severe Tire Damage. */
abort ();
}
emit_insn (GEN_FCN (icode)
(gen_rtx_SUBREG (fieldmode, op0,
(bitnum % BITS_PER_WORD) / BITS_PER_UNIT
+ (offset * UNITS_PER_WORD)),
value));
return value;
}
/* Handle fields bigger than a word. */
if (bitsize > BITS_PER_WORD)
{
/* Here we transfer the words of the field
in the order least significant first.
This is because the most significant word is the one which may
be less than full.
However, only do that if the value is not BLKmode. */
unsigned int backwards = WORDS_BIG_ENDIAN && fieldmode != BLKmode;
unsigned int nwords = (bitsize + (BITS_PER_WORD - 1)) / BITS_PER_WORD;
unsigned int i;
/* This is the mode we must force value to, so that there will be enough
subwords to extract. Note that fieldmode will often (always?) be
VOIDmode, because that is what store_field uses to indicate that this
is a bit field, but passing VOIDmode to operand_subword_force will
result in an abort. */
fieldmode = GET_MODE (value);
if (fieldmode == VOIDmode)
fieldmode = smallest_mode_for_size (nwords * BITS_PER_WORD, MODE_INT);
for (i = 0; i < nwords; i++)
{
/* If I is 0, use the low-order word in both field and target;
if I is 1, use the next to lowest word; and so on. */
unsigned int wordnum = (backwards ? nwords - i - 1 : i);
unsigned int bit_offset = (backwards
? MAX ((int) bitsize - ((int) i + 1)
* BITS_PER_WORD,
0)
: (int) i * BITS_PER_WORD);
store_bit_field (op0, MIN (BITS_PER_WORD,
bitsize - i * BITS_PER_WORD),
bitnum + bit_offset, word_mode,
operand_subword_force (value, wordnum, fieldmode),
total_size);
}
return value;
}
/* From here on we can assume that the field to be stored in is
a full-word (whatever type that is), since it is shorter than a word. */
/* OFFSET is the number of words or bytes (UNIT says which)
from STR_RTX to the first word or byte containing part of the field. */
if (GET_CODE (op0) != MEM)
{
if (offset != 0
|| GET_MODE_SIZE (GET_MODE (op0)) > UNITS_PER_WORD)
{
if (GET_CODE (op0) != REG)
{
/* Since this is a destination (lvalue), we can't copy it to a
pseudo. We can trivially remove a SUBREG that does not
change the size of the operand. Such a SUBREG may have been
added above. Otherwise, abort. */
if (GET_CODE (op0) == SUBREG
&& (GET_MODE_SIZE (GET_MODE (op0))
== GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0)))))
op0 = SUBREG_REG (op0);
else
abort ();
}
op0 = gen_rtx_SUBREG (mode_for_size (BITS_PER_WORD, MODE_INT, 0),
op0, (offset * UNITS_PER_WORD));
}
offset = 0;
}
else
op0 = protect_from_queue (op0, 1);
/* If VALUE is a floating-point mode, access it as an integer of the
corresponding size. This can occur on a machine with 64 bit registers
that uses SFmode for float. This can also occur for unaligned float
structure fields. */
orig_value = value;
if (GET_MODE_CLASS (GET_MODE (value)) != MODE_INT
&& GET_MODE_CLASS (GET_MODE (value)) != MODE_PARTIAL_INT)
value = gen_lowpart ((GET_MODE (value) == VOIDmode
? word_mode : int_mode_for_mode (GET_MODE (value))),
value);
/* Now OFFSET is nonzero only if OP0 is memory
and is therefore always measured in bytes. */
if (HAVE_insv
&& GET_MODE (value) != BLKmode
&& !(bitsize == 1 && GET_CODE (value) == CONST_INT)
/* Ensure insv's size is wide enough for this field. */
&& (GET_MODE_BITSIZE (op_mode) >= bitsize)
&& ! ((GET_CODE (op0) == REG || GET_CODE (op0) == SUBREG)
&& (bitsize + bitpos > GET_MODE_BITSIZE (op_mode))))
{
int xbitpos = bitpos;
rtx value1;
rtx xop0 = op0;
rtx last = get_last_insn ();
rtx pat;
enum machine_mode maxmode = mode_for_extraction (EP_insv, 3);
int save_volatile_ok = volatile_ok;
volatile_ok = 1;
/* If this machine's insv can only insert into a register, copy OP0
into a register and save it back later. */
/* This used to check flag_force_mem, but that was a serious
de-optimization now that flag_force_mem is enabled by -O2. */
if (GET_CODE (op0) == MEM
&& ! ((*insn_data[(int) CODE_FOR_insv].operand[0].predicate)
(op0, VOIDmode)))
{
rtx tempreg;
enum machine_mode bestmode;
/* Get the mode to use for inserting into this field. If OP0 is
BLKmode, get the smallest mode consistent with the alignment. If
OP0 is a non-BLKmode object that is no wider than MAXMODE, use its
mode. Otherwise, use the smallest mode containing the field. */
if (GET_MODE (op0) == BLKmode
|| GET_MODE_SIZE (GET_MODE (op0)) > GET_MODE_SIZE (maxmode))
bestmode
= get_best_mode (bitsize, bitnum, MEM_ALIGN (op0), maxmode,
MEM_VOLATILE_P (op0));
else
bestmode = GET_MODE (op0);
if (bestmode == VOIDmode
|| (SLOW_UNALIGNED_ACCESS (bestmode, MEM_ALIGN (op0))
&& GET_MODE_BITSIZE (bestmode) > MEM_ALIGN (op0)))
goto insv_loses;
/* Adjust address to point to the containing unit of that mode.
Compute offset as multiple of this unit, counting in bytes. */
unit = GET_MODE_BITSIZE (bestmode);
offset = (bitnum / unit) * GET_MODE_SIZE (bestmode);
bitpos = bitnum % unit;
op0 = adjust_address (op0, bestmode, offset);
/* Fetch that unit, store the bitfield in it, then store
the unit. */
tempreg = copy_to_reg (op0);
store_bit_field (tempreg, bitsize, bitpos, fieldmode, orig_value,
total_size);
emit_move_insn (op0, tempreg);
return value;
}
volatile_ok = save_volatile_ok;
/* Add OFFSET into OP0's address. */
if (GET_CODE (xop0) == MEM)
xop0 = adjust_address (xop0, byte_mode, offset);
/* If xop0 is a register, we need it in MAXMODE
to make it acceptable to the format of insv. */
if (GET_CODE (xop0) == SUBREG)
/* We can't just change the mode, because this might clobber op0,
and we will need the original value of op0 if insv fails. */
xop0 = gen_rtx_SUBREG (maxmode, SUBREG_REG (xop0), SUBREG_BYTE (xop0));
if (GET_CODE (xop0) == REG && GET_MODE (xop0) != maxmode)
xop0 = gen_rtx_SUBREG (maxmode, xop0, 0);
/* On big-endian machines, we count bits from the most significant.
If the bit field insn does not, we must invert. */
if (BITS_BIG_ENDIAN != BYTES_BIG_ENDIAN)
xbitpos = unit - bitsize - xbitpos;
/* We have been counting XBITPOS within UNIT.
Count instead within the size of the register. */
if (BITS_BIG_ENDIAN && GET_CODE (xop0) != MEM)
xbitpos += GET_MODE_BITSIZE (maxmode) - unit;
unit = GET_MODE_BITSIZE (maxmode);
/* Convert VALUE to maxmode (which insv insn wants) in VALUE1. */
value1 = value;
if (GET_MODE (value) != maxmode)
{
if (GET_MODE_BITSIZE (GET_MODE (value)) >= bitsize)
{
/* Optimization: Don't bother really extending VALUE
if it has all the bits we will actually use. However,
if we must narrow it, be sure we do it correctly. */
if (GET_MODE_SIZE (GET_MODE (value)) < GET_MODE_SIZE (maxmode))
{
rtx tmp;
tmp = simplify_subreg (maxmode, value1, GET_MODE (value), 0);
if (! tmp)
tmp = simplify_gen_subreg (maxmode,
force_reg (GET_MODE (value),
value1),
GET_MODE (value), 0);
value1 = tmp;
}
else
value1 = gen_lowpart (maxmode, value1);
}
else if (GET_CODE (value) == CONST_INT)
value1 = gen_int_mode (INTVAL (value), maxmode);
else if (!CONSTANT_P (value))
/* Parse phase is supposed to make VALUE's data type
match that of the component reference, which is a type
at least as wide as the field; so VALUE should have
a mode that corresponds to that type. */
abort ();
}
/* If this machine's insv insists on a register,
get VALUE1 into a register. */
if (! ((*insn_data[(int) CODE_FOR_insv].operand[3].predicate)
(value1, maxmode)))
value1 = force_reg (maxmode, value1);
pat = gen_insv (xop0, GEN_INT (bitsize), GEN_INT (xbitpos), value1);
if (pat)
emit_insn (pat);
else
{
delete_insns_since (last);
store_fixed_bit_field (op0, offset, bitsize, bitpos, value);
}
}
else
insv_loses:
/* Insv is not available; store using shifts and boolean ops. */
store_fixed_bit_field (op0, offset, bitsize, bitpos, value);
return value;
}
/* Use shifts and boolean operations to store VALUE
into a bit field of width BITSIZE
in a memory location specified by OP0 except offset by OFFSET bytes.
(OFFSET must be 0 if OP0 is a register.)
The field starts at position BITPOS within the byte.
(If OP0 is a register, it may be a full word or a narrower mode,
but BITPOS still counts within a full word,
which is significant on bigendian machines.)
Note that protect_from_queue has already been done on OP0 and VALUE. */
static void
store_fixed_bit_field (rtx op0, unsigned HOST_WIDE_INT offset,
unsigned HOST_WIDE_INT bitsize,
unsigned HOST_WIDE_INT bitpos, rtx value)
{
enum machine_mode mode;
unsigned int total_bits = BITS_PER_WORD;
rtx subtarget, temp;
int all_zero = 0;
int all_one = 0;
/* There is a case not handled here:
a structure with a known alignment of just a halfword
and a field split across two aligned halfwords within the structure.
Or likewise a structure with a known alignment of just a byte
and a field split across two bytes.
Such cases are not supposed to be able to occur. */
if (GET_CODE (op0) == REG || GET_CODE (op0) == SUBREG)
{
if (offset != 0)
abort ();
/* Special treatment for a bit field split across two registers. */
if (bitsize + bitpos > BITS_PER_WORD)
{
store_split_bit_field (op0, bitsize, bitpos, value);
return;
}
}
else
{
/* Get the proper mode to use for this field. We want a mode that
includes the entire field. If such a mode would be larger than
a word, we won't be doing the extraction the normal way.
We don't want a mode bigger than the destination. */
mode = GET_MODE (op0);
if (GET_MODE_BITSIZE (mode) == 0
|| GET_MODE_BITSIZE (mode) > GET_MODE_BITSIZE (word_mode))
mode = word_mode;
mode = get_best_mode (bitsize, bitpos + offset * BITS_PER_UNIT,
MEM_ALIGN (op0), mode, MEM_VOLATILE_P (op0));
if (mode == VOIDmode)
{
/* The only way this should occur is if the field spans word
boundaries. */
store_split_bit_field (op0, bitsize, bitpos + offset * BITS_PER_UNIT,
value);
return;
}
total_bits = GET_MODE_BITSIZE (mode);
/* Make sure bitpos is valid for the chosen mode. Adjust BITPOS to
be in the range 0 to total_bits-1, and put any excess bytes in
OFFSET. */
if (bitpos >= total_bits)
{
offset += (bitpos / total_bits) * (total_bits / BITS_PER_UNIT);
bitpos -= ((bitpos / total_bits) * (total_bits / BITS_PER_UNIT)
* BITS_PER_UNIT);
}
/* Get ref to an aligned byte, halfword, or word containing the field.
Adjust BITPOS to be position within a word,
and OFFSET to be the offset of that word.
Then alter OP0 to refer to that word. */
bitpos += (offset % (total_bits / BITS_PER_UNIT)) * BITS_PER_UNIT;
offset -= (offset % (total_bits / BITS_PER_UNIT));
op0 = adjust_address (op0, mode, offset);
}
mode = GET_MODE (op0);
/* Now MODE is either some integral mode for a MEM as OP0,
or is a full-word for a REG as OP0. TOTAL_BITS corresponds.
The bit field is contained entirely within OP0.
BITPOS is the starting bit number within OP0.
(OP0's mode may actually be narrower than MODE.) */
if (BYTES_BIG_ENDIAN)
/* BITPOS is the distance between our msb
and that of the containing datum.
Convert it to the distance from the lsb. */
bitpos = total_bits - bitsize - bitpos;
/* Now BITPOS is always the distance between our lsb
and that of OP0. */
/* Shift VALUE left by BITPOS bits. If VALUE is not constant,
we must first convert its mode to MODE. */
if (GET_CODE (value) == CONST_INT)
{
HOST_WIDE_INT v = INTVAL (value);
if (bitsize < HOST_BITS_PER_WIDE_INT)
v &= ((HOST_WIDE_INT) 1 << bitsize) - 1;
if (v == 0)
all_zero = 1;
else if ((bitsize < HOST_BITS_PER_WIDE_INT
&& v == ((HOST_WIDE_INT) 1 << bitsize) - 1)
|| (bitsize == HOST_BITS_PER_WIDE_INT && v == -1))
all_one = 1;
value = lshift_value (mode, value, bitpos, bitsize);
}
else
{
int must_and = (GET_MODE_BITSIZE (GET_MODE (value)) != bitsize
&& bitpos + bitsize != GET_MODE_BITSIZE (mode));
if (GET_MODE (value) != mode)
{
if ((GET_CODE (value) == REG || GET_CODE (value) == SUBREG)
&& GET_MODE_SIZE (mode) < GET_MODE_SIZE (GET_MODE (value)))
value = gen_lowpart (mode, value);
else
value = convert_to_mode (mode, value, 1);
}
if (must_and)
value = expand_binop (mode, and_optab, value,
mask_rtx (mode, 0, bitsize, 0),
NULL_RTX, 1, OPTAB_LIB_WIDEN);
if (bitpos > 0)
value = expand_shift (LSHIFT_EXPR, mode, value,
build_int_2 (bitpos, 0), NULL_RTX, 1);
}
/* Now clear the chosen bits in OP0,
except that if VALUE is -1 we need not bother. */
subtarget = (GET_CODE (op0) == REG || ! flag_force_mem) ? op0 : 0;
if (! all_one)
{
temp = expand_binop (mode, and_optab, op0,
mask_rtx (mode, bitpos, bitsize, 1),
subtarget, 1, OPTAB_LIB_WIDEN);
subtarget = temp;
}
else
temp = op0;
/* Now logical-or VALUE into OP0, unless it is zero. */
if (! all_zero)
temp = expand_binop (mode, ior_optab, temp, value,
subtarget, 1, OPTAB_LIB_WIDEN);
if (op0 != temp)
emit_move_insn (op0, temp);
}
/* Store a bit field that is split across multiple accessible memory objects.
OP0 is the REG, SUBREG or MEM rtx for the first of the objects.
BITSIZE is the field width; BITPOS the position of its first bit
(within the word).
VALUE is the value to store.
This does not yet handle fields wider than BITS_PER_WORD. */
static void
store_split_bit_field (rtx op0, unsigned HOST_WIDE_INT bitsize,
unsigned HOST_WIDE_INT bitpos, rtx value)
{
unsigned int unit;
unsigned int bitsdone = 0;
/* Make sure UNIT isn't larger than BITS_PER_WORD, we can only handle that
much at a time. */
if (GET_CODE (op0) == REG || GET_CODE (op0) == SUBREG)
unit = BITS_PER_WORD;
else
unit = MIN (MEM_ALIGN (op0), BITS_PER_WORD);
/* If VALUE is a constant other than a CONST_INT, get it into a register in
WORD_MODE. If we can do this using gen_lowpart_common, do so. Note
that VALUE might be a floating-point constant. */
if (CONSTANT_P (value) && GET_CODE (value) != CONST_INT)
{
rtx word = gen_lowpart_common (word_mode, value);
if (word && (value != word))
value = word;
else
value = gen_lowpart_common (word_mode,
force_reg (GET_MODE (value) != VOIDmode
? GET_MODE (value)
: word_mode, value));
}
else if (GET_CODE (value) == ADDRESSOF)
value = copy_to_reg (value);
while (bitsdone < bitsize)
{
unsigned HOST_WIDE_INT thissize;
rtx part, word;
unsigned HOST_WIDE_INT thispos;
unsigned HOST_WIDE_INT offset;
offset = (bitpos + bitsdone) / unit;
thispos = (bitpos + bitsdone) % unit;
/* THISSIZE must not overrun a word boundary. Otherwise,
store_fixed_bit_field will call us again, and we will mutually
recurse forever. */
thissize = MIN (bitsize - bitsdone, BITS_PER_WORD);
thissize = MIN (thissize, unit - thispos);
if (BYTES_BIG_ENDIAN)
{
int total_bits;
/* We must do an endian conversion exactly the same way as it is
done in extract_bit_field, so that the two calls to
extract_fixed_bit_field will have comparable arguments. */
if (GET_CODE (value) != MEM || GET_MODE (value) == BLKmode)
total_bits = BITS_PER_WORD;
else
total_bits = GET_MODE_BITSIZE (GET_MODE (value));
/* Fetch successively less significant portions. */
if (GET_CODE (value) == CONST_INT)
part = GEN_INT (((unsigned HOST_WIDE_INT) (INTVAL (value))
>> (bitsize - bitsdone - thissize))
& (((HOST_WIDE_INT) 1 << thissize) - 1));
else
/* The args are chosen so that the last part includes the
lsb. Give extract_bit_field the value it needs (with
endianness compensation) to fetch the piece we want. */
part = extract_fixed_bit_field (word_mode, value, 0, thissize,
total_bits - bitsize + bitsdone,
NULL_RTX, 1);
}
else
{
/* Fetch successively more significant portions. */
if (GET_CODE (value) == CONST_INT)
part = GEN_INT (((unsigned HOST_WIDE_INT) (INTVAL (value))
>> bitsdone)
& (((HOST_WIDE_INT) 1 << thissize) - 1));
else
part = extract_fixed_bit_field (word_mode, value, 0, thissize,
bitsdone, NULL_RTX, 1);
}
/* If OP0 is a register, then handle OFFSET here.
When handling multiword bitfields, extract_bit_field may pass
down a word_mode SUBREG of a larger REG for a bitfield that actually
crosses a word boundary. Thus, for a SUBREG, we must find
the current word starting from the base register. */
if (GET_CODE (op0) == SUBREG)
{
int word_offset = (SUBREG_BYTE (op0) / UNITS_PER_WORD) + offset;
word = operand_subword_force (SUBREG_REG (op0), word_offset,
GET_MODE (SUBREG_REG (op0)));
offset = 0;
}
else if (GET_CODE (op0) == REG)
{
word = operand_subword_force (op0, offset, GET_MODE (op0));
offset = 0;
}
else
word = op0;
/* OFFSET is in UNITs, and UNIT is in bits.
store_fixed_bit_field wants offset in bytes. */
store_fixed_bit_field (word, offset * unit / BITS_PER_UNIT, thissize,
thispos, part);
bitsdone += thissize;
}
}
/* Generate code to extract a byte-field from STR_RTX
containing BITSIZE bits, starting at BITNUM,
and put it in TARGET if possible (if TARGET is nonzero).
Regardless of TARGET, we return the rtx for where the value is placed.
It may be a QUEUED.
STR_RTX is the structure containing the byte (a REG or MEM).
UNSIGNEDP is nonzero if this is an unsigned bit field.
MODE is the natural mode of the field value once extracted.
TMODE is the mode the caller would like the value to have;
but the value may be returned with type MODE instead.
TOTAL_SIZE is the size in bytes of the containing structure,
or -1 if varying.
If a TARGET is specified and we can store in it at no extra cost,
we do so, and return TARGET.
Otherwise, we return a REG of mode TMODE or MODE, with TMODE preferred
if they are equally easy. */
rtx
extract_bit_field (rtx str_rtx, unsigned HOST_WIDE_INT bitsize,
unsigned HOST_WIDE_INT bitnum, int unsignedp, rtx target,
enum machine_mode mode, enum machine_mode tmode,
HOST_WIDE_INT total_size)
{
unsigned int unit
= (GET_CODE (str_rtx) == MEM) ? BITS_PER_UNIT : BITS_PER_WORD;
unsigned HOST_WIDE_INT offset = bitnum / unit;
unsigned HOST_WIDE_INT bitpos = bitnum % unit;
rtx op0 = str_rtx;
rtx spec_target = target;
rtx spec_target_subreg = 0;
enum machine_mode int_mode;
enum machine_mode extv_mode = mode_for_extraction (EP_extv, 0);
enum machine_mode extzv_mode = mode_for_extraction (EP_extzv, 0);
enum machine_mode mode1;
int byte_offset;
/* Discount the part of the structure before the desired byte.
We need to know how many bytes are safe to reference after it. */
if (total_size >= 0)
total_size -= (bitpos / BIGGEST_ALIGNMENT
* (BIGGEST_ALIGNMENT / BITS_PER_UNIT));
if (tmode == VOIDmode)
tmode = mode;
while (GET_CODE (op0) == SUBREG)
{
bitpos += SUBREG_BYTE (op0) * BITS_PER_UNIT;
if (bitpos > unit)
{
offset += (bitpos / unit);
bitpos %= unit;
}
op0 = SUBREG_REG (op0);
}
if (GET_CODE (op0) == REG
&& mode == GET_MODE (op0)
&& bitnum == 0
&& bitsize == GET_MODE_BITSIZE (GET_MODE (op0)))
{
/* We're trying to extract a full register from itself. */
return op0;
}
/* Use vec_extract patterns for extracting parts of vectors whenever
available. */
if (VECTOR_MODE_P (GET_MODE (op0))
&& GET_CODE (op0) != MEM
&& (vec_extract_optab->handlers[(int)GET_MODE (op0)].insn_code
!= CODE_FOR_nothing)
&& ((bitsize + bitnum) / GET_MODE_BITSIZE (GET_MODE_INNER (GET_MODE (op0)))
== bitsize / GET_MODE_BITSIZE (GET_MODE_INNER (GET_MODE (op0)))))
{
enum machine_mode outermode = GET_MODE (op0);
enum machine_mode innermode = GET_MODE_INNER (outermode);
int icode = (int) vec_extract_optab->handlers[(int) outermode].insn_code;
int pos = bitnum / GET_MODE_BITSIZE (innermode);
rtx rtxpos = GEN_INT (pos);
rtx src = op0;
rtx dest = NULL, pat, seq;
enum machine_mode mode0 = insn_data[icode].operand[0].mode;
enum machine_mode mode1 = insn_data[icode].operand[1].mode;
enum machine_mode mode2 = insn_data[icode].operand[2].mode;
if (innermode == tmode || innermode == mode)
dest = target;
if (!dest)
dest = gen_reg_rtx (innermode);
start_sequence ();
if (! (*insn_data[icode].operand[0].predicate) (dest, mode0))
dest = copy_to_mode_reg (mode0, dest);
if (! (*insn_data[icode].operand[1].predicate) (src, mode1))
src = copy_to_mode_reg (mode1, src);
if (! (*insn_data[icode].operand[2].predicate) (rtxpos, mode2))
rtxpos = copy_to_mode_reg (mode1, rtxpos);
/* We could handle this, but we should always be called with a pseudo
for our targets and all insns should take them as outputs. */
if (! (*insn_data[icode].operand[0].predicate) (dest, mode0)
|| ! (*insn_data[icode].operand[1].predicate) (src, mode1)
|| ! (*insn_data[icode].operand[2].predicate) (rtxpos, mode2))
abort ();
pat = GEN_FCN (icode) (dest, src, rtxpos);
seq = get_insns ();
end_sequence ();
if (pat)
{
emit_insn (seq);
emit_insn (pat);
return dest;
}
}
/* Make sure we are playing with integral modes. Pun with subregs
if we aren't. */
{
enum machine_mode imode = int_mode_for_mode (GET_MODE (op0));
if (imode != GET_MODE (op0))
{
if (GET_CODE (op0) == MEM)
op0 = adjust_address (op0, imode, 0);
else if (imode != BLKmode)
op0 = gen_lowpart (imode, op0);
else
abort ();
}
}
/* We may be accessing data outside the field, which means
we can alias adjacent data. */
if (GET_CODE (op0) == MEM)
{
op0 = shallow_copy_rtx (op0);
set_mem_alias_set (op0, 0);
set_mem_expr (op0, 0);
}
/* Extraction of a full-word or multi-word value from a structure
in a register or aligned memory can be done with just a SUBREG.
A subword value in the least significant part of a register
can also be extracted with a SUBREG. For this, we need the
byte offset of the value in op0. */
byte_offset = bitpos / BITS_PER_UNIT + offset * UNITS_PER_WORD;
/* If OP0 is a register, BITPOS must count within a word.
But as we have it, it counts within whatever size OP0 now has.
On a bigendian machine, these are not the same, so convert. */
if (BYTES_BIG_ENDIAN
&& GET_CODE (op0) != MEM
&& unit > GET_MODE_BITSIZE (GET_MODE (op0)))
bitpos += unit - GET_MODE_BITSIZE (GET_MODE (op0));
/* ??? We currently assume TARGET is at least as big as BITSIZE.
If that's wrong, the solution is to test for it and set TARGET to 0
if needed. */
/* Only scalar integer modes can be converted via subregs. There is an
additional problem for FP modes here in that they can have a precision
which is different from the size. mode_for_size uses precision, but
we want a mode based on the size, so we must avoid calling it for FP
modes. */
mode1 = (SCALAR_INT_MODE_P (tmode)
? mode_for_size (bitsize, GET_MODE_CLASS (tmode), 0)
: mode);
if (((bitsize >= BITS_PER_WORD && bitsize == GET_MODE_BITSIZE (mode)
&& bitpos % BITS_PER_WORD == 0)
|| (mode1 != BLKmode
/* ??? The big endian test here is wrong. This is correct
if the value is in a register, and if mode_for_size is not
the same mode as op0. This causes us to get unnecessarily
inefficient code from the Thumb port when -mbig-endian. */
&& (BYTES_BIG_ENDIAN
? bitpos + bitsize == BITS_PER_WORD
: bitpos == 0)))
&& ((GET_CODE (op0) != MEM
&& TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (mode),
GET_MODE_BITSIZE (GET_MODE (op0)))
&& GET_MODE_SIZE (mode1) != 0
&& byte_offset % GET_MODE_SIZE (mode1) == 0)
|| (GET_CODE (op0) == MEM
&& (! SLOW_UNALIGNED_ACCESS (mode, MEM_ALIGN (op0))
|| (offset * BITS_PER_UNIT % bitsize == 0
&& MEM_ALIGN (op0) % bitsize == 0)))))
{
if (mode1 != GET_MODE (op0))
{
if (GET_CODE (op0) == SUBREG)
{
if (GET_MODE (SUBREG_REG (op0)) == mode1
|| GET_MODE_CLASS (mode1) == MODE_INT
|| GET_MODE_CLASS (mode1) == MODE_PARTIAL_INT)
op0 = SUBREG_REG (op0);
else
/* Else we've got some float mode source being extracted into
a different float mode destination -- this combination of
subregs results in Severe Tire Damage. */
goto no_subreg_mode_swap;
}
if (GET_CODE (op0) == REG)
op0 = gen_rtx_SUBREG (mode1, op0, byte_offset);
else
op0 = adjust_address (op0, mode1, offset);
}
if (mode1 != mode)
return convert_to_mode (tmode, op0, unsignedp);
return op0;
}
no_subreg_mode_swap:
/* Handle fields bigger than a word. */
if (bitsize > BITS_PER_WORD)
{
/* Here we transfer the words of the field
in the order least significant first.
This is because the most significant word is the one which may
be less than full. */
unsigned int nwords = (bitsize + (BITS_PER_WORD - 1)) / BITS_PER_WORD;
unsigned int i;
if (target == 0 || GET_CODE (target) != REG)
target = gen_reg_rtx (mode);
/* Indicate for flow that the entire target reg is being set. */
emit_insn (gen_rtx_CLOBBER (VOIDmode, target));
for (i = 0; i < nwords; i++)
{
/* If I is 0, use the low-order word in both field and target;
if I is 1, use the next to lowest word; and so on. */
/* Word number in TARGET to use. */
unsigned int wordnum
= (WORDS_BIG_ENDIAN
? GET_MODE_SIZE (GET_MODE (target)) / UNITS_PER_WORD - i - 1
: i);
/* Offset from start of field in OP0. */
unsigned int bit_offset = (WORDS_BIG_ENDIAN
? MAX (0, ((int) bitsize - ((int) i + 1)
* (int) BITS_PER_WORD))
: (int) i * BITS_PER_WORD);
rtx target_part = operand_subword (target, wordnum, 1, VOIDmode);
rtx result_part
= extract_bit_field (op0, MIN (BITS_PER_WORD,
bitsize - i * BITS_PER_WORD),
bitnum + bit_offset, 1, target_part, mode,
word_mode, total_size);
if (target_part == 0)
abort ();
if (result_part != target_part)
emit_move_insn (target_part, result_part);
}
if (unsignedp)
{
/* Unless we've filled TARGET, the upper regs in a multi-reg value
need to be zero'd out. */
if (GET_MODE_SIZE (GET_MODE (target)) > nwords * UNITS_PER_WORD)
{
unsigned int i, total_words;
total_words = GET_MODE_SIZE (GET_MODE (target)) / UNITS_PER_WORD;
for (i = nwords; i < total_words; i++)
emit_move_insn
(operand_subword (target,
WORDS_BIG_ENDIAN ? total_words - i - 1 : i,
1, VOIDmode),
const0_rtx);
}
return target;
}
/* Signed bit field: sign-extend with two arithmetic shifts. */
target = expand_shift (LSHIFT_EXPR, mode, target,
build_int_2 (GET_MODE_BITSIZE (mode) - bitsize, 0),
NULL_RTX, 0);
return expand_shift (RSHIFT_EXPR, mode, target,
build_int_2 (GET_MODE_BITSIZE (mode) - bitsize, 0),
NULL_RTX, 0);
}
/* From here on we know the desired field is smaller than a word. */
/* Check if there is a correspondingly-sized integer field, so we can
safely extract it as one size of integer, if necessary; then
truncate or extend to the size that is wanted; then use SUBREGs or
convert_to_mode to get one of the modes we really wanted. */
int_mode = int_mode_for_mode (tmode);
if (int_mode == BLKmode)
int_mode = int_mode_for_mode (mode);
if (int_mode == BLKmode)
abort (); /* Should probably push op0 out to memory and then
do a load. */
/* OFFSET is the number of words or bytes (UNIT says which)
from STR_RTX to the first word or byte containing part of the field. */
if (GET_CODE (op0) != MEM)
{
if (offset != 0
|| GET_MODE_SIZE (GET_MODE (op0)) > UNITS_PER_WORD)
{
if (GET_CODE (op0) != REG)
op0 = copy_to_reg (op0);
op0 = gen_rtx_SUBREG (mode_for_size (BITS_PER_WORD, MODE_INT, 0),
op0, (offset * UNITS_PER_WORD));
}
offset = 0;
}
else
op0 = protect_from_queue (str_rtx, 1);
/* Now OFFSET is nonzero only for memory operands. */
if (unsignedp)
{
if (HAVE_extzv
&& (GET_MODE_BITSIZE (extzv_mode) >= bitsize)
&& ! ((GET_CODE (op0) == REG || GET_CODE (op0) == SUBREG)
&& (bitsize + bitpos > GET_MODE_BITSIZE (extzv_mode))))
{
unsigned HOST_WIDE_INT xbitpos = bitpos, xoffset = offset;
rtx bitsize_rtx, bitpos_rtx;
rtx last = get_last_insn ();
rtx xop0 = op0;
rtx xtarget = target;
rtx xspec_target = spec_target;
rtx xspec_target_subreg = spec_target_subreg;
rtx pat;
enum machine_mode maxmode = mode_for_extraction (EP_extzv, 0);
if (GET_CODE (xop0) == MEM)
{
int save_volatile_ok = volatile_ok;
volatile_ok = 1;
/* Is the memory operand acceptable? */
if (! ((*insn_data[(int) CODE_FOR_extzv].operand[1].predicate)
(xop0, GET_MODE (xop0))))
{
/* No, load into a reg and extract from there. */
enum machine_mode bestmode;
/* Get the mode to use for inserting into this field. If
OP0 is BLKmode, get the smallest mode consistent with the
alignment. If OP0 is a non-BLKmode object that is no
wider than MAXMODE, use its mode. Otherwise, use the
smallest mode containing the field. */
if (GET_MODE (xop0) == BLKmode
|| (GET_MODE_SIZE (GET_MODE (op0))
> GET_MODE_SIZE (maxmode)))
bestmode = get_best_mode (bitsize, bitnum,
MEM_ALIGN (xop0), maxmode,
MEM_VOLATILE_P (xop0));
else
bestmode = GET_MODE (xop0);
if (bestmode == VOIDmode
|| (SLOW_UNALIGNED_ACCESS (bestmode, MEM_ALIGN (xop0))
&& GET_MODE_BITSIZE (bestmode) > MEM_ALIGN (xop0)))
goto extzv_loses;
/* Compute offset as multiple of this unit,
counting in bytes. */
unit = GET_MODE_BITSIZE (bestmode);
xoffset = (bitnum / unit) * GET_MODE_SIZE (bestmode);
xbitpos = bitnum % unit;
xop0 = adjust_address (xop0, bestmode, xoffset);
/* Fetch it to a register in that size. */
xop0 = force_reg (bestmode, xop0);
/* XBITPOS counts within UNIT, which is what is expected. */
}
else
/* Get ref to first byte containing part of the field. */
xop0 = adjust_address (xop0, byte_mode, xoffset);
volatile_ok = save_volatile_ok;
}
/* If op0 is a register, we need it in MAXMODE (which is usually
SImode). to make it acceptable to the format of extzv. */
if (GET_CODE (xop0) == SUBREG && GET_MODE (xop0) != maxmode)
goto extzv_loses;
if (GET_CODE (xop0) == REG && GET_MODE (xop0) != maxmode)
xop0 = gen_rtx_SUBREG (maxmode, xop0, 0);
/* On big-endian machines, we count bits from the most significant.
If the bit field insn does not, we must invert. */
if (BITS_BIG_ENDIAN != BYTES_BIG_ENDIAN)
xbitpos = unit - bitsize - xbitpos;
/* Now convert from counting within UNIT to counting in MAXMODE. */
if (BITS_BIG_ENDIAN && GET_CODE (xop0) != MEM)
xbitpos += GET_MODE_BITSIZE (maxmode) - unit;
unit = GET_MODE_BITSIZE (maxmode);
if (xtarget == 0
|| (flag_force_mem && GET_CODE (xtarget) == MEM))
xtarget = xspec_target = gen_reg_rtx (tmode);
if (GET_MODE (xtarget) != maxmode)
{
if (GET_CODE (xtarget) == REG)
{
int wider = (GET_MODE_SIZE (maxmode)
> GET_MODE_SIZE (GET_MODE (xtarget)));
xtarget = gen_lowpart (maxmode, xtarget);
if (wider)
xspec_target_subreg = xtarget;
}
else
xtarget = gen_reg_rtx (maxmode);
}
/* If this machine's extzv insists on a register target,
make sure we have one. */
if (! ((*insn_data[(int) CODE_FOR_extzv].operand[0].predicate)
(xtarget, maxmode)))
xtarget = gen_reg_rtx (maxmode);
bitsize_rtx = GEN_INT (bitsize);
bitpos_rtx = GEN_INT (xbitpos);
pat = gen_extzv (protect_from_queue (xtarget, 1),
xop0, bitsize_rtx, bitpos_rtx);
if (pat)
{
emit_insn (pat);
target = xtarget;
spec_target = xspec_target;
spec_target_subreg = xspec_target_subreg;
}
else
{
delete_insns_since (last);
target = extract_fixed_bit_field (int_mode, op0, offset, bitsize,
bitpos, target, 1);
}
}
else
extzv_loses:
target = extract_fixed_bit_field (int_mode, op0, offset, bitsize,
bitpos, target, 1);
}
else
{
if (HAVE_extv
&& (GET_MODE_BITSIZE (extv_mode) >= bitsize)
&& ! ((GET_CODE (op0) == REG || GET_CODE (op0) == SUBREG)
&& (bitsize + bitpos > GET_MODE_BITSIZE (extv_mode))))
{
int xbitpos = bitpos, xoffset = offset;
rtx bitsize_rtx, bitpos_rtx;
rtx last = get_last_insn ();
rtx xop0 = op0, xtarget = target;
rtx xspec_target = spec_target;
rtx xspec_target_subreg = spec_target_subreg;
rtx pat;
enum machine_mode maxmode = mode_for_extraction (EP_extv, 0);
if (GET_CODE (xop0) == MEM)
{
/* Is the memory operand acceptable? */
if (! ((*insn_data[(int) CODE_FOR_extv].operand[1].predicate)
(xop0, GET_MODE (xop0))))
{
/* No, load into a reg and extract from there. */
enum machine_mode bestmode;
/* Get the mode to use for inserting into this field. If
OP0 is BLKmode, get the smallest mode consistent with the
alignment. If OP0 is a non-BLKmode object that is no
wider than MAXMODE, use its mode. Otherwise, use the
smallest mode containing the field. */
if (GET_MODE (xop0) == BLKmode
|| (GET_MODE_SIZE (GET_MODE (op0))
> GET_MODE_SIZE (maxmode)))
bestmode = get_best_mode (bitsize, bitnum,
MEM_ALIGN (xop0), maxmode,
MEM_VOLATILE_P (xop0));
else
bestmode = GET_MODE (xop0);
if (bestmode == VOIDmode
|| (SLOW_UNALIGNED_ACCESS (bestmode, MEM_ALIGN (xop0))
&& GET_MODE_BITSIZE (bestmode) > MEM_ALIGN (xop0)))
goto extv_loses;
/* Compute offset as multiple of this unit,
counting in bytes. */
unit = GET_MODE_BITSIZE (bestmode);
xoffset = (bitnum / unit) * GET_MODE_SIZE (bestmode);
xbitpos = bitnum % unit;
xop0 = adjust_address (xop0, bestmode, xoffset);
/* Fetch it to a register in that size. */
xop0 = force_reg (bestmode, xop0);
/* XBITPOS counts within UNIT, which is what is expected. */
}
else
/* Get ref to first byte containing part of the field. */
xop0 = adjust_address (xop0, byte_mode, xoffset);
}
/* If op0 is a register, we need it in MAXMODE (which is usually
SImode) to make it acceptable to the format of extv. */
if (GET_CODE (xop0) == SUBREG && GET_MODE (xop0) != maxmode)
goto extv_loses;
if (GET_CODE (xop0) == REG && GET_MODE (xop0) != maxmode)
xop0 = gen_rtx_SUBREG (maxmode, xop0, 0);
/* On big-endian machines, we count bits from the most significant.
If the bit field insn does not, we must invert. */
if (BITS_BIG_ENDIAN != BYTES_BIG_ENDIAN)
xbitpos = unit - bitsize - xbitpos;
/* XBITPOS counts within a size of UNIT.
Adjust to count within a size of MAXMODE. */
if (BITS_BIG_ENDIAN && GET_CODE (xop0) != MEM)
xbitpos += (GET_MODE_BITSIZE (maxmode) - unit);
unit = GET_MODE_BITSIZE (maxmode);
if (xtarget == 0
|| (flag_force_mem && GET_CODE (xtarget) == MEM))
xtarget = xspec_target = gen_reg_rtx (tmode);
if (GET_MODE (xtarget) != maxmode)
{
if (GET_CODE (xtarget) == REG)
{
int wider = (GET_MODE_SIZE (maxmode)
> GET_MODE_SIZE (GET_MODE (xtarget)));
xtarget = gen_lowpart (maxmode, xtarget);
if (wider)
xspec_target_subreg = xtarget;
}
else
xtarget = gen_reg_rtx (maxmode);
}
/* If this machine's extv insists on a register target,
make sure we have one. */
if (! ((*insn_data[(int) CODE_FOR_extv].operand[0].predicate)
(xtarget, maxmode)))
xtarget = gen_reg_rtx (maxmode);
bitsize_rtx = GEN_INT (bitsize);
bitpos_rtx = GEN_INT (xbitpos);
pat = gen_extv (protect_from_queue (xtarget, 1),
xop0, bitsize_rtx, bitpos_rtx);
if (pat)
{
emit_insn (pat);
target = xtarget;
spec_target = xspec_target;
spec_target_subreg = xspec_target_subreg;
}
else
{
delete_insns_since (last);
target = extract_fixed_bit_field (int_mode, op0, offset, bitsize,
bitpos, target, 0);
}
}
else
extv_loses:
target = extract_fixed_bit_field (int_mode, op0, offset, bitsize,
bitpos, target, 0);
}
if (target == spec_target)
return target;
if (target == spec_target_subreg)
return spec_target;
if (GET_MODE (target) != tmode && GET_MODE (target) != mode)
{
/* If the target mode is floating-point, first convert to the
integer mode of that size and then access it as a floating-point
value via a SUBREG. */
if (GET_MODE_CLASS (tmode) != MODE_INT
&& GET_MODE_CLASS (tmode) != MODE_PARTIAL_INT)
{
target = convert_to_mode (mode_for_size (GET_MODE_BITSIZE (tmode),
MODE_INT, 0),
target, unsignedp);
return gen_lowpart (tmode, target);
}
else
return convert_to_mode (tmode, target, unsignedp);
}
return target;
}
/* Extract a bit field using shifts and boolean operations
Returns an rtx to represent the value.
OP0 addresses a register (word) or memory (byte).
BITPOS says which bit within the word or byte the bit field starts in.
OFFSET says how many bytes farther the bit field starts;
it is 0 if OP0 is a register.
BITSIZE says how many bits long the bit field is.
(If OP0 is a register, it may be narrower than a full word,
but BITPOS still counts within a full word,
which is significant on bigendian machines.)
UNSIGNEDP is nonzero for an unsigned bit field (don't sign-extend value).
If TARGET is nonzero, attempts to store the value there
and return TARGET, but this is not guaranteed.
If TARGET is not used, create a pseudo-reg of mode TMODE for the value. */
static rtx
extract_fixed_bit_field (enum machine_mode tmode, rtx op0,
unsigned HOST_WIDE_INT offset,
unsigned HOST_WIDE_INT bitsize,
unsigned HOST_WIDE_INT bitpos, rtx target,
int unsignedp)
{
unsigned int total_bits = BITS_PER_WORD;
enum machine_mode mode;
if (GET_CODE (op0) == SUBREG || GET_CODE (op0) == REG)
{
/* Special treatment for a bit field split across two registers. */
if (bitsize + bitpos > BITS_PER_WORD)
return extract_split_bit_field (op0, bitsize, bitpos, unsignedp);
}
else
{
/* Get the proper mode to use for this field. We want a mode that
includes the entire field. If such a mode would be larger than
a word, we won't be doing the extraction the normal way. */
mode = get_best_mode (bitsize, bitpos + offset * BITS_PER_UNIT,
MEM_ALIGN (op0), word_mode, MEM_VOLATILE_P (op0));
if (mode == VOIDmode)
/* The only way this should occur is if the field spans word
boundaries. */
return extract_split_bit_field (op0, bitsize,
bitpos + offset * BITS_PER_UNIT,
unsignedp);
total_bits = GET_MODE_BITSIZE (mode);
/* Make sure bitpos is valid for the chosen mode. Adjust BITPOS to
be in the range 0 to total_bits-1, and put any excess bytes in
OFFSET. */
if (bitpos >= total_bits)
{
offset += (bitpos / total_bits) * (total_bits / BITS_PER_UNIT);
bitpos -= ((bitpos / total_bits) * (total_bits / BITS_PER_UNIT)
* BITS_PER_UNIT);
}
/* Get ref to an aligned byte, halfword, or word containing the field.
Adjust BITPOS to be position within a word,
and OFFSET to be the offset of that word.
Then alter OP0 to refer to that word. */
bitpos += (offset % (total_bits / BITS_PER_UNIT)) * BITS_PER_UNIT;
offset -= (offset % (total_bits / BITS_PER_UNIT));
op0 = adjust_address (op0, mode, offset);
}
mode = GET_MODE (op0);
if (BYTES_BIG_ENDIAN)
/* BITPOS is the distance between our msb and that of OP0.
Convert it to the distance from the lsb. */
bitpos = total_bits - bitsize - bitpos;
/* Now BITPOS is always the distance between the field's lsb and that of OP0.
We have reduced the big-endian case to the little-endian case. */
if (unsignedp)
{
if (bitpos)
{
/* If the field does not already start at the lsb,
shift it so it does. */
tree amount = build_int_2 (bitpos, 0);
/* Maybe propagate the target for the shift. */
/* But not if we will return it--could confuse integrate.c. */
rtx subtarget = (target != 0 && GET_CODE (target) == REG
&& !REG_FUNCTION_VALUE_P (target)
? target : 0);
if (tmode != mode) subtarget = 0;
op0 = expand_shift (RSHIFT_EXPR, mode, op0, amount, subtarget, 1);
}
/* Convert the value to the desired mode. */
if (mode != tmode)
op0 = convert_to_mode (tmode, op0, 1);
/* Unless the msb of the field used to be the msb when we shifted,
mask out the upper bits. */
if (GET_MODE_BITSIZE (mode) != bitpos + bitsize)
return expand_binop (GET_MODE (op0), and_optab, op0,
mask_rtx (GET_MODE (op0), 0, bitsize, 0),
target, 1, OPTAB_LIB_WIDEN);
return op0;
}
/* To extract a signed bit-field, first shift its msb to the msb of the word,
then arithmetic-shift its lsb to the lsb of the word. */
op0 = force_reg (mode, op0);
if (mode != tmode)
target = 0;
/* Find the narrowest integer mode that contains the field. */
for (mode = GET_CLASS_NARROWEST_MODE (MODE_INT); mode != VOIDmode;
mode = GET_MODE_WIDER_MODE (mode))
if (GET_MODE_BITSIZE (mode) >= bitsize + bitpos)
{
op0 = convert_to_mode (mode, op0, 0);
break;
}
if (GET_MODE_BITSIZE (mode) != (bitsize + bitpos))
{
tree amount
= build_int_2 (GET_MODE_BITSIZE (mode) - (bitsize + bitpos), 0);
/* Maybe propagate the target for the shift. */
/* But not if we will return the result--could confuse integrate.c. */
rtx subtarget = (target != 0 && GET_CODE (target) == REG
&& ! REG_FUNCTION_VALUE_P (target)
? target : 0);
op0 = expand_shift (LSHIFT_EXPR, mode, op0, amount, subtarget, 1);
}
return expand_shift (RSHIFT_EXPR, mode, op0,
build_int_2 (GET_MODE_BITSIZE (mode) - bitsize, 0),
target, 0);
}
/* Return a constant integer (CONST_INT or CONST_DOUBLE) mask value
of mode MODE with BITSIZE ones followed by BITPOS zeros, or the
complement of that if COMPLEMENT. The mask is truncated if
necessary to the width of mode MODE. The mask is zero-extended if
BITSIZE+BITPOS is too small for MODE. */
static rtx
mask_rtx (enum machine_mode mode, int bitpos, int bitsize, int complement)
{
HOST_WIDE_INT masklow, maskhigh;
if (bitsize == 0)
masklow = 0;
else if (bitpos < HOST_BITS_PER_WIDE_INT)
masklow = (HOST_WIDE_INT) -1 << bitpos;
else
masklow = 0;
if (bitpos + bitsize < HOST_BITS_PER_WIDE_INT)
masklow &= ((unsigned HOST_WIDE_INT) -1
>> (HOST_BITS_PER_WIDE_INT - bitpos - bitsize));
if (bitpos <= HOST_BITS_PER_WIDE_INT)
maskhigh = -1;
else
maskhigh = (HOST_WIDE_INT) -1 << (bitpos - HOST_BITS_PER_WIDE_INT);
if (bitsize == 0)
maskhigh = 0;
else if (bitpos + bitsize > HOST_BITS_PER_WIDE_INT)
maskhigh &= ((unsigned HOST_WIDE_INT) -1
>> (2 * HOST_BITS_PER_WIDE_INT - bitpos - bitsize));
else
maskhigh = 0;
if (complement)
{
maskhigh = ~maskhigh;
masklow = ~masklow;
}
return immed_double_const (masklow, maskhigh, mode);
}
/* Return a constant integer (CONST_INT or CONST_DOUBLE) rtx with the value
VALUE truncated to BITSIZE bits and then shifted left BITPOS bits. */
static rtx
lshift_value (enum machine_mode mode, rtx value, int bitpos, int bitsize)
{
unsigned HOST_WIDE_INT v = INTVAL (value);
HOST_WIDE_INT low, high;
if (bitsize < HOST_BITS_PER_WIDE_INT)
v &= ~((HOST_WIDE_INT) -1 << bitsize);
if (bitpos < HOST_BITS_PER_WIDE_INT)
{
low = v << bitpos;
high = (bitpos > 0 ? (v >> (HOST_BITS_PER_WIDE_INT - bitpos)) : 0);
}
else
{
low = 0;
high = v << (bitpos - HOST_BITS_PER_WIDE_INT);
}
return immed_double_const (low, high, mode);
}
/* Extract a bit field that is split across two words
and return an RTX for the result.
OP0 is the REG, SUBREG or MEM rtx for the first of the two words.
BITSIZE is the field width; BITPOS, position of its first bit, in the word.
UNSIGNEDP is 1 if should zero-extend the contents; else sign-extend. */
static rtx
extract_split_bit_field (rtx op0, unsigned HOST_WIDE_INT bitsize,
unsigned HOST_WIDE_INT bitpos, int unsignedp)
{
unsigned int unit;
unsigned int bitsdone = 0;
rtx result = NULL_RTX;
int first = 1;
/* Make sure UNIT isn't larger than BITS_PER_WORD, we can only handle that
much at a time. */
if (GET_CODE (op0) == REG || GET_CODE (op0) == SUBREG)
unit = BITS_PER_WORD;
else
unit = MIN (MEM_ALIGN (op0), BITS_PER_WORD);
while (bitsdone < bitsize)
{
unsigned HOST_WIDE_INT thissize;
rtx part, word;
unsigned HOST_WIDE_INT thispos;
unsigned HOST_WIDE_INT offset;
offset = (bitpos + bitsdone) / unit;
thispos = (bitpos + bitsdone) % unit;
/* THISSIZE must not overrun a word boundary. Otherwise,
extract_fixed_bit_field will call us again, and we will mutually
recurse forever. */
thissize = MIN (bitsize - bitsdone, BITS_PER_WORD);
thissize = MIN (thissize, unit - thispos);
/* If OP0 is a register, then handle OFFSET here.
When handling multiword bitfields, extract_bit_field may pass
down a word_mode SUBREG of a larger REG for a bitfield that actually
crosses a word boundary. Thus, for a SUBREG, we must find
the current word starting from the base register. */
if (GET_CODE (op0) == SUBREG)
{
int word_offset = (SUBREG_BYTE (op0) / UNITS_PER_WORD) + offset;
word = operand_subword_force (SUBREG_REG (op0), word_offset,
GET_MODE (SUBREG_REG (op0)));
offset = 0;
}
else if (GET_CODE (op0) == REG)
{
word = operand_subword_force (op0, offset, GET_MODE (op0));
offset = 0;
}
else
word = op0;
/* Extract the parts in bit-counting order,
whose meaning is determined by BYTES_PER_UNIT.
OFFSET is in UNITs, and UNIT is in bits.
extract_fixed_bit_field wants offset in bytes. */
part = extract_fixed_bit_field (word_mode, word,
offset * unit / BITS_PER_UNIT,
thissize, thispos, 0, 1);
bitsdone += thissize;
/* Shift this part into place for the result. */
if (BYTES_BIG_ENDIAN)
{
if (bitsize != bitsdone)
part = expand_shift (LSHIFT_EXPR, word_mode, part,
build_int_2 (bitsize - bitsdone, 0), 0, 1);
}
else
{
if (bitsdone != thissize)
part = expand_shift (LSHIFT_EXPR, word_mode, part,
build_int_2 (bitsdone - thissize, 0), 0, 1);
}
if (first)
result = part;
else
/* Combine the parts with bitwise or. This works
because we extracted each part as an unsigned bit field. */
result = expand_binop (word_mode, ior_optab, part, result, NULL_RTX, 1,
OPTAB_LIB_WIDEN);
first = 0;
}
/* Unsigned bit field: we are done. */
if (unsignedp)
return result;
/* Signed bit field: sign-extend with two arithmetic shifts. */
result = expand_shift (LSHIFT_EXPR, word_mode, result,
build_int_2 (BITS_PER_WORD - bitsize, 0),
NULL_RTX, 0);
return expand_shift (RSHIFT_EXPR, word_mode, result,
build_int_2 (BITS_PER_WORD - bitsize, 0), NULL_RTX, 0);
}
/* Add INC into TARGET. */
void
expand_inc (rtx target, rtx inc)
{
rtx value = expand_binop (GET_MODE (target), add_optab,
target, inc,
target, 0, OPTAB_LIB_WIDEN);
if (value != target)
emit_move_insn (target, value);
}
/* Subtract DEC from TARGET. */
void
expand_dec (rtx target, rtx dec)
{
rtx value = expand_binop (GET_MODE (target), sub_optab,
target, dec,
target, 0, OPTAB_LIB_WIDEN);
if (value != target)
emit_move_insn (target, value);
}
/* Output a shift instruction for expression code CODE,
with SHIFTED being the rtx for the value to shift,
and AMOUNT the tree for the amount to shift by.
Store the result in the rtx TARGET, if that is convenient.
If UNSIGNEDP is nonzero, do a logical shift; otherwise, arithmetic.
Return the rtx for where the value is. */
rtx
expand_shift (enum tree_code code, enum machine_mode mode, rtx shifted,
tree amount, rtx target, int unsignedp)
{
rtx op1, temp = 0;
int left = (code == LSHIFT_EXPR || code == LROTATE_EXPR);
int rotate = (code == LROTATE_EXPR || code == RROTATE_EXPR);
int try;
/* Previously detected shift-counts computed by NEGATE_EXPR
and shifted in the other direction; but that does not work
on all machines. */
op1 = expand_expr (amount, NULL_RTX, VOIDmode, 0);
#ifdef SHIFT_COUNT_TRUNCATED
if (SHIFT_COUNT_TRUNCATED)
{
if (GET_CODE (op1) == CONST_INT
&& ((unsigned HOST_WIDE_INT) INTVAL (op1) >=
(unsigned HOST_WIDE_INT) GET_MODE_BITSIZE (mode)))
op1 = GEN_INT ((unsigned HOST_WIDE_INT) INTVAL (op1)
% GET_MODE_BITSIZE (mode));
else if (GET_CODE (op1) == SUBREG
&& subreg_lowpart_p (op1))
op1 = SUBREG_REG (op1);
}
#endif
if (op1 == const0_rtx)
return shifted;
for (try = 0; temp == 0 && try < 3; try++)
{
enum optab_methods methods;
if (try == 0)
methods = OPTAB_DIRECT;
else if (try == 1)
methods = OPTAB_WIDEN;
else
methods = OPTAB_LIB_WIDEN;
if (rotate)
{
/* Widening does not work for rotation. */
if (methods == OPTAB_WIDEN)
continue;
else if (methods == OPTAB_LIB_WIDEN)
{
/* If we have been unable to open-code this by a rotation,
do it as the IOR of two shifts. I.e., to rotate A
by N bits, compute (A << N) | ((unsigned) A >> (C - N))
where C is the bitsize of A.
It is theoretically possible that the target machine might
not be able to perform either shift and hence we would
be making two libcalls rather than just the one for the
shift (similarly if IOR could not be done). We will allow
this extremely unlikely lossage to avoid complicating the
code below. */
rtx subtarget = target == shifted ? 0 : target;
rtx temp1;
tree type = TREE_TYPE (amount);
tree new_amount = make_tree (type, op1);
tree other_amount
= fold (build (MINUS_EXPR, type,
convert (type,
build_int_2 (GET_MODE_BITSIZE (mode),
0)),
amount));
shifted = force_reg (mode, shifted);
temp = expand_shift (left ? LSHIFT_EXPR : RSHIFT_EXPR,
mode, shifted, new_amount, subtarget, 1);
temp1 = expand_shift (left ? RSHIFT_EXPR : LSHIFT_EXPR,
mode, shifted, other_amount, 0, 1);
return expand_binop (mode, ior_optab, temp, temp1, target,
unsignedp, methods);
}
temp = expand_binop (mode,
left ? rotl_optab : rotr_optab,
shifted, op1, target, unsignedp, methods);
/* If we don't have the rotate, but we are rotating by a constant
that is in range, try a rotate in the opposite direction. */
if (temp == 0 && GET_CODE (op1) == CONST_INT
&& INTVAL (op1) > 0
&& (unsigned int) INTVAL (op1) < GET_MODE_BITSIZE (mode))
temp = expand_binop (mode,
left ? rotr_optab : rotl_optab,
shifted,
GEN_INT (GET_MODE_BITSIZE (mode)
- INTVAL (op1)),
target, unsignedp, methods);
}
else if (unsignedp)
temp = expand_binop (mode,
left ? ashl_optab : lshr_optab,
shifted, op1, target, unsignedp, methods);
/* Do arithmetic shifts.
Also, if we are going to widen the operand, we can just as well
use an arithmetic right-shift instead of a logical one. */
if (temp == 0 && ! rotate
&& (! unsignedp || (! left && methods == OPTAB_WIDEN)))
{
enum optab_methods methods1 = methods;
/* If trying to widen a log shift to an arithmetic shift,
don't accept an arithmetic shift of the same size. */
if (unsignedp)
methods1 = OPTAB_MUST_WIDEN;
/* Arithmetic shift */
temp = expand_binop (mode,
left ? ashl_optab : ashr_optab,
shifted, op1, target, unsignedp, methods1);
}
/* We used to try extzv here for logical right shifts, but that was
only useful for one machine, the VAX, and caused poor code
generation there for lshrdi3, so the code was deleted and a
define_expand for lshrsi3 was added to vax.md. */
}
if (temp == 0)
abort ();
return temp;
}
enum alg_code { alg_zero, alg_m, alg_shift,
alg_add_t_m2, alg_sub_t_m2,
alg_add_factor, alg_sub_factor,
alg_add_t2_m, alg_sub_t2_m,
alg_add, alg_subtract, alg_factor, alg_shiftop };
/* This structure records a sequence of operations.
`ops' is the number of operations recorded.
`cost' is their total cost.
The operations are stored in `op' and the corresponding
logarithms of the integer coefficients in `log'.
These are the operations:
alg_zero total := 0;
alg_m total := multiplicand;
alg_shift total := total * coeff
alg_add_t_m2 total := total + multiplicand * coeff;
alg_sub_t_m2 total := total - multiplicand * coeff;
alg_add_factor total := total * coeff + total;
alg_sub_factor total := total * coeff - total;
alg_add_t2_m total := total * coeff + multiplicand;
alg_sub_t2_m total := total * coeff - multiplicand;
The first operand must be either alg_zero or alg_m. */
struct algorithm
{
short cost;
short ops;
/* The size of the OP and LOG fields are not directly related to the
word size, but the worst-case algorithms will be if we have few
consecutive ones or zeros, i.e., a multiplicand like 10101010101...
In that case we will generate shift-by-2, add, shift-by-2, add,...,
in total wordsize operations. */
enum alg_code op[MAX_BITS_PER_WORD];
char log[MAX_BITS_PER_WORD];
};
static void synth_mult (struct algorithm *, unsigned HOST_WIDE_INT, int);
static unsigned HOST_WIDE_INT choose_multiplier (unsigned HOST_WIDE_INT, int,
int, unsigned HOST_WIDE_INT *,
int *, int *);
static unsigned HOST_WIDE_INT invert_mod2n (unsigned HOST_WIDE_INT, int);
/* Compute and return the best algorithm for multiplying by T.
The algorithm must cost less than cost_limit
If retval.cost >= COST_LIMIT, no algorithm was found and all
other field of the returned struct are undefined. */
static void
synth_mult (struct algorithm *alg_out, unsigned HOST_WIDE_INT t,
int cost_limit)
{
int m;
struct algorithm *alg_in, *best_alg;
int cost;
unsigned HOST_WIDE_INT q;
/* Indicate that no algorithm is yet found. If no algorithm
is found, this value will be returned and indicate failure. */
alg_out->cost = cost_limit;
if (cost_limit <= 0)
return;
/* t == 1 can be done in zero cost. */
if (t == 1)
{
alg_out->ops = 1;
alg_out->cost = 0;
alg_out->op[0] = alg_m;
return;
}
/* t == 0 sometimes has a cost. If it does and it exceeds our limit,
fail now. */
if (t == 0)
{
if (zero_cost >= cost_limit)
return;
else
{
alg_out->ops = 1;
alg_out->cost = zero_cost;
alg_out->op[0] = alg_zero;
return;
}
}
/* We'll be needing a couple extra algorithm structures now. */
alg_in = alloca (sizeof (struct algorithm));
best_alg = alloca (sizeof (struct algorithm));
/* If we have a group of zero bits at the low-order part of T, try
multiplying by the remaining bits and then doing a shift. */
if ((t & 1) == 0)
{
m = floor_log2 (t & -t); /* m = number of low zero bits */
if (m < BITS_PER_WORD)
{
q = t >> m;
cost = shift_cost[m];
synth_mult (alg_in, q, cost_limit - cost);
cost += alg_in->cost;
if (cost < cost_limit)
{
struct algorithm *x;
x = alg_in, alg_in = best_alg, best_alg = x;
best_alg->log[best_alg->ops] = m;
best_alg->op[best_alg->ops] = alg_shift;
cost_limit = cost;
}
}
}
/* If we have an odd number, add or subtract one. */
if ((t & 1) != 0)
{
unsigned HOST_WIDE_INT w;
for (w = 1; (w & t) != 0; w <<= 1)
;
/* If T was -1, then W will be zero after the loop. This is another
case where T ends with ...111. Handling this with (T + 1) and
subtract 1 produces slightly better code and results in algorithm
selection much faster than treating it like the ...0111 case
below. */
if (w == 0
|| (w > 2
/* Reject the case where t is 3.
Thus we prefer addition in that case. */
&& t != 3))
{
/* T ends with ...111. Multiply by (T + 1) and subtract 1. */
cost = add_cost;
synth_mult (alg_in, t + 1, cost_limit - cost);
cost += alg_in->cost;
if (cost < cost_limit)
{
struct algorithm *x;
x = alg_in, alg_in = best_alg, best_alg = x;
best_alg->log[best_alg->ops] = 0;
best_alg->op[best_alg->ops] = alg_sub_t_m2;
cost_limit = cost;
}
}
else
{
/* T ends with ...01 or ...011. Multiply by (T - 1) and add 1. */
cost = add_cost;
synth_mult (alg_in, t - 1, cost_limit - cost);
cost += alg_in->cost;
if (cost < cost_limit)
{
struct algorithm *x;
x = alg_in, alg_in = best_alg, best_alg = x;
best_alg->log[best_alg->ops] = 0;
best_alg->op[best_alg->ops] = alg_add_t_m2;
cost_limit = cost;
}
}
}
/* Look for factors of t of the form
t = q(2**m +- 1), 2 <= m <= floor(log2(t - 1)).
If we find such a factor, we can multiply by t using an algorithm that
multiplies by q, shift the result by m and add/subtract it to itself.
We search for large factors first and loop down, even if large factors
are less probable than small; if we find a large factor we will find a
good sequence quickly, and therefore be able to prune (by decreasing
COST_LIMIT) the search. */
for (m = floor_log2 (t - 1); m >= 2; m--)
{
unsigned HOST_WIDE_INT d;
d = ((unsigned HOST_WIDE_INT) 1 << m) + 1;
if (t % d == 0 && t > d && m < BITS_PER_WORD)
{
cost = MIN (shiftadd_cost[m], add_cost + shift_cost[m]);
synth_mult (alg_in, t / d, cost_limit - cost);
cost += alg_in->cost;
if (cost < cost_limit)
{
struct algorithm *x;
x = alg_in, alg_in = best_alg, best_alg = x;
best_alg->log[best_alg->ops] = m;
best_alg->op[best_alg->ops] = alg_add_factor;
cost_limit = cost;
}
/* Other factors will have been taken care of in the recursion. */
break;
}
d = ((unsigned HOST_WIDE_INT) 1 << m) - 1;
if (t % d == 0 && t > d && m < BITS_PER_WORD)
{
cost = MIN (shiftsub_cost[m], add_cost + shift_cost[m]);
synth_mult (alg_in, t / d, cost_limit - cost);
cost += alg_in->cost;
if (cost < cost_limit)
{
struct algorithm *x;
x = alg_in, alg_in = best_alg, best_alg = x;
best_alg->log[best_alg->ops] = m;
best_alg->op[best_alg->ops] = alg_sub_factor;
cost_limit = cost;
}
break;
}
}
/* Try shift-and-add (load effective address) instructions,
i.e. do a*3, a*5, a*9. */
if ((t & 1) != 0)
{
q = t - 1;
q = q & -q;
m = exact_log2 (q);
if (m >= 0 && m < BITS_PER_WORD)
{
cost = shiftadd_cost[m];
synth_mult (alg_in, (t - 1) >> m, cost_limit - cost);
cost += alg_in->cost;
if (cost < cost_limit)
{
struct algorithm *x;
x = alg_in, alg_in = best_alg, best_alg = x;
best_alg->log[best_alg->ops] = m;
best_alg->op[best_alg->ops] = alg_add_t2_m;
cost_limit = cost;
}
}
q = t + 1;
q = q & -q;
m = exact_log2 (q);
if (m >= 0 && m < BITS_PER_WORD)
{
cost = shiftsub_cost[m];
synth_mult (alg_in, (t + 1) >> m, cost_limit - cost);
cost += alg_in->cost;
if (cost < cost_limit)
{
struct algorithm *x;
x = alg_in, alg_in = best_alg, best_alg = x;
best_alg->log[best_alg->ops] = m;
best_alg->op[best_alg->ops] = alg_sub_t2_m;
cost_limit = cost;
}
}
}
/* If cost_limit has not decreased since we stored it in alg_out->cost,
we have not found any algorithm. */
if (cost_limit == alg_out->cost)
return;
/* If we are getting a too long sequence for `struct algorithm'
to record, make this search fail. */
if (best_alg->ops == MAX_BITS_PER_WORD)
return;
/* Copy the algorithm from temporary space to the space at alg_out.
We avoid using structure assignment because the majority of
best_alg is normally undefined, and this is a critical function. */
alg_out->ops = best_alg->ops + 1;
alg_out->cost = cost_limit;
memcpy (alg_out->op, best_alg->op,
alg_out->ops * sizeof *alg_out->op);
memcpy (alg_out->log, best_alg->log,
alg_out->ops * sizeof *alg_out->log);
}
/* Perform a multiplication and return an rtx for the result.
MODE is mode of value; OP0 and OP1 are what to multiply (rtx's);
TARGET is a suggestion for where to store the result (an rtx).
We check specially for a constant integer as OP1.
If you want this check for OP0 as well, then before calling
you should swap the two operands if OP0 would be constant. */
rtx
expand_mult (enum machine_mode mode, rtx op0, rtx op1, rtx target,
int unsignedp)
{
rtx const_op1 = op1;
/* synth_mult does an `unsigned int' multiply. As long as the mode is
less than or equal in size to `unsigned int' this doesn't matter.
If the mode is larger than `unsigned int', then synth_mult works only
if the constant value exactly fits in an `unsigned int' without any
truncation. This means that multiplying by negative values does
not work; results are off by 2^32 on a 32 bit machine. */
/* If we are multiplying in DImode, it may still be a win
to try to work with shifts and adds. */
if (GET_CODE (op1) == CONST_DOUBLE
&& GET_MODE_CLASS (GET_MODE (op1)) == MODE_INT
&& HOST_BITS_PER_INT >= BITS_PER_WORD
&& CONST_DOUBLE_HIGH (op1) == 0)
const_op1 = GEN_INT (CONST_DOUBLE_LOW (op1));
else if (HOST_BITS_PER_INT < GET_MODE_BITSIZE (mode)
&& GET_CODE (op1) == CONST_INT
&& INTVAL (op1) < 0)
const_op1 = 0;
/* We used to test optimize here, on the grounds that it's better to
produce a smaller program when -O is not used.
But this causes such a terrible slowdown sometimes
that it seems better to use synth_mult always. */
if (const_op1 && GET_CODE (const_op1) == CONST_INT
&& (unsignedp || ! flag_trapv))
{
struct algorithm alg;
struct algorithm alg2;
HOST_WIDE_INT val = INTVAL (op1);
HOST_WIDE_INT val_so_far;
rtx insn;
int mult_cost;
enum {basic_variant, negate_variant, add_variant} variant = basic_variant;
/* op0 must be register to make mult_cost match the precomputed
shiftadd_cost array. */
op0 = force_reg (mode, op0);
/* Try to do the computation three ways: multiply by the negative of OP1
and then negate, do the multiplication directly, or do multiplication
by OP1 - 1. */
mult_cost = rtx_cost (gen_rtx_MULT (mode, op0, op1), SET);
mult_cost = MIN (12 * add_cost, mult_cost);
synth_mult (&alg, val, mult_cost);
/* This works only if the inverted value actually fits in an
`unsigned int' */
if (HOST_BITS_PER_INT >= GET_MODE_BITSIZE (mode))
{
synth_mult (&alg2, - val,
(alg.cost < mult_cost ? alg.cost : mult_cost) - negate_cost);
if (alg2.cost + negate_cost < alg.cost)
alg = alg2, variant = negate_variant;
}
/* This proves very useful for division-by-constant. */
synth_mult (&alg2, val - 1,
(alg.cost < mult_cost ? alg.cost : mult_cost) - add_cost);
if (alg2.cost + add_cost < alg.cost)
alg = alg2, variant = add_variant;
if (alg.cost < mult_cost)
{
/* We found something cheaper than a multiply insn. */
int opno;
rtx accum, tem;
enum machine_mode nmode;
op0 = protect_from_queue (op0, 0);
/* Avoid referencing memory over and over.
For speed, but also for correctness when mem is volatile. */
if (GET_CODE (op0) == MEM)
op0 = force_reg (mode, op0);
/* ACCUM starts out either as OP0 or as a zero, depending on
the first operation. */
if (alg.op[0] == alg_zero)
{
accum = copy_to_mode_reg (mode, const0_rtx);
val_so_far = 0;
}
else if (alg.op[0] == alg_m)
{
accum = copy_to_mode_reg (mode, op0);
val_so_far = 1;
}
else
abort ();
for (opno = 1; opno < alg.ops; opno++)
{
int log = alg.log[opno];
int preserve = preserve_subexpressions_p ();
rtx shift_subtarget = preserve ? 0 : accum;
rtx add_target
= (opno == alg.ops - 1 && target != 0 && variant != add_variant
&& ! preserve)
? target : 0;
rtx accum_target = preserve ? 0 : accum;
switch (alg.op[opno])
{
case alg_shift:
accum = expand_shift (LSHIFT_EXPR, mode, accum,
build_int_2 (log, 0), NULL_RTX, 0);
val_so_far <<= log;
break;
case alg_add_t_m2:
tem = expand_shift (LSHIFT_EXPR, mode, op0,
build_int_2 (log, 0), NULL_RTX, 0);
accum = force_operand (gen_rtx_PLUS (mode, accum, tem),
add_target
? add_target : accum_target);
val_so_far += (HOST_WIDE_INT) 1 << log;
break;
case alg_sub_t_m2:
tem = expand_shift (LSHIFT_EXPR, mode, op0,
build_int_2 (log, 0), NULL_RTX, 0);
accum = force_operand (gen_rtx_MINUS (mode, accum, tem),
add_target
? add_target : accum_target);
val_so_far -= (HOST_WIDE_INT) 1 << log;
break;
case alg_add_t2_m:
accum = expand_shift (LSHIFT_EXPR, mode, accum,
build_int_2 (log, 0), shift_subtarget,
0);
accum = force_operand (gen_rtx_PLUS (mode, accum, op0),
add_target
? add_target : accum_target);
val_so_far = (val_so_far << log) + 1;
break;
case alg_sub_t2_m:
accum = expand_shift (LSHIFT_EXPR, mode, accum,
build_int_2 (log, 0), shift_subtarget,
0);
accum = force_operand (gen_rtx_MINUS (mode, accum, op0),
add_target
? add_target : accum_target);
val_so_far = (val_so_far << log) - 1;
break;
case alg_add_factor:
tem = expand_shift (LSHIFT_EXPR, mode, accum,
build_int_2 (log, 0), NULL_RTX, 0);
accum = force_operand (gen_rtx_PLUS (mode, accum, tem),
add_target
? add_target : accum_target);
val_so_far += val_so_far << log;
break;
case alg_sub_factor:
tem = expand_shift (LSHIFT_EXPR, mode, accum,
build_int_2 (log, 0), NULL_RTX, 0);
accum = force_operand (gen_rtx_MINUS (mode, tem, accum),
(add_target ? add_target
: preserve ? 0 : tem));
val_so_far = (val_so_far << log) - val_so_far;
break;
default:
abort ();
}
/* Write a REG_EQUAL note on the last insn so that we can cse
multiplication sequences. Note that if ACCUM is a SUBREG,
we've set the inner register and must properly indicate
that. */
tem = op0, nmode = mode;
if (GET_CODE (accum) == SUBREG)
{
nmode = GET_MODE (SUBREG_REG (accum));
tem = gen_lowpart (nmode, op0);
}
insn = get_last_insn ();
set_unique_reg_note (insn,
REG_EQUAL,
gen_rtx_MULT (nmode, tem,
GEN_INT (val_so_far)));
}
if (variant == negate_variant)
{
val_so_far = - val_so_far;
accum = expand_unop (mode, neg_optab, accum, target, 0);
}
else if (variant == add_variant)
{
val_so_far = val_so_far + 1;
accum = force_operand (gen_rtx_PLUS (mode, accum, op0), target);
}
if (val != val_so_far)
abort ();
return accum;
}
}
if (GET_CODE (op0) == CONST_DOUBLE)
{
rtx temp = op0;
op0 = op1;
op1 = temp;
}
/* Expand x*2.0 as x+x. */
if (GET_CODE (op1) == CONST_DOUBLE
&& GET_MODE_CLASS (mode) == MODE_FLOAT)
{
REAL_VALUE_TYPE d;
REAL_VALUE_FROM_CONST_DOUBLE (d, op1);
if (REAL_VALUES_EQUAL (d, dconst2))
{
op0 = force_reg (GET_MODE (op0), op0);
return expand_binop (mode, add_optab, op0, op0,
target, unsignedp, OPTAB_LIB_WIDEN);
}
}
/* This used to use umul_optab if unsigned, but for non-widening multiply
there is no difference between signed and unsigned. */
op0 = expand_binop (mode,
! unsignedp
&& flag_trapv && (GET_MODE_CLASS(mode) == MODE_INT)
? smulv_optab : smul_optab,
op0, op1, target, unsignedp, OPTAB_LIB_WIDEN);
if (op0 == 0)
abort ();
return op0;
}
/* Return the smallest n such that 2**n >= X. */
int
ceil_log2 (unsigned HOST_WIDE_INT x)
{
return floor_log2 (x - 1) + 1;
}
/* Choose a minimal N + 1 bit approximation to 1/D that can be used to
replace division by D, and put the least significant N bits of the result
in *MULTIPLIER_PTR and return the most significant bit.
The width of operations is N (should be <= HOST_BITS_PER_WIDE_INT), the
needed precision is in PRECISION (should be <= N).
PRECISION should be as small as possible so this function can choose
multiplier more freely.
The rounded-up logarithm of D is placed in *lgup_ptr. A shift count that
is to be used for a final right shift is placed in *POST_SHIFT_PTR.
Using this function, x/D will be equal to (x * m) >> (*POST_SHIFT_PTR),
where m is the full HOST_BITS_PER_WIDE_INT + 1 bit multiplier. */
static
unsigned HOST_WIDE_INT
choose_multiplier (unsigned HOST_WIDE_INT d, int n, int precision,
unsigned HOST_WIDE_INT *multiplier_ptr,
int *post_shift_ptr, int *lgup_ptr)
{
HOST_WIDE_INT mhigh_hi, mlow_hi;
unsigned HOST_WIDE_INT mhigh_lo, mlow_lo;
int lgup, post_shift;
int pow, pow2;
unsigned HOST_WIDE_INT nl, dummy1;
HOST_WIDE_INT nh, dummy2;
/* lgup = ceil(log2(divisor)); */
lgup = ceil_log2 (d);
if (lgup > n)
abort ();
pow = n + lgup;
pow2 = n + lgup - precision;
if (pow == 2 * HOST_BITS_PER_WIDE_INT)
{
/* We could handle this with some effort, but this case is much better
handled directly with a scc insn, so rely on caller using that. */
abort ();
}
/* mlow = 2^(N + lgup)/d */
if (pow >= HOST_BITS_PER_WIDE_INT)
{
nh = (HOST_WIDE_INT) 1 << (pow - HOST_BITS_PER_WIDE_INT);
nl = 0;
}
else
{
nh = 0;
nl = (unsigned HOST_WIDE_INT) 1 << pow;
}
div_and_round_double (TRUNC_DIV_EXPR, 1, nl, nh, d, (HOST_WIDE_INT) 0,
&mlow_lo, &mlow_hi, &dummy1, &dummy2);
/* mhigh = (2^(N + lgup) + 2^N + lgup - precision)/d */
if (pow2 >= HOST_BITS_PER_WIDE_INT)
nh |= (HOST_WIDE_INT) 1 << (pow2 - HOST_BITS_PER_WIDE_INT);
else
nl |= (unsigned HOST_WIDE_INT) 1 << pow2;
div_and_round_double (TRUNC_DIV_EXPR, 1, nl, nh, d, (HOST_WIDE_INT) 0,
&mhigh_lo, &mhigh_hi, &dummy1, &dummy2);
if (mhigh_hi && nh - d >= d)
abort ();
if (mhigh_hi > 1 || mlow_hi > 1)
abort ();
/* Assert that mlow < mhigh. */
if (! (mlow_hi < mhigh_hi || (mlow_hi == mhigh_hi && mlow_lo < mhigh_lo)))
abort ();
/* If precision == N, then mlow, mhigh exceed 2^N
(but they do not exceed 2^(N+1)). */
/* Reduce to lowest terms. */
for (post_shift = lgup; post_shift > 0; post_shift--)
{
unsigned HOST_WIDE_INT ml_lo = (mlow_hi << (HOST_BITS_PER_WIDE_INT - 1)) | (mlow_lo >> 1);
unsigned HOST_WIDE_INT mh_lo = (mhigh_hi << (HOST_BITS_PER_WIDE_INT - 1)) | (mhigh_lo >> 1);
if (ml_lo >= mh_lo)
break;
mlow_hi = 0;
mlow_lo = ml_lo;
mhigh_hi = 0;
mhigh_lo = mh_lo;
}
*post_shift_ptr = post_shift;
*lgup_ptr = lgup;
if (n < HOST_BITS_PER_WIDE_INT)
{
unsigned HOST_WIDE_INT mask = ((unsigned HOST_WIDE_INT) 1 << n) - 1;
*multiplier_ptr = mhigh_lo & mask;
return mhigh_lo >= mask;
}
else
{
*multiplier_ptr = mhigh_lo;
return mhigh_hi;
}
}
/* Compute the inverse of X mod 2**n, i.e., find Y such that X * Y is
congruent to 1 (mod 2**N). */
static unsigned HOST_WIDE_INT
invert_mod2n (unsigned HOST_WIDE_INT x, int n)
{
/* Solve x*y == 1 (mod 2^n), where x is odd. Return y. */
/* The algorithm notes that the choice y = x satisfies
x*y == 1 mod 2^3, since x is assumed odd.
Each iteration doubles the number of bits of significance in y. */
unsigned HOST_WIDE_INT mask;
unsigned HOST_WIDE_INT y = x;
int nbit = 3;
mask = (n == HOST_BITS_PER_WIDE_INT
? ~(unsigned HOST_WIDE_INT) 0
: ((unsigned HOST_WIDE_INT) 1 << n) - 1);
while (nbit < n)
{
y = y * (2 - x*y) & mask; /* Modulo 2^N */
nbit *= 2;
}
return y;
}
/* Emit code to adjust ADJ_OPERAND after multiplication of wrong signedness
flavor of OP0 and OP1. ADJ_OPERAND is already the high half of the
product OP0 x OP1. If UNSIGNEDP is nonzero, adjust the signed product
to become unsigned, if UNSIGNEDP is zero, adjust the unsigned product to
become signed.
The result is put in TARGET if that is convenient.
MODE is the mode of operation. */
rtx
expand_mult_highpart_adjust (enum machine_mode mode, rtx adj_operand, rtx op0,
rtx op1, rtx target, int unsignedp)
{
rtx tem;
enum rtx_code adj_code = unsignedp ? PLUS : MINUS;
tem = expand_shift (RSHIFT_EXPR, mode, op0,
build_int_2 (GET_MODE_BITSIZE (mode) - 1, 0),
NULL_RTX, 0);
tem = expand_and (mode, tem, op1, NULL_RTX);
adj_operand
= force_operand (gen_rtx_fmt_ee (adj_code, mode, adj_operand, tem),
adj_operand);
tem = expand_shift (RSHIFT_EXPR, mode, op1,
build_int_2 (GET_MODE_BITSIZE (mode) - 1, 0),
NULL_RTX, 0);
tem = expand_and (mode, tem, op0, NULL_RTX);
target = force_operand (gen_rtx_fmt_ee (adj_code, mode, adj_operand, tem),
target);
return target;
}
/* Emit code to multiply OP0 and CNST1, putting the high half of the result
in TARGET if that is convenient, and return where the result is. If the
operation can not be performed, 0 is returned.
MODE is the mode of operation and result.
UNSIGNEDP nonzero means unsigned multiply.
MAX_COST is the total allowed cost for the expanded RTL. */
rtx
expand_mult_highpart (enum machine_mode mode, rtx op0,
unsigned HOST_WIDE_INT cnst1, rtx target,
int unsignedp, int max_cost)
{
enum machine_mode wider_mode = GET_MODE_WIDER_MODE (mode);
optab mul_highpart_optab;
optab moptab;
rtx tem;
int size = GET_MODE_BITSIZE (mode);
rtx op1, wide_op1;
/* We can't support modes wider than HOST_BITS_PER_INT. */
if (size > HOST_BITS_PER_WIDE_INT)
abort ();
op1 = gen_int_mode (cnst1, mode);
wide_op1
= immed_double_const (cnst1,
(unsignedp
? (HOST_WIDE_INT) 0
: -(cnst1 >> (HOST_BITS_PER_WIDE_INT - 1))),
wider_mode);
/* expand_mult handles constant multiplication of word_mode
or narrower. It does a poor job for large modes. */
if (size < BITS_PER_WORD
&& mul_cost[(int) wider_mode] + shift_cost[size-1] < max_cost)
{
/* We have to do this, since expand_binop doesn't do conversion for
multiply. Maybe change expand_binop to handle widening multiply? */
op0 = convert_to_mode (wider_mode, op0, unsignedp);
/* We know that this can't have signed overflow, so pretend this is
an unsigned multiply. */
tem = expand_mult (wider_mode, op0, wide_op1, NULL_RTX, 0);
tem = expand_shift (RSHIFT_EXPR, wider_mode, tem,
build_int_2 (size, 0), NULL_RTX, 1);
return convert_modes (mode, wider_mode, tem, unsignedp);
}
if (target == 0)
target = gen_reg_rtx (mode);
/* Firstly, try using a multiplication insn that only generates the needed
high part of the product, and in the sign flavor of unsignedp. */
if (mul_highpart_cost[(int) mode] < max_cost)
{
mul_highpart_optab = unsignedp ? umul_highpart_optab : smul_highpart_optab;
target = expand_binop (mode, mul_highpart_optab,
op0, op1, target, unsignedp, OPTAB_DIRECT);
if (target)
return target;
}
/* Secondly, same as above, but use sign flavor opposite of unsignedp.
Need to adjust the result after the multiplication. */
if (size - 1 < BITS_PER_WORD
&& (mul_highpart_cost[(int) mode] + 2 * shift_cost[size-1] + 4 * add_cost
< max_cost))
{
mul_highpart_optab = unsignedp ? smul_highpart_optab : umul_highpart_optab;
target = expand_binop (mode, mul_highpart_optab,
op0, op1, target, unsignedp, OPTAB_DIRECT);
if (target)
/* We used the wrong signedness. Adjust the result. */
return expand_mult_highpart_adjust (mode, target, op0,
op1, target, unsignedp);
}
/* Try widening multiplication. */
moptab = unsignedp ? umul_widen_optab : smul_widen_optab;
if (moptab->handlers[(int) wider_mode].insn_code != CODE_FOR_nothing
&& mul_widen_cost[(int) wider_mode] < max_cost)
{
op1 = force_reg (mode, op1);
goto try;
}
/* Try widening the mode and perform a non-widening multiplication. */
moptab = smul_optab;
if (smul_optab->handlers[(int) wider_mode].insn_code != CODE_FOR_nothing
&& size - 1 < BITS_PER_WORD
&& mul_cost[(int) wider_mode] + shift_cost[size-1] < max_cost)
{
op1 = wide_op1;
goto try;
}
/* Try widening multiplication of opposite signedness, and adjust. */
moptab = unsignedp ? smul_widen_optab : umul_widen_optab;
if (moptab->handlers[(int) wider_mode].insn_code != CODE_FOR_nothing
&& size - 1 < BITS_PER_WORD
&& (mul_widen_cost[(int) wider_mode]
+ 2 * shift_cost[size-1] + 4 * add_cost < max_cost))
{
rtx regop1 = force_reg (mode, op1);
tem = expand_binop (wider_mode, moptab, op0, regop1,
NULL_RTX, ! unsignedp, OPTAB_WIDEN);
if (tem != 0)
{
/* Extract the high half of the just generated product. */
tem = expand_shift (RSHIFT_EXPR, wider_mode, tem,
build_int_2 (size, 0), NULL_RTX, 1);
tem = convert_modes (mode, wider_mode, tem, unsignedp);
/* We used the wrong signedness. Adjust the result. */
return expand_mult_highpart_adjust (mode, tem, op0, op1,
target, unsignedp);
}
}
return 0;
try:
/* Pass NULL_RTX as target since TARGET has wrong mode. */
tem = expand_binop (wider_mode, moptab, op0, op1,
NULL_RTX, unsignedp, OPTAB_WIDEN);
if (tem == 0)
return 0;
/* Extract the high half of the just generated product. */
if (mode == word_mode)
{
return gen_highpart (mode, tem);
}
else
{
tem = expand_shift (RSHIFT_EXPR, wider_mode, tem,
build_int_2 (size, 0), NULL_RTX, 1);
return convert_modes (mode, wider_mode, tem, unsignedp);
}
}
/* Emit the code to divide OP0 by OP1, putting the result in TARGET
if that is convenient, and returning where the result is.
You may request either the quotient or the remainder as the result;
specify REM_FLAG nonzero to get the remainder.
CODE is the expression code for which kind of division this is;
it controls how rounding is done. MODE is the machine mode to use.
UNSIGNEDP nonzero means do unsigned division. */
/* ??? For CEIL_MOD_EXPR, can compute incorrect remainder with ANDI
and then correct it by or'ing in missing high bits
if result of ANDI is nonzero.
For ROUND_MOD_EXPR, can use ANDI and then sign-extend the result.
This could optimize to a bfexts instruction.
But C doesn't use these operations, so their optimizations are
left for later. */
/* ??? For modulo, we don't actually need the highpart of the first product,
the low part will do nicely. And for small divisors, the second multiply
can also be a low-part only multiply or even be completely left out.
E.g. to calculate the remainder of a division by 3 with a 32 bit
multiply, multiply with 0x55555556 and extract the upper two bits;
the result is exact for inputs up to 0x1fffffff.
The input range can be reduced by using cross-sum rules.
For odd divisors >= 3, the following table gives right shift counts
so that if a number is shifted by an integer multiple of the given
amount, the remainder stays the same:
2, 4, 3, 6, 10, 12, 4, 8, 18, 6, 11, 20, 18, 0, 5, 10, 12, 0, 12, 20,
14, 12, 23, 21, 8, 0, 20, 18, 0, 0, 6, 12, 0, 22, 0, 18, 20, 30, 0, 0,
0, 8, 0, 11, 12, 10, 36, 0, 30, 0, 0, 12, 0, 0, 0, 0, 44, 12, 24, 0,
20, 0, 7, 14, 0, 18, 36, 0, 0, 46, 60, 0, 42, 0, 15, 24, 20, 0, 0, 33,
0, 20, 0, 0, 18, 0, 60, 0, 0, 0, 0, 0, 40, 18, 0, 0, 12
Cross-sum rules for even numbers can be derived by leaving as many bits
to the right alone as the divisor has zeros to the right.
E.g. if x is an unsigned 32 bit number:
(x mod 12) == (((x & 1023) + ((x >> 8) & ~3)) * 0x15555558 >> 2 * 3) >> 28
*/
#define EXACT_POWER_OF_2_OR_ZERO_P(x) (((x) & ((x) - 1)) == 0)
rtx
expand_divmod (int rem_flag, enum tree_code code, enum machine_mode mode,
rtx op0, rtx op1, rtx target, int unsignedp)
{
enum machine_mode compute_mode;
rtx tquotient;
rtx quotient = 0, remainder = 0;
rtx last;
int size;
rtx insn, set;
optab optab1, optab2;
int op1_is_constant, op1_is_pow2 = 0;
int max_cost, extra_cost;
static HOST_WIDE_INT last_div_const = 0;
static HOST_WIDE_INT ext_op1;
op1_is_constant = GET_CODE (op1) == CONST_INT;
if (op1_is_constant)
{
ext_op1 = INTVAL (op1);
if (unsignedp)
ext_op1 &= GET_MODE_MASK (mode);
op1_is_pow2 = ((EXACT_POWER_OF_2_OR_ZERO_P (ext_op1)
|| (! unsignedp && EXACT_POWER_OF_2_OR_ZERO_P (-ext_op1))));
}
/*
This is the structure of expand_divmod: