/* 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.
+ 1999, 2000, 2001, 2002, 2003, 2004, 2005 Free Software Foundation, Inc.
This file is part of GCC.
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. */
+Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA
+02110-1301, USA. */
#include "config.h"
static rtx expand_smod_pow2 (enum machine_mode, rtx, HOST_WIDE_INT);
static rtx expand_sdiv_pow2 (enum machine_mode, rtx, HOST_WIDE_INT);
+/* Test whether a value is zero of a power of two. */
+#define EXACT_POWER_OF_2_OR_ZERO_P(x) (((x) & ((x) - 1)) == 0)
+
/* 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
memset (&all, 0, sizeof all);
PUT_CODE (&all.reg, REG);
- REGNO (&all.reg) = 10000;
+ /* Avoid using hard regs in ways which may be unsupported. */
+ REGNO (&all.reg) = LAST_VIRTUAL_REGISTER + 1;
PUT_CODE (&all.plus, PLUS);
XEXP (&all.plus, 0) = &all.reg;
{
unsigned int unit
= (MEM_P (str_rtx)) ? BITS_PER_UNIT : BITS_PER_WORD;
- unsigned HOST_WIDE_INT offset = bitnum / unit;
- unsigned HOST_WIDE_INT bitpos = bitnum % unit;
+ unsigned HOST_WIDE_INT offset, bitpos;
rtx op0 = str_rtx;
int byte_offset;
rtx orig_value;
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. */
+ bitnum += SUBREG_BYTE (op0) * BITS_PER_UNIT;
op0 = SUBREG_REG (op0);
}
+ /* No action is needed if the target is a register and if the field
+ lies completely outside that register. This can occur if the source
+ code contains an out-of-bounds access to a small array. */
+ if (REG_P (op0) && bitnum >= GET_MODE_BITSIZE (GET_MODE (op0)))
+ return value;
+
/* Use vec_set patterns for inserting parts of vectors whenever
available. */
if (VECTOR_MODE_P (GET_MODE (op0))
}
}
- 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.
done with a simple store. For targets that support fast unaligned
memory, any naturally sized, unit aligned field can be done directly. */
+ offset = bitnum / unit;
+ bitpos = bitnum % unit;
byte_offset = (bitnum % BITS_PER_WORD) / BITS_PER_UNIT
+ (offset * UNITS_PER_WORD);
{
if (GET_MODE (op0) != fieldmode)
{
- if (GET_CODE (op0) == SUBREG)
- {
- /* 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. */
- gcc_assert (GET_MODE (SUBREG_REG (op0)) == fieldmode
- || GET_MODE_CLASS (fieldmode) == MODE_INT
- || GET_MODE_CLASS (fieldmode) == MODE_PARTIAL_INT);
- op0 = SUBREG_REG (op0);
- }
- if (REG_P (op0))
- op0 = gen_rtx_SUBREG (fieldmode, op0, byte_offset);
- else
+ if (MEM_P (op0))
op0 = adjust_address (op0, fieldmode, offset);
+ else
+ op0 = simplify_gen_subreg (fieldmode, op0, GET_MODE (op0),
+ byte_offset);
}
emit_move_insn (op0, value);
return value;
/* 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. */
+ is a bit field, but passing VOIDmode to operand_subword_force
+ is not allowed. */
fieldmode = GET_MODE (value);
if (fieldmode == VOIDmode)
fieldmode = smallest_mode_for_size (nwords * BITS_PER_WORD, MODE_INT);
{
if (!REG_P (op0))
{
- /* 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. */
+ /* Since this is a destination (lvalue), we can't copy
+ it to a pseudo. We can remove a SUBREG that does not
+ change the size of the operand. Such a SUBREG may
+ have been added above. */
gcc_assert (GET_CODE (op0) == SUBREG
&& (GET_MODE_SIZE (GET_MODE (op0))
== GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0)))));
offset = 0;
}
- /* 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. */
+ /* If VALUE has a floating-point or complex 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. It can also
+ occur for unaligned float or complex fields. */
orig_value = value;
- if (GET_MODE_CLASS (GET_MODE (value)) != MODE_INT
+ if (GET_MODE (value) != VOIDmode
+ && 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);
+ {
+ value = gen_reg_rtx (int_mode_for_mode (GET_MODE (value)));
+ emit_move_insn (gen_lowpart (GET_MODE (orig_value), value), orig_value);
+ }
/* Now OFFSET is nonzero only if OP0 is memory
and is therefore always measured in bytes. */
/* 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 (MEM_P (op0)
&& ! ((*insn_data[(int) CODE_FOR_insv].operand[0].predicate)
(op0, VOIDmode)))
/* Now clear the chosen bits in OP0,
except that if VALUE is -1 we need not bother. */
- subtarget = (REG_P (op0) || ! flag_force_mem) ? op0 : 0;
+ subtarget = op0;
if (! all_one)
{
{
unsigned int unit
= (MEM_P (str_rtx)) ? BITS_PER_UNIT : BITS_PER_WORD;
- unsigned HOST_WIDE_INT offset = bitnum / unit;
- unsigned HOST_WIDE_INT bitpos = bitnum % unit;
+ unsigned HOST_WIDE_INT offset, bitpos;
rtx op0 = str_rtx;
rtx spec_target = target;
rtx spec_target_subreg = 0;
while (GET_CODE (op0) == SUBREG)
{
- bitpos += SUBREG_BYTE (op0) * BITS_PER_UNIT;
- if (bitpos > unit)
- {
- offset += (bitpos / unit);
- bitpos %= unit;
- }
+ bitnum += SUBREG_BYTE (op0) * BITS_PER_UNIT;
op0 = SUBREG_REG (op0);
}
+ /* If we have an out-of-bounds access to a register, just return an
+ uninitialized register of the required mode. This can occur if the
+ source code contains an out-of-bounds access to a small array. */
+ if (REG_P (op0) && bitnum >= GET_MODE_BITSIZE (GET_MODE (op0)))
+ return gen_reg_rtx (tmode);
+
if (REG_P (op0)
&& mode == GET_MODE (op0)
&& bitnum == 0
enum machine_mode imode = int_mode_for_mode (GET_MODE (op0));
if (imode != GET_MODE (op0))
{
- op0 = gen_lowpart (imode, op0);
+ if (MEM_P (op0))
+ op0 = adjust_address (op0, imode, 0);
+ else
+ {
+ gcc_assert (imode != BLKmode);
+ op0 = gen_lowpart (imode, op0);
- /* If we got a SUBREG, force it into a register since we aren't going
- to be able to do another SUBREG on it. */
- if (GET_CODE (op0) == SUBREG)
- op0 = force_reg (imode, op0);
+ /* If we got a SUBREG, force it into a register since we
+ aren't going to be able to do another SUBREG on it. */
+ if (GET_CODE (op0) == SUBREG)
+ op0 = force_reg (imode, op0);
+ }
}
}
can also be extracted with a SUBREG. For this, we need the
byte offset of the value in op0. */
+ bitpos = bitnum % unit;
+ offset = bitnum / unit;
byte_offset = bitpos / BITS_PER_UNIT + offset * UNITS_PER_WORD;
/* If OP0 is a register, BITPOS must count within a word.
unit = GET_MODE_BITSIZE (maxmode);
- if (xtarget == 0
- || (flag_force_mem && MEM_P (xtarget)))
+ if (xtarget == 0)
xtarget = xspec_target = gen_reg_rtx (tmode);
if (GET_MODE (xtarget) != maxmode)
unit = GET_MODE_BITSIZE (maxmode);
- if (xtarget == 0
- || (flag_force_mem && MEM_P (xtarget)))
+ if (xtarget == 0)
xtarget = xspec_target = gen_reg_rtx (tmode);
if (GET_MODE (xtarget) != maxmode)
return spec_target;
if (GET_MODE (target) != tmode && GET_MODE (target) != mode)
{
- /* If the target mode is complex, then extract the two scalar elements
- from the value now. Creating (subreg:SC (reg:DI) 0), as we would do
- with the clause below, will cause gen_realpart or gen_imagpart to
- fail, since those functions must return lvalues. */
- if (COMPLEX_MODE_P (tmode))
- {
- rtx realpart, imagpart;
- enum machine_mode itmode = GET_MODE_INNER (tmode);
-
- target = convert_to_mode (mode_for_size (GET_MODE_BITSIZE (tmode),
- MODE_INT, 0),
- target, unsignedp);
-
- realpart = extract_bit_field (target, GET_MODE_BITSIZE (itmode), 0,
- unsignedp, NULL, itmode, itmode);
- imagpart = extract_bit_field (target, GET_MODE_BITSIZE (itmode),
- GET_MODE_BITSIZE (itmode), unsignedp,
- NULL, itmode, itmode);
-
- return gen_rtx_CONCAT (tmode, realpart, imagpart);
- }
-
/* If the target mode is not a scalar integral, first convert to the
integer mode of that size and then access it as a floating-point
value via a SUBREG. */
tree type = TREE_TYPE (amount);
tree new_amount = make_tree (type, op1);
tree other_amount
- = fold (build2 (MINUS_EXPR, type, convert
- (type, build_int_cst
- (NULL_TREE, GET_MODE_BITSIZE (mode))),
- amount));
+ = fold_build2 (MINUS_EXPR, type,
+ build_int_cst (type, GET_MODE_BITSIZE (mode)),
+ amount);
shifted = force_reg (mode, shifted);
temp = expand_shift (left ? LSHIFT_EXPR : RSHIFT_EXPR,
- mode, shifted, new_amount, subtarget, 1);
+ mode, shifted, new_amount, 0, 1);
temp1 = expand_shift (left ? RSHIFT_EXPR : LSHIFT_EXPR,
- mode, shifted, other_amount, 0, 1);
+ mode, shifted, other_amount, subtarget, 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,
return temp;
}
\f
-enum alg_code { alg_unknown, 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 };
+enum alg_code {
+ alg_unknown,
+ 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_impossible
+};
/* This structure holds the "cost" of a multiply sequence. The
"cost" field holds the total rtx_cost of every operator in the
/* The best multiplication algorithm for t. */
enum alg_code alg;
+
+ /* The cost of multiplication if ALG_CODE is not alg_impossible.
+ Otherwise, the cost within which multiplication by T is
+ impossible. */
+ struct mult_cost cost;
};
/* The number of cache/hash entries. */
static rtx expand_mult_const (enum machine_mode, rtx, HOST_WIDE_INT, rtx,
const struct algorithm *, enum mult_variant);
static unsigned HOST_WIDE_INT choose_multiplier (unsigned HOST_WIDE_INT, int,
- int, unsigned HOST_WIDE_INT *,
- int *, int *);
+ int, rtx *, int *, int *);
static unsigned HOST_WIDE_INT invert_mod2n (unsigned HOST_WIDE_INT, int);
static rtx extract_high_half (enum machine_mode, rtx);
+static rtx expand_mult_highpart (enum machine_mode, rtx, rtx, rtx, int, int);
static rtx expand_mult_highpart_optab (enum machine_mode, rtx, rtx, rtx,
int, int);
/* Compute and return the best algorithm for multiplying by T.
&& alg_hash[hash_index].mode == mode
&& alg_hash[hash_index].alg != alg_unknown)
{
- cache_hit = true;
cache_alg = alg_hash[hash_index].alg;
- switch (cache_alg)
+
+ if (cache_alg == alg_impossible)
{
- case alg_shift:
- goto do_alg_shift;
+ /* The cache tells us that it's impossible to synthesize
+ multiplication by T within alg_hash[hash_index].cost. */
+ if (!CHEAPER_MULT_COST (&alg_hash[hash_index].cost, cost_limit))
+ /* COST_LIMIT is at least as restrictive as the one
+ recorded in the hash table, in which case we have no
+ hope of synthesizing a multiplication. Just
+ return. */
+ return;
+
+ /* If we get here, COST_LIMIT is less restrictive than the
+ one recorded in the hash table, so we may be able to
+ synthesize a multiplication. Proceed as if we didn't
+ have the cache entry. */
+ }
+ else
+ {
+ if (CHEAPER_MULT_COST (cost_limit, &alg_hash[hash_index].cost))
+ /* The cached algorithm shows that this multiplication
+ requires more cost than COST_LIMIT. Just return. This
+ way, we don't clobber this cache entry with
+ alg_impossible but retain useful information. */
+ return;
- case alg_add_t_m2:
- case alg_sub_t_m2:
- goto do_alg_addsub_t_m2;
+ cache_hit = true;
- case alg_add_factor:
- case alg_sub_factor:
- goto do_alg_addsub_factor;
+ switch (cache_alg)
+ {
+ case alg_shift:
+ goto do_alg_shift;
- case alg_add_t2_m:
- goto do_alg_add_t2_m;
+ case alg_add_t_m2:
+ case alg_sub_t_m2:
+ goto do_alg_addsub_t_m2;
- case alg_sub_t2_m:
- goto do_alg_sub_t2_m;
+ case alg_add_factor:
+ case alg_sub_factor:
+ goto do_alg_addsub_factor;
- default:
- gcc_unreachable ();
+ case alg_add_t2_m:
+ goto do_alg_add_t2_m;
+
+ case alg_sub_t2_m:
+ goto do_alg_sub_t2_m;
+
+ default:
+ gcc_unreachable ();
+ }
}
}
op_latency = add_cost[mode];
new_limit.cost = best_cost.cost - op_cost;
- new_limit.cost = best_cost.cost - op_latency;
+ new_limit.latency = best_cost.latency - op_latency;
synth_mult (alg_in, t / d, &new_limit, mode);
alg_in->cost.cost += op_cost;
done:
/* If best_cost has not decreased, we have not found any algorithm. */
if (!CHEAPER_MULT_COST (&best_cost, cost_limit))
- return;
+ {
+ /* We failed to find an algorithm. Record alg_impossible for
+ this case (that is, <T, MODE, COST_LIMIT>) so that next time
+ we are asked to find an algorithm for T within the same or
+ lower COST_LIMIT, we can immediately return to the
+ caller. */
+ alg_hash[hash_index].t = t;
+ alg_hash[hash_index].mode = mode;
+ alg_hash[hash_index].alg = alg_impossible;
+ alg_hash[hash_index].cost = *cost_limit;
+ return;
+ }
/* Cache the result. */
if (!cache_hit)
alg_hash[hash_index].t = t;
alg_hash[hash_index].mode = mode;
alg_hash[hash_index].alg = best_alg->op[best_alg->ops];
+ alg_hash[hash_index].cost.cost = best_cost.cost;
+ alg_hash[hash_index].cost.latency = best_cost.latency;
}
/* If we are getting a too long sequence for `struct algorithm'
expand_mult (enum machine_mode mode, rtx op0, rtx op1, rtx target,
int unsignedp)
{
- rtx const_op1 = op1;
enum mult_variant variant;
struct algorithm algorithm;
+ int max_cost;
- /* 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
+ /* Handling const0_rtx here allows us to use zero as a rogue value for
+ coeff below. */
+ if (op1 == const0_rtx)
+ return const0_rtx;
+ if (op1 == const1_rtx)
+ return op0;
+ if (op1 == constm1_rtx)
+ return expand_unop (mode,
+ GET_MODE_CLASS (mode) == MODE_INT
+ && !unsignedp && flag_trapv
+ ? negv_optab : neg_optab,
+ op0, target, 0);
+
+ /* These are the operations that are potentially turned into a sequence
+ of shifts and additions. */
+ if (SCALAR_INT_MODE_P (mode)
&& (unsignedp || !flag_trapv))
{
- int mult_cost = rtx_cost (gen_rtx_MULT (mode, op0, op1), SET);
+ HOST_WIDE_INT coeff = 0;
+ rtx fake_reg = gen_raw_REG (mode, LAST_VIRTUAL_REGISTER + 1);
+
+ /* 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 (choose_mult_variant (mode, INTVAL (const_op1), &algorithm, &variant,
- mult_cost))
- return expand_mult_const (mode, op0, INTVAL (const_op1), target,
- &algorithm, variant);
+ if (GET_CODE (op1) == CONST_INT)
+ {
+ /* Attempt to handle multiplication of DImode values by negative
+ coefficients, by performing the multiplication by a positive
+ multiplier and then inverting the result. */
+ if (INTVAL (op1) < 0
+ && GET_MODE_BITSIZE (mode) > HOST_BITS_PER_WIDE_INT)
+ {
+ /* Its safe to use -INTVAL (op1) even for INT_MIN, as the
+ result is interpreted as an unsigned coefficient.
+ Exclude cost of op0 from max_cost to match the cost
+ calculation of the synth_mult. */
+ max_cost = rtx_cost (gen_rtx_MULT (mode, fake_reg, op1), SET)
+ - neg_cost[mode];
+ if (max_cost > 0
+ && choose_mult_variant (mode, -INTVAL (op1), &algorithm,
+ &variant, max_cost))
+ {
+ rtx temp = expand_mult_const (mode, op0, -INTVAL (op1),
+ NULL_RTX, &algorithm,
+ variant);
+ return expand_unop (mode, neg_optab, temp, target, 0);
+ }
+ }
+ else coeff = INTVAL (op1);
+ }
+ else if (GET_CODE (op1) == CONST_DOUBLE)
+ {
+ /* If we are multiplying in DImode, it may still be a win
+ to try to work with shifts and adds. */
+ if (CONST_DOUBLE_HIGH (op1) == 0)
+ coeff = CONST_DOUBLE_LOW (op1);
+ else if (CONST_DOUBLE_LOW (op1) == 0
+ && EXACT_POWER_OF_2_OR_ZERO_P (CONST_DOUBLE_HIGH (op1)))
+ {
+ int shift = floor_log2 (CONST_DOUBLE_HIGH (op1))
+ + HOST_BITS_PER_WIDE_INT;
+ return expand_shift (LSHIFT_EXPR, mode, op0,
+ build_int_cst (NULL_TREE, shift),
+ target, unsignedp);
+ }
+ }
+
+ /* 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 always
+ use synth_mult. */
+ if (coeff != 0)
+ {
+ /* Special case powers of two. */
+ if (EXACT_POWER_OF_2_OR_ZERO_P (coeff))
+ return expand_shift (LSHIFT_EXPR, mode, op0,
+ build_int_cst (NULL_TREE, floor_log2 (coeff)),
+ target, unsignedp);
+
+ /* Exclude cost of op0 from max_cost to match the cost
+ calculation of the synth_mult. */
+ max_cost = rtx_cost (gen_rtx_MULT (mode, fake_reg, op1), SET);
+ if (choose_mult_variant (mode, coeff, &algorithm, &variant,
+ max_cost))
+ return expand_mult_const (mode, op0, coeff, target,
+ &algorithm, variant);
+ }
}
if (GET_CODE (op0) == CONST_DOUBLE)
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)
+ rtx *multiplier_ptr, int *post_shift_ptr, int *lgup_ptr)
{
HOST_WIDE_INT mhigh_hi, mlow_hi;
unsigned HOST_WIDE_INT mhigh_lo, mlow_lo;
if (n < HOST_BITS_PER_WIDE_INT)
{
unsigned HOST_WIDE_INT mask = ((unsigned HOST_WIDE_INT) 1 << n) - 1;
- *multiplier_ptr = mhigh_lo & mask;
+ *multiplier_ptr = GEN_INT (mhigh_lo & mask);
return mhigh_lo >= mask;
}
else
{
- *multiplier_ptr = mhigh_lo;
+ *multiplier_ptr = GEN_INT (mhigh_lo);
return mhigh_hi;
}
}
}
/* Try widening the mode and perform a non-widening multiplication. */
- moptab = smul_optab;
if (smul_optab->handlers[wider_mode].insn_code != CODE_FOR_nothing
&& size - 1 < BITS_PER_WORD
&& mul_cost[wider_mode] + shift_cost[mode][size-1] < max_cost)
{
- tem = expand_binop (wider_mode, moptab, op0, op1, 0,
+ rtx insns, wop0, wop1;
+
+ /* We need to widen the operands, for example to ensure the
+ constant multiplier is correctly sign or zero extended.
+ Use a sequence to clean-up any instructions emitted by
+ the conversions if things don't work out. */
+ start_sequence ();
+ wop0 = convert_modes (wider_mode, mode, op0, unsignedp);
+ wop1 = convert_modes (wider_mode, mode, op1, unsignedp);
+ tem = expand_binop (wider_mode, smul_optab, wop0, wop1, 0,
unsignedp, OPTAB_WIDEN);
+ insns = get_insns ();
+ end_sequence ();
+
if (tem)
- return extract_high_half (mode, tem);
+ {
+ emit_insn (insns);
+ return extract_high_half (mode, tem);
+ }
}
/* Try widening multiplication of opposite signedness, and adjust. */
return 0;
}
-/* 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.
+/* Emit code to multiply OP0 and OP1 (where OP1 is an integer constant),
+ 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.
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)
+static rtx
+expand_mult_highpart (enum machine_mode mode, rtx op0, rtx op1,
+ rtx target, int unsignedp, int max_cost)
{
enum machine_mode wider_mode = GET_MODE_WIDER_MODE (mode);
+ unsigned HOST_WIDE_INT cnst1;
int extra_cost;
bool sign_adjust = false;
enum mult_variant variant;
struct algorithm alg;
- rtx op1, tem;
+ rtx tem;
/* We can't support modes wider than HOST_BITS_PER_INT. */
gcc_assert (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT);
- op1 = gen_int_mode (cnst1, wider_mode);
- cnst1 &= GET_MODE_MASK (mode);
+ cnst1 = INTVAL (op1) & GET_MODE_MASK (mode);
/* We can't optimize modes wider than BITS_PER_WORD.
??? We might be able to perform double-word arithmetic if
(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)
|| optab2->handlers[compute_mode].libfunc)
break;
- /* If we still couldn't find a mode, use MODE, but we'll probably abort
- in expand_binop. */
+ /* If we still couldn't find a mode, use MODE, but expand_binop will
+ probably die. */
if (compute_mode == VOIDmode)
compute_mode = mode;
{
if (unsignedp)
{
- unsigned HOST_WIDE_INT mh, ml;
+ unsigned HOST_WIDE_INT mh;
int pre_shift, post_shift;
int dummy;
+ rtx ml;
unsigned HOST_WIDE_INT d = (INTVAL (op1)
& GET_MODE_MASK (compute_mode));
{
unsigned HOST_WIDE_INT ml;
int lgup, post_shift;
+ rtx mlr;
HOST_WIDE_INT d = INTVAL (op1);
unsigned HOST_WIDE_INT abs_d = d >= 0 ? d : -d;
else if (size <= HOST_BITS_PER_WIDE_INT)
{
choose_multiplier (abs_d, size, size - 1,
- &ml, &post_shift, &lgup);
+ &mlr, &post_shift, &lgup);
+ ml = (unsigned HOST_WIDE_INT) INTVAL (mlr);
if (ml < (unsigned HOST_WIDE_INT) 1 << (size - 1))
{
rtx t1, t2, t3;
extra_cost = (shift_cost[compute_mode][post_shift]
+ shift_cost[compute_mode][size - 1]
+ add_cost[compute_mode]);
- t1 = expand_mult_highpart (compute_mode, op0, ml,
+ t1 = expand_mult_highpart (compute_mode, op0, mlr,
NULL_RTX, 0,
max_cost - extra_cost);
if (t1 == 0)
goto fail1;
ml |= (~(unsigned HOST_WIDE_INT) 0) << (size - 1);
+ mlr = gen_int_mode (ml, compute_mode);
extra_cost = (shift_cost[compute_mode][post_shift]
+ shift_cost[compute_mode][size - 1]
+ 2 * add_cost[compute_mode]);
- t1 = expand_mult_highpart (compute_mode, op0, ml,
+ t1 = expand_mult_highpart (compute_mode, op0, mlr,
NULL_RTX, 0,
max_cost - extra_cost);
if (t1 == 0)
/* We will come here only for signed operations. */
if (op1_is_constant && HOST_BITS_PER_WIDE_INT >= size)
{
- unsigned HOST_WIDE_INT mh, ml;
+ unsigned HOST_WIDE_INT mh;
int pre_shift, lgup, post_shift;
HOST_WIDE_INT d = INTVAL (op1);
+ rtx ml;
if (d > 0)
{
}
case PLUS:
- return fold (build2 (PLUS_EXPR, type, make_tree (type, XEXP (x, 0)),
- make_tree (type, XEXP (x, 1))));
+ return fold_build2 (PLUS_EXPR, type, make_tree (type, XEXP (x, 0)),
+ make_tree (type, XEXP (x, 1)));
case MINUS:
- return fold (build2 (MINUS_EXPR, type, make_tree (type, XEXP (x, 0)),
- make_tree (type, XEXP (x, 1))));
+ return fold_build2 (MINUS_EXPR, type, make_tree (type, XEXP (x, 0)),
+ make_tree (type, XEXP (x, 1)));
case NEG:
- return fold (build1 (NEGATE_EXPR, type, make_tree (type, XEXP (x, 0))));
+ return fold_build1 (NEGATE_EXPR, type, make_tree (type, XEXP (x, 0)));
case MULT:
- return fold (build2 (MULT_EXPR, type, make_tree (type, XEXP (x, 0)),
- make_tree (type, XEXP (x, 1))));
+ return fold_build2 (MULT_EXPR, type, make_tree (type, XEXP (x, 0)),
+ make_tree (type, XEXP (x, 1)));
case ASHIFT:
- return fold (build2 (LSHIFT_EXPR, type, make_tree (type, XEXP (x, 0)),
- make_tree (type, XEXP (x, 1))));
+ return fold_build2 (LSHIFT_EXPR, type, make_tree (type, XEXP (x, 0)),
+ make_tree (type, XEXP (x, 1)));
case LSHIFTRT:
t = lang_hooks.types.unsigned_type (type);
- return fold (convert (type,
- build2 (RSHIFT_EXPR, t,
- make_tree (t, XEXP (x, 0)),
- make_tree (type, XEXP (x, 1)))));
+ return fold_convert (type, build2 (RSHIFT_EXPR, t,
+ make_tree (t, XEXP (x, 0)),
+ make_tree (type, XEXP (x, 1))));
case ASHIFTRT:
t = lang_hooks.types.signed_type (type);
- return fold (convert (type,
- build2 (RSHIFT_EXPR, t,
- make_tree (t, XEXP (x, 0)),
- make_tree (type, XEXP (x, 1)))));
+ return fold_convert (type, build2 (RSHIFT_EXPR, t,
+ make_tree (t, XEXP (x, 0)),
+ make_tree (type, XEXP (x, 1))));
case DIV:
if (TREE_CODE (type) != REAL_TYPE)
else
t = type;
- return fold (convert (type,
- build2 (TRUNC_DIV_EXPR, t,
- make_tree (t, XEXP (x, 0)),
- make_tree (t, XEXP (x, 1)))));
+ return fold_convert (type, build2 (TRUNC_DIV_EXPR, t,
+ make_tree (t, XEXP (x, 0)),
+ make_tree (t, XEXP (x, 1))));
case UDIV:
t = lang_hooks.types.unsigned_type (type);
- return fold (convert (type,
- build2 (TRUNC_DIV_EXPR, t,
- make_tree (t, XEXP (x, 0)),
- make_tree (t, XEXP (x, 1)))));
+ return fold_convert (type, build2 (TRUNC_DIV_EXPR, t,
+ make_tree (t, XEXP (x, 0)),
+ make_tree (t, XEXP (x, 1))));
case SIGN_EXTEND:
case ZERO_EXTEND:
t = lang_hooks.types.type_for_mode (GET_MODE (XEXP (x, 0)),
GET_CODE (x) == ZERO_EXTEND);
- return fold (convert (type, make_tree (t, XEXP (x, 0))));
+ return fold_convert (type, make_tree (t, XEXP (x, 0)));
default:
t = build_decl (VAR_DECL, NULL_TREE, type);
/* Note that we do *not* use SET_DECL_RTL here, because we do not
want set_decl_rtl to go adjusting REG_ATTRS for this temporary. */
- t->decl.rtl = x;
+ t->decl_with_rtl.rtl = x;
return t;
}
add_type = (GET_MODE (add) == VOIDmode ? mult_type
: lang_hooks.types.type_for_mode (GET_MODE (add), unsignedp));
- result = fold (build2 (PLUS_EXPR, mult_type,
- fold (build2 (MULT_EXPR, mult_type,
- make_tree (mult_type, x),
- make_tree (mult_type, mult))),
- make_tree (add_type, add)));
+ result = fold_build2 (PLUS_EXPR, mult_type,
+ fold_build2 (MULT_EXPR, mult_type,
+ make_tree (mult_type, x),
+ make_tree (mult_type, mult)),
+ make_tree (add_type, add));
return TREE_CONSTANT_OVERFLOW (result);
}
tree add_type = (GET_MODE (add) == VOIDmode
? type: lang_hooks.types.type_for_mode (GET_MODE (add),
unsignedp));
- tree result = fold (build2 (PLUS_EXPR, type,
- fold (build2 (MULT_EXPR, type,
- make_tree (type, x),
- make_tree (type, mult))),
- make_tree (add_type, add)));
+ tree result = fold_build2 (PLUS_EXPR, type,
+ fold_build2 (MULT_EXPR, type,
+ make_tree (type, x),
+ make_tree (type, mult)),
+ make_tree (add_type, add));
return expand_expr (result, target, VOIDmode, 0);
}
The algorithm is based on the code in expr.c:do_jump.
- Note that this does not perform a general comparison. Only variants
- generated within expmed.c are correctly handled, others abort (but could
- be handled if needed). */
+ Note that this does not perform a general comparison. Only
+ variants generated within expmed.c are correctly handled, others
+ could be handled if needed. */
static void
do_cmp_and_jump (rtx arg1, rtx arg2, enum rtx_code op, enum machine_mode mode,
OSDN Git Service
