/* 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, 2005, 2006, 2007
+ 1999, 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009
Free Software Foundation, Inc.
This file is part of GCC.
Usually, this will mean that the MD file will emit non-branch
sequences. */
-static bool sdiv_pow2_cheap[NUM_MACHINE_MODES];
-static bool smod_pow2_cheap[NUM_MACHINE_MODES];
+static bool sdiv_pow2_cheap[2][NUM_MACHINE_MODES];
+static bool smod_pow2_cheap[2][NUM_MACHINE_MODES];
#ifndef SLOW_UNALIGNED_ACCESS
#define SLOW_UNALIGNED_ACCESS(MODE, ALIGN) STRICT_ALIGNMENT
/* Cost of various pieces of RTL. Note that some of these are indexed by
shift count and some by mode. */
-static int zero_cost;
-static int add_cost[NUM_MACHINE_MODES];
-static int neg_cost[NUM_MACHINE_MODES];
-static int shift_cost[NUM_MACHINE_MODES][MAX_BITS_PER_WORD];
-static int shiftadd_cost[NUM_MACHINE_MODES][MAX_BITS_PER_WORD];
-static int shiftsub_cost[NUM_MACHINE_MODES][MAX_BITS_PER_WORD];
-static int mul_cost[NUM_MACHINE_MODES];
-static int sdiv_cost[NUM_MACHINE_MODES];
-static int udiv_cost[NUM_MACHINE_MODES];
-static int mul_widen_cost[NUM_MACHINE_MODES];
-static int mul_highpart_cost[NUM_MACHINE_MODES];
+static int zero_cost[2];
+static int add_cost[2][NUM_MACHINE_MODES];
+static int neg_cost[2][NUM_MACHINE_MODES];
+static int shift_cost[2][NUM_MACHINE_MODES][MAX_BITS_PER_WORD];
+static int shiftadd_cost[2][NUM_MACHINE_MODES][MAX_BITS_PER_WORD];
+static int shiftsub0_cost[2][NUM_MACHINE_MODES][MAX_BITS_PER_WORD];
+static int shiftsub1_cost[2][NUM_MACHINE_MODES][MAX_BITS_PER_WORD];
+static int mul_cost[2][NUM_MACHINE_MODES];
+static int sdiv_cost[2][NUM_MACHINE_MODES];
+static int udiv_cost[2][NUM_MACHINE_MODES];
+static int mul_widen_cost[2][NUM_MACHINE_MODES];
+static int mul_highpart_cost[2][NUM_MACHINE_MODES];
void
init_expmed (void)
struct rtx_def shift; rtunion shift_fld1;
struct rtx_def shift_mult; rtunion shift_mult_fld1;
struct rtx_def shift_add; rtunion shift_add_fld1;
- struct rtx_def shift_sub; rtunion shift_sub_fld1;
+ struct rtx_def shift_sub0; rtunion shift_sub0_fld1;
+ struct rtx_def shift_sub1; rtunion shift_sub1_fld1;
} all;
rtx pow2[MAX_BITS_PER_WORD];
rtx cint[MAX_BITS_PER_WORD];
int m, n;
enum machine_mode mode, wider_mode;
+ int speed;
- zero_cost = rtx_cost (const0_rtx, 0);
for (m = 1; m < MAX_BITS_PER_WORD; m++)
{
pow2[m] = GEN_INT ((HOST_WIDE_INT) 1 << m);
cint[m] = GEN_INT (m);
}
-
memset (&all, 0, sizeof all);
PUT_CODE (&all.reg, REG);
XEXP (&all.shift_add, 0) = &all.shift_mult;
XEXP (&all.shift_add, 1) = &all.reg;
- PUT_CODE (&all.shift_sub, MINUS);
- XEXP (&all.shift_sub, 0) = &all.shift_mult;
- XEXP (&all.shift_sub, 1) = &all.reg;
+ PUT_CODE (&all.shift_sub0, MINUS);
+ XEXP (&all.shift_sub0, 0) = &all.shift_mult;
+ XEXP (&all.shift_sub0, 1) = &all.reg;
- for (mode = GET_CLASS_NARROWEST_MODE (MODE_INT);
- mode != VOIDmode;
- mode = GET_MODE_WIDER_MODE (mode))
+ PUT_CODE (&all.shift_sub1, MINUS);
+ XEXP (&all.shift_sub1, 0) = &all.reg;
+ XEXP (&all.shift_sub1, 1) = &all.shift_mult;
+
+ for (speed = 0; speed < 2; speed++)
{
- PUT_MODE (&all.reg, mode);
- PUT_MODE (&all.plus, mode);
- PUT_MODE (&all.neg, mode);
- PUT_MODE (&all.mult, mode);
- PUT_MODE (&all.sdiv, mode);
- PUT_MODE (&all.udiv, mode);
- PUT_MODE (&all.sdiv_32, mode);
- PUT_MODE (&all.smod_32, mode);
- PUT_MODE (&all.wide_trunc, mode);
- PUT_MODE (&all.shift, mode);
- PUT_MODE (&all.shift_mult, mode);
- PUT_MODE (&all.shift_add, mode);
- PUT_MODE (&all.shift_sub, mode);
-
- add_cost[mode] = rtx_cost (&all.plus, SET);
- neg_cost[mode] = rtx_cost (&all.neg, SET);
- mul_cost[mode] = rtx_cost (&all.mult, SET);
- sdiv_cost[mode] = rtx_cost (&all.sdiv, SET);
- udiv_cost[mode] = rtx_cost (&all.udiv, SET);
-
- sdiv_pow2_cheap[mode] = (rtx_cost (&all.sdiv_32, SET)
- <= 2 * add_cost[mode]);
- smod_pow2_cheap[mode] = (rtx_cost (&all.smod_32, SET)
- <= 4 * add_cost[mode]);
-
- wider_mode = GET_MODE_WIDER_MODE (mode);
- if (wider_mode != VOIDmode)
- {
- PUT_MODE (&all.zext, wider_mode);
- PUT_MODE (&all.wide_mult, wider_mode);
- PUT_MODE (&all.wide_lshr, wider_mode);
- XEXP (&all.wide_lshr, 1) = GEN_INT (GET_MODE_BITSIZE (mode));
+ crtl->maybe_hot_insn_p = speed;
+ zero_cost[speed] = rtx_cost (const0_rtx, SET, speed);
- mul_widen_cost[wider_mode] = rtx_cost (&all.wide_mult, SET);
- mul_highpart_cost[mode] = rtx_cost (&all.wide_trunc, SET);
- }
+ for (mode = GET_CLASS_NARROWEST_MODE (MODE_INT);
+ mode != VOIDmode;
+ mode = GET_MODE_WIDER_MODE (mode))
+ {
+ PUT_MODE (&all.reg, mode);
+ PUT_MODE (&all.plus, mode);
+ PUT_MODE (&all.neg, mode);
+ PUT_MODE (&all.mult, mode);
+ PUT_MODE (&all.sdiv, mode);
+ PUT_MODE (&all.udiv, mode);
+ PUT_MODE (&all.sdiv_32, mode);
+ PUT_MODE (&all.smod_32, mode);
+ PUT_MODE (&all.wide_trunc, mode);
+ PUT_MODE (&all.shift, mode);
+ PUT_MODE (&all.shift_mult, mode);
+ PUT_MODE (&all.shift_add, mode);
+ PUT_MODE (&all.shift_sub0, mode);
+ PUT_MODE (&all.shift_sub1, mode);
+
+ add_cost[speed][mode] = rtx_cost (&all.plus, SET, speed);
+ neg_cost[speed][mode] = rtx_cost (&all.neg, SET, speed);
+ mul_cost[speed][mode] = rtx_cost (&all.mult, SET, speed);
+ sdiv_cost[speed][mode] = rtx_cost (&all.sdiv, SET, speed);
+ udiv_cost[speed][mode] = rtx_cost (&all.udiv, SET, speed);
+
+ sdiv_pow2_cheap[speed][mode] = (rtx_cost (&all.sdiv_32, SET, speed)
+ <= 2 * add_cost[speed][mode]);
+ smod_pow2_cheap[speed][mode] = (rtx_cost (&all.smod_32, SET, speed)
+ <= 4 * add_cost[speed][mode]);
+
+ wider_mode = GET_MODE_WIDER_MODE (mode);
+ if (wider_mode != VOIDmode)
+ {
+ PUT_MODE (&all.zext, wider_mode);
+ PUT_MODE (&all.wide_mult, wider_mode);
+ PUT_MODE (&all.wide_lshr, wider_mode);
+ XEXP (&all.wide_lshr, 1) = GEN_INT (GET_MODE_BITSIZE (mode));
+
+ mul_widen_cost[speed][wider_mode]
+ = rtx_cost (&all.wide_mult, SET, speed);
+ mul_highpart_cost[speed][mode]
+ = rtx_cost (&all.wide_trunc, SET, speed);
+ }
- shift_cost[mode][0] = 0;
- shiftadd_cost[mode][0] = shiftsub_cost[mode][0] = add_cost[mode];
+ shift_cost[speed][mode][0] = 0;
+ shiftadd_cost[speed][mode][0] = shiftsub0_cost[speed][mode][0]
+ = shiftsub1_cost[speed][mode][0] = add_cost[speed][mode];
- n = MIN (MAX_BITS_PER_WORD, GET_MODE_BITSIZE (mode));
- for (m = 1; m < n; m++)
- {
- XEXP (&all.shift, 1) = cint[m];
- XEXP (&all.shift_mult, 1) = pow2[m];
+ n = MIN (MAX_BITS_PER_WORD, GET_MODE_BITSIZE (mode));
+ for (m = 1; m < n; m++)
+ {
+ XEXP (&all.shift, 1) = cint[m];
+ XEXP (&all.shift_mult, 1) = pow2[m];
- shift_cost[mode][m] = rtx_cost (&all.shift, SET);
- shiftadd_cost[mode][m] = rtx_cost (&all.shift_add, SET);
- shiftsub_cost[mode][m] = rtx_cost (&all.shift_sub, SET);
+ shift_cost[speed][mode][m] = rtx_cost (&all.shift, SET, speed);
+ shiftadd_cost[speed][mode][m] = rtx_cost (&all.shift_add, SET, speed);
+ shiftsub0_cost[speed][mode][m] = rtx_cost (&all.shift_sub0, SET, speed);
+ shiftsub1_cost[speed][mode][m] = rtx_cost (&all.shift_sub1, SET, speed);
+ }
}
}
+ default_rtl_profile ();
}
/* Return an rtx representing minus the value of X.
!= CODE_FOR_nothing))
{
int icode = optab_handler (movstrict_optab, fieldmode)->insn_code;
+ rtx insn;
+ rtx start = get_last_insn ();
+ rtx arg0 = op0;
/* Get appropriate low part of the value being stored. */
if (GET_CODE (value) == CONST_INT || REG_P (value))
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);
+ arg0 = SUBREG_REG (op0);
}
- emit_insn (GEN_FCN (icode)
- (gen_rtx_SUBREG (fieldmode, op0,
+ insn = (GEN_FCN (icode)
+ (gen_rtx_SUBREG (fieldmode, arg0,
(bitnum % BITS_PER_WORD) / BITS_PER_UNIT
+ (offset * UNITS_PER_WORD)),
value));
-
- return true;
+ if (insn)
+ {
+ emit_insn (insn);
+ return true;
+ }
+ delete_insns_since (start);
}
/* Handle fields bigger than a word. */
if (pat)
{
emit_insn (pat);
+
+ /* If the mode of the insertion is wider than the mode of the
+ target register we created a paradoxical subreg for the
+ target. Truncate the paradoxical subreg of the target to
+ itself properly. */
+ if (!TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (GET_MODE (op0)),
+ GET_MODE_BITSIZE (op_mode))
+ && (REG_P (xop0)
+ || GET_CODE (xop0) == SUBREG))
+ convert_move (op0, xop0, true);
return true;
}
delete_insns_since (last);
&& bitpos + bitsize != GET_MODE_BITSIZE (mode));
if (GET_MODE (value) != mode)
- {
- if ((REG_P (value) || 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);
- }
+ value = convert_to_mode (mode, value, 1);
if (must_and)
value = expand_binop (mode, and_optab, value,
for (; new_mode != VOIDmode ; new_mode = GET_MODE_WIDER_MODE (new_mode))
if (GET_MODE_NUNITS (new_mode) == nunits
- && GET_MODE_INNER (new_mode) == tmode
+ && GET_MODE_SIZE (new_mode) == GET_MODE_SIZE (GET_MODE (op0))
&& targetm.vector_mode_supported_p (new_mode))
break;
if (new_mode != VOIDmode)
{
if (MEM_P (op0))
op0 = adjust_address (op0, imode, 0);
- else
+ else if (imode != BLKmode)
{
- gcc_assert (imode != BLKmode);
op0 = gen_lowpart (imode, op0);
/* If we got a SUBREG, force it into a register since we
if (GET_CODE (op0) == SUBREG)
op0 = force_reg (imode, op0);
}
+ else if (REG_P (op0))
+ {
+ rtx reg, subreg;
+ imode = smallest_mode_for_size (GET_MODE_BITSIZE (GET_MODE (op0)),
+ MODE_INT);
+ reg = gen_reg_rtx (imode);
+ subreg = gen_lowpart_SUBREG (GET_MODE (op0), reg);
+ emit_move_insn (subreg, op0);
+ op0 = reg;
+ bitnum += SUBREG_BYTE (subreg) * BITS_PER_UNIT;
+ }
+ else
+ {
+ rtx mem = assign_stack_temp (GET_MODE (op0),
+ GET_MODE_SIZE (GET_MODE (op0)), 0);
+ emit_move_insn (mem, op0);
+ op0 = adjust_address (mem, BLKmode, 0);
+ }
}
}
? bitpos + bitsize == BITS_PER_WORD
: bitpos == 0)))
&& ((!MEM_P (op0)
- && TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (mode),
+ && TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (mode1),
GET_MODE_BITSIZE (GET_MODE (op0)))
&& GET_MODE_SIZE (mode1) != 0
&& byte_offset % GET_MODE_SIZE (mode1) == 0)
target = gen_reg_rtx (mode);
/* Indicate for flow that the entire target reg is being set. */
- emit_insn (gen_rtx_CLOBBER (VOIDmode, target));
+ emit_clobber (target);
for (i = 0; i < nwords; i++)
{
if (GET_MODE (xtarget) != ext_mode)
{
- if (REG_P (xtarget))
+ /* Don't use LHS paradoxical subreg if explicit truncation is needed
+ between the mode of the extraction (word_mode) and the target
+ mode. Instead, create a temporary and use convert_move to set
+ the target. */
+ if (REG_P (xtarget)
+ && TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (GET_MODE (xtarget)),
+ GET_MODE_BITSIZE (ext_mode)))
{
xtarget = gen_lowpart (ext_mode, xtarget);
if (GET_MODE_SIZE (ext_mode)
return src;
if (CONSTANT_P (src))
- return simplify_gen_subreg (mode, src, src_mode,
- subreg_lowpart_offset (mode, src_mode));
+ {
+ /* simplify_gen_subreg can't be used here, as if simplify_subreg
+ fails, it will happily create (subreg (symbol_ref)) or similar
+ invalid SUBREGs. */
+ unsigned int byte = subreg_lowpart_offset (mode, src_mode);
+ rtx ret = simplify_subreg (mode, src, src_mode, byte);
+ if (ret)
+ return ret;
+
+ if (GET_MODE (src) == VOIDmode
+ || !validate_subreg (mode, src_mode, src, byte))
+ return NULL_RTX;
+
+ src = force_reg (GET_MODE (src), src);
+ return gen_rtx_SUBREG (mode, src, byte);
+ }
if (GET_MODE_CLASS (mode) == MODE_CC || GET_MODE_CLASS (src_mode) == MODE_CC)
return NULL_RTX;
optab lrotate_optab = rotl_optab;
optab rrotate_optab = rotr_optab;
enum machine_mode op1_mode;
- int try;
+ int attempt;
+ bool speed = optimize_insn_for_speed_p ();
op1 = expand_normal (amount);
op1_mode = GET_MODE (op1);
op1 = GEN_INT ((unsigned HOST_WIDE_INT) INTVAL (op1)
% GET_MODE_BITSIZE (mode));
else if (GET_CODE (op1) == SUBREG
- && subreg_lowpart_p (op1))
+ && subreg_lowpart_p (op1)
+ && INTEGRAL_MODE_P (GET_MODE (SUBREG_REG (op1))))
op1 = SUBREG_REG (op1);
}
&& INTVAL (op1) > 0
&& INTVAL (op1) < GET_MODE_BITSIZE (mode)
&& INTVAL (op1) < MAX_BITS_PER_WORD
- && shift_cost[mode][INTVAL (op1)] > INTVAL (op1) * add_cost[mode]
- && shift_cost[mode][INTVAL (op1)] != MAX_COST)
+ && shift_cost[speed][mode][INTVAL (op1)] > INTVAL (op1) * add_cost[speed][mode]
+ && shift_cost[speed][mode][INTVAL (op1)] != MAX_COST)
{
int i;
for (i = 0; i < INTVAL (op1); i++)
return shifted;
}
- for (try = 0; temp == 0 && try < 3; try++)
+ for (attempt = 0; temp == 0 && attempt < 3; attempt++)
{
enum optab_methods methods;
- if (try == 0)
+ if (attempt == 0)
methods = OPTAB_DIRECT;
- else if (try == 1)
+ else if (attempt == 1)
methods = OPTAB_WIDEN;
else
methods = OPTAB_LIB_WIDEN;
Otherwise, the cost within which multiplication by T is
impossible. */
struct mult_cost cost;
+
+ /* OPtimized for speed? */
+ bool speed;
};
/* The number of cache/hash entries. */
struct mult_cost best_cost;
struct mult_cost new_limit;
int op_cost, op_latency;
+ unsigned HOST_WIDE_INT orig_t = t;
unsigned HOST_WIDE_INT q;
int maxm = MIN (BITS_PER_WORD, GET_MODE_BITSIZE (mode));
int hash_index;
bool cache_hit = false;
enum alg_code cache_alg = alg_zero;
+ bool speed = optimize_insn_for_speed_p ();
/* Indicate that no algorithm is yet found. If no algorithm
is found, this value will be returned and indicate failure. */
fail now. */
if (t == 0)
{
- if (MULT_COST_LESS (cost_limit, zero_cost))
+ if (MULT_COST_LESS (cost_limit, zero_cost[speed]))
return;
else
{
alg_out->ops = 1;
- alg_out->cost.cost = zero_cost;
- alg_out->cost.latency = zero_cost;
+ alg_out->cost.cost = zero_cost[speed];
+ alg_out->cost.latency = zero_cost[speed];
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));
+ alg_in = XALLOCA (struct algorithm);
+ best_alg = XALLOCA (struct algorithm);
best_cost = *cost_limit;
/* Compute the hash index. */
- hash_index = (t ^ (unsigned int) mode) % NUM_ALG_HASH_ENTRIES;
+ hash_index = (t ^ (unsigned int) mode ^ (speed * 256)) % NUM_ALG_HASH_ENTRIES;
/* See if we already know what to do for T. */
if (alg_hash[hash_index].t == t
&& alg_hash[hash_index].mode == mode
+ && alg_hash[hash_index].mode == mode
+ && alg_hash[hash_index].speed == speed
&& alg_hash[hash_index].alg != alg_unknown)
{
cache_alg = alg_hash[hash_index].alg;
q = t >> m;
/* The function expand_shift will choose between a shift and
a sequence of additions, so the observed cost is given as
- MIN (m * add_cost[mode], shift_cost[mode][m]). */
- op_cost = m * add_cost[mode];
- if (shift_cost[mode][m] < op_cost)
- op_cost = shift_cost[mode][m];
+ MIN (m * add_cost[speed][mode], shift_cost[speed][mode][m]). */
+ op_cost = m * add_cost[speed][mode];
+ if (shift_cost[speed][mode][m] < op_cost)
+ op_cost = shift_cost[speed][mode][m];
new_limit.cost = best_cost.cost - op_cost;
new_limit.latency = best_cost.latency - op_cost;
synth_mult (alg_in, q, &new_limit, mode);
best_alg->log[best_alg->ops] = m;
best_alg->op[best_alg->ops] = alg_shift;
}
+
+ /* See if treating ORIG_T as a signed number yields a better
+ sequence. Try this sequence only for a negative ORIG_T
+ as it would be useless for a non-negative ORIG_T. */
+ if ((HOST_WIDE_INT) orig_t < 0)
+ {
+ /* Shift ORIG_T as follows because a right shift of a
+ negative-valued signed type is implementation
+ defined. */
+ q = ~(~orig_t >> m);
+ /* The function expand_shift will choose between a shift
+ and a sequence of additions, so the observed cost is
+ given as MIN (m * add_cost[speed][mode],
+ shift_cost[speed][mode][m]). */
+ op_cost = m * add_cost[speed][mode];
+ if (shift_cost[speed][mode][m] < op_cost)
+ op_cost = shift_cost[speed][mode][m];
+ new_limit.cost = best_cost.cost - op_cost;
+ new_limit.latency = best_cost.latency - op_cost;
+ synth_mult (alg_in, q, &new_limit, mode);
+
+ alg_in->cost.cost += op_cost;
+ alg_in->cost.latency += op_cost;
+ if (CHEAPER_MULT_COST (&alg_in->cost, &best_cost))
+ {
+ struct algorithm *x;
+ best_cost = alg_in->cost;
+ 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;
+ }
+ }
}
if (cache_hit)
goto done;
{
/* T ends with ...111. Multiply by (T + 1) and subtract 1. */
- op_cost = add_cost[mode];
+ op_cost = add_cost[speed][mode];
new_limit.cost = best_cost.cost - op_cost;
new_limit.latency = best_cost.latency - op_cost;
synth_mult (alg_in, t + 1, &new_limit, mode);
{
/* T ends with ...01 or ...011. Multiply by (T - 1) and add 1. */
- op_cost = add_cost[mode];
+ op_cost = add_cost[speed][mode];
new_limit.cost = best_cost.cost - op_cost;
new_limit.latency = best_cost.latency - op_cost;
synth_mult (alg_in, t - 1, &new_limit, mode);
best_alg->op[best_alg->ops] = alg_add_t_m2;
}
}
+
+ /* We may be able to calculate a * -7, a * -15, a * -31, etc
+ quickly with a - a * n for some appropriate constant n. */
+ m = exact_log2 (-orig_t + 1);
+ if (m >= 0 && m < maxm)
+ {
+ op_cost = shiftsub1_cost[speed][mode][m];
+ new_limit.cost = best_cost.cost - op_cost;
+ new_limit.latency = best_cost.latency - op_cost;
+ synth_mult (alg_in, (unsigned HOST_WIDE_INT) (-orig_t + 1) >> m, &new_limit, mode);
+
+ alg_in->cost.cost += op_cost;
+ alg_in->cost.latency += op_cost;
+ if (CHEAPER_MULT_COST (&alg_in->cost, &best_cost))
+ {
+ struct algorithm *x;
+ best_cost = alg_in->cost;
+ 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_t_m2;
+ }
+ }
+
if (cache_hit)
goto done;
}
equal to its cost, otherwise assume that on superscalar
hardware the shift may be executed concurrently with the
earlier steps in the algorithm. */
- op_cost = add_cost[mode] + shift_cost[mode][m];
- if (shiftadd_cost[mode][m] < op_cost)
+ op_cost = add_cost[speed][mode] + shift_cost[speed][mode][m];
+ if (shiftadd_cost[speed][mode][m] < op_cost)
{
- op_cost = shiftadd_cost[mode][m];
+ op_cost = shiftadd_cost[speed][mode][m];
op_latency = op_cost;
}
else
- op_latency = add_cost[mode];
+ op_latency = add_cost[speed][mode];
new_limit.cost = best_cost.cost - op_cost;
new_limit.latency = best_cost.latency - op_latency;
equal to it's cost, otherwise assume that on superscalar
hardware the shift may be executed concurrently with the
earlier steps in the algorithm. */
- op_cost = add_cost[mode] + shift_cost[mode][m];
- if (shiftsub_cost[mode][m] < op_cost)
+ op_cost = add_cost[speed][mode] + shift_cost[speed][mode][m];
+ if (shiftsub0_cost[speed][mode][m] < op_cost)
{
- op_cost = shiftsub_cost[mode][m];
+ op_cost = shiftsub0_cost[speed][mode][m];
op_latency = op_cost;
}
else
- op_latency = add_cost[mode];
+ op_latency = add_cost[speed][mode];
new_limit.cost = best_cost.cost - op_cost;
new_limit.latency = best_cost.latency - op_latency;
m = exact_log2 (q);
if (m >= 0 && m < maxm)
{
- op_cost = shiftadd_cost[mode][m];
+ op_cost = shiftadd_cost[speed][mode][m];
new_limit.cost = best_cost.cost - op_cost;
new_limit.latency = best_cost.latency - op_cost;
synth_mult (alg_in, (t - 1) >> m, &new_limit, mode);
m = exact_log2 (q);
if (m >= 0 && m < maxm)
{
- op_cost = shiftsub_cost[mode][m];
+ op_cost = shiftsub0_cost[speed][mode][m];
new_limit.cost = best_cost.cost - op_cost;
new_limit.latency = best_cost.latency - op_cost;
synth_mult (alg_in, (t + 1) >> m, &new_limit, mode);
caller. */
alg_hash[hash_index].t = t;
alg_hash[hash_index].mode = mode;
+ alg_hash[hash_index].speed = speed;
alg_hash[hash_index].alg = alg_impossible;
alg_hash[hash_index].cost = *cost_limit;
return;
{
alg_hash[hash_index].t = t;
alg_hash[hash_index].mode = mode;
+ alg_hash[hash_index].speed = speed;
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;
struct algorithm alg2;
struct mult_cost limit;
int op_cost;
+ bool speed = optimize_insn_for_speed_p ();
/* Fail quickly for impossible bounds. */
if (mult_cost < 0)
/* Ensure that mult_cost provides a reasonable upper bound.
Any constant multiplication can be performed with less
than 2 * bits additions. */
- op_cost = 2 * GET_MODE_BITSIZE (mode) * add_cost[mode];
+ op_cost = 2 * GET_MODE_BITSIZE (mode) * add_cost[speed][mode];
if (mult_cost > op_cost)
mult_cost = op_cost;
`unsigned int' */
if (HOST_BITS_PER_INT >= GET_MODE_BITSIZE (mode))
{
- op_cost = neg_cost[mode];
+ op_cost = neg_cost[speed][mode];
if (MULT_COST_LESS (&alg->cost, mult_cost))
{
limit.cost = alg->cost.cost - op_cost;
}
/* This proves very useful for division-by-constant. */
- op_cost = add_cost[mode];
+ op_cost = add_cost[speed][mode];
if (MULT_COST_LESS (&alg->cost, mult_cost))
{
limit.cost = alg->cost.cost - op_cost;
enum mult_variant variant;
struct algorithm algorithm;
int max_cost;
+ bool speed = optimize_insn_for_speed_p ();
/* Handling const0_rtx here allows us to use zero as a rogue value for
coeff below. */
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];
+ max_cost = rtx_cost (gen_rtx_MULT (mode, fake_reg, op1), SET, speed)
+ - neg_cost[speed][mode];
if (max_cost > 0
&& choose_mult_variant (mode, -INTVAL (op1), &algorithm,
&variant, max_cost))
{
/* 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)
+ if (CONST_DOUBLE_HIGH (op1) == 0
+ && CONST_DOUBLE_LOW (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)))
/* 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);
+ max_cost = rtx_cost (gen_rtx_MULT (mode, fake_reg, op1), SET, speed);
if (choose_mult_variant (mode, coeff, &algorithm, &variant,
max_cost))
return expand_mult_const (mode, op0, coeff, target,
optab moptab;
rtx tem;
int size;
+ bool speed = optimize_insn_for_speed_p ();
gcc_assert (!SCALAR_FLOAT_MODE_P (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[mode] < max_cost)
+ if (mul_highpart_cost[speed][mode] < max_cost)
{
moptab = unsignedp ? umul_highpart_optab : smul_highpart_optab;
tem = expand_binop (mode, moptab, op0, narrow_op1, 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[mode] + 2 * shift_cost[mode][size-1]
- + 4 * add_cost[mode] < max_cost))
+ && (mul_highpart_cost[speed][mode] + 2 * shift_cost[speed][mode][size-1]
+ + 4 * add_cost[speed][mode] < max_cost))
{
moptab = unsignedp ? smul_highpart_optab : umul_highpart_optab;
tem = expand_binop (mode, moptab, op0, narrow_op1, target,
/* Try widening multiplication. */
moptab = unsignedp ? umul_widen_optab : smul_widen_optab;
if (optab_handler (moptab, wider_mode)->insn_code != CODE_FOR_nothing
- && mul_widen_cost[wider_mode] < max_cost)
+ && mul_widen_cost[speed][wider_mode] < max_cost)
{
tem = expand_binop (wider_mode, moptab, op0, narrow_op1, 0,
unsignedp, OPTAB_WIDEN);
/* Try widening the mode and perform a non-widening multiplication. */
if (optab_handler (smul_optab, wider_mode)->insn_code != CODE_FOR_nothing
&& size - 1 < BITS_PER_WORD
- && mul_cost[wider_mode] + shift_cost[mode][size-1] < max_cost)
+ && mul_cost[speed][wider_mode] + shift_cost[speed][mode][size-1] < max_cost)
{
rtx insns, wop0, wop1;
moptab = unsignedp ? smul_widen_optab : umul_widen_optab;
if (optab_handler (moptab, wider_mode)->insn_code != CODE_FOR_nothing
&& size - 1 < BITS_PER_WORD
- && (mul_widen_cost[wider_mode] + 2 * shift_cost[mode][size-1]
- + 4 * add_cost[mode] < max_cost))
+ && (mul_widen_cost[speed][wider_mode] + 2 * shift_cost[speed][mode][size-1]
+ + 4 * add_cost[speed][mode] < max_cost))
{
tem = expand_binop (wider_mode, moptab, op0, narrow_op1,
NULL_RTX, ! unsignedp, OPTAB_WIDEN);
enum mult_variant variant;
struct algorithm alg;
rtx tem;
+ bool speed = optimize_insn_for_speed_p ();
gcc_assert (!SCALAR_FLOAT_MODE_P (mode));
/* We can't support modes wider than HOST_BITS_PER_INT. */
return expand_mult_highpart_optab (mode, op0, op1, target,
unsignedp, max_cost);
- extra_cost = shift_cost[mode][GET_MODE_BITSIZE (mode) - 1];
+ extra_cost = shift_cost[speed][mode][GET_MODE_BITSIZE (mode) - 1];
/* Check whether we try to multiply by a negative constant. */
if (!unsignedp && ((cnst1 >> (GET_MODE_BITSIZE (mode) - 1)) & 1))
{
sign_adjust = true;
- extra_cost += add_cost[mode];
+ extra_cost += add_cost[speed][mode];
}
/* See whether shift/add multiplication is cheap enough. */
result = gen_reg_rtx (mode);
/* Avoid conditional branches when they're expensive. */
- if (BRANCH_COST >= 2
- && !optimize_size)
+ if (BRANCH_COST (optimize_insn_for_speed_p (), false) >= 2
+ && optimize_insn_for_speed_p ())
{
rtx signmask = emit_store_flag (result, LT, op0, const0_rtx,
mode, 0, -1);
temp = gen_rtx_LSHIFTRT (mode, result, shift);
if (optab_handler (lshr_optab, mode)->insn_code == CODE_FOR_nothing
- || rtx_cost (temp, SET) > COSTS_N_INSNS (2))
+ || rtx_cost (temp, SET, optimize_insn_for_speed_p ()) > COSTS_N_INSNS (2))
{
temp = expand_binop (mode, xor_optab, op0, signmask,
NULL_RTX, 1, OPTAB_LIB_WIDEN);
logd = floor_log2 (d);
shift = build_int_cst (NULL_TREE, logd);
- if (d == 2 && BRANCH_COST >= 1)
+ if (d == 2
+ && BRANCH_COST (optimize_insn_for_speed_p (),
+ false) >= 1)
{
temp = gen_reg_rtx (mode);
temp = emit_store_flag (temp, LT, op0, const0_rtx, mode, 0, 1);
}
#ifdef HAVE_conditional_move
- if (BRANCH_COST >= 2)
+ if (BRANCH_COST (optimize_insn_for_speed_p (), false)
+ >= 2)
{
rtx temp2;
}
#endif
- if (BRANCH_COST >= 2)
+ if (BRANCH_COST (optimize_insn_for_speed_p (),
+ false) >= 2)
{
int ushift = GET_MODE_BITSIZE (mode) - logd;
temp = gen_reg_rtx (mode);
temp = emit_store_flag (temp, LT, op0, const0_rtx, mode, 0, -1);
- if (shift_cost[mode][ushift] > COSTS_N_INSNS (1))
+ if (shift_cost[optimize_insn_for_speed_p ()][mode][ushift] > COSTS_N_INSNS (1))
temp = expand_binop (mode, and_optab, temp, GEN_INT (d - 1),
NULL_RTX, 0, OPTAB_LIB_WIDEN);
else
int max_cost, extra_cost;
static HOST_WIDE_INT last_div_const = 0;
static HOST_WIDE_INT ext_op1;
+ bool speed = optimize_insn_for_speed_p ();
op1_is_constant = GET_CODE (op1) == CONST_INT;
if (op1_is_constant)
/* Only deduct something for a REM if the last divide done was
for a different constant. Then set the constant of the last
divide. */
- max_cost = unsignedp ? udiv_cost[compute_mode] : sdiv_cost[compute_mode];
+ max_cost = unsignedp ? udiv_cost[speed][compute_mode] : sdiv_cost[speed][compute_mode];
if (rem_flag && ! (last_div_const != 0 && op1_is_constant
&& INTVAL (op1) == last_div_const))
- max_cost -= mul_cost[compute_mode] + add_cost[compute_mode];
+ max_cost -= mul_cost[speed][compute_mode] + add_cost[speed][compute_mode];
last_div_const = ! rem_flag && op1_is_constant ? INTVAL (op1) : 0;
{
/* Most significant bit of divisor is set; emit an scc
insn. */
- quotient = emit_store_flag (tquotient, GEU, op0, op1,
- compute_mode, 1, 1);
- if (quotient == 0)
- goto fail1;
+ quotient = emit_store_flag_force (tquotient, GEU, op0, op1,
+ compute_mode, 1, 1);
}
else
{
goto fail1;
extra_cost
- = (shift_cost[compute_mode][post_shift - 1]
- + shift_cost[compute_mode][1]
- + 2 * add_cost[compute_mode]);
+ = (shift_cost[speed][compute_mode][post_shift - 1]
+ + shift_cost[speed][compute_mode][1]
+ + 2 * add_cost[speed][compute_mode]);
t1 = expand_mult_highpart (compute_mode, op0, ml,
NULL_RTX, 1,
max_cost - extra_cost);
build_int_cst (NULL_TREE, pre_shift),
NULL_RTX, 1);
extra_cost
- = (shift_cost[compute_mode][pre_shift]
- + shift_cost[compute_mode][post_shift]);
+ = (shift_cost[speed][compute_mode][pre_shift]
+ + shift_cost[speed][compute_mode][post_shift]);
t2 = expand_mult_highpart (compute_mode, t1, ml,
NULL_RTX, 1,
max_cost - extra_cost);
goto fail1;
}
else if (EXACT_POWER_OF_2_OR_ZERO_P (d)
- && (rem_flag ? smod_pow2_cheap[compute_mode]
- : sdiv_pow2_cheap[compute_mode])
+ && (rem_flag ? smod_pow2_cheap[speed][compute_mode]
+ : sdiv_pow2_cheap[speed][compute_mode])
/* We assume that cheap metric is true if the
optab has an expander for this mode. */
&& ((optab_handler ((rem_flag ? smod_optab
return gen_lowpart (mode, remainder);
}
- if (sdiv_pow2_cheap[compute_mode]
+ if (sdiv_pow2_cheap[speed][compute_mode]
&& ((optab_handler (sdiv_optab, compute_mode)->insn_code
!= CODE_FOR_nothing)
|| (optab_handler (sdivmod_optab, compute_mode)->insn_code
|| size - 1 >= BITS_PER_WORD)
goto fail1;
- extra_cost = (shift_cost[compute_mode][post_shift]
- + shift_cost[compute_mode][size - 1]
- + add_cost[compute_mode]);
+ extra_cost = (shift_cost[speed][compute_mode][post_shift]
+ + shift_cost[speed][compute_mode][size - 1]
+ + add_cost[speed][compute_mode]);
t1 = expand_mult_highpart (compute_mode, op0, mlr,
NULL_RTX, 0,
max_cost - extra_cost);
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]);
+ extra_cost = (shift_cost[speed][compute_mode][post_shift]
+ + shift_cost[speed][compute_mode][size - 1]
+ + 2 * add_cost[speed][compute_mode]);
t1 = expand_mult_highpart (compute_mode, op0, mlr,
NULL_RTX, 0,
max_cost - extra_cost);
NULL_RTX, 0);
t2 = expand_binop (compute_mode, xor_optab, op0, t1,
NULL_RTX, 0, OPTAB_WIDEN);
- extra_cost = (shift_cost[compute_mode][post_shift]
- + shift_cost[compute_mode][size - 1]
- + 2 * add_cost[compute_mode]);
+ extra_cost = (shift_cost[speed][compute_mode][post_shift]
+ + shift_cost[speed][compute_mode][size - 1]
+ + 2 * add_cost[speed][compute_mode]);
t3 = expand_mult_highpart (compute_mode, t2, ml,
NULL_RTX, 1,
max_cost - extra_cost);
enum insn_code icode;
enum machine_mode compare_mode;
enum machine_mode target_mode = GET_MODE (target);
+ enum mode_class mclass;
rtx tem;
- rtx last = get_last_insn ();
+ rtx last;
rtx pattern, comparison;
if (unsignedp)
return op0;
}
- icode = setcc_gen_code[(int) code];
-
- if (icode != CODE_FOR_nothing)
+ mclass = GET_MODE_CLASS (mode);
+ for (compare_mode = mode; compare_mode != VOIDmode;
+ compare_mode = GET_MODE_WIDER_MODE (compare_mode))
{
- insn_operand_predicate_fn pred;
-
- /* We think we may be able to do this with a scc insn. Emit the
- comparison and then the scc insn. */
-
- do_pending_stack_adjust ();
- last = get_last_insn ();
-
- comparison
- = compare_from_rtx (op0, op1, code, unsignedp, mode, NULL_RTX);
- if (CONSTANT_P (comparison))
- {
- switch (GET_CODE (comparison))
- {
- case CONST_INT:
- if (comparison == const0_rtx)
- return const0_rtx;
- break;
-
-#ifdef FLOAT_STORE_FLAG_VALUE
- case CONST_DOUBLE:
- if (comparison == CONST0_RTX (GET_MODE (comparison)))
- return const0_rtx;
- break;
-#endif
- default:
- gcc_unreachable ();
- }
-
- if (normalizep == 1)
- return const1_rtx;
- if (normalizep == -1)
- return constm1_rtx;
- return const_true_rtx;
- }
-
- /* The code of COMPARISON may not match CODE if compare_from_rtx
- decided to swap its operands and reverse the original code.
-
- We know that compare_from_rtx returns either a CONST_INT or
- a new comparison code, so it is safe to just extract the
- code from COMPARISON. */
- code = GET_CODE (comparison);
-
- /* Get a reference to the target in the proper mode for this insn. */
- compare_mode = insn_data[(int) icode].operand[0].mode;
- subtarget = target;
- pred = insn_data[(int) icode].operand[0].predicate;
- if (optimize || ! (*pred) (subtarget, compare_mode))
- subtarget = gen_reg_rtx (compare_mode);
-
- pattern = GEN_FCN (icode) (subtarget);
- if (pattern)
- {
- emit_insn (pattern);
- return emit_store_flag_1 (target, subtarget, compare_mode,
- normalizep);
- }
- }
- else
- {
- /* We don't have an scc insn, so try a cstore insn. */
-
- for (compare_mode = mode; compare_mode != VOIDmode;
- compare_mode = GET_MODE_WIDER_MODE (compare_mode))
- {
- icode = optab_handler (cstore_optab, compare_mode)->insn_code;
- if (icode != CODE_FOR_nothing)
- break;
- }
-
- if (icode != CODE_FOR_nothing)
+ enum machine_mode optab_mode = mclass == MODE_CC ? CCmode : compare_mode;
+ icode = optab_handler (cstore_optab, optab_mode)->insn_code;
+ if (icode != CODE_FOR_nothing)
{
+ rtx x, y;
enum machine_mode result_mode
= insn_data[(int) icode].operand[0].mode;
- rtx cstore_op0 = op0;
- rtx cstore_op1 = op1;
do_pending_stack_adjust ();
last = get_last_insn ();
- if (compare_mode != mode)
+ x = prepare_operand (icode, op0, 2, mode, compare_mode, unsignedp);
+ y = prepare_operand (icode, op1, 3, mode, compare_mode, unsignedp);
+ comparison = gen_rtx_fmt_ee (code, result_mode, x, y);
+ if (!x || !y
+ || !insn_data[icode].operand[2].predicate
+ (x, insn_data[icode].operand[2].mode)
+ || !insn_data[icode].operand[3].predicate
+ (y, insn_data[icode].operand[3].mode)
+ || !insn_data[icode].operand[1].predicate (comparison, VOIDmode))
{
- cstore_op0 = convert_modes (compare_mode, mode, cstore_op0,
- unsignedp);
- cstore_op1 = convert_modes (compare_mode, mode, cstore_op1,
- unsignedp);
+ delete_insns_since (last);
+ continue;
}
-
- if (!insn_data[(int) icode].operand[2].predicate (cstore_op0,
- compare_mode))
- cstore_op0 = copy_to_mode_reg (compare_mode, cstore_op0);
-
- if (!insn_data[(int) icode].operand[3].predicate (cstore_op1,
- compare_mode))
- cstore_op1 = copy_to_mode_reg (compare_mode, cstore_op1);
- comparison = gen_rtx_fmt_ee (code, result_mode, cstore_op0,
- cstore_op1);
subtarget = target;
-
if (optimize || !(insn_data[(int) icode].operand[0].predicate
(subtarget, result_mode)))
subtarget = gen_reg_rtx (result_mode);
- pattern = GEN_FCN (icode) (subtarget, comparison, cstore_op0,
- cstore_op1);
+ pattern = GEN_FCN (icode) (subtarget, comparison, x, y);
if (pattern)
{
return emit_store_flag_1 (target, subtarget, result_mode,
normalizep);
}
+
+ delete_insns_since (last);
+ break;
}
}
- delete_insns_since (last);
+ last = get_last_insn ();
/* If optimizing, use different pseudo registers for each insn, instead
of reusing the same pseudo. This leads to better CSE, but slows
comparison with zero. Don't do any of these cases if branches are
very cheap. */
- if (BRANCH_COST > 0
+ if (BRANCH_COST (optimize_insn_for_speed_p (),
+ false) > 0
&& GET_MODE_CLASS (mode) == MODE_INT && (code == EQ || code == NE)
&& op1 != const0_rtx)
{
do LE and GT if branches are expensive since they are expensive on
2-operand machines. */
- if (BRANCH_COST == 0
+ if (BRANCH_COST (optimize_insn_for_speed_p (),
+ false) == 0
|| GET_MODE_CLASS (mode) != MODE_INT || op1 != const0_rtx
|| (code != EQ && code != NE
- && (BRANCH_COST <= 1 || (code != LE && code != GT))))
+ && (BRANCH_COST (optimize_insn_for_speed_p (),
+ false) <= 1 || (code != LE && code != GT))))
return 0;
/* See what we need to return. We can only return a 1, -1, or the
that "or", which is an extra insn, so we only handle EQ if branches
are expensive. */
- if (tem == 0 && (code == NE || BRANCH_COST > 1))
+ if (tem == 0
+ && (code == NE
+ || BRANCH_COST (optimize_insn_for_speed_p (),
+ false) > 1))
{
if (rtx_equal_p (subtarget, op0))
subtarget = 0;