/* 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 Free Software Foundation, Inc.
+ 1999, 2000, 2001, 2002, 2003, 2004 Free Software Foundation, Inc.
This file is part of GCC.
Usually, this will mean that the MD file will emit non-branch
sequences. */
-static int sdiv_pow2_cheap, smod_pow2_cheap;
+static int sdiv_pow2_cheap[NUM_MACHINE_MODES];
+static int smod_pow2_cheap[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 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 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 div_cost[NUM_MACHINE_MODES];
static int mul_widen_cost[NUM_MACHINE_MODES];
init_expmed (void)
{
rtx reg, shift_insn, shiftadd_insn, shiftsub_insn;
+ rtx shift_pat, shiftadd_pat, shiftsub_pat;
+ rtx pow2[MAX_BITS_PER_WORD];
+ rtx cint[MAX_BITS_PER_WORD];
int dummy;
- int m;
+ int m, n;
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);
+ pow2[m] = GEN_INT ((HOST_WIDE_INT) 1 << m);
+ cint[m] = GEN_INT (m);
}
- 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);
+ add_cost[mode] = rtx_cost (gen_rtx_PLUS (mode, reg, reg), SET);
+ neg_cost[mode] = rtx_cost (gen_rtx_NEG (mode, reg), SET);
+ div_cost[mode] = rtx_cost (gen_rtx_UDIV (mode, reg, reg), SET);
+ mul_cost[mode] = rtx_cost (gen_rtx_MULT (mode, reg, reg), SET);
+
+ sdiv_pow2_cheap[mode]
+ = (rtx_cost (gen_rtx_DIV (mode, reg, GEN_INT (32)), SET)
+ <= 2 * add_cost[mode]);
+ smod_pow2_cheap[mode]
+ = (rtx_cost (gen_rtx_MOD (mode, reg, GEN_INT (32)), SET)
+ <= 2 * add_cost[mode]);
+
wider_mode = GET_MODE_WIDER_MODE (mode);
if (wider_mode != VOIDmode)
{
- mul_widen_cost[(int) wider_mode]
+ mul_widen_cost[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]
+ mul_highpart_cost[mode]
= rtx_cost (gen_rtx_TRUNCATE
(mode,
gen_rtx_LSHIFTRT (wider_mode,
GEN_INT (GET_MODE_BITSIZE (mode)))),
SET);
}
+
+ shift_insn = emit_insn (gen_rtx_SET (VOIDmode, reg,
+ gen_rtx_ASHIFT (mode, reg,
+ const0_rtx)));
+
+ shiftadd_insn
+ = emit_insn (gen_rtx_SET (VOIDmode, reg,
+ gen_rtx_PLUS (mode,
+ gen_rtx_MULT (mode,
+ reg,
+ const0_rtx),
+ reg)));
+
+ shiftsub_insn
+ = emit_insn (gen_rtx_SET (VOIDmode, reg,
+ gen_rtx_MINUS (mode,
+ gen_rtx_MULT (mode,
+ reg,
+ const0_rtx),
+ reg)));
+
+ shift_pat = PATTERN (shift_insn);
+ shiftadd_pat = PATTERN (shiftadd_insn);
+ shiftsub_pat = PATTERN (shiftsub_insn);
+
+ shift_cost[mode][0] = 0;
+ shiftadd_cost[mode][0] = shiftsub_cost[mode][0] = add_cost[mode];
+
+ n = MIN (MAX_BITS_PER_WORD, GET_MODE_BITSIZE (mode));
+ for (m = 1; m < n; m++)
+ {
+ shift_cost[mode][m] = 32000;
+ XEXP (SET_SRC (shift_pat), 1) = cint[m];
+ if (recog (shift_pat, shift_insn, &dummy) >= 0)
+ shift_cost[mode][m] = rtx_cost (SET_SRC (shift_pat), SET);
+
+ shiftadd_cost[mode][m] = 32000;
+ XEXP (XEXP (SET_SRC (shiftadd_pat), 0), 1) = pow2[m];
+ if (recog (shiftadd_pat, shiftadd_insn, &dummy) >= 0)
+ shiftadd_cost[mode][m] = rtx_cost (SET_SRC (shiftadd_pat), SET);
+
+ shiftsub_cost[mode][m] = 32000;
+ XEXP (XEXP (SET_SRC (shiftsub_pat), 0), 1) = pow2[m];
+ if (recog (shiftsub_pat, shiftsub_insn, &dummy) >= 0)
+ shiftsub_cost[mode][m] = rtx_cost (SET_SRC (shiftsub_pat), SET);
+ }
}
end_sequence ();
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[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[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;
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.
-
- It's okay if the requested bitsize is greater than fieldmode's
- bitsize; that just means the mode has padding bits. */
+ 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)
+ && 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))
subregs results in Severe Tire Damage. */
abort ();
}
- if (GET_CODE (op0) == REG)
+ if (REG_P (op0))
op0 = gen_rtx_SUBREG (fieldmode, op0, byte_offset);
else
op0 = adjust_address (op0, fieldmode, offset);
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
+ && (movstrict_optab->handlers[fieldmode].insn_code
!= CODE_FOR_nothing))
{
- int icode = movstrict_optab->handlers[(int) fieldmode].insn_code;
+ int icode = movstrict_optab->handlers[fieldmode].insn_code;
/* Get appropriate low part of the value being stored. */
- if (GET_CODE (value) == CONST_INT || GET_CODE (value) == REG)
+ if (GET_CODE (value) == CONST_INT || REG_P (value))
value = gen_lowpart (fieldmode, value);
else if (!(GET_CODE (value) == SYMBOL_REF
|| GET_CODE (value) == LABEL_REF
if (offset != 0
|| GET_MODE_SIZE (GET_MODE (op0)) > UNITS_PER_WORD)
{
- if (GET_CODE (op0) != REG)
+ 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
&& !(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)
+ && ! ((REG_P (op0) || GET_CODE (op0) == SUBREG)
&& (bitsize + bitpos > GET_MODE_BITSIZE (op_mode))))
{
int xbitpos = bitpos;
/* 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)
+ if (REG_P (xop0) && GET_MODE (xop0) != maxmode)
xop0 = gen_rtx_SUBREG (maxmode, xop0, 0);
/* On big-endian machines, we count bits from the most significant.
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 (REG_P (op0) || GET_CODE (op0) == SUBREG)
{
if (offset != 0)
abort ();
if (GET_MODE (value) != mode)
{
- if ((GET_CODE (value) == REG || GET_CODE (value) == SUBREG)
+ if ((REG_P (value) || GET_CODE (value) == SUBREG)
&& GET_MODE_SIZE (mode) < GET_MODE_SIZE (GET_MODE (value)))
value = gen_lowpart (mode, value);
else
/* 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;
+ subtarget = (REG_P (op0) || ! flag_force_mem) ? op0 : 0;
if (! all_one)
{
/* 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)
+ if (REG_P (op0) || GET_CODE (op0) == SUBREG)
unit = BITS_PER_WORD;
else
unit = MIN (MEM_ALIGN (op0), BITS_PER_WORD);
GET_MODE (SUBREG_REG (op0)));
offset = 0;
}
- else if (GET_CODE (op0) == REG)
+ else if (REG_P (op0))
{
word = operand_subword_force (op0, offset, GET_MODE (op0));
offset = 0;
op0 = SUBREG_REG (op0);
}
- if (GET_CODE (op0) == REG
+ if (REG_P (op0)
&& mode == GET_MODE (op0)
&& bitnum == 0
- && bitsize >= GET_MODE_BITSIZE (GET_MODE (op0)))
+ && bitsize == GET_MODE_BITSIZE (GET_MODE (op0)))
{
- /* We're trying to extract a full register from itself.
- (If the requested bitsize is greater than the bitsize of op0,
- that just means op0's mode has padding bits.) */
+ /* 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[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[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. */
{
If that's wrong, the solution is to test for it and set TARGET to 0
if needed. */
- mode1 = (VECTOR_MODE_P (tmode)
- ? mode
- : mode_for_size (bitsize, GET_MODE_CLASS (tmode), 0));
+ /* 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)
- || (mode_for_size (bitsize, GET_MODE_CLASS (tmode), 0) != BLKmode
+ || (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
subregs results in Severe Tire Damage. */
goto no_subreg_mode_swap;
}
- if (GET_CODE (op0) == REG)
+ if (REG_P (op0))
op0 = gen_rtx_SUBREG (mode1, op0, byte_offset);
else
op0 = adjust_address (op0, mode1, offset);
unsigned int nwords = (bitsize + (BITS_PER_WORD - 1)) / BITS_PER_WORD;
unsigned int i;
- if (target == 0 || GET_CODE (target) != REG)
+ if (target == 0 || !REG_P (target))
target = gen_reg_rtx (mode);
/* Indicate for flow that the entire target reg is being set. */
if (offset != 0
|| GET_MODE_SIZE (GET_MODE (op0)) > UNITS_PER_WORD)
{
- if (GET_CODE (op0) != REG)
+ if (!REG_P (op0))
op0 = copy_to_reg (op0);
op0 = gen_rtx_SUBREG (mode_for_size (BITS_PER_WORD, MODE_INT, 0),
op0, (offset * UNITS_PER_WORD));
{
if (HAVE_extzv
&& (GET_MODE_BITSIZE (extzv_mode) >= bitsize)
- && ! ((GET_CODE (op0) == REG || GET_CODE (op0) == SUBREG)
+ && ! ((REG_P (op0) || GET_CODE (op0) == SUBREG)
&& (bitsize + bitpos > GET_MODE_BITSIZE (extzv_mode))))
{
unsigned HOST_WIDE_INT xbitpos = bitpos, xoffset = offset;
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)
+ if (REG_P (xop0) && GET_MODE (xop0) != maxmode)
xop0 = gen_rtx_SUBREG (maxmode, xop0, 0);
/* On big-endian machines, we count bits from the most significant.
if (GET_MODE (xtarget) != maxmode)
{
- if (GET_CODE (xtarget) == REG)
+ if (REG_P (xtarget))
{
int wider = (GET_MODE_SIZE (maxmode)
> GET_MODE_SIZE (GET_MODE (xtarget)));
{
if (HAVE_extv
&& (GET_MODE_BITSIZE (extv_mode) >= bitsize)
- && ! ((GET_CODE (op0) == REG || GET_CODE (op0) == SUBREG)
+ && ! ((REG_P (op0) || GET_CODE (op0) == SUBREG)
&& (bitsize + bitpos > GET_MODE_BITSIZE (extv_mode))))
{
int xbitpos = bitpos, xoffset = offset;
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)
+ if (REG_P (xop0) && GET_MODE (xop0) != maxmode)
xop0 = gen_rtx_SUBREG (maxmode, xop0, 0);
/* On big-endian machines, we count bits from the most significant.
if (GET_MODE (xtarget) != maxmode)
{
- if (GET_CODE (xtarget) == REG)
+ if (REG_P (xtarget))
{
int wider = (GET_MODE_SIZE (maxmode)
> GET_MODE_SIZE (GET_MODE (xtarget)));
unsigned int total_bits = BITS_PER_WORD;
enum machine_mode mode;
- if (GET_CODE (op0) == SUBREG || GET_CODE (op0) == REG)
+ if (GET_CODE (op0) == SUBREG || REG_P (op0))
{
/* Special treatment for a bit field split across two registers. */
if (bitsize + bitpos > BITS_PER_WORD)
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);
+ rtx subtarget = (target != 0 && REG_P (target) ? target : 0);
if (tmode != mode) subtarget = 0;
op0 = expand_shift (RSHIFT_EXPR, mode, op0, amount, subtarget, 1);
}
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);
+ rtx subtarget = (target != 0 && REG_P (target) ? target : 0);
op0 = expand_shift (LSHIFT_EXPR, mode, op0, amount, subtarget, 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)
+ if (REG_P (op0) || GET_CODE (op0) == SUBREG)
unit = BITS_PER_WORD;
else
unit = MIN (MEM_ALIGN (op0), BITS_PER_WORD);
GET_MODE (SUBREG_REG (op0)));
offset = 0;
}
- else if (GET_CODE (op0) == REG)
+ else if (REG_P (op0))
{
word = operand_subword_force (op0, offset, GET_MODE (op0));
offset = 0;
op1 = expand_expr (amount, NULL_RTX, VOIDmode, 0);
-#ifdef SHIFT_COUNT_TRUNCATED
if (SHIFT_COUNT_TRUNCATED)
{
if (GET_CODE (op1) == CONST_INT
&& subreg_lowpart_p (op1))
op1 = SUBREG_REG (op1);
}
-#endif
if (op1 == const0_rtx)
return shifted;
char log[MAX_BITS_PER_WORD];
};
-static void synth_mult (struct algorithm *, unsigned HOST_WIDE_INT, int);
+/* Indicates the type of fixup needed after a constant multiplication.
+ BASIC_VARIANT means no fixup is needed, NEGATE_VARIANT means that
+ the result should be negated, and ADD_VARIANT means that the
+ multiplicand should be added to the result. */
+enum mult_variant {basic_variant, negate_variant, add_variant};
+
+static void synth_mult (struct algorithm *, unsigned HOST_WIDE_INT,
+ int, enum machine_mode mode);
+static bool choose_mult_variant (enum machine_mode, HOST_WIDE_INT,
+ struct algorithm *, enum mult_variant *, int);
+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 *);
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_optab (enum machine_mode, rtx, rtx, rtx,
+ 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. */
+ other field of the returned struct are undefined.
+ MODE is the machine mode of the multiplication. */
static void
synth_mult (struct algorithm *alg_out, unsigned HOST_WIDE_INT t,
- int cost_limit)
+ int cost_limit, enum machine_mode mode)
{
int m;
struct algorithm *alg_in, *best_alg;
int cost;
unsigned HOST_WIDE_INT q;
+ int maxm = MIN (BITS_PER_WORD, GET_MODE_BITSIZE (mode));
/* Indicate that no algorithm is yet found. If no algorithm
is found, this value will be returned and indicate failure. */
if (cost_limit <= 0)
return;
+ /* Restrict the bits of "t" to the multiplication's mode. */
+ t &= GET_MODE_MASK (mode);
+
/* t == 1 can be done in zero cost. */
if (t == 1)
{
if ((t & 1) == 0)
{
m = floor_log2 (t & -t); /* m = number of low zero bits */
- if (m < BITS_PER_WORD)
+ if (m < maxm)
{
q = t >> m;
- cost = shift_cost[m];
- synth_mult (alg_in, q, cost_limit - cost);
+ cost = shift_cost[mode][m];
+ synth_mult (alg_in, q, cost_limit - cost, mode);
cost += alg_in->cost;
if (cost < cost_limit)
{
/* T ends with ...111. Multiply by (T + 1) and subtract 1. */
- cost = add_cost;
- synth_mult (alg_in, t + 1, cost_limit - cost);
+ cost = add_cost[mode];
+ synth_mult (alg_in, t + 1, cost_limit - cost, mode);
cost += alg_in->cost;
if (cost < cost_limit)
{
/* 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 = add_cost[mode];
+ synth_mult (alg_in, t - 1, cost_limit - cost, mode);
cost += alg_in->cost;
if (cost < cost_limit)
unsigned HOST_WIDE_INT d;
d = ((unsigned HOST_WIDE_INT) 1 << m) + 1;
- if (t % d == 0 && t > d && m < BITS_PER_WORD)
+ if (t % d == 0 && t > d && m < maxm)
{
- cost = MIN (shiftadd_cost[m], add_cost + shift_cost[m]);
- synth_mult (alg_in, t / d, cost_limit - cost);
+ cost = add_cost[mode] + shift_cost[mode][m];
+ if (shiftadd_cost[mode][m] < cost)
+ cost = shiftadd_cost[mode][m];
+ synth_mult (alg_in, t / d, cost_limit - cost, mode);
cost += alg_in->cost;
if (cost < cost_limit)
}
d = ((unsigned HOST_WIDE_INT) 1 << m) - 1;
- if (t % d == 0 && t > d && m < BITS_PER_WORD)
+ if (t % d == 0 && t > d && m < maxm)
{
- cost = MIN (shiftsub_cost[m], add_cost + shift_cost[m]);
- synth_mult (alg_in, t / d, cost_limit - cost);
+ cost = add_cost[mode] + shift_cost[mode][m];
+ if (shiftsub_cost[mode][m] < cost)
+ cost = shiftsub_cost[mode][m];
+ synth_mult (alg_in, t / d, cost_limit - cost, mode);
cost += alg_in->cost;
if (cost < cost_limit)
q = t - 1;
q = q & -q;
m = exact_log2 (q);
- if (m >= 0 && m < BITS_PER_WORD)
+ if (m >= 0 && m < maxm)
{
- cost = shiftadd_cost[m];
- synth_mult (alg_in, (t - 1) >> m, cost_limit - cost);
+ cost = shiftadd_cost[mode][m];
+ synth_mult (alg_in, (t - 1) >> m, cost_limit - cost, mode);
cost += alg_in->cost;
if (cost < cost_limit)
q = t + 1;
q = q & -q;
m = exact_log2 (q);
- if (m >= 0 && m < BITS_PER_WORD)
+ if (m >= 0 && m < maxm)
{
- cost = shiftsub_cost[m];
- synth_mult (alg_in, (t + 1) >> m, cost_limit - cost);
+ cost = shiftsub_cost[mode][m];
+ synth_mult (alg_in, (t + 1) >> m, cost_limit - cost, mode);
cost += alg_in->cost;
if (cost < cost_limit)
alg_out->ops * sizeof *alg_out->log);
}
\f
+/* Find the cheapest way of multiplying a value of mode MODE by VAL.
+ Try three variations:
+
+ - a shift/add sequence based on VAL itself
+ - a shift/add sequence based on -VAL, followed by a negation
+ - a shift/add sequence based on VAL - 1, followed by an addition.
+
+ Return true if the cheapest of these cost less than MULT_COST,
+ describing the algorithm in *ALG and final fixup in *VARIANT. */
+
+static bool
+choose_mult_variant (enum machine_mode mode, HOST_WIDE_INT val,
+ struct algorithm *alg, enum mult_variant *variant,
+ int mult_cost)
+{
+ struct algorithm alg2;
+
+ *variant = basic_variant;
+ synth_mult (alg, val, mult_cost, mode);
+
+ /* 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, MIN (alg->cost, mult_cost) - neg_cost[mode],
+ mode);
+ alg2.cost += neg_cost[mode];
+ if (alg2.cost < alg->cost)
+ *alg = alg2, *variant = negate_variant;
+ }
+
+ /* This proves very useful for division-by-constant. */
+ synth_mult (&alg2, val - 1, MIN (alg->cost, mult_cost) - add_cost[mode],
+ mode);
+ alg2.cost += add_cost[mode];
+ if (alg2.cost < alg->cost)
+ *alg = alg2, *variant = add_variant;
+
+ return alg->cost < mult_cost;
+}
+
+/* A subroutine of expand_mult, used for constant multiplications.
+ Multiply OP0 by VAL in mode MODE, storing the result in TARGET if
+ convenient. Use the shift/add sequence described by ALG and apply
+ the final fixup specified by VARIANT. */
+
+static rtx
+expand_mult_const (enum machine_mode mode, rtx op0, HOST_WIDE_INT val,
+ rtx target, const struct algorithm *alg,
+ enum mult_variant variant)
+{
+ HOST_WIDE_INT val_so_far;
+ rtx insn, accum, tem;
+ int opno;
+ enum machine_mode nmode;
+
+ /* op0 must be register to make mult_cost match the precomputed
+ shiftadd_cost array. */
+ 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);
+ }
+
+ /* Compare only the bits of val and val_so_far that are significant
+ in the result mode, to avoid sign-/zero-extension confusion. */
+ val &= GET_MODE_MASK (mode);
+ val_so_far &= GET_MODE_MASK (mode);
+ if (val != val_so_far)
+ abort ();
+
+ return accum;
+}
+
/* 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).
int unsignedp)
{
rtx const_op1 = op1;
+ enum mult_variant variant;
+ struct algorithm algorithm;
/* 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.
that it seems better to use synth_mult always. */
if (const_op1 && GET_CODE (const_op1) == CONST_INT
- && (unsignedp || ! flag_trapv))
+ && (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 ();
- }
+ int mult_cost = rtx_cost (gen_rtx_MULT (mode, op0, op1), SET);
+ mult_cost = MIN (12 * add_cost[mode], mult_cost);
- /* 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 (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 (op0) == CONST_DOUBLE)
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.
+/* Subroutine of expand_mult_highpart. Return the MODE high part of OP. */
- MODE is the mode of operation and result.
+static rtx
+extract_high_half (enum machine_mode mode, rtx op)
+{
+ enum machine_mode wider_mode;
- UNSIGNEDP nonzero means unsigned multiply.
+ if (mode == word_mode)
+ return gen_highpart (mode, op);
- MAX_COST is the total allowed cost for the expanded RTL. */
+ wider_mode = GET_MODE_WIDER_MODE (mode);
+ op = expand_shift (RSHIFT_EXPR, wider_mode, op,
+ build_int_2 (GET_MODE_BITSIZE (mode), 0), 0, 1);
+ return convert_modes (mode, wider_mode, op, 0);
+}
-rtx
-expand_mult_highpart (enum machine_mode mode, rtx op0,
- unsigned HOST_WIDE_INT cnst1, rtx target,
- int unsignedp, int max_cost)
+/* Like expand_mult_highpart, but only consider using a multiplication
+ optab. OP1 is an rtx for the constant operand. */
+
+static rtx
+expand_mult_highpart_optab (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);
- optab mul_highpart_optab;
+ rtx narrow_op1 = gen_int_mode (INTVAL (op1), mode);
+ enum machine_mode wider_mode;
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);
- }
+ int size;
- if (target == 0)
- target = gen_reg_rtx (mode);
+ wider_mode = GET_MODE_WIDER_MODE (mode);
+ size = GET_MODE_BITSIZE (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)
+ if (mul_highpart_cost[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;
+ moptab = unsignedp ? umul_highpart_optab : smul_highpart_optab;
+ tem = expand_binop (mode, moptab, op0, narrow_op1, target,
+ unsignedp, OPTAB_DIRECT);
+ if (tem)
+ return tem;
}
/* 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_cost[mode] + 2 * shift_cost[mode][size-1]
+ + 4 * add_cost[mode] < 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)
+ moptab = unsignedp ? smul_highpart_optab : umul_highpart_optab;
+ tem = expand_binop (mode, moptab, op0, narrow_op1, target,
+ unsignedp, OPTAB_DIRECT);
+ if (tem)
/* We used the wrong signedness. Adjust the result. */
- return expand_mult_highpart_adjust (mode, target, op0,
- op1, target, unsignedp);
+ return expand_mult_highpart_adjust (mode, tem, op0, narrow_op1,
+ tem, 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)
+ if (moptab->handlers[wider_mode].insn_code != CODE_FOR_nothing
+ && mul_widen_cost[wider_mode] < max_cost)
{
- op1 = force_reg (mode, op1);
- goto try;
+ tem = expand_binop (wider_mode, moptab, op0, narrow_op1, 0,
+ unsignedp, OPTAB_WIDEN);
+ if (tem)
+ return extract_high_half (mode, tem);
}
/* 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
+ if (smul_optab->handlers[wider_mode].insn_code != CODE_FOR_nothing
&& size - 1 < BITS_PER_WORD
- && mul_cost[(int) wider_mode] + shift_cost[size-1] < max_cost)
+ && mul_cost[wider_mode] + shift_cost[mode][size-1] < max_cost)
{
- op1 = wide_op1;
- goto try;
+ tem = expand_binop (wider_mode, moptab, op0, op1, 0,
+ unsignedp, OPTAB_WIDEN);
+ if (tem)
+ return extract_high_half (mode, tem);
}
/* 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
+ if (moptab->handlers[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))
+ && (mul_widen_cost[wider_mode] + 2 * shift_cost[mode][size-1]
+ + 4 * add_cost[mode] < max_cost))
{
- rtx regop1 = force_reg (mode, op1);
- tem = expand_binop (wider_mode, moptab, op0, regop1,
+ tem = expand_binop (wider_mode, moptab, op0, narrow_op1,
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);
+ tem = extract_high_half (mode, tem);
/* We used the wrong signedness. Adjust the result. */
- return expand_mult_highpart_adjust (mode, tem, op0, op1,
+ return expand_mult_highpart_adjust (mode, tem, op0, narrow_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;
+/* 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.
- /* Extract the high half of the just generated product. */
- if (mode == word_mode)
+ 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);
+ int extra_cost;
+ bool sign_adjust = false;
+ enum mult_variant variant;
+ struct algorithm alg;
+ rtx op1, tem;
+
+ /* We can't support modes wider than HOST_BITS_PER_INT. */
+ if (GET_MODE_BITSIZE (mode) > HOST_BITS_PER_WIDE_INT)
+ abort ();
+
+ op1 = gen_int_mode (cnst1, wider_mode);
+ cnst1 &= GET_MODE_MASK (mode);
+
+ /* We can't optimize modes wider than BITS_PER_WORD.
+ ??? We might be able to perform double-word arithmetic if
+ mode == word_mode, however all the cost calculations in
+ synth_mult etc. assume single-word operations. */
+ if (GET_MODE_BITSIZE (wider_mode) > BITS_PER_WORD)
+ return expand_mult_highpart_optab (mode, op0, op1, target,
+ unsignedp, max_cost);
+
+ extra_cost = shift_cost[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))
{
- return gen_highpart (mode, tem);
+ sign_adjust = true;
+ extra_cost += add_cost[mode];
}
- else
+
+ /* See whether shift/add multiplication is cheap enough. */
+ if (choose_mult_variant (wider_mode, cnst1, &alg, &variant,
+ max_cost - extra_cost))
{
- tem = expand_shift (RSHIFT_EXPR, wider_mode, tem,
- build_int_2 (size, 0), NULL_RTX, 1);
- return convert_modes (mode, wider_mode, tem, unsignedp);
+ /* See whether the specialized multiplication optabs are
+ cheaper than the shift/add version. */
+ tem = expand_mult_highpart_optab (mode, op0, op1, target,
+ unsignedp, alg.cost + extra_cost);
+ if (tem)
+ return tem;
+
+ tem = convert_to_mode (wider_mode, op0, unsignedp);
+ tem = expand_mult_const (wider_mode, tem, cnst1, 0, &alg, variant);
+ tem = extract_high_half (mode, tem);
+
+ /* Adjust result for signedness. */
+ if (sign_adjust)
+ tem = force_operand (gen_rtx_MINUS (mode, tem, op0), tem);
+
+ return tem;
}
+ return expand_mult_highpart_optab (mode, op0, op1, target,
+ unsignedp, max_cost);
}
\f
/* Emit the code to divide OP0 by OP1, putting the result in TARGET
for (compute_mode = mode; compute_mode != VOIDmode;
compute_mode = GET_MODE_WIDER_MODE (compute_mode))
- if (optab1->handlers[(int) compute_mode].insn_code != CODE_FOR_nothing
- || optab2->handlers[(int) compute_mode].insn_code != CODE_FOR_nothing)
+ if (optab1->handlers[compute_mode].insn_code != CODE_FOR_nothing
+ || optab2->handlers[compute_mode].insn_code != CODE_FOR_nothing)
break;
if (compute_mode == VOIDmode)
for (compute_mode = mode; compute_mode != VOIDmode;
compute_mode = GET_MODE_WIDER_MODE (compute_mode))
- if (optab1->handlers[(int) compute_mode].libfunc
- || optab2->handlers[(int) compute_mode].libfunc)
+ if (optab1->handlers[compute_mode].libfunc
+ || optab2->handlers[compute_mode].libfunc)
break;
/* If we still couldn't find a mode, use MODE, but we'll probably abort
/* 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 = div_cost[(int) compute_mode]
+ max_cost = div_cost[compute_mode]
- (rem_flag && ! (last_div_const != 0 && op1_is_constant
&& INTVAL (op1) == last_div_const)
- ? mul_cost[(int) compute_mode] + add_cost : 0);
+ ? mul_cost[compute_mode] + add_cost[compute_mode]
+ : 0);
last_div_const = ! rem_flag && op1_is_constant ? INTVAL (op1) : 0;
if (post_shift - 1 >= BITS_PER_WORD)
goto fail1;
- extra_cost = (shift_cost[post_shift - 1]
- + shift_cost[1] + 2 * add_cost);
+ extra_cost
+ = (shift_cost[compute_mode][post_shift - 1]
+ + shift_cost[compute_mode][1]
+ + 2 * add_cost[compute_mode]);
t1 = expand_mult_highpart (compute_mode, op0, ml,
NULL_RTX, 1,
max_cost - extra_cost);
t1 = expand_shift (RSHIFT_EXPR, compute_mode, op0,
build_int_2 (pre_shift, 0),
NULL_RTX, 1);
- extra_cost = (shift_cost[pre_shift]
- + shift_cost[post_shift]);
+ extra_cost
+ = (shift_cost[compute_mode][pre_shift]
+ + shift_cost[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 : sdiv_pow2_cheap)
+ && (rem_flag ? smod_pow2_cheap[compute_mode]
+ : sdiv_pow2_cheap[compute_mode])
/* ??? The cheap metric is computed only for
word_mode. If this operation is wider, this may
not be so. Assume true if the optab has an
expander for this mode. */
&& (((rem_flag ? smod_optab : sdiv_optab)
- ->handlers[(int) compute_mode].insn_code
+ ->handlers[compute_mode].insn_code
!= CODE_FOR_nothing)
- || (sdivmod_optab->handlers[(int) compute_mode]
+ || (sdivmod_optab->handlers[compute_mode]
.insn_code != CODE_FOR_nothing)))
;
else if (EXACT_POWER_OF_2_OR_ZERO_P (abs_d))
|| size - 1 >= BITS_PER_WORD)
goto fail1;
- extra_cost = (shift_cost[post_shift]
- + shift_cost[size - 1] + add_cost);
+ 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,
NULL_RTX, 0,
max_cost - extra_cost);
goto fail1;
ml |= (~(unsigned HOST_WIDE_INT) 0) << (size - 1);
- extra_cost = (shift_cost[post_shift]
- + shift_cost[size - 1] + 2 * add_cost);
+ 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,
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[post_shift]
- + shift_cost[size - 1] + 2 * add_cost);
+ extra_cost = (shift_cost[compute_mode][post_shift]
+ + shift_cost[compute_mode][size - 1]
+ + 2 * add_cost[compute_mode]);
t3 = expand_mult_highpart (compute_mode, t2, ml,
NULL_RTX, 1,
max_cost - extra_cost);
if (rem_flag)
{
remainder
- = GET_CODE (target) == REG ? target : gen_reg_rtx (compute_mode);
+ = REG_P (target) ? target : gen_reg_rtx (compute_mode);
quotient = gen_reg_rtx (compute_mode);
}
else
{
quotient
- = GET_CODE (target) == REG ? target : gen_reg_rtx (compute_mode);
+ = REG_P (target) ? target : gen_reg_rtx (compute_mode);
remainder = gen_reg_rtx (compute_mode);
}
if (rem_flag)
{
- remainder = (GET_CODE (target) == REG
+ remainder = (REG_P (target)
? target : gen_reg_rtx (compute_mode));
quotient = gen_reg_rtx (compute_mode);
}
else
{
- quotient = (GET_CODE (target) == REG
+ quotient = (REG_P (target)
? target : gen_reg_rtx (compute_mode));
remainder = gen_reg_rtx (compute_mode);
}
target = gen_reg_rtx (compute_mode);
if (rem_flag)
{
- remainder= (GET_CODE (target) == REG
+ remainder= (REG_P (target)
? target : gen_reg_rtx (compute_mode));
quotient = gen_reg_rtx (compute_mode);
}
else
{
- quotient = (GET_CODE (target) == REG
+ quotient = (REG_P (target)
? target : gen_reg_rtx (compute_mode));
remainder = gen_reg_rtx (compute_mode);
}
= sign_expand_binop (compute_mode, umod_optab, smod_optab,
op0, op1, target,
unsignedp,
- ((optab2->handlers[(int) compute_mode].insn_code
+ ((optab2->handlers[compute_mode].insn_code
!= CODE_FOR_nothing)
? OPTAB_DIRECT : OPTAB_WIDEN));
if (remainder == 0)
= sign_expand_binop (compute_mode, udiv_optab, sdiv_optab,
op0, op1, rem_flag ? NULL_RTX : target,
unsignedp,
- ((optab2->handlers[(int) compute_mode].insn_code
+ ((optab2->handlers[compute_mode].insn_code
!= CODE_FOR_nothing)
? OPTAB_DIRECT : OPTAB_WIDEN));
{
case CONST_INT:
t = build_int_2 (INTVAL (x),
- (TREE_UNSIGNED (type)
- && (GET_MODE_BITSIZE (TYPE_MODE (type)) < HOST_BITS_PER_WIDE_INT))
+ (TYPE_UNSIGNED (type)
+ && (GET_MODE_BITSIZE (TYPE_MODE (type))
+ < HOST_BITS_PER_WIDE_INT))
|| INTVAL (x) >= 0 ? 0 : -1);
TREE_TYPE (t) = type;
return t;
make_tree (type, XEXP (x, 1))));
case LSHIFTRT:
- t = (*lang_hooks.types.unsigned_type) (type);
+ t = lang_hooks.types.unsigned_type (type);
return fold (convert (type,
build (RSHIFT_EXPR, t,
make_tree (t, XEXP (x, 0)),
make_tree (type, XEXP (x, 1)))));
case ASHIFTRT:
- t = (*lang_hooks.types.signed_type) (type);
+ t = lang_hooks.types.signed_type (type);
return fold (convert (type,
build (RSHIFT_EXPR, t,
make_tree (t, XEXP (x, 0)),
case DIV:
if (TREE_CODE (type) != REAL_TYPE)
- t = (*lang_hooks.types.signed_type) (type);
+ t = lang_hooks.types.signed_type (type);
else
t = type;
make_tree (t, XEXP (x, 0)),
make_tree (t, XEXP (x, 1)))));
case UDIV:
- t = (*lang_hooks.types.unsigned_type) (type);
+ t = lang_hooks.types.unsigned_type (type);
return fold (convert (type,
build (TRUNC_DIV_EXPR, t,
make_tree (t, XEXP (x, 0)),
case SIGN_EXTEND:
case ZERO_EXTEND:
- t = (*lang_hooks.types.type_for_mode) (GET_MODE (XEXP (x, 0)),
- GET_CODE (x) == 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))));
default:
{
tree type, mult_type, add_type, result;
- type = (*lang_hooks.types.type_for_mode) (mode, unsignedp);
+ type = lang_hooks.types.type_for_mode (mode, unsignedp);
/* In order to get a proper overflow indication from an unsigned
type, we have to pretend that it's a sizetype. */
}
add_type = (GET_MODE (add) == VOIDmode ? mult_type
- : (*lang_hooks.types.type_for_mode) (GET_MODE (add), unsignedp));
+ : lang_hooks.types.type_for_mode (GET_MODE (add), unsignedp));
result = fold (build (PLUS_EXPR, mult_type,
fold (build (MULT_EXPR, mult_type,
expand_mult_add (rtx x, rtx target, rtx mult, rtx add, enum machine_mode mode,
int unsignedp)
{
- tree type = (*lang_hooks.types.type_for_mode) (mode, unsignedp);
+ tree type = lang_hooks.types.type_for_mode (mode, unsignedp);
tree add_type = (GET_MODE (add) == VOIDmode
- ? type: (*lang_hooks.types.type_for_mode) (GET_MODE (add),
- unsignedp));
+ ? type: lang_hooks.types.type_for_mode (GET_MODE (add),
+ unsignedp));
tree result = fold (build (PLUS_EXPR, type,
fold (build (MULT_EXPR, type,
make_tree (type, x),
comparison
= compare_from_rtx (op0, op1, code, unsignedp, mode, NULL_RTX);
- if (GET_CODE (comparison) == CONST_INT)
- return (comparison == const0_rtx ? const0_rtx
- : normalizep == 1 ? const1_rtx
- : normalizep == -1 ? constm1_rtx
- : const_true_rtx);
+ if (CONSTANT_P (comparison))
+ {
+ if (GET_CODE (comparison) == CONST_INT)
+ {
+ if (comparison == const0_rtx)
+ return const0_rtx;
+ }
+#ifdef FLOAT_STORE_FLAG_VALUE
+ else if (GET_CODE (comparison) == CONST_DOUBLE)
+ {
+ if (comparison == CONST0_RTX (GET_MODE (comparison)))
+ return const0_rtx;
+ }
+#endif
+ else
+ abort ();
+ 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.
that is compensated by the subsequent overflow when subtracting
one / negating. */
- if (abs_optab->handlers[(int) mode].insn_code != CODE_FOR_nothing)
+ if (abs_optab->handlers[mode].insn_code != CODE_FOR_nothing)
tem = expand_unop (mode, abs_optab, op0, subtarget, 1);
- else if (ffs_optab->handlers[(int) mode].insn_code != CODE_FOR_nothing)
+ else if (ffs_optab->handlers[mode].insn_code != CODE_FOR_nothing)
tem = expand_unop (mode, ffs_optab, op0, subtarget, 1);
else if (GET_MODE_SIZE (mode) < UNITS_PER_WORD)
{
/* If this failed, we have to do this with set/compare/jump/set code. */
- if (GET_CODE (target) != REG
+ if (!REG_P (target)
|| reg_mentioned_p (target, op0) || reg_mentioned_p (target, op1))
target = gen_reg_rtx (GET_MODE (target));