/* 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, 2006, 2007
+ Free Software Foundation, Inc.
This file is part of GCC.
GCC is free software; you can redistribute it and/or modify it under
the terms of the GNU General Public License as published by the Free
-Software Foundation; either version 2, or (at your option) any later
+Software Foundation; either version 3, or (at your option) any later
version.
GCC is distributed in the hope that it will be useful, but WITHOUT ANY
for more details.
You should have received a copy of the GNU General Public License
-along with GCC; see the file COPYING. If not, write to the Free
-Software Foundation, 59 Temple Place - Suite 330, Boston, MA
-02111-1307, USA. */
+along with GCC; see the file COPYING3. If not see
+<http://www.gnu.org/licenses/>. */
#include "config.h"
#include "real.h"
#include "recog.h"
#include "langhooks.h"
+#include "df.h"
+#include "target.h"
static void store_fixed_bit_field (rtx, unsigned HOST_WIDE_INT,
unsigned HOST_WIDE_INT,
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
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 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];
struct rtx_def reg; rtunion reg_fld[2];
struct rtx_def plus; rtunion plus_fld1;
struct rtx_def neg;
- struct rtx_def udiv; rtunion udiv_fld1;
struct rtx_def mult; rtunion mult_fld1;
- struct rtx_def div; rtunion div_fld1;
- struct rtx_def mod; rtunion mod_fld1;
+ struct rtx_def sdiv; rtunion sdiv_fld1;
+ struct rtx_def udiv; rtunion udiv_fld1;
struct rtx_def zext;
+ struct rtx_def sdiv_32; rtunion sdiv_32_fld1;
+ struct rtx_def smod_32; rtunion smod_32_fld1;
struct rtx_def wide_mult; rtunion wide_mult_fld1;
struct rtx_def wide_lshr; rtunion wide_lshr_fld1;
struct rtx_def wide_trunc;
memset (&all, 0, sizeof all);
PUT_CODE (&all.reg, REG);
- REGNO (&all.reg) = 10000;
+ /* Avoid using hard regs in ways which may be unsupported. */
+ SET_REGNO (&all.reg, LAST_VIRTUAL_REGISTER + 1);
PUT_CODE (&all.plus, PLUS);
XEXP (&all.plus, 0) = &all.reg;
PUT_CODE (&all.neg, NEG);
XEXP (&all.neg, 0) = &all.reg;
- PUT_CODE (&all.udiv, UDIV);
- XEXP (&all.udiv, 0) = &all.reg;
- XEXP (&all.udiv, 1) = &all.reg;
-
PUT_CODE (&all.mult, MULT);
XEXP (&all.mult, 0) = &all.reg;
XEXP (&all.mult, 1) = &all.reg;
- PUT_CODE (&all.div, DIV);
- XEXP (&all.div, 0) = &all.reg;
- XEXP (&all.div, 1) = 32 < MAX_BITS_PER_WORD ? cint[32] : GEN_INT (32);
+ PUT_CODE (&all.sdiv, DIV);
+ XEXP (&all.sdiv, 0) = &all.reg;
+ XEXP (&all.sdiv, 1) = &all.reg;
+
+ PUT_CODE (&all.udiv, UDIV);
+ XEXP (&all.udiv, 0) = &all.reg;
+ XEXP (&all.udiv, 1) = &all.reg;
+
+ PUT_CODE (&all.sdiv_32, DIV);
+ XEXP (&all.sdiv_32, 0) = &all.reg;
+ XEXP (&all.sdiv_32, 1) = 32 < MAX_BITS_PER_WORD ? cint[32] : GEN_INT (32);
- PUT_CODE (&all.mod, MOD);
- XEXP (&all.mod, 0) = &all.reg;
- XEXP (&all.mod, 1) = XEXP (&all.div, 1);
+ PUT_CODE (&all.smod_32, MOD);
+ XEXP (&all.smod_32, 0) = &all.reg;
+ XEXP (&all.smod_32, 1) = XEXP (&all.sdiv_32, 1);
PUT_CODE (&all.zext, ZERO_EXTEND);
XEXP (&all.zext, 0) = &all.reg;
PUT_MODE (&all.reg, mode);
PUT_MODE (&all.plus, mode);
PUT_MODE (&all.neg, mode);
- PUT_MODE (&all.udiv, mode);
PUT_MODE (&all.mult, mode);
- PUT_MODE (&all.div, mode);
- PUT_MODE (&all.mod, 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);
add_cost[mode] = rtx_cost (&all.plus, SET);
neg_cost[mode] = rtx_cost (&all.neg, SET);
- div_cost[mode] = rtx_cost (&all.udiv, 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.div, SET) <= 2 * add_cost[mode]);
- smod_pow2_cheap[mode] = (rtx_cost (&all.mod, SET) <= 4 * add_cost[mode]);
+ 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)
return data->operand[opno].mode;
}
-\f
-/* Generate code to store value from rtx VALUE
- into a bit-field within structure STR_RTX
- containing BITSIZE bits starting at bit BITNUM.
- FIELDMODE is the machine-mode of the FIELD_DECL node for this field.
- ALIGN is the alignment that STR_RTX is known to have.
- TOTAL_SIZE is the size of the structure in bytes, or -1 if varying. */
+/* Return true if X, of mode MODE, matches the predicate for operand
+ OPNO of instruction ICODE. Allow volatile memories, regardless of
+ the ambient volatile_ok setting. */
-/* ??? Note that there are two different ideas here for how
- to determine the size to count bits within, for a register.
- One is BITS_PER_WORD, and the other is the size of operand 3
- of the insv pattern.
+static bool
+check_predicate_volatile_ok (enum insn_code icode, int opno,
+ rtx x, enum machine_mode mode)
+{
+ bool save_volatile_ok, result;
- If operand 3 of the insv pattern is VOIDmode, then we will use BITS_PER_WORD
- else, we use the mode of operand 3. */
+ save_volatile_ok = volatile_ok;
+ result = insn_data[(int) icode].operand[opno].predicate (x, mode);
+ volatile_ok = save_volatile_ok;
+ return result;
+}
+\f
+/* A subroutine of store_bit_field, with the same arguments. Return true
+ if the operation could be implemented.
-rtx
-store_bit_field (rtx str_rtx, unsigned HOST_WIDE_INT bitsize,
- unsigned HOST_WIDE_INT bitnum, enum machine_mode fieldmode,
- rtx value)
+ If FALLBACK_P is true, fall back to store_fixed_bit_field if we have
+ no other way of implementing the operation. If FALLBACK_P is false,
+ return false instead. */
+
+static bool
+store_bit_field_1 (rtx str_rtx, unsigned HOST_WIDE_INT bitsize,
+ unsigned HOST_WIDE_INT bitnum, enum machine_mode fieldmode,
+ rtx value, bool fallback_p)
{
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. */
+ int inner_mode_size = GET_MODE_SIZE (GET_MODE (SUBREG_REG (op0)));
+ int outer_mode_size = GET_MODE_SIZE (GET_MODE (op0));
+
+ byte_offset = 0;
+
+ /* Paradoxical subregs need special handling on big endian machines. */
+ if (SUBREG_BYTE (op0) == 0 && inner_mode_size < outer_mode_size)
+ {
+ int difference = inner_mode_size - outer_mode_size;
+
+ if (WORDS_BIG_ENDIAN)
+ byte_offset += (difference / UNITS_PER_WORD) * UNITS_PER_WORD;
+ if (BYTES_BIG_ENDIAN)
+ byte_offset += difference % UNITS_PER_WORD;
+ }
+ else
+ byte_offset = SUBREG_BYTE (op0);
+
+ bitnum += byte_offset * 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 true;
+
/* Use vec_set patterns for inserting parts of vectors whenever
available. */
if (VECTOR_MODE_P (GET_MODE (op0))
&& !MEM_P (op0)
- && (vec_set_optab->handlers[GET_MODE (op0)].insn_code
+ && (optab_handler (vec_set_optab, 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)))
{
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 icode = (int) optab_handler (vec_set_optab, outermode)->insn_code;
int pos = bitnum / GET_MODE_BITSIZE (innermode);
rtx rtxpos = GEN_INT (pos);
rtx src = value;
{
emit_insn (seq);
emit_insn (pat);
- return dest;
+ return true;
}
}
- 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);
|| (offset * BITS_PER_UNIT % bitsize == 0
&& MEM_ALIGN (op0) % GET_MODE_BITSIZE (fieldmode) == 0))))
{
- 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
- op0 = adjust_address (op0, fieldmode, offset);
- }
+ if (MEM_P (op0))
+ op0 = adjust_address (op0, fieldmode, offset);
+ else if (GET_MODE (op0) != fieldmode)
+ op0 = simplify_gen_subreg (fieldmode, op0, GET_MODE (op0),
+ byte_offset);
emit_move_insn (op0, value);
- return value;
+ return true;
}
/* Make sure we are playing with integral modes. Pun with subregs
if (!MEM_P (op0)
&& (BYTES_BIG_ENDIAN ? bitpos + bitsize == unit : bitpos == 0)
&& bitsize == GET_MODE_BITSIZE (fieldmode)
- && (movstrict_optab->handlers[fieldmode].insn_code
+ && (optab_handler (movstrict_optab, fieldmode)->insn_code
!= CODE_FOR_nothing))
{
- int icode = movstrict_optab->handlers[fieldmode].insn_code;
+ int icode = optab_handler (movstrict_optab, fieldmode)->insn_code;
/* Get appropriate low part of the value being stored. */
if (GET_CODE (value) == CONST_INT || REG_P (value))
+ (offset * UNITS_PER_WORD)),
value));
- return value;
+ return true;
}
/* Handle fields bigger than a word. */
unsigned int backwards = WORDS_BIG_ENDIAN && fieldmode != BLKmode;
unsigned int nwords = (bitsize + (BITS_PER_WORD - 1)) / BITS_PER_WORD;
unsigned int i;
+ rtx last;
/* 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);
+ last = get_last_insn ();
for (i = 0; i < nwords; i++)
{
/* If I is 0, use the low-order word in both field and target;
* BITS_PER_WORD,
0)
: (int) i * BITS_PER_WORD);
+ rtx value_word = operand_subword_force (value, wordnum, fieldmode);
- store_bit_field (op0, MIN (BITS_PER_WORD,
- bitsize - i * BITS_PER_WORD),
- bitnum + bit_offset, word_mode,
- operand_subword_force (value, wordnum, fieldmode));
+ if (!store_bit_field_1 (op0, MIN (BITS_PER_WORD,
+ bitsize - i * BITS_PER_WORD),
+ bitnum + bit_offset, word_mode,
+ value_word, fallback_p))
+ {
+ delete_insns_since (last);
+ return false;
+ }
}
- return value;
+ return true;
}
/* From here on we can assume that the field to be stored in is
{
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 (HAVE_insv
&& GET_MODE (value) != BLKmode
- && !(bitsize == 1 && GET_CODE (value) == CONST_INT)
- /* Ensure insv's size is wide enough for this field. */
- && (GET_MODE_BITSIZE (op_mode) >= bitsize)
+ && bitsize > 0
+ && GET_MODE_BITSIZE (op_mode) >= bitsize
&& ! ((REG_P (op0) || GET_CODE (op0) == SUBREG)
- && (bitsize + bitpos > GET_MODE_BITSIZE (op_mode))))
+ && (bitsize + bitpos > GET_MODE_BITSIZE (op_mode)))
+ && insn_data[CODE_FOR_insv].operand[1].predicate (GEN_INT (bitsize),
+ VOIDmode)
+ && check_predicate_volatile_ok (CODE_FOR_insv, 0, op0, VOIDmode))
{
int xbitpos = bitpos;
rtx value1;
rtx xop0 = op0;
rtx last = get_last_insn ();
rtx pat;
- enum machine_mode maxmode = mode_for_extraction (EP_insv, 3);
- int save_volatile_ok = volatile_ok;
-
- volatile_ok = 1;
-
- /* If this machine's insv can only insert into a register, copy OP0
- into a register and save it back later. */
- /* This used to check flag_force_mem, but that was a serious
- de-optimization now that flag_force_mem is enabled by -O2. */
- if (MEM_P (op0)
- && ! ((*insn_data[(int) CODE_FOR_insv].operand[0].predicate)
- (op0, VOIDmode)))
- {
- rtx tempreg;
- enum machine_mode bestmode;
-
- /* Get the mode to use for inserting into this field. If OP0 is
- BLKmode, get the smallest mode consistent with the alignment. If
- OP0 is a non-BLKmode object that is no wider than MAXMODE, use its
- mode. Otherwise, use the smallest mode containing the field. */
-
- if (GET_MODE (op0) == BLKmode
- || GET_MODE_SIZE (GET_MODE (op0)) > GET_MODE_SIZE (maxmode))
- bestmode
- = get_best_mode (bitsize, bitnum, MEM_ALIGN (op0), maxmode,
- MEM_VOLATILE_P (op0));
- else
- bestmode = GET_MODE (op0);
-
- if (bestmode == VOIDmode
- || (SLOW_UNALIGNED_ACCESS (bestmode, MEM_ALIGN (op0))
- && GET_MODE_BITSIZE (bestmode) > MEM_ALIGN (op0)))
- goto insv_loses;
-
- /* Adjust address to point to the containing unit of that mode.
- Compute offset as multiple of this unit, counting in bytes. */
- unit = GET_MODE_BITSIZE (bestmode);
- offset = (bitnum / unit) * GET_MODE_SIZE (bestmode);
- bitpos = bitnum % unit;
- op0 = adjust_address (op0, bestmode, offset);
-
- /* Fetch that unit, store the bitfield in it, then store
- the unit. */
- tempreg = copy_to_reg (op0);
- store_bit_field (tempreg, bitsize, bitpos, fieldmode, orig_value);
- emit_move_insn (op0, tempreg);
- return value;
- }
- volatile_ok = save_volatile_ok;
/* Add OFFSET into OP0's address. */
if (MEM_P (xop0))
xop0 = adjust_address (xop0, byte_mode, offset);
- /* If xop0 is a register, we need it in MAXMODE
+ /* If xop0 is a register, we need it in OP_MODE
to make it acceptable to the format of insv. */
if (GET_CODE (xop0) == SUBREG)
/* We can't just change the mode, because this might clobber op0,
and we will need the original value of op0 if insv fails. */
- xop0 = gen_rtx_SUBREG (maxmode, SUBREG_REG (xop0), SUBREG_BYTE (xop0));
- if (REG_P (xop0) && GET_MODE (xop0) != maxmode)
- xop0 = gen_rtx_SUBREG (maxmode, xop0, 0);
+ xop0 = gen_rtx_SUBREG (op_mode, SUBREG_REG (xop0), SUBREG_BYTE (xop0));
+ if (REG_P (xop0) && GET_MODE (xop0) != op_mode)
+ xop0 = gen_rtx_SUBREG (op_mode, xop0, 0);
/* On big-endian machines, we count bits from the most significant.
If the bit field insn does not, we must invert. */
/* We have been counting XBITPOS within UNIT.
Count instead within the size of the register. */
if (BITS_BIG_ENDIAN && !MEM_P (xop0))
- xbitpos += GET_MODE_BITSIZE (maxmode) - unit;
+ xbitpos += GET_MODE_BITSIZE (op_mode) - unit;
- unit = GET_MODE_BITSIZE (maxmode);
+ unit = GET_MODE_BITSIZE (op_mode);
- /* Convert VALUE to maxmode (which insv insn wants) in VALUE1. */
+ /* Convert VALUE to op_mode (which insv insn wants) in VALUE1. */
value1 = value;
- if (GET_MODE (value) != maxmode)
+ if (GET_MODE (value) != op_mode)
{
if (GET_MODE_BITSIZE (GET_MODE (value)) >= bitsize)
{
if it has all the bits we will actually use. However,
if we must narrow it, be sure we do it correctly. */
- if (GET_MODE_SIZE (GET_MODE (value)) < GET_MODE_SIZE (maxmode))
+ if (GET_MODE_SIZE (GET_MODE (value)) < GET_MODE_SIZE (op_mode))
{
rtx tmp;
- tmp = simplify_subreg (maxmode, value1, GET_MODE (value), 0);
+ tmp = simplify_subreg (op_mode, value1, GET_MODE (value), 0);
if (! tmp)
- tmp = simplify_gen_subreg (maxmode,
+ tmp = simplify_gen_subreg (op_mode,
force_reg (GET_MODE (value),
value1),
GET_MODE (value), 0);
value1 = tmp;
}
else
- value1 = gen_lowpart (maxmode, value1);
+ value1 = gen_lowpart (op_mode, value1);
}
else if (GET_CODE (value) == CONST_INT)
- value1 = gen_int_mode (INTVAL (value), maxmode);
+ value1 = gen_int_mode (INTVAL (value), op_mode);
else
/* Parse phase is supposed to make VALUE's data type
match that of the component reference, which is a type
/* If this machine's insv insists on a register,
get VALUE1 into a register. */
if (! ((*insn_data[(int) CODE_FOR_insv].operand[3].predicate)
- (value1, maxmode)))
- value1 = force_reg (maxmode, value1);
+ (value1, op_mode)))
+ value1 = force_reg (op_mode, value1);
pat = gen_insv (xop0, GEN_INT (bitsize), GEN_INT (xbitpos), value1);
if (pat)
- emit_insn (pat);
+ {
+ emit_insn (pat);
+ return true;
+ }
+ delete_insns_since (last);
+ }
+
+ /* If OP0 is a memory, try copying it to a register and seeing if a
+ cheap register alternative is available. */
+ if (HAVE_insv && MEM_P (op0))
+ {
+ enum machine_mode bestmode;
+
+ /* Get the mode to use for inserting into this field. If OP0 is
+ BLKmode, get the smallest mode consistent with the alignment. If
+ OP0 is a non-BLKmode object that is no wider than OP_MODE, use its
+ mode. Otherwise, use the smallest mode containing the field. */
+
+ if (GET_MODE (op0) == BLKmode
+ || (op_mode != MAX_MACHINE_MODE
+ && GET_MODE_SIZE (GET_MODE (op0)) > GET_MODE_SIZE (op_mode)))
+ bestmode = get_best_mode (bitsize, bitnum, MEM_ALIGN (op0),
+ (op_mode == MAX_MACHINE_MODE
+ ? VOIDmode : op_mode),
+ MEM_VOLATILE_P (op0));
else
+ bestmode = GET_MODE (op0);
+
+ if (bestmode != VOIDmode
+ && GET_MODE_SIZE (bestmode) >= GET_MODE_SIZE (fieldmode)
+ && !(SLOW_UNALIGNED_ACCESS (bestmode, MEM_ALIGN (op0))
+ && GET_MODE_BITSIZE (bestmode) > MEM_ALIGN (op0)))
{
+ rtx last, tempreg, xop0;
+ unsigned HOST_WIDE_INT xoffset, xbitpos;
+
+ last = get_last_insn ();
+
+ /* Adjust address to point to the containing unit of
+ that mode. Compute the offset as a multiple of this unit,
+ counting in bytes. */
+ unit = GET_MODE_BITSIZE (bestmode);
+ xoffset = (bitnum / unit) * GET_MODE_SIZE (bestmode);
+ xbitpos = bitnum % unit;
+ xop0 = adjust_address (op0, bestmode, xoffset);
+
+ /* Fetch that unit, store the bitfield in it, then store
+ the unit. */
+ tempreg = copy_to_reg (xop0);
+ if (store_bit_field_1 (tempreg, bitsize, xbitpos,
+ fieldmode, orig_value, false))
+ {
+ emit_move_insn (xop0, tempreg);
+ return true;
+ }
delete_insns_since (last);
- store_fixed_bit_field (op0, offset, bitsize, bitpos, value);
}
}
- else
- insv_loses:
- /* Insv is not available; store using shifts and boolean ops. */
- store_fixed_bit_field (op0, offset, bitsize, bitpos, value);
- return value;
+
+ if (!fallback_p)
+ return false;
+
+ store_fixed_bit_field (op0, offset, bitsize, bitpos, value);
+ return true;
+}
+
+/* Generate code to store value from rtx VALUE
+ into a bit-field within structure STR_RTX
+ containing BITSIZE bits starting at bit BITNUM.
+ FIELDMODE is the machine-mode of the FIELD_DECL node for this field. */
+
+void
+store_bit_field (rtx str_rtx, unsigned HOST_WIDE_INT bitsize,
+ unsigned HOST_WIDE_INT bitnum, enum machine_mode fieldmode,
+ rtx value)
+{
+ if (!store_bit_field_1 (str_rtx, bitsize, bitnum, fieldmode, value, true))
+ gcc_unreachable ();
}
\f
/* Use shifts and boolean operations to store VALUE
{
enum machine_mode mode;
unsigned int total_bits = BITS_PER_WORD;
- rtx subtarget, temp;
+ rtx temp;
int all_zero = 0;
int all_one = 0;
/* Now clear the chosen bits in OP0,
except that if VALUE is -1 we need not bother. */
+ /* We keep the intermediates in registers to allow CSE to combine
+ consecutive bitfield assignments. */
- subtarget = (REG_P (op0) || ! flag_force_mem) ? op0 : 0;
+ temp = force_reg (mode, op0);
if (! all_one)
{
- temp = expand_binop (mode, and_optab, op0,
+ temp = expand_binop (mode, and_optab, temp,
mask_rtx (mode, bitpos, bitsize, 1),
- subtarget, 1, OPTAB_LIB_WIDEN);
- subtarget = temp;
+ NULL_RTX, 1, OPTAB_LIB_WIDEN);
+ temp = force_reg (mode, temp);
}
- else
- temp = op0;
/* Now logical-or VALUE into OP0, unless it is zero. */
if (! all_zero)
- temp = expand_binop (mode, ior_optab, temp, value,
- subtarget, 1, OPTAB_LIB_WIDEN);
+ {
+ temp = expand_binop (mode, ior_optab, temp, value,
+ NULL_RTX, 1, OPTAB_LIB_WIDEN);
+ temp = force_reg (mode, temp);
+ }
+
if (op0 != temp)
emit_move_insn (op0, temp);
}
}
}
\f
-/* Generate code to extract a byte-field from STR_RTX
- containing BITSIZE bits, starting at BITNUM,
- and put it in TARGET if possible (if TARGET is nonzero).
- Regardless of TARGET, we return the rtx for where the value is placed.
+/* A subroutine of extract_bit_field_1 that converts return value X
+ to either MODE or TMODE. MODE, TMODE and UNSIGNEDP are arguments
+ to extract_bit_field. */
- STR_RTX is the structure containing the byte (a REG or MEM).
- UNSIGNEDP is nonzero if this is an unsigned bit field.
- MODE is the natural mode of the field value once extracted.
- TMODE is the mode the caller would like the value to have;
- but the value may be returned with type MODE instead.
+static rtx
+convert_extracted_bit_field (rtx x, enum machine_mode mode,
+ enum machine_mode tmode, bool unsignedp)
+{
+ if (GET_MODE (x) == tmode || GET_MODE (x) == mode)
+ return x;
- TOTAL_SIZE is the size in bytes of the containing structure,
- or -1 if varying.
+ /* If the x 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. */
+ if (!SCALAR_INT_MODE_P (tmode))
+ {
+ enum machine_mode smode;
- If a TARGET is specified and we can store in it at no extra cost,
- we do so, and return TARGET.
- Otherwise, we return a REG of mode TMODE or MODE, with TMODE preferred
- if they are equally easy. */
+ smode = mode_for_size (GET_MODE_BITSIZE (tmode), MODE_INT, 0);
+ x = convert_to_mode (smode, x, unsignedp);
+ x = force_reg (smode, x);
+ return gen_lowpart (tmode, x);
+ }
-rtx
-extract_bit_field (rtx str_rtx, unsigned HOST_WIDE_INT bitsize,
- unsigned HOST_WIDE_INT bitnum, int unsignedp, rtx target,
- enum machine_mode mode, enum machine_mode tmode)
+ return convert_to_mode (tmode, x, unsignedp);
+}
+
+/* A subroutine of extract_bit_field, with the same arguments.
+ If FALLBACK_P is true, fall back to extract_fixed_bit_field
+ if we can find no other means of implementing the operation.
+ if FALLBACK_P is false, return NULL instead. */
+
+static rtx
+extract_bit_field_1 (rtx str_rtx, unsigned HOST_WIDE_INT bitsize,
+ unsigned HOST_WIDE_INT bitnum, int unsignedp, rtx target,
+ enum machine_mode mode, enum machine_mode tmode,
+ bool fallback_p)
{
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;
enum machine_mode int_mode;
- enum machine_mode extv_mode = mode_for_extraction (EP_extv, 0);
- enum machine_mode extzv_mode = mode_for_extraction (EP_extzv, 0);
+ enum machine_mode ext_mode;
enum machine_mode mode1;
+ enum insn_code icode;
int byte_offset;
if (tmode == VOIDmode)
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
return op0;
}
+ /* See if we can get a better vector mode before extracting. */
+ if (VECTOR_MODE_P (GET_MODE (op0))
+ && !MEM_P (op0)
+ && GET_MODE_INNER (GET_MODE (op0)) != tmode)
+ {
+ enum machine_mode new_mode;
+ int nunits = GET_MODE_NUNITS (GET_MODE (op0));
+
+ if (GET_MODE_CLASS (tmode) == MODE_FLOAT)
+ new_mode = MIN_MODE_VECTOR_FLOAT;
+ else if (GET_MODE_CLASS (tmode) == MODE_FRACT)
+ new_mode = MIN_MODE_VECTOR_FRACT;
+ else if (GET_MODE_CLASS (tmode) == MODE_UFRACT)
+ new_mode = MIN_MODE_VECTOR_UFRACT;
+ else if (GET_MODE_CLASS (tmode) == MODE_ACCUM)
+ new_mode = MIN_MODE_VECTOR_ACCUM;
+ else if (GET_MODE_CLASS (tmode) == MODE_UACCUM)
+ new_mode = MIN_MODE_VECTOR_UACCUM;
+ else
+ new_mode = MIN_MODE_VECTOR_INT;
+
+ 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
+ && targetm.vector_mode_supported_p (new_mode))
+ break;
+ if (new_mode != VOIDmode)
+ op0 = gen_lowpart (new_mode, op0);
+ }
+
/* Use vec_extract patterns for extracting parts of vectors whenever
available. */
if (VECTOR_MODE_P (GET_MODE (op0))
&& !MEM_P (op0)
- && (vec_extract_optab->handlers[GET_MODE (op0)].insn_code
+ && (optab_handler (vec_extract_optab, GET_MODE (op0))->insn_code
!= CODE_FOR_nothing)
&& ((bitnum + bitsize - 1) / 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_extract_optab->handlers[outermode].insn_code;
+ int icode = (int) optab_handler (vec_extract_optab, outermode)->insn_code;
unsigned HOST_WIDE_INT pos = bitnum / GET_MODE_BITSIZE (innermode);
rtx rtxpos = GEN_INT (pos);
rtx src = op0;
{
emit_insn (seq);
emit_insn (pat);
+ if (mode0 != mode)
+ return gen_lowpart (tmode, dest);
return dest;
}
}
{
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);
}
}
}
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.
{
if (mode1 != GET_MODE (op0))
{
- if (GET_CODE (op0) == SUBREG)
+ if (MEM_P (op0))
+ op0 = adjust_address (op0, mode1, offset);
+ else
{
- if (GET_MODE (SUBREG_REG (op0)) == mode1
- || GET_MODE_CLASS (mode1) == MODE_INT
- || GET_MODE_CLASS (mode1) == MODE_PARTIAL_INT)
- op0 = SUBREG_REG (op0);
- else
- /* Else we've got some float mode source being extracted into
- a different float mode destination -- this combination of
- subregs results in Severe Tire Damage. */
+ rtx sub = simplify_gen_subreg (mode1, op0, GET_MODE (op0),
+ byte_offset);
+ if (sub == NULL)
goto no_subreg_mode_swap;
+ op0 = sub;
}
- if (REG_P (op0))
- op0 = gen_rtx_SUBREG (mode1, op0, byte_offset);
- else
- op0 = adjust_address (op0, mode1, offset);
}
if (mode1 != mode)
return convert_to_mode (tmode, op0, unsignedp);
}
/* Now OFFSET is nonzero only for memory operands. */
-
- if (unsignedp)
+ ext_mode = mode_for_extraction (unsignedp ? EP_extzv : EP_extv, 0);
+ icode = unsignedp ? CODE_FOR_extzv : CODE_FOR_extv;
+ if (ext_mode != MAX_MACHINE_MODE
+ && bitsize > 0
+ && GET_MODE_BITSIZE (ext_mode) >= bitsize
+ /* If op0 is a register, we need it in EXT_MODE to make it
+ acceptable to the format of ext(z)v. */
+ && !(GET_CODE (op0) == SUBREG && GET_MODE (op0) != ext_mode)
+ && !((REG_P (op0) || GET_CODE (op0) == SUBREG)
+ && (bitsize + bitpos > GET_MODE_BITSIZE (ext_mode)))
+ && check_predicate_volatile_ok (icode, 1, op0, GET_MODE (op0)))
{
- if (HAVE_extzv
- && (GET_MODE_BITSIZE (extzv_mode) >= bitsize)
- && ! ((REG_P (op0) || GET_CODE (op0) == SUBREG)
- && (bitsize + bitpos > GET_MODE_BITSIZE (extzv_mode))))
- {
- unsigned HOST_WIDE_INT xbitpos = bitpos, xoffset = offset;
- rtx bitsize_rtx, bitpos_rtx;
- rtx last = get_last_insn ();
- rtx xop0 = op0;
- rtx xtarget = target;
- rtx xspec_target = spec_target;
- rtx xspec_target_subreg = spec_target_subreg;
- rtx pat;
- enum machine_mode maxmode = mode_for_extraction (EP_extzv, 0);
-
- if (MEM_P (xop0))
- {
- int save_volatile_ok = volatile_ok;
- volatile_ok = 1;
-
- /* Is the memory operand acceptable? */
- if (! ((*insn_data[(int) CODE_FOR_extzv].operand[1].predicate)
- (xop0, GET_MODE (xop0))))
- {
- /* No, load into a reg and extract from there. */
- enum machine_mode bestmode;
-
- /* Get the mode to use for inserting into this field. If
- OP0 is BLKmode, get the smallest mode consistent with the
- alignment. If OP0 is a non-BLKmode object that is no
- wider than MAXMODE, use its mode. Otherwise, use the
- smallest mode containing the field. */
-
- if (GET_MODE (xop0) == BLKmode
- || (GET_MODE_SIZE (GET_MODE (op0))
- > GET_MODE_SIZE (maxmode)))
- bestmode = get_best_mode (bitsize, bitnum,
- MEM_ALIGN (xop0), maxmode,
- MEM_VOLATILE_P (xop0));
- else
- bestmode = GET_MODE (xop0);
-
- if (bestmode == VOIDmode
- || (SLOW_UNALIGNED_ACCESS (bestmode, MEM_ALIGN (xop0))
- && GET_MODE_BITSIZE (bestmode) > MEM_ALIGN (xop0)))
- goto extzv_loses;
-
- /* Compute offset as multiple of this unit,
- counting in bytes. */
- unit = GET_MODE_BITSIZE (bestmode);
- xoffset = (bitnum / unit) * GET_MODE_SIZE (bestmode);
- xbitpos = bitnum % unit;
- xop0 = adjust_address (xop0, bestmode, xoffset);
-
- /* Fetch it to a register in that size. */
- xop0 = force_reg (bestmode, xop0);
-
- /* XBITPOS counts within UNIT, which is what is expected. */
- }
- else
- /* Get ref to first byte containing part of the field. */
- xop0 = adjust_address (xop0, byte_mode, xoffset);
-
- volatile_ok = save_volatile_ok;
- }
+ unsigned HOST_WIDE_INT xbitpos = bitpos, xoffset = offset;
+ rtx bitsize_rtx, bitpos_rtx;
+ rtx last = get_last_insn ();
+ rtx xop0 = op0;
+ rtx xtarget = target;
+ rtx xspec_target = target;
+ rtx xspec_target_subreg = 0;
+ rtx pat;
- /* If op0 is a register, we need it in MAXMODE (which is usually
- SImode). to make it acceptable to the format of extzv. */
- if (GET_CODE (xop0) == SUBREG && GET_MODE (xop0) != maxmode)
- goto extzv_loses;
- if (REG_P (xop0) && GET_MODE (xop0) != maxmode)
- xop0 = gen_rtx_SUBREG (maxmode, xop0, 0);
+ /* If op0 is a register, we need it in EXT_MODE to make it
+ acceptable to the format of ext(z)v. */
+ if (REG_P (xop0) && GET_MODE (xop0) != ext_mode)
+ xop0 = gen_rtx_SUBREG (ext_mode, xop0, 0);
+ if (MEM_P (xop0))
+ /* Get ref to first byte containing part of the field. */
+ xop0 = adjust_address (xop0, byte_mode, xoffset);
- /* On big-endian machines, we count bits from the most significant.
- If the bit field insn does not, we must invert. */
- if (BITS_BIG_ENDIAN != BYTES_BIG_ENDIAN)
- xbitpos = unit - bitsize - xbitpos;
+ /* On big-endian machines, we count bits from the most significant.
+ If the bit field insn does not, we must invert. */
+ if (BITS_BIG_ENDIAN != BYTES_BIG_ENDIAN)
+ xbitpos = unit - bitsize - xbitpos;
- /* Now convert from counting within UNIT to counting in MAXMODE. */
- if (BITS_BIG_ENDIAN && !MEM_P (xop0))
- xbitpos += GET_MODE_BITSIZE (maxmode) - unit;
+ /* Now convert from counting within UNIT to counting in EXT_MODE. */
+ if (BITS_BIG_ENDIAN && !MEM_P (xop0))
+ xbitpos += GET_MODE_BITSIZE (ext_mode) - unit;
- unit = GET_MODE_BITSIZE (maxmode);
+ unit = GET_MODE_BITSIZE (ext_mode);
- if (xtarget == 0
- || (flag_force_mem && MEM_P (xtarget)))
- xtarget = xspec_target = gen_reg_rtx (tmode);
+ if (xtarget == 0)
+ xtarget = xspec_target = gen_reg_rtx (tmode);
- if (GET_MODE (xtarget) != maxmode)
+ if (GET_MODE (xtarget) != ext_mode)
+ {
+ if (REG_P (xtarget))
{
- if (REG_P (xtarget))
- {
- int wider = (GET_MODE_SIZE (maxmode)
- > GET_MODE_SIZE (GET_MODE (xtarget)));
- xtarget = gen_lowpart (maxmode, xtarget);
- if (wider)
- xspec_target_subreg = xtarget;
- }
- else
- xtarget = gen_reg_rtx (maxmode);
+ xtarget = gen_lowpart (ext_mode, xtarget);
+ if (GET_MODE_SIZE (ext_mode)
+ > GET_MODE_SIZE (GET_MODE (xspec_target)))
+ xspec_target_subreg = xtarget;
}
+ else
+ xtarget = gen_reg_rtx (ext_mode);
+ }
- /* If this machine's extzv insists on a register target,
- make sure we have one. */
- if (! ((*insn_data[(int) CODE_FOR_extzv].operand[0].predicate)
- (xtarget, maxmode)))
- xtarget = gen_reg_rtx (maxmode);
+ /* If this machine's ext(z)v insists on a register target,
+ make sure we have one. */
+ if (!insn_data[(int) icode].operand[0].predicate (xtarget, ext_mode))
+ xtarget = gen_reg_rtx (ext_mode);
- bitsize_rtx = GEN_INT (bitsize);
- bitpos_rtx = GEN_INT (xbitpos);
+ bitsize_rtx = GEN_INT (bitsize);
+ bitpos_rtx = GEN_INT (xbitpos);
- pat = gen_extzv (xtarget, xop0, bitsize_rtx, bitpos_rtx);
- if (pat)
- {
- emit_insn (pat);
- target = xtarget;
- spec_target = xspec_target;
- spec_target_subreg = xspec_target_subreg;
- }
- else
- {
- delete_insns_since (last);
- target = extract_fixed_bit_field (int_mode, op0, offset, bitsize,
- bitpos, target, 1);
- }
+ pat = (unsignedp
+ ? gen_extzv (xtarget, xop0, bitsize_rtx, bitpos_rtx)
+ : gen_extv (xtarget, xop0, bitsize_rtx, bitpos_rtx));
+ if (pat)
+ {
+ emit_insn (pat);
+ if (xtarget == xspec_target)
+ return xtarget;
+ if (xtarget == xspec_target_subreg)
+ return xspec_target;
+ return convert_extracted_bit_field (xtarget, mode, tmode, unsignedp);
}
- else
- extzv_loses:
- target = extract_fixed_bit_field (int_mode, op0, offset, bitsize,
- bitpos, target, 1);
+ delete_insns_since (last);
}
- else
- {
- if (HAVE_extv
- && (GET_MODE_BITSIZE (extv_mode) >= bitsize)
- && ! ((REG_P (op0) || GET_CODE (op0) == SUBREG)
- && (bitsize + bitpos > GET_MODE_BITSIZE (extv_mode))))
- {
- int xbitpos = bitpos, xoffset = offset;
- rtx bitsize_rtx, bitpos_rtx;
- rtx last = get_last_insn ();
- rtx xop0 = op0, xtarget = target;
- rtx xspec_target = spec_target;
- rtx xspec_target_subreg = spec_target_subreg;
- rtx pat;
- enum machine_mode maxmode = mode_for_extraction (EP_extv, 0);
-
- if (MEM_P (xop0))
- {
- /* Is the memory operand acceptable? */
- if (! ((*insn_data[(int) CODE_FOR_extv].operand[1].predicate)
- (xop0, GET_MODE (xop0))))
- {
- /* No, load into a reg and extract from there. */
- enum machine_mode bestmode;
-
- /* Get the mode to use for inserting into this field. If
- OP0 is BLKmode, get the smallest mode consistent with the
- alignment. If OP0 is a non-BLKmode object that is no
- wider than MAXMODE, use its mode. Otherwise, use the
- smallest mode containing the field. */
-
- if (GET_MODE (xop0) == BLKmode
- || (GET_MODE_SIZE (GET_MODE (op0))
- > GET_MODE_SIZE (maxmode)))
- bestmode = get_best_mode (bitsize, bitnum,
- MEM_ALIGN (xop0), maxmode,
- MEM_VOLATILE_P (xop0));
- else
- bestmode = GET_MODE (xop0);
-
- if (bestmode == VOIDmode
- || (SLOW_UNALIGNED_ACCESS (bestmode, MEM_ALIGN (xop0))
- && GET_MODE_BITSIZE (bestmode) > MEM_ALIGN (xop0)))
- goto extv_loses;
-
- /* Compute offset as multiple of this unit,
- counting in bytes. */
- unit = GET_MODE_BITSIZE (bestmode);
- xoffset = (bitnum / unit) * GET_MODE_SIZE (bestmode);
- xbitpos = bitnum % unit;
- xop0 = adjust_address (xop0, bestmode, xoffset);
-
- /* Fetch it to a register in that size. */
- xop0 = force_reg (bestmode, xop0);
-
- /* XBITPOS counts within UNIT, which is what is expected. */
- }
- else
- /* Get ref to first byte containing part of the field. */
- xop0 = adjust_address (xop0, byte_mode, xoffset);
- }
-
- /* If op0 is a register, we need it in MAXMODE (which is usually
- SImode) to make it acceptable to the format of extv. */
- if (GET_CODE (xop0) == SUBREG && GET_MODE (xop0) != maxmode)
- goto extv_loses;
- if (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 the bit field insn does not, we must invert. */
- if (BITS_BIG_ENDIAN != BYTES_BIG_ENDIAN)
- xbitpos = unit - bitsize - xbitpos;
- /* XBITPOS counts within a size of UNIT.
- Adjust to count within a size of MAXMODE. */
- if (BITS_BIG_ENDIAN && !MEM_P (xop0))
- xbitpos += (GET_MODE_BITSIZE (maxmode) - unit);
+ /* If OP0 is a memory, try copying it to a register and seeing if a
+ cheap register alternative is available. */
+ if (ext_mode != MAX_MACHINE_MODE && MEM_P (op0))
+ {
+ enum machine_mode bestmode;
+
+ /* Get the mode to use for inserting into this field. If
+ OP0 is BLKmode, get the smallest mode consistent with the
+ alignment. If OP0 is a non-BLKmode object that is no
+ wider than EXT_MODE, use its mode. Otherwise, use the
+ smallest mode containing the field. */
+
+ if (GET_MODE (op0) == BLKmode
+ || (ext_mode != MAX_MACHINE_MODE
+ && GET_MODE_SIZE (GET_MODE (op0)) > GET_MODE_SIZE (ext_mode)))
+ bestmode = get_best_mode (bitsize, bitnum, MEM_ALIGN (op0),
+ (ext_mode == MAX_MACHINE_MODE
+ ? VOIDmode : ext_mode),
+ MEM_VOLATILE_P (op0));
+ else
+ bestmode = GET_MODE (op0);
- unit = GET_MODE_BITSIZE (maxmode);
+ if (bestmode != VOIDmode
+ && !(SLOW_UNALIGNED_ACCESS (bestmode, MEM_ALIGN (op0))
+ && GET_MODE_BITSIZE (bestmode) > MEM_ALIGN (op0)))
+ {
+ unsigned HOST_WIDE_INT xoffset, xbitpos;
- if (xtarget == 0
- || (flag_force_mem && MEM_P (xtarget)))
- xtarget = xspec_target = gen_reg_rtx (tmode);
+ /* Compute the offset as a multiple of this unit,
+ counting in bytes. */
+ unit = GET_MODE_BITSIZE (bestmode);
+ xoffset = (bitnum / unit) * GET_MODE_SIZE (bestmode);
+ xbitpos = bitnum % unit;
- if (GET_MODE (xtarget) != maxmode)
+ /* Make sure the register is big enough for the whole field. */
+ if (xoffset * BITS_PER_UNIT + unit
+ >= offset * BITS_PER_UNIT + bitsize)
{
- if (REG_P (xtarget))
- {
- int wider = (GET_MODE_SIZE (maxmode)
- > GET_MODE_SIZE (GET_MODE (xtarget)));
- xtarget = gen_lowpart (maxmode, xtarget);
- if (wider)
- xspec_target_subreg = xtarget;
- }
- else
- xtarget = gen_reg_rtx (maxmode);
- }
+ rtx last, result, xop0;
- /* If this machine's extv insists on a register target,
- make sure we have one. */
- if (! ((*insn_data[(int) CODE_FOR_extv].operand[0].predicate)
- (xtarget, maxmode)))
- xtarget = gen_reg_rtx (maxmode);
+ last = get_last_insn ();
- bitsize_rtx = GEN_INT (bitsize);
- bitpos_rtx = GEN_INT (xbitpos);
+ /* Fetch it to a register in that size. */
+ xop0 = adjust_address (op0, bestmode, xoffset);
+ xop0 = force_reg (bestmode, xop0);
+ result = extract_bit_field_1 (xop0, bitsize, xbitpos,
+ unsignedp, target,
+ mode, tmode, false);
+ if (result)
+ return result;
- pat = gen_extv (xtarget, xop0, bitsize_rtx, bitpos_rtx);
- if (pat)
- {
- emit_insn (pat);
- target = xtarget;
- spec_target = xspec_target;
- spec_target_subreg = xspec_target_subreg;
- }
- else
- {
delete_insns_since (last);
- target = extract_fixed_bit_field (int_mode, op0, offset, bitsize,
- bitpos, target, 0);
}
}
- else
- extv_loses:
- target = extract_fixed_bit_field (int_mode, op0, offset, bitsize,
- bitpos, target, 0);
}
- if (target == spec_target)
- return target;
- if (target == spec_target_subreg)
- return spec_target;
- if (GET_MODE (target) != tmode && GET_MODE (target) != mode)
- {
- /* If the target mode is floating-point, first convert to the
- integer mode of that size and then access it as a floating-point
- value via a SUBREG. */
- if (GET_MODE_CLASS (tmode) != MODE_INT
- && GET_MODE_CLASS (tmode) != MODE_PARTIAL_INT)
- {
- target = convert_to_mode (mode_for_size (GET_MODE_BITSIZE (tmode),
- MODE_INT, 0),
- target, unsignedp);
- return gen_lowpart (tmode, target);
- }
- else
- return convert_to_mode (tmode, target, unsignedp);
- }
- return target;
+
+ if (!fallback_p)
+ return NULL;
+
+ target = extract_fixed_bit_field (int_mode, op0, offset, bitsize,
+ bitpos, target, unsignedp);
+ return convert_extracted_bit_field (target, mode, tmode, unsignedp);
+}
+
+/* Generate code to extract a byte-field from STR_RTX
+ containing BITSIZE bits, starting at BITNUM,
+ and put it in TARGET if possible (if TARGET is nonzero).
+ Regardless of TARGET, we return the rtx for where the value is placed.
+
+ STR_RTX is the structure containing the byte (a REG or MEM).
+ UNSIGNEDP is nonzero if this is an unsigned bit field.
+ MODE is the natural mode of the field value once extracted.
+ TMODE is the mode the caller would like the value to have;
+ but the value may be returned with type MODE instead.
+
+ If a TARGET is specified and we can store in it at no extra cost,
+ we do so, and return TARGET.
+ Otherwise, we return a REG of mode TMODE or MODE, with TMODE preferred
+ if they are equally easy. */
+
+rtx
+extract_bit_field (rtx str_rtx, unsigned HOST_WIDE_INT bitsize,
+ unsigned HOST_WIDE_INT bitnum, int unsignedp, rtx target,
+ enum machine_mode mode, enum machine_mode tmode)
+{
+ return extract_bit_field_1 (str_rtx, bitsize, bitnum, unsignedp,
+ target, mode, tmode, true);
}
\f
/* Extract a bit field using shifts and boolean operations
unsigned int sign_shift_up, sign_shift_dn;
rtx base, a1, a2, v1, v2, comb, shift, result, start;
- /* Choose a mode that will fit BITSIZE. */
+ /* Choose a mode that will fit BITSIZE. */
mode = smallest_mode_for_size (bitsize, MODE_INT);
m_size = GET_MODE_SIZE (mode);
m_bitsize = GET_MODE_BITSIZE (mode);
/* At the end, we'll need an additional shift to deal with sign/zero
extension. By default this will be a left+right shift of the
- appropriate size. But we may be able to elimitate one of them. */
+ appropriate size. But we may be able to eliminate one of them. */
sign_shift_up = sign_shift_dn = m_bitsize - bitsize;
if (STRICT_ALIGNMENT)
{
base = plus_constant (XEXP (op0, 0), bitpos / BITS_PER_UNIT);
- base = force_operand (base, NULL);
bitpos %= BITS_PER_UNIT;
- /* Force alignment of the address; load two sequential values. */
- a1 = expand_simple_binop (Pmode, AND, base,
+ /* We load two values to be concatenate. There's an edge condition
+ that bears notice -- an aligned value at the end of a page can
+ only load one value lest we segfault. So the two values we load
+ are at "base & -size" and "(base + size - 1) & -size". If base
+ is unaligned, the addresses will be aligned and sequential; if
+ base is aligned, the addresses will both be equal to base. */
+
+ a1 = expand_simple_binop (Pmode, AND, force_operand (base, NULL),
GEN_INT (-(HOST_WIDE_INT)m_size),
NULL, true, OPTAB_LIB_WIDEN);
mark_reg_pointer (a1, m_bitsize);
set_mem_align (v1, m_bitsize);
v1 = force_reg (mode, validize_mem (v1));
- a2 = plus_constant (a1, GET_MODE_SIZE (mode));
+ a2 = plus_constant (base, GET_MODE_SIZE (mode) - 1);
+ a2 = expand_simple_binop (Pmode, AND, force_operand (a2, NULL),
+ GEN_INT (-(HOST_WIDE_INT)m_size),
+ NULL, true, OPTAB_LIB_WIDEN);
v2 = gen_rtx_MEM (mode, a2);
set_mem_align (v2, m_bitsize);
v2 = force_reg (mode, validize_mem (v2));
else
{
unit = MIN (MEM_ALIGN (op0), BITS_PER_WORD);
- if (bitsize / unit > 2)
+ if (0 && bitsize / unit > 2)
{
rtx tmp = extract_force_align_mem_bit_field (op0, bitsize, bitpos,
unsignedp);
and shifted in the other direction; but that does not work
on all machines. */
- op1 = expand_expr (amount, NULL_RTX, VOIDmode, 0);
+ op1 = expand_normal (amount);
if (SHIFT_COUNT_TRUNCATED)
{
&& GET_CODE (op1) == CONST_INT
&& INTVAL (op1) > 0
&& INTVAL (op1) < GET_MODE_BITSIZE (mode)
- && shift_cost[mode][INTVAL (op1)] > INTVAL (op1) * add_cost[mode])
+ && INTVAL (op1) < MAX_BITS_PER_WORD
+ && shift_cost[mode][INTVAL (op1)] > INTVAL (op1) * add_cost[mode]
+ && shift_cost[mode][INTVAL (op1)] != MAX_COST)
{
int i;
for (i = 0; i < INTVAL (op1); i++)
code below. */
rtx subtarget = target == shifted ? 0 : target;
+ tree new_amount, other_amount;
rtx temp1;
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));
+ if (GET_MODE (op1) != TYPE_MODE (type)
+ && GET_MODE (op1) != VOIDmode)
+ op1 = convert_to_mode (TYPE_MODE (type), op1, 1);
+ new_amount = make_tree (type, op1);
+ other_amount
+ = fold_build2 (MINUS_EXPR, type,
+ build_int_cst (type, GET_MODE_BITSIZE (mode)),
+ new_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_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
char log[MAX_BITS_PER_WORD];
};
+/* The entry for our multiplication cache/hash table. */
+struct alg_hash_entry {
+ /* The number we are multiplying by. */
+ unsigned HOST_WIDE_INT t;
+
+ /* The mode in which we are multiplying something by T. */
+ enum machine_mode mode;
+
+ /* 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. */
+#if HOST_BITS_PER_WIDE_INT == 64
+#define NUM_ALG_HASH_ENTRIES 1031
+#else
+#define NUM_ALG_HASH_ENTRIES 307
+#endif
+
+/* Each entry of ALG_HASH caches alg_code for some integer. This is
+ actually a hash table. If we have a collision, that the older
+ entry is kicked out. */
+static struct alg_hash_entry alg_hash[NUM_ALG_HASH_ENTRIES];
+
/* 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
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.
int op_cost, op_latency;
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;
/* Indicate that no algorithm is yet found. If no algorithm
is found, this value will be returned and indicate failure. */
best_alg = alloca (sizeof (struct algorithm));
best_cost = *cost_limit;
+ /* Compute the hash index. */
+ hash_index = (t ^ (unsigned int) mode) % 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].alg != alg_unknown)
+ {
+ cache_alg = alg_hash[hash_index].alg;
+
+ if (cache_alg == alg_impossible)
+ {
+ /* 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;
+
+ cache_hit = true;
+
+ switch (cache_alg)
+ {
+ case alg_shift:
+ goto do_alg_shift;
+
+ case alg_add_t_m2:
+ case alg_sub_t_m2:
+ goto do_alg_addsub_t_m2;
+
+ case alg_add_factor:
+ case alg_sub_factor:
+ goto do_alg_addsub_factor;
+
+ 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 ();
+ }
+ }
+ }
+
/* If we have a group of zero bits at the low-order part of T, try
multiplying by the remaining bits and then doing a shift. */
if ((t & 1) == 0)
{
+ do_alg_shift:
m = floor_log2 (t & -t); /* m = number of low zero bits */
if (m < maxm)
{
best_alg->op[best_alg->ops] = alg_shift;
}
}
+ if (cache_hit)
+ goto done;
}
/* If we have an odd number, add or subtract one. */
{
unsigned HOST_WIDE_INT w;
+ do_alg_addsub_t_m2:
for (w = 1; (w & t) != 0; w <<= 1)
;
/* If T was -1, then W will be zero after the loop. This is another
best_alg->op[best_alg->ops] = alg_add_t_m2;
}
}
+ if (cache_hit)
+ goto done;
}
/* Look for factors of t of the form
good sequence quickly, and therefore be able to prune (by decreasing
COST_LIMIT) the search. */
+ do_alg_addsub_factor:
for (m = floor_log2 (t - 1); m >= 2; m--)
{
unsigned HOST_WIDE_INT d;
d = ((unsigned HOST_WIDE_INT) 1 << m) + 1;
- if (t % d == 0 && t > d && m < maxm)
+ if (t % d == 0 && t > d && m < maxm
+ && (!cache_hit || cache_alg == alg_add_factor))
{
/* If the target has a cheap shift-and-add instruction use
that in preference to a shift insn followed by an add insn.
Assume that the shift-and-add is "atomic" with a latency
- equal to it's cost, otherwise assume that on superscalar
+ 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];
}
d = ((unsigned HOST_WIDE_INT) 1 << m) - 1;
- if (t % d == 0 && t > d && m < maxm)
+ if (t % d == 0 && t > d && m < maxm
+ && (!cache_hit || cache_alg == alg_sub_factor))
{
/* If the target has a cheap shift-and-subtract insn use
that in preference to a shift insn followed by a sub insn.
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;
break;
}
}
+ if (cache_hit)
+ goto done;
/* Try shift-and-add (load effective address) instructions,
i.e. do a*3, a*5, a*9. */
if ((t & 1) != 0)
{
+ do_alg_add_t2_m:
q = t - 1;
q = q & -q;
m = exact_log2 (q);
best_alg->op[best_alg->ops] = alg_add_t2_m;
}
}
+ if (cache_hit)
+ goto done;
+ do_alg_sub_t2_m:
q = t + 1;
q = q & -q;
m = exact_log2 (q);
best_alg->op[best_alg->ops] = alg_sub_t2_m;
}
}
+ if (cache_hit)
+ goto done;
}
+ 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'
to record, make this search fail. */
struct mult_cost limit;
int op_cost;
+ /* Fail quickly for impossible bounds. */
+ if (mult_cost < 0)
+ return false;
+
+ /* 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];
+ if (mult_cost > op_cost)
+ mult_cost = op_cost;
+
*variant = basic_variant;
limit.cost = mult_cost;
limit.latency = mult_cost;
int opno;
enum machine_mode nmode;
- /* Avoid referencing memory over and over.
- For speed, but also for correctness when mem is volatile. */
- if (MEM_P (op0))
- op0 = force_reg (mode, op0);
+ /* Avoid referencing memory over and over and invalid sharing
+ on SUBREGs. */
+ op0 = force_reg (mode, op0);
/* ACCUM starts out either as OP0 or as a zero, depending on
the first operation. */
insn = get_last_insn ();
set_unique_reg_note (insn, REG_EQUAL,
- gen_rtx_MULT (nmode, tem, GEN_INT (val_so_far)));
+ gen_rtx_MULT (nmode, tem,
+ GEN_INT (val_so_far)));
}
if (variant == negate_variant)
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)
/* Expand x*2.0 as x+x. */
if (GET_CODE (op1) == CONST_DOUBLE
- && GET_MODE_CLASS (mode) == MODE_FLOAT)
+ && SCALAR_FLOAT_MODE_P (mode))
{
REAL_VALUE_TYPE d;
REAL_VALUE_FROM_CONST_DOUBLE (d, op1);
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;
}
}
if (mode == word_mode)
return gen_highpart (mode, op);
+ gcc_assert (!SCALAR_FLOAT_MODE_P (mode));
+
wider_mode = GET_MODE_WIDER_MODE (mode);
op = expand_shift (RSHIFT_EXPR, wider_mode, op,
build_int_cst (NULL_TREE, GET_MODE_BITSIZE (mode)), 0, 1);
rtx tem;
int size;
+ gcc_assert (!SCALAR_FLOAT_MODE_P (mode));
+
wider_mode = GET_MODE_WIDER_MODE (mode);
size = GET_MODE_BITSIZE (mode);
/* Try widening multiplication. */
moptab = unsignedp ? umul_widen_optab : smul_widen_optab;
- if (moptab->handlers[wider_mode].insn_code != CODE_FOR_nothing
+ if (optab_handler (moptab, wider_mode)->insn_code != CODE_FOR_nothing
&& mul_widen_cost[wider_mode] < max_cost)
{
tem = expand_binop (wider_mode, moptab, op0, narrow_op1, 0,
}
/* Try widening the mode and perform a non-widening multiplication. */
- moptab = smul_optab;
- if (smul_optab->handlers[wider_mode].insn_code != CODE_FOR_nothing
+ 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)
{
- 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. */
moptab = unsignedp ? smul_widen_optab : umul_widen_optab;
- if (moptab->handlers[wider_mode].insn_code != CODE_FOR_nothing
+ 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))
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;
+ gcc_assert (!SCALAR_FLOAT_MODE_P (mode));
/* 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
use a LSHIFTRT, 1 ADD, 1 SUB and an AND. */
temp = gen_rtx_LSHIFTRT (mode, result, shift);
- if (lshr_optab->handlers[mode].insn_code == CODE_FOR_nothing
+ if (optab_handler (lshr_optab, mode)->insn_code == CODE_FOR_nothing
|| rtx_cost (temp, SET) > COSTS_N_INSNS (2))
{
temp = expand_binop (mode, xor_optab, op0, signmask,
(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)
for (compute_mode = mode; compute_mode != VOIDmode;
compute_mode = GET_MODE_WIDER_MODE (compute_mode))
- if (optab1->handlers[compute_mode].insn_code != CODE_FOR_nothing
- || optab2->handlers[compute_mode].insn_code != CODE_FOR_nothing)
+ if (optab_handler (optab1, compute_mode)->insn_code != CODE_FOR_nothing
+ || optab_handler (optab2, 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[compute_mode].libfunc
- || optab2->handlers[compute_mode].libfunc)
+ if (optab_libfunc (optab1, compute_mode)
+ || optab_libfunc (optab2, compute_mode))
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;
/* 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[compute_mode]
- - (rem_flag && ! (last_div_const != 0 && op1_is_constant
- && INTVAL (op1) == last_div_const)
- ? mul_cost[compute_mode] + add_cost[compute_mode]
- : 0);
+ max_cost = unsignedp ? udiv_cost[compute_mode] : sdiv_cost[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];
last_div_const = ! rem_flag && op1_is_constant ? INTVAL (op1) : 0;
{
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;
+ unsigned HOST_WIDE_INT abs_d;
+
+ /* Since d might be INT_MIN, we have to cast to
+ unsigned HOST_WIDE_INT before negating to avoid
+ undefined signed overflow. */
+ abs_d = (d >= 0
+ ? (unsigned HOST_WIDE_INT) d
+ : - (unsigned HOST_WIDE_INT) d);
/* n rem d = n rem -d */
if (rem_flag && d < 0)
: sdiv_pow2_cheap[compute_mode])
/* We assume that cheap metric is true if the
optab has an expander for this mode. */
- && (((rem_flag ? smod_optab : sdiv_optab)
- ->handlers[compute_mode].insn_code
+ && ((optab_handler ((rem_flag ? smod_optab
+ : sdiv_optab),
+ compute_mode)->insn_code
!= CODE_FOR_nothing)
- || (sdivmod_optab->handlers[compute_mode]
- .insn_code != CODE_FOR_nothing)))
+ || (optab_handler(sdivmod_optab,
+ compute_mode)
+ ->insn_code != CODE_FOR_nothing)))
;
else if (EXACT_POWER_OF_2_OR_ZERO_P (abs_d))
{
}
if (sdiv_pow2_cheap[compute_mode]
- && ((sdiv_optab->handlers[compute_mode].insn_code
+ && ((optab_handler (sdiv_optab, compute_mode)->insn_code
!= CODE_FOR_nothing)
- || (sdivmod_optab->handlers[compute_mode].insn_code
+ || (optab_handler (sdivmod_optab, compute_mode)->insn_code
!= CODE_FOR_nothing)))
quotient = expand_divmod (0, TRUNC_DIV_EXPR,
compute_mode, op0,
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)
{
= sign_expand_binop (compute_mode, umod_optab, smod_optab,
op0, op1, target,
unsignedp,
- ((optab2->handlers[compute_mode].insn_code
+ ((optab_handler (optab2, 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[compute_mode].insn_code
+ ((optab_handler (optab2, compute_mode)->insn_code
!= CODE_FOR_nothing)
? OPTAB_DIRECT : OPTAB_WIDEN));
case CONST_VECTOR:
{
- int i, units;
- rtx elt;
+ int units = CONST_VECTOR_NUNITS (x);
+ tree itype = TREE_TYPE (type);
tree t = NULL_TREE;
+ int i;
- units = CONST_VECTOR_NUNITS (x);
/* Build a tree with vector elements. */
for (i = units - 1; i >= 0; --i)
{
- elt = CONST_VECTOR_ELT (x, i);
- t = tree_cons (NULL_TREE, make_tree (type, elt), t);
+ rtx elt = CONST_VECTOR_ELT (x, i);
+ t = tree_cons (NULL_TREE, make_tree (itype, elt), t);
}
return build_vector (type, t);
}
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)))));
+ t = unsigned_type_for (type);
+ 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)))));
+ t = signed_type_for (type);
+ 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)
- t = lang_hooks.types.signed_type (type);
+ t = signed_type_for (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)))));
+ t = unsigned_type_for (type);
+ 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)));
+
+ case CONST:
+ return make_tree (type, XEXP (x, 0));
+
+ case SYMBOL_REF:
+ t = SYMBOL_REF_DECL (x);
+ if (t)
+ return fold_convert (type, build_fold_addr_expr (t));
+ /* else fall through. */
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;
}
}
-
-/* Check whether the multiplication X * MULT + ADD overflows.
- X, MULT and ADD must be CONST_*.
- MODE is the machine mode for the computation.
- X and MULT must have mode MODE. ADD may have a different mode.
- So can X (defaults to same as MODE).
- UNSIGNEDP is nonzero to do unsigned multiplication. */
-
-bool
-const_mult_add_overflow_p (rtx x, rtx mult, rtx add,
- enum machine_mode mode, int unsignedp)
-{
- tree type, mult_type, add_type, result;
-
- 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. */
- mult_type = type;
- if (unsignedp)
- {
- /* FIXME:It would be nice if we could step directly from this
- type to its sizetype equivalent. */
- mult_type = build_distinct_type_copy (type);
- TYPE_IS_SIZETYPE (mult_type) = 1;
- }
-
- 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)));
-
- return TREE_CONSTANT_OVERFLOW (result);
-}
-
-/* Return an rtx representing the value of X * MULT + ADD.
- TARGET is a suggestion for where to store the result (an rtx).
- MODE is the machine mode for the computation.
- X and MULT must have mode MODE. ADD may have a different mode.
- So can X (defaults to same as MODE).
- UNSIGNEDP is nonzero to do unsigned multiplication.
- This may emit insns. */
-
-rtx
-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 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)));
-
- return expand_expr (result, target, VOIDmode, 0);
-}
\f
/* Compute the logical-and of OP0 and OP1, storing it in TARGET
and returning TARGET.
return target;
}
\f
+/* Helper function for emit_store_flag. */
+static rtx
+emit_store_flag_1 (rtx target, rtx subtarget, enum machine_mode mode,
+ int normalizep)
+{
+ rtx op0;
+ enum machine_mode target_mode = GET_MODE (target);
+
+ /* If we are converting to a wider mode, first convert to
+ TARGET_MODE, then normalize. This produces better combining
+ opportunities on machines that have a SIGN_EXTRACT when we are
+ testing a single bit. This mostly benefits the 68k.
+
+ If STORE_FLAG_VALUE does not have the sign bit set when
+ interpreted in MODE, we can do this conversion as unsigned, which
+ is usually more efficient. */
+ if (GET_MODE_SIZE (target_mode) > GET_MODE_SIZE (mode))
+ {
+ convert_move (target, subtarget,
+ (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT)
+ && 0 == (STORE_FLAG_VALUE
+ & ((HOST_WIDE_INT) 1
+ << (GET_MODE_BITSIZE (mode) -1))));
+ op0 = target;
+ mode = target_mode;
+ }
+ else
+ op0 = subtarget;
+
+ /* If we want to keep subexpressions around, don't reuse our last
+ target. */
+ if (optimize)
+ subtarget = 0;
+
+ /* Now normalize to the proper value in MODE. Sometimes we don't
+ have to do anything. */
+ if (normalizep == 0 || normalizep == STORE_FLAG_VALUE)
+ ;
+ /* STORE_FLAG_VALUE might be the most negative number, so write
+ the comparison this way to avoid a compiler-time warning. */
+ else if (- normalizep == STORE_FLAG_VALUE)
+ op0 = expand_unop (mode, neg_optab, op0, subtarget, 0);
+
+ /* We don't want to use STORE_FLAG_VALUE < 0 below since this makes
+ it hard to use a value of just the sign bit due to ANSI integer
+ constant typing rules. */
+ else if (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
+ && (STORE_FLAG_VALUE
+ & ((HOST_WIDE_INT) 1 << (GET_MODE_BITSIZE (mode) - 1))))
+ op0 = expand_shift (RSHIFT_EXPR, mode, op0,
+ size_int (GET_MODE_BITSIZE (mode) - 1), subtarget,
+ normalizep == 1);
+ else
+ {
+ gcc_assert (STORE_FLAG_VALUE & 1);
+
+ op0 = expand_and (mode, op0, const1_rtx, subtarget);
+ if (normalizep == -1)
+ op0 = expand_unop (mode, neg_optab, op0, op0, 0);
+ }
+
+ /* If we were converting to a smaller mode, do the conversion now. */
+ if (target_mode != mode)
+ {
+ convert_move (target, op0, 0);
+ return target;
+ }
+ else
+ return op0;
+}
+
/* Emit a store-flags instruction for comparison CODE on OP0 and OP1
and storing in TARGET. Normally return TARGET.
Return 0 if that cannot be done.
{
rtx op00, op01, op0both;
- /* Do a logical OR or AND of the two words and compare the result. */
+ /* Do a logical OR or AND of the two words and compare the
+ result. */
op00 = simplify_gen_subreg (word_mode, op0, mode, 0);
op01 = simplify_gen_subreg (word_mode, op0, mode, UNITS_PER_WORD);
op0both = expand_binop (word_mode,
op1 == const0_rtx ? ior_optab : and_optab,
- op00, op01, NULL_RTX, unsignedp, OPTAB_DIRECT);
+ op00, op01, NULL_RTX, unsignedp,
+ OPTAB_DIRECT);
if (op0both != 0)
return emit_store_flag (target, code, op0both, op1, word_mode,
/* If testing the sign bit, can just test on high word. */
op0h = simplify_gen_subreg (word_mode, op0, mode,
- subreg_highpart_offset (word_mode, mode));
+ subreg_highpart_offset (word_mode,
+ mode));
return emit_store_flag (target, code, op0h, op1, word_mode,
unsignedp, normalizep);
}
}
- /* From now on, we won't change CODE, so set ICODE now. */
- icode = setcc_gen_code[(int) code];
-
/* If this is A < 0 or A >= 0, we can do this by taking the ones
complement of A (for GE) and shifting the sign bit to the low bit. */
if (op1 == const0_rtx && (code == LT || code == GE)
&& (normalizep || STORE_FLAG_VALUE == 1
|| (GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
&& ((STORE_FLAG_VALUE & GET_MODE_MASK (mode))
- == (unsigned HOST_WIDE_INT) 1 << (GET_MODE_BITSIZE (mode) - 1)))))
+ == ((unsigned HOST_WIDE_INT) 1
+ << (GET_MODE_BITSIZE (mode) - 1))))))
{
subtarget = target;
return op0;
}
+ icode = setcc_gen_code[(int) code];
+
if (icode != CODE_FOR_nothing)
{
insn_operand_predicate_fn pred;
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 we are converting to a wider mode, first convert to
- TARGET_MODE, then normalize. This produces better combining
- opportunities on machines that have a SIGN_EXTRACT when we are
- testing a single bit. This mostly benefits the 68k.
+ if (icode != CODE_FOR_nothing)
+ {
+ enum machine_mode result_mode
+ = insn_data[(int) icode].operand[0].mode;
+ rtx cstore_op0 = op0;
+ rtx cstore_op1 = op1;
- If STORE_FLAG_VALUE does not have the sign bit set when
- interpreted in COMPARE_MODE, we can do this conversion as
- unsigned, which is usually more efficient. */
- if (GET_MODE_SIZE (target_mode) > GET_MODE_SIZE (compare_mode))
+ do_pending_stack_adjust ();
+ last = get_last_insn ();
+
+ if (compare_mode != mode)
{
- convert_move (target, subtarget,
- (GET_MODE_BITSIZE (compare_mode)
- <= HOST_BITS_PER_WIDE_INT)
- && 0 == (STORE_FLAG_VALUE
- & ((HOST_WIDE_INT) 1
- << (GET_MODE_BITSIZE (compare_mode) -1))));
- op0 = target;
- compare_mode = target_mode;
+ cstore_op0 = convert_modes (compare_mode, mode, cstore_op0,
+ unsignedp);
+ cstore_op1 = convert_modes (compare_mode, mode, cstore_op1,
+ unsignedp);
}
- else
- op0 = subtarget;
+
+ if (!insn_data[(int) icode].operand[2].predicate (cstore_op0,
+ compare_mode))
+ cstore_op0 = copy_to_mode_reg (compare_mode, cstore_op0);
- /* If we want to keep subexpressions around, don't reuse our
- last target. */
+ if (!insn_data[(int) icode].operand[3].predicate (cstore_op1,
+ compare_mode))
+ cstore_op1 = copy_to_mode_reg (compare_mode, cstore_op1);
- if (optimize)
- subtarget = 0;
+ comparison = gen_rtx_fmt_ee (code, result_mode, cstore_op0,
+ cstore_op1);
+ subtarget = target;
- /* Now normalize to the proper value in COMPARE_MODE. Sometimes
- we don't have to do anything. */
- if (normalizep == 0 || normalizep == STORE_FLAG_VALUE)
- ;
- /* STORE_FLAG_VALUE might be the most negative number, so write
- the comparison this way to avoid a compiler-time warning. */
- else if (- normalizep == STORE_FLAG_VALUE)
- op0 = expand_unop (compare_mode, neg_optab, op0, subtarget, 0);
-
- /* We don't want to use STORE_FLAG_VALUE < 0 below since this
- makes it hard to use a value of just the sign bit due to
- ANSI integer constant typing rules. */
- else if (GET_MODE_BITSIZE (compare_mode) <= HOST_BITS_PER_WIDE_INT
- && (STORE_FLAG_VALUE
- & ((HOST_WIDE_INT) 1
- << (GET_MODE_BITSIZE (compare_mode) - 1))))
- op0 = expand_shift (RSHIFT_EXPR, compare_mode, op0,
- size_int (GET_MODE_BITSIZE (compare_mode) - 1),
- subtarget, normalizep == 1);
- else
- {
- gcc_assert (STORE_FLAG_VALUE & 1);
-
- op0 = expand_and (compare_mode, op0, const1_rtx, subtarget);
- if (normalizep == -1)
- op0 = expand_unop (compare_mode, neg_optab, op0, op0, 0);
- }
+ 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);
- /* If we were converting to a smaller mode, do the
- conversion now. */
- if (target_mode != compare_mode)
+ if (pattern)
{
- convert_move (target, op0, 0);
- return target;
+ emit_insn (pattern);
+ return emit_store_flag_1 (target, subtarget, result_mode,
+ normalizep);
}
- else
- return op0;
}
}
that is compensated by the subsequent overflow when subtracting
one / negating. */
- if (abs_optab->handlers[mode].insn_code != CODE_FOR_nothing)
+ if (optab_handler (abs_optab, mode)->insn_code != CODE_FOR_nothing)
tem = expand_unop (mode, abs_optab, op0, subtarget, 1);
- else if (ffs_optab->handlers[mode].insn_code != CODE_FOR_nothing)
+ else if (optab_handler (ffs_optab, mode)->insn_code != CODE_FOR_nothing)
tem = expand_unop (mode, ffs_optab, op0, subtarget, 1);
else if (GET_MODE_SIZE (mode) < UNITS_PER_WORD)
{
}
\f
/* Perform possibly multi-word comparison and conditional jump to LABEL
- if ARG1 OP ARG2 true where ARG1 and ARG2 are of mode MODE
-
- 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). */
+ if ARG1 OP ARG2 true where ARG1 and ARG2 are of mode MODE. This is
+ now a thin wrapper around do_compare_rtx_and_jump. */
static void
do_cmp_and_jump (rtx arg1, rtx arg2, enum rtx_code op, enum machine_mode mode,
rtx label)
{
- /* If this mode is an integer too wide to compare properly,
- compare word by word. Rely on cse to optimize constant cases. */
-
- if (GET_MODE_CLASS (mode) == MODE_INT
- && ! can_compare_p (op, mode, ccp_jump))
- {
- rtx label2 = gen_label_rtx ();
-
- switch (op)
- {
- case LTU:
- do_jump_by_parts_greater_rtx (mode, 1, arg2, arg1, label2, label);
- break;
-
- case LEU:
- do_jump_by_parts_greater_rtx (mode, 1, arg1, arg2, label, label2);
- break;
-
- case LT:
- do_jump_by_parts_greater_rtx (mode, 0, arg2, arg1, label2, label);
- break;
-
- case GT:
- do_jump_by_parts_greater_rtx (mode, 0, arg1, arg2, label2, label);
- break;
-
- case GE:
- do_jump_by_parts_greater_rtx (mode, 0, arg2, arg1, label, label2);
- break;
-
- /* do_jump_by_parts_equality_rtx compares with zero. Luckily
- that's the only equality operations we do */
- case EQ:
- gcc_assert (arg2 == const0_rtx && mode == GET_MODE(arg1));
- do_jump_by_parts_equality_rtx (arg1, label2, label);
- break;
-
- case NE:
- gcc_assert (arg2 == const0_rtx && mode == GET_MODE(arg1));
- do_jump_by_parts_equality_rtx (arg1, label, label2);
- break;
-
- default:
- gcc_unreachable ();
- }
-
- emit_label (label2);
- }
- else
- emit_cmp_and_jump_insns (arg1, arg2, op, NULL_RTX, mode, 0, label);
+ int unsignedp = (op == LTU || op == LEU || op == GTU || op == GEU);
+ do_compare_rtx_and_jump (arg1, arg2, op, unsignedp, mode,
+ NULL_RTX, NULL_RTX, label);
}
OSDN Git Service
