blob: fe263ba04d3353ce2f49143b3ab8c6275f37f858 [file] [log] [blame]
/* Output routines for GCC for ARM/RISCiX.
Copyright (C) 1991, 93, 94, 95, 96, 1997 Free Software Foundation, Inc.
Contributed by Pieter `Tiggr' Schoenmakers (rcpieter@win.tue.nl)
and Martin Simmons (@harleqn.co.uk).
More major hacks by Richard Earnshaw (rwe11@cl.cam.ac.uk)
This file is part of GNU CC.
GNU CC is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2, or (at your option)
any later version.
GNU CC is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with GNU CC; see the file COPYING. If not, write to
the Free Software Foundation, 59 Temple Place - Suite 330,
Boston, MA 02111-1307, USA. */
#include <stdio.h>
#include <string.h>
#include "assert.h"
#include "config.h"
#include "rtl.h"
#include "regs.h"
#include "hard-reg-set.h"
#include "real.h"
#include "insn-config.h"
#include "conditions.h"
#include "insn-flags.h"
#include "output.h"
#include "insn-attr.h"
#include "flags.h"
#include "reload.h"
#include "tree.h"
#include "expr.h"
/* The maximum number of insns skipped which will be conditionalised if
possible. */
#define MAX_INSNS_SKIPPED 5
/* Some function declarations. */
extern FILE *asm_out_file;
static HOST_WIDE_INT int_log2 PROTO ((HOST_WIDE_INT));
static char *output_multi_immediate PROTO ((rtx *, char *, char *, int,
HOST_WIDE_INT));
static int arm_gen_constant PROTO ((enum rtx_code, enum machine_mode,
HOST_WIDE_INT, rtx, rtx, int, int));
static int arm_naked_function_p PROTO ((tree));
static void init_fpa_table PROTO ((void));
static enum machine_mode select_dominance_cc_mode PROTO ((enum rtx_code, rtx,
rtx, HOST_WIDE_INT));
static HOST_WIDE_INT add_constant PROTO ((rtx, enum machine_mode));
static void dump_table PROTO ((rtx));
static int fixit PROTO ((rtx, enum machine_mode, int));
static rtx find_barrier PROTO ((rtx, int));
static int broken_move PROTO ((rtx));
static char *fp_const_from_val PROTO ((REAL_VALUE_TYPE *));
static int eliminate_lr2ip PROTO ((rtx *));
static char *shift_op PROTO ((rtx, HOST_WIDE_INT *));
static int pattern_really_clobbers_lr PROTO ((rtx));
static int function_really_clobbers_lr PROTO ((rtx));
static void emit_multi_reg_push PROTO ((int));
static void emit_sfm PROTO ((int, int));
static enum arm_cond_code get_arm_condition_code PROTO ((rtx));
/* Define the information needed to generate branch insns. This is
stored from the compare operation. */
rtx arm_compare_op0, arm_compare_op1;
int arm_compare_fp;
/* What type of cpu are we compiling for? */
enum processor_type arm_cpu;
/* What type of floating point are we tuning for? */
enum floating_point_type arm_fpu;
/* What type of floating point instructions are available? */
enum floating_point_type arm_fpu_arch;
/* What program mode is the cpu running in? 26-bit mode or 32-bit mode */
enum prog_mode_type arm_prgmode;
/* Set by the -mfp=... option */
char *target_fp_name = NULL;
/* Nonzero if this is an "M" variant of the processor. */
int arm_fast_multiply = 0;
/* Nonzero if this chip supports the ARM Architecture 4 extensions */
int arm_arch4 = 0;
/* Set to the features we should tune the code for (multiply speed etc). */
int tune_flags = 0;
/* In case of a PRE_INC, POST_INC, PRE_DEC, POST_DEC memory reference, we
must report the mode of the memory reference from PRINT_OPERAND to
PRINT_OPERAND_ADDRESS. */
enum machine_mode output_memory_reference_mode;
/* Nonzero if the prologue must setup `fp'. */
int current_function_anonymous_args;
/* The register number to be used for the PIC offset register. */
int arm_pic_register = 9;
/* Location counter of .text segment. */
int arm_text_location = 0;
/* Set to one if we think that lr is only saved because of subroutine calls,
but all of these can be `put after' return insns */
int lr_save_eliminated;
/* Set to 1 when a return insn is output, this means that the epilogue
is not needed. */
static int return_used_this_function;
static int arm_constant_limit = 3;
/* For an explanation of these variables, see final_prescan_insn below. */
int arm_ccfsm_state;
enum arm_cond_code arm_current_cc;
rtx arm_target_insn;
int arm_target_label;
/* The condition codes of the ARM, and the inverse function. */
char *arm_condition_codes[] =
{
"eq", "ne", "cs", "cc", "mi", "pl", "vs", "vc",
"hi", "ls", "ge", "lt", "gt", "le", "al", "nv"
};
static enum arm_cond_code get_arm_condition_code ();
/* Initialization code */
struct arm_cpu_select arm_select[4] =
{
/* switch name, tune arch */
{ (char *)0, "--with-cpu=", 1, 1 },
{ (char *)0, "-mcpu=", 1, 1 },
{ (char *)0, "-march=", 0, 1 },
{ (char *)0, "-mtune=", 1, 0 },
};
#define FL_CO_PROC 0x01 /* Has external co-processor bus */
#define FL_FAST_MULT 0x02 /* Fast multiply */
#define FL_MODE26 0x04 /* 26-bit mode support */
#define FL_MODE32 0x08 /* 32-bit mode support */
#define FL_ARCH4 0x10 /* Architecture rel 4 */
#define FL_THUMB 0x20 /* Thumb aware */
struct processors
{
char *name;
enum processor_type type;
unsigned int flags;
};
/* Not all of these give usefully different compilation alternatives,
but there is no simple way of generalizing them. */
static struct processors all_procs[] =
{
{"arm2", PROCESSOR_ARM2, FL_CO_PROC | FL_MODE26},
{"arm250", PROCESSOR_ARM2, FL_CO_PROC | FL_MODE26},
{"arm3", PROCESSOR_ARM2, FL_CO_PROC | FL_MODE26},
{"arm6", PROCESSOR_ARM6, FL_CO_PROC | FL_MODE32 | FL_MODE26},
{"arm600", PROCESSOR_ARM6, FL_CO_PROC | FL_MODE32 | FL_MODE26},
{"arm610", PROCESSOR_ARM6, FL_MODE32 | FL_MODE26},
{"arm7", PROCESSOR_ARM7, FL_CO_PROC | FL_MODE32 | FL_MODE26},
/* arm7m doesn't exist on its own, only in conjuction with D, (and I), but
those don't alter the code, so it is sometimes known as the arm7m */
{"arm7m", PROCESSOR_ARM7, (FL_CO_PROC | FL_FAST_MULT | FL_MODE32
| FL_MODE26)},
{"arm7dm", PROCESSOR_ARM7, (FL_CO_PROC | FL_FAST_MULT | FL_MODE32
| FL_MODE26)},
{"arm7dmi", PROCESSOR_ARM7, (FL_CO_PROC | FL_FAST_MULT | FL_MODE32
| FL_MODE26)},
{"arm700", PROCESSOR_ARM7, FL_CO_PROC | FL_MODE32 | FL_MODE26},
{"arm710", PROCESSOR_ARM7, FL_MODE32 | FL_MODE26},
{"arm7100", PROCESSOR_ARM7, FL_MODE32 | FL_MODE26},
{"arm7500", PROCESSOR_ARM7, FL_MODE32 | FL_MODE26},
/* Doesn't really have an external co-proc, but does have embedded fpu */
{"arm7500fe", PROCESSOR_ARM7, FL_CO_PROC | FL_MODE32 | FL_MODE26},
{"arm7tdmi", PROCESSOR_ARM7, (FL_CO_PROC | FL_FAST_MULT | FL_MODE32
| FL_ARCH4 | FL_THUMB)},
{"arm8", PROCESSOR_ARM8, (FL_FAST_MULT | FL_MODE32 | FL_MODE26
| FL_ARCH4)},
{"arm810", PROCESSOR_ARM8, (FL_FAST_MULT | FL_MODE32 | FL_MODE26
| FL_ARCH4)},
{"strongarm", PROCESSOR_STARM, (FL_FAST_MULT | FL_MODE32 | FL_MODE26
| FL_ARCH4)},
{"strongarm110", PROCESSOR_STARM, (FL_FAST_MULT | FL_MODE32 | FL_MODE26
| FL_ARCH4)},
{"armv2", PROCESSOR_NONE, FL_CO_PROC | FL_MODE26},
{"armv2a", PROCESSOR_NONE, FL_CO_PROC | FL_MODE26},
{"armv3", PROCESSOR_NONE, FL_CO_PROC | FL_MODE32 | FL_MODE26},
{"armv3m", PROCESSOR_NONE, (FL_CO_PROC | FL_FAST_MULT | FL_MODE32
| FL_MODE26)},
{"armv4", PROCESSOR_NONE, (FL_CO_PROC | FL_FAST_MULT | FL_MODE32
| FL_MODE26 | FL_ARCH4)},
/* Strictly, FL_MODE26 is a permitted option for v4t, but there are no
implementations that support it, so we will leave it out for now. */
{"armv4t", PROCESSOR_NONE, (FL_CO_PROC | FL_FAST_MULT | FL_MODE32
| FL_ARCH4)},
{NULL, 0, 0}
};
/* Fix up any incompatible options that the user has specified.
This has now turned into a maze. */
void
arm_override_options ()
{
int arm_thumb_aware = 0;
int flags = 0;
int i;
struct arm_cpu_select *ptr;
static struct cpu_default {
int cpu;
char *name;
} cpu_defaults[] = {
{ TARGET_CPU_arm2, "arm2" },
{ TARGET_CPU_arm6, "arm6" },
{ TARGET_CPU_arm610, "arm610" },
{ TARGET_CPU_arm7dm, "arm7dm" },
{ TARGET_CPU_arm7500fe, "arm7500fe" },
{ TARGET_CPU_arm7tdmi, "arm7tdmi" },
{ TARGET_CPU_arm8, "arm8" },
{ TARGET_CPU_arm810, "arm810" },
{ TARGET_CPU_strongarm, "strongarm" },
{ 0, 0 }
};
struct cpu_default *def;
/* Set the default. */
for (def = &cpu_defaults[0]; def->name; ++def)
if (def->cpu == TARGET_CPU_DEFAULT)
break;
if (! def->name)
abort ();
arm_select[0].string = def->name;
for (i = 0; i < sizeof (arm_select) / sizeof (arm_select[0]); i++)
{
ptr = &arm_select[i];
if (ptr->string != (char *)0 && ptr->string[0] != '\0')
{
struct processors *sel;
for (sel = all_procs; sel->name != NULL; sel++)
if (! strcmp (ptr->string, sel->name))
{
/* -march= is the only flag that can take an architecture
type, so if we match when the tune bit is set, the
option was invalid. */
if (ptr->set_tune_p)
{
if (sel->type == PROCESSOR_NONE)
continue; /* Its an architecture, not a cpu */
arm_cpu = sel->type;
tune_flags = sel->flags;
}
if (ptr->set_arch_p)
flags = sel->flags;
break;
}
if (sel->name == NULL)
error ("bad value (%s) for %s switch", ptr->string, ptr->name);
}
}
if (write_symbols != NO_DEBUG && flag_omit_frame_pointer)
warning ("-g with -fomit-frame-pointer may not give sensible debugging");
if (TARGET_POKE_FUNCTION_NAME)
target_flags |= ARM_FLAG_APCS_FRAME;
if (TARGET_6)
warning ("Option '-m6' deprecated. Use: '-mapcs-32' or -mcpu=<proc>");
if (TARGET_3)
warning ("Option '-m3' deprecated. Use: '-mapcs-26' or -mcpu=<proc>");
if (TARGET_APCS_REENT && flag_pic)
fatal ("-fpic and -mapcs-reent are incompatible");
if (TARGET_APCS_REENT)
warning ("APCS reentrant code not supported.");
/* If stack checking is disabled, we can use r10 as the PIC register,
which keeps r9 available. */
if (flag_pic && ! TARGET_APCS_STACK)
arm_pic_register = 10;
/* Well, I'm about to have a go, but pic is NOT going to be compatible
with APCS reentrancy, since that requires too much support in the
assembler and linker, and the ARMASM assembler seems to lack some
required directives. */
if (flag_pic)
warning ("Position independent code not supported. Ignored");
if (TARGET_APCS_FLOAT)
warning ("Passing floating point arguments in fp regs not yet supported");
if (TARGET_APCS_STACK && ! TARGET_APCS)
{
warning ("-mapcs-stack-check incompatible with -mno-apcs-frame");
target_flags |= ARM_FLAG_APCS_FRAME;
}
/* Default is to tune for an FPA */
arm_fpu = FP_HARD;
/* Default value for floating point code... if no co-processor
bus, then schedule for emulated floating point. Otherwise,
assume the user has an FPA.
Note: this does not prevent use of floating point instructions,
-msoft-float does that. */
if (tune_flags & FL_CO_PROC == 0)
arm_fpu = FP_SOFT3;
arm_fast_multiply = (flags & FL_FAST_MULT) != 0;
arm_arch4 = (flags & FL_ARCH4) != 0;
arm_thumb_aware = (flags & FL_THUMB) != 0;
if (target_fp_name)
{
if (strcmp (target_fp_name, "2") == 0)
arm_fpu_arch = FP_SOFT2;
else if (strcmp (target_fp_name, "3") == 0)
arm_fpu_arch = FP_HARD;
else
fatal ("Invalid floating point emulation option: -mfpe=%s",
target_fp_name);
}
else
arm_fpu_arch = FP_DEFAULT;
if (TARGET_THUMB_INTERWORK && ! arm_thumb_aware)
{
warning ("This processor variant does not support Thumb interworking");
target_flags &= ~ARM_FLAG_THUMB;
}
if (TARGET_FPE && arm_fpu != FP_HARD)
arm_fpu = FP_SOFT2;
/* For arm2/3 there is no need to do any scheduling if there is only
a floating point emulator, or we are doing software floating-point. */
if ((TARGET_SOFT_FLOAT || arm_fpu != FP_HARD) && arm_cpu == PROCESSOR_ARM2)
flag_schedule_insns = flag_schedule_insns_after_reload = 0;
arm_prog_mode = TARGET_APCS_32 ? PROG_MODE_PROG32 : PROG_MODE_PROG26;
}
/* Return 1 if it is possible to return using a single instruction */
int
use_return_insn ()
{
int regno;
if (!reload_completed ||current_function_pretend_args_size
|| current_function_anonymous_args
|| (get_frame_size () && !(TARGET_APCS || frame_pointer_needed)))
return 0;
/* Can't be done if interworking with Thumb, and any registers have been
stacked */
if (TARGET_THUMB_INTERWORK)
for (regno = 0; regno < 16; regno++)
if (regs_ever_live[regno] && ! call_used_regs[regno])
return 0;
/* Can't be done if any of the FPU regs are pushed, since this also
requires an insn */
for (regno = 16; regno < 24; regno++)
if (regs_ever_live[regno] && ! call_used_regs[regno])
return 0;
/* If a function is naked, don't use the "return" insn. */
if (arm_naked_function_p (current_function_decl))
return 0;
return 1;
}
/* Return TRUE if int I is a valid immediate ARM constant. */
int
const_ok_for_arm (i)
HOST_WIDE_INT i;
{
unsigned HOST_WIDE_INT mask = ~0xFF;
/* Fast return for 0 and powers of 2 */
if ((i & (i - 1)) == 0)
return TRUE;
do
{
if ((i & mask & (unsigned HOST_WIDE_INT) 0xffffffff) == 0)
return TRUE;
mask =
(mask << 2) | ((mask & (unsigned HOST_WIDE_INT) 0xffffffff)
>> (32 - 2)) | ~((unsigned HOST_WIDE_INT) 0xffffffff);
} while (mask != ~0xFF);
return FALSE;
}
/* Return true if I is a valid constant for the operation CODE. */
int
const_ok_for_op (i, code, mode)
HOST_WIDE_INT i;
enum rtx_code code;
enum machine_mode mode;
{
if (const_ok_for_arm (i))
return 1;
switch (code)
{
case PLUS:
return const_ok_for_arm (ARM_SIGN_EXTEND (-i));
case MINUS: /* Should only occur with (MINUS I reg) => rsb */
case XOR:
case IOR:
return 0;
case AND:
return const_ok_for_arm (ARM_SIGN_EXTEND (~i));
default:
abort ();
}
}
/* Emit a sequence of insns to handle a large constant.
CODE is the code of the operation required, it can be any of SET, PLUS,
IOR, AND, XOR, MINUS;
MODE is the mode in which the operation is being performed;
VAL is the integer to operate on;
SOURCE is the other operand (a register, or a null-pointer for SET);
SUBTARGETS means it is safe to create scratch registers if that will
either produce a simpler sequence, or we will want to cse the values.
Return value is the number of insns emitted. */
int
arm_split_constant (code, mode, val, target, source, subtargets)
enum rtx_code code;
enum machine_mode mode;
HOST_WIDE_INT val;
rtx target;
rtx source;
int subtargets;
{
if (subtargets || code == SET
|| (GET_CODE (target) == REG && GET_CODE (source) == REG
&& REGNO (target) != REGNO (source)))
{
rtx temp;
if (arm_gen_constant (code, mode, val, target, source, 1, 0)
> arm_constant_limit + (code != SET))
{
if (code == SET)
{
/* Currently SET is the only monadic value for CODE, all
the rest are diadic. */
emit_insn (gen_rtx (SET, VOIDmode, target, GEN_INT (val)));
return 1;
}
else
{
rtx temp = subtargets ? gen_reg_rtx (mode) : target;
emit_insn (gen_rtx (SET, VOIDmode, temp, GEN_INT (val)));
/* For MINUS, the value is subtracted from, since we never
have subtraction of a constant. */
if (code == MINUS)
emit_insn (gen_rtx (SET, VOIDmode, target,
gen_rtx (code, mode, temp, source)));
else
emit_insn (gen_rtx (SET, VOIDmode, target,
gen_rtx (code, mode, source, temp)));
return 2;
}
}
}
return arm_gen_constant (code, mode, val, target, source, subtargets, 1);
}
/* As above, but extra parameter GENERATE which, if clear, suppresses
RTL generation. */
int
arm_gen_constant (code, mode, val, target, source, subtargets, generate)
enum rtx_code code;
enum machine_mode mode;
HOST_WIDE_INT val;
rtx target;
rtx source;
int subtargets;
int generate;
{
int can_add = 0;
int can_invert = 0;
int can_negate = 0;
int can_negate_initial = 0;
int can_shift = 0;
int i;
int num_bits_set = 0;
int set_sign_bit_copies = 0;
int clear_sign_bit_copies = 0;
int clear_zero_bit_copies = 0;
int set_zero_bit_copies = 0;
int insns = 0;
rtx new_src;
unsigned HOST_WIDE_INT temp1, temp2;
unsigned HOST_WIDE_INT remainder = val & 0xffffffff;
/* find out which operations are safe for a given CODE. Also do a quick
check for degenerate cases; these can occur when DImode operations
are split. */
switch (code)
{
case SET:
can_invert = 1;
can_shift = 1;
can_negate = 1;
break;
case PLUS:
can_negate = 1;
can_negate_initial = 1;
break;
case IOR:
if (remainder == 0xffffffff)
{
if (generate)
emit_insn (gen_rtx (SET, VOIDmode, target,
GEN_INT (ARM_SIGN_EXTEND (val))));
return 1;
}
if (remainder == 0)
{
if (reload_completed && rtx_equal_p (target, source))
return 0;
if (generate)
emit_insn (gen_rtx (SET, VOIDmode, target, source));
return 1;
}
break;
case AND:
if (remainder == 0)
{
if (generate)
emit_insn (gen_rtx (SET, VOIDmode, target, const0_rtx));
return 1;
}
if (remainder == 0xffffffff)
{
if (reload_completed && rtx_equal_p (target, source))
return 0;
if (generate)
emit_insn (gen_rtx (SET, VOIDmode, target, source));
return 1;
}
can_invert = 1;
break;
case XOR:
if (remainder == 0)
{
if (reload_completed && rtx_equal_p (target, source))
return 0;
if (generate)
emit_insn (gen_rtx (SET, VOIDmode, target, source));
return 1;
}
if (remainder == 0xffffffff)
{
if (generate)
emit_insn (gen_rtx (SET, VOIDmode, target,
gen_rtx (NOT, mode, source)));
return 1;
}
/* We don't know how to handle this yet below. */
abort ();
case MINUS:
/* We treat MINUS as (val - source), since (source - val) is always
passed as (source + (-val)). */
if (remainder == 0)
{
if (generate)
emit_insn (gen_rtx (SET, VOIDmode, target,
gen_rtx (NEG, mode, source)));
return 1;
}
if (const_ok_for_arm (val))
{
if (generate)
emit_insn (gen_rtx (SET, VOIDmode, target,
gen_rtx (MINUS, mode, GEN_INT (val), source)));
return 1;
}
can_negate = 1;
break;
default:
abort ();
}
/* If we can do it in one insn get out quickly */
if (const_ok_for_arm (val)
|| (can_negate_initial && const_ok_for_arm (-val))
|| (can_invert && const_ok_for_arm (~val)))
{
if (generate)
emit_insn (gen_rtx (SET, VOIDmode, target,
(source ? gen_rtx (code, mode, source,
GEN_INT (val))
: GEN_INT (val))));
return 1;
}
/* Calculate a few attributes that may be useful for specific
optimizations. */
for (i = 31; i >= 0; i--)
{
if ((remainder & (1 << i)) == 0)
clear_sign_bit_copies++;
else
break;
}
for (i = 31; i >= 0; i--)
{
if ((remainder & (1 << i)) != 0)
set_sign_bit_copies++;
else
break;
}
for (i = 0; i <= 31; i++)
{
if ((remainder & (1 << i)) == 0)
clear_zero_bit_copies++;
else
break;
}
for (i = 0; i <= 31; i++)
{
if ((remainder & (1 << i)) != 0)
set_zero_bit_copies++;
else
break;
}
switch (code)
{
case SET:
/* See if we can do this by sign_extending a constant that is known
to be negative. This is a good, way of doing it, since the shift
may well merge into a subsequent insn. */
if (set_sign_bit_copies > 1)
{
if (const_ok_for_arm
(temp1 = ARM_SIGN_EXTEND (remainder
<< (set_sign_bit_copies - 1))))
{
if (generate)
{
new_src = subtargets ? gen_reg_rtx (mode) : target;
emit_insn (gen_rtx (SET, VOIDmode, new_src,
GEN_INT (temp1)));
emit_insn (gen_ashrsi3 (target, new_src,
GEN_INT (set_sign_bit_copies - 1)));
}
return 2;
}
/* For an inverted constant, we will need to set the low bits,
these will be shifted out of harm's way. */
temp1 |= (1 << (set_sign_bit_copies - 1)) - 1;
if (const_ok_for_arm (~temp1))
{
if (generate)
{
new_src = subtargets ? gen_reg_rtx (mode) : target;
emit_insn (gen_rtx (SET, VOIDmode, new_src,
GEN_INT (temp1)));
emit_insn (gen_ashrsi3 (target, new_src,
GEN_INT (set_sign_bit_copies - 1)));
}
return 2;
}
}
/* See if we can generate this by setting the bottom (or the top)
16 bits, and then shifting these into the other half of the
word. We only look for the simplest cases, to do more would cost
too much. Be careful, however, not to generate this when the
alternative would take fewer insns. */
if (val & 0xffff0000)
{
temp1 = remainder & 0xffff0000;
temp2 = remainder & 0x0000ffff;
/* Overlaps outside this range are best done using other methods. */
for (i = 9; i < 24; i++)
{
if ((((temp2 | (temp2 << i)) & 0xffffffff) == remainder)
&& ! const_ok_for_arm (temp2))
{
insns = arm_gen_constant (code, mode, temp2,
new_src = (subtargets
? gen_reg_rtx (mode)
: target),
source, subtargets, generate);
source = new_src;
if (generate)
emit_insn (gen_rtx (SET, VOIDmode, target,
gen_rtx (IOR, mode,
gen_rtx (ASHIFT, mode, source,
GEN_INT (i)),
source)));
return insns + 1;
}
}
/* Don't duplicate cases already considered. */
for (i = 17; i < 24; i++)
{
if (((temp1 | (temp1 >> i)) == remainder)
&& ! const_ok_for_arm (temp1))
{
insns = arm_gen_constant (code, mode, temp1,
new_src = (subtargets
? gen_reg_rtx (mode)
: target),
source, subtargets, generate);
source = new_src;
if (generate)
emit_insn (gen_rtx (SET, VOIDmode, target,
gen_rtx (IOR, mode,
gen_rtx (LSHIFTRT, mode,
source, GEN_INT (i)),
source)));
return insns + 1;
}
}
}
break;
case IOR:
case XOR:
/* If we have IOR or XOR, and the constant can be loaded in a
single instruction, and we can find a temporary to put it in,
then this can be done in two instructions instead of 3-4. */
if (subtargets
|| (reload_completed && ! reg_mentioned_p (target, source)))
{
if (const_ok_for_arm (ARM_SIGN_EXTEND (~ val)))
{
if (generate)
{
rtx sub = subtargets ? gen_reg_rtx (mode) : target;
emit_insn (gen_rtx (SET, VOIDmode, sub, GEN_INT (val)));
emit_insn (gen_rtx (SET, VOIDmode, target,
gen_rtx (code, mode, source, sub)));
}
return 2;
}
}
if (code == XOR)
break;
if (set_sign_bit_copies > 8
&& (val & (-1 << (32 - set_sign_bit_copies))) == val)
{
if (generate)
{
rtx sub = subtargets ? gen_reg_rtx (mode) : target;
rtx shift = GEN_INT (set_sign_bit_copies);
emit_insn (gen_rtx (SET, VOIDmode, sub,
gen_rtx (NOT, mode,
gen_rtx (ASHIFT, mode, source,
shift))));
emit_insn (gen_rtx (SET, VOIDmode, target,
gen_rtx (NOT, mode,
gen_rtx (LSHIFTRT, mode, sub,
shift))));
}
return 2;
}
if (set_zero_bit_copies > 8
&& (remainder & ((1 << set_zero_bit_copies) - 1)) == remainder)
{
if (generate)
{
rtx sub = subtargets ? gen_reg_rtx (mode) : target;
rtx shift = GEN_INT (set_zero_bit_copies);
emit_insn (gen_rtx (SET, VOIDmode, sub,
gen_rtx (NOT, mode,
gen_rtx (LSHIFTRT, mode, source,
shift))));
emit_insn (gen_rtx (SET, VOIDmode, target,
gen_rtx (NOT, mode,
gen_rtx (ASHIFT, mode, sub,
shift))));
}
return 2;
}
if (const_ok_for_arm (temp1 = ARM_SIGN_EXTEND (~ val)))
{
if (generate)
{
rtx sub = subtargets ? gen_reg_rtx (mode) : target;
emit_insn (gen_rtx (SET, VOIDmode, sub,
gen_rtx (NOT, mode, source)));
source = sub;
if (subtargets)
sub = gen_reg_rtx (mode);
emit_insn (gen_rtx (SET, VOIDmode, sub,
gen_rtx (AND, mode, source,
GEN_INT (temp1))));
emit_insn (gen_rtx (SET, VOIDmode, target,
gen_rtx (NOT, mode, sub)));
}
return 3;
}
break;
case AND:
/* See if two shifts will do 2 or more insn's worth of work. */
if (clear_sign_bit_copies >= 16 && clear_sign_bit_copies < 24)
{
HOST_WIDE_INT shift_mask = ((0xffffffff
<< (32 - clear_sign_bit_copies))
& 0xffffffff);
rtx new_source;
rtx shift;
if ((remainder | shift_mask) != 0xffffffff)
{
if (generate)
{
new_source = subtargets ? gen_reg_rtx (mode) : target;
insns = arm_gen_constant (AND, mode, remainder | shift_mask,
new_source, source, subtargets, 1);
source = new_source;
}
else
insns = arm_gen_constant (AND, mode, remainder | shift_mask,
new_source, source, subtargets, 0);
}
if (generate)
{
shift = GEN_INT (clear_sign_bit_copies);
new_source = subtargets ? gen_reg_rtx (mode) : target;
emit_insn (gen_ashlsi3 (new_source, source, shift));
emit_insn (gen_lshrsi3 (target, new_source, shift));
}
return insns + 2;
}
if (clear_zero_bit_copies >= 16 && clear_zero_bit_copies < 24)
{
HOST_WIDE_INT shift_mask = (1 << clear_zero_bit_copies) - 1;
rtx new_source;
rtx shift;
if ((remainder | shift_mask) != 0xffffffff)
{
if (generate)
{
new_source = subtargets ? gen_reg_rtx (mode) : target;
insns = arm_gen_constant (AND, mode, remainder | shift_mask,
new_source, source, subtargets, 1);
source = new_source;
}
else
insns = arm_gen_constant (AND, mode, remainder | shift_mask,
new_source, source, subtargets, 0);
}
if (generate)
{
shift = GEN_INT (clear_zero_bit_copies);
new_source = subtargets ? gen_reg_rtx (mode) : target;
emit_insn (gen_lshrsi3 (new_source, source, shift));
emit_insn (gen_ashlsi3 (target, new_source, shift));
}
return insns + 2;
}
break;
default:
break;
}
for (i = 0; i < 32; i++)
if (remainder & (1 << i))
num_bits_set++;
if (code == AND || (can_invert && num_bits_set > 16))
remainder = (~remainder) & 0xffffffff;
else if (code == PLUS && num_bits_set > 16)
remainder = (-remainder) & 0xffffffff;
else
{
can_invert = 0;
can_negate = 0;
}
/* Now try and find a way of doing the job in either two or three
instructions.
We start by looking for the largest block of zeros that are aligned on
a 2-bit boundary, we then fill up the temps, wrapping around to the
top of the word when we drop off the bottom.
In the worst case this code should produce no more than four insns. */
{
int best_start = 0;
int best_consecutive_zeros = 0;
for (i = 0; i < 32; i += 2)
{
int consecutive_zeros = 0;
if (! (remainder & (3 << i)))
{
while ((i < 32) && ! (remainder & (3 << i)))
{
consecutive_zeros += 2;
i += 2;
}
if (consecutive_zeros > best_consecutive_zeros)
{
best_consecutive_zeros = consecutive_zeros;
best_start = i - consecutive_zeros;
}
i -= 2;
}
}
/* Now start emitting the insns, starting with the one with the highest
bit set: we do this so that the smallest number will be emitted last;
this is more likely to be combinable with addressing insns. */
i = best_start;
do
{
int end;
if (i <= 0)
i += 32;
if (remainder & (3 << (i - 2)))
{
end = i - 8;
if (end < 0)
end += 32;
temp1 = remainder & ((0x0ff << end)
| ((i < end) ? (0xff >> (32 - end)) : 0));
remainder &= ~temp1;
if (code == SET)
{
if (generate)
emit_insn (gen_rtx (SET, VOIDmode,
new_src = (subtargets
? gen_reg_rtx (mode)
: target),
GEN_INT (can_invert ? ~temp1 : temp1)));
can_invert = 0;
code = PLUS;
}
else if (code == MINUS)
{
if (generate)
emit_insn (gen_rtx (SET, VOIDmode,
new_src = (subtargets
? gen_reg_rtx (mode)
: target),
gen_rtx (code, mode, GEN_INT (temp1),
source)));
code = PLUS;
}
else
{
if (generate)
emit_insn (gen_rtx (SET, VOIDmode,
new_src = (remainder
? (subtargets
? gen_reg_rtx (mode)
: target)
: target),
gen_rtx (code, mode, source,
GEN_INT (can_invert ? ~temp1
: (can_negate
? -temp1
: temp1)))));
}
insns++;
source = new_src;
i -= 6;
}
i -= 2;
} while (remainder);
}
return insns;
}
/* Canonicalize a comparison so that we are more likely to recognize it.
This can be done for a few constant compares, where we can make the
immediate value easier to load. */
enum rtx_code
arm_canonicalize_comparison (code, op1)
enum rtx_code code;
rtx *op1;
{
HOST_WIDE_INT i = INTVAL (*op1);
switch (code)
{
case EQ:
case NE:
return code;
case GT:
case LE:
if (i != (1 << (HOST_BITS_PER_WIDE_INT - 1) - 1)
&& (const_ok_for_arm (i+1) || const_ok_for_arm (- (i+1))))
{
*op1 = GEN_INT (i+1);
return code == GT ? GE : LT;
}
break;
case GE:
case LT:
if (i != (1 << (HOST_BITS_PER_WIDE_INT - 1))
&& (const_ok_for_arm (i-1) || const_ok_for_arm (- (i-1))))
{
*op1 = GEN_INT (i-1);
return code == GE ? GT : LE;
}
break;
case GTU:
case LEU:
if (i != ~0
&& (const_ok_for_arm (i+1) || const_ok_for_arm (- (i+1))))
{
*op1 = GEN_INT (i + 1);
return code == GTU ? GEU : LTU;
}
break;
case GEU:
case LTU:
if (i != 0
&& (const_ok_for_arm (i - 1) || const_ok_for_arm (- (i - 1))))
{
*op1 = GEN_INT (i - 1);
return code == GEU ? GTU : LEU;
}
break;
default:
abort ();
}
return code;
}
/* Handle aggregates that are not laid out in a BLKmode element.
This is a sub-element of RETURN_IN_MEMORY. */
int
arm_return_in_memory (type)
tree type;
{
if (TREE_CODE (type) == RECORD_TYPE)
{
tree field;
/* For a struct, we can return in a register if every element was a
bit-field. */
for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
if (TREE_CODE (field) != FIELD_DECL
|| ! DECL_BIT_FIELD_TYPE (field))
return 1;
return 0;
}
else if (TREE_CODE (type) == UNION_TYPE)
{
tree field;
/* Unions can be returned in registers if every element is
integral, or can be returned in an integer register. */
for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
{
if (TREE_CODE (field) != FIELD_DECL
|| (AGGREGATE_TYPE_P (TREE_TYPE (field))
&& RETURN_IN_MEMORY (TREE_TYPE (field)))
|| FLOAT_TYPE_P (TREE_TYPE (field)))
return 1;
}
return 0;
}
/* XXX Not sure what should be done for other aggregates, so put them in
memory. */
return 1;
}
int
legitimate_pic_operand_p (x)
rtx x;
{
if (CONSTANT_P (x) && flag_pic
&& (GET_CODE (x) == SYMBOL_REF
|| (GET_CODE (x) == CONST
&& GET_CODE (XEXP (x, 0)) == PLUS
&& GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF)))
return 0;
return 1;
}
rtx
legitimize_pic_address (orig, mode, reg)
rtx orig;
enum machine_mode mode;
rtx reg;
{
if (GET_CODE (orig) == SYMBOL_REF)
{
rtx pic_ref, address;
rtx insn;
int subregs = 0;
if (reg == 0)
{
if (reload_in_progress || reload_completed)
abort ();
else
reg = gen_reg_rtx (Pmode);
subregs = 1;
}
#ifdef AOF_ASSEMBLER
/* The AOF assembler can generate relocations for these directly, and
understands that the PIC register has to be added into the offset.
*/
insn = emit_insn (gen_pic_load_addr_based (reg, orig));
#else
if (subregs)
address = gen_reg_rtx (Pmode);
else
address = reg;
emit_insn (gen_pic_load_addr (address, orig));
pic_ref = gen_rtx (MEM, Pmode,
gen_rtx (PLUS, Pmode, pic_offset_table_rtx, address));
RTX_UNCHANGING_P (pic_ref) = 1;
insn = emit_move_insn (reg, pic_ref);
#endif
current_function_uses_pic_offset_table = 1;
/* Put a REG_EQUAL note on this insn, so that it can be optimized
by loop. */
REG_NOTES (insn) = gen_rtx (EXPR_LIST, REG_EQUAL, orig,
REG_NOTES (insn));
return reg;
}
else if (GET_CODE (orig) == CONST)
{
rtx base, offset;
if (GET_CODE (XEXP (orig, 0)) == PLUS
&& XEXP (XEXP (orig, 0), 0) == pic_offset_table_rtx)
return orig;
if (reg == 0)
{
if (reload_in_progress || reload_completed)
abort ();
else
reg = gen_reg_rtx (Pmode);
}
if (GET_CODE (XEXP (orig, 0)) == PLUS)
{
base = legitimize_pic_address (XEXP (XEXP (orig, 0), 0), Pmode, reg);
offset = legitimize_pic_address (XEXP (XEXP (orig, 0), 1), Pmode,
base == reg ? 0 : reg);
}
else
abort ();
if (GET_CODE (offset) == CONST_INT)
{
/* The base register doesn't really matter, we only want to
test the index for the appropriate mode. */
GO_IF_LEGITIMATE_INDEX (mode, 0, offset, win);
if (! reload_in_progress && ! reload_completed)
offset = force_reg (Pmode, offset);
else
abort ();
win:
if (GET_CODE (offset) == CONST_INT)
return plus_constant_for_output (base, INTVAL (offset));
}
if (GET_MODE_SIZE (mode) > 4
&& (GET_MODE_CLASS (mode) == MODE_INT
|| TARGET_SOFT_FLOAT))
{
emit_insn (gen_addsi3 (reg, base, offset));
return reg;
}
return gen_rtx (PLUS, Pmode, base, offset);
}
else if (GET_CODE (orig) == LABEL_REF)
current_function_uses_pic_offset_table = 1;
return orig;
}
static rtx pic_rtx;
int
is_pic(x)
rtx x;
{
if (x == pic_rtx)
return 1;
return 0;
}
void
arm_finalize_pic ()
{
#ifndef AOF_ASSEMBLER
rtx l1, pic_tmp, pic_tmp2, seq;
rtx global_offset_table;
if (current_function_uses_pic_offset_table == 0)
return;
if (! flag_pic)
abort ();
start_sequence ();
l1 = gen_label_rtx ();
global_offset_table = gen_rtx (SYMBOL_REF, Pmode, "_GLOBAL_OFFSET_TABLE_");
pic_tmp = gen_rtx (CONST, VOIDmode,
gen_rtx (PLUS, Pmode,
gen_rtx (LABEL_REF, VOIDmode, l1),
GEN_INT (8)));
pic_tmp2 = gen_rtx (CONST, VOIDmode,
gen_rtx (PLUS, Pmode,
global_offset_table,
pc_rtx));
pic_rtx = gen_rtx (CONST, Pmode,
gen_rtx (MINUS, Pmode, pic_tmp2, pic_tmp));
emit_insn (gen_pic_load_addr (pic_offset_table_rtx, pic_rtx));
emit_jump_insn (gen_pic_add_dot_plus_eight(l1, pic_offset_table_rtx));
emit_label (l1);
seq = gen_sequence ();
end_sequence ();
emit_insn_after (seq, get_insns ());
/* Need to emit this whether or not we obey regdecls,
since setjmp/longjmp can cause life info to screw up. */
emit_insn (gen_rtx (USE, VOIDmode, pic_offset_table_rtx));
#endif /* AOF_ASSEMBLER */
}
#define REG_OR_SUBREG_REG(X) \
(GET_CODE (X) == REG \
|| (GET_CODE (X) == SUBREG && GET_CODE (SUBREG_REG (X)) == REG))
#define REG_OR_SUBREG_RTX(X) \
(GET_CODE (X) == REG ? (X) : SUBREG_REG (X))
#define ARM_FRAME_RTX(X) \
((X) == frame_pointer_rtx || (X) == stack_pointer_rtx \
|| (X) == arg_pointer_rtx)
int
arm_rtx_costs (x, code, outer_code)
rtx x;
enum rtx_code code, outer_code;
{
enum machine_mode mode = GET_MODE (x);
enum rtx_code subcode;
int extra_cost;
switch (code)
{
case MEM:
/* Memory costs quite a lot for the first word, but subsequent words
load at the equivalent of a single insn each. */
return (10 + 4 * ((GET_MODE_SIZE (mode) - 1) / UNITS_PER_WORD)
+ (CONSTANT_POOL_ADDRESS_P (x) ? 4 : 0));
case DIV:
case MOD:
return 100;
case ROTATE:
if (mode == SImode && GET_CODE (XEXP (x, 1)) == REG)
return 4;
/* Fall through */
case ROTATERT:
if (mode != SImode)
return 8;
/* Fall through */
case ASHIFT: case LSHIFTRT: case ASHIFTRT:
if (mode == DImode)
return (8 + (GET_CODE (XEXP (x, 1)) == CONST_INT ? 0 : 8)
+ ((GET_CODE (XEXP (x, 0)) == REG
|| (GET_CODE (XEXP (x, 0)) == SUBREG
&& GET_CODE (SUBREG_REG (XEXP (x, 0))) == REG))
? 0 : 8));
return (1 + ((GET_CODE (XEXP (x, 0)) == REG
|| (GET_CODE (XEXP (x, 0)) == SUBREG
&& GET_CODE (SUBREG_REG (XEXP (x, 0))) == REG))
? 0 : 4)
+ ((GET_CODE (XEXP (x, 1)) == REG
|| (GET_CODE (XEXP (x, 1)) == SUBREG
&& GET_CODE (SUBREG_REG (XEXP (x, 1))) == REG)
|| (GET_CODE (XEXP (x, 1)) == CONST_INT))
? 0 : 4));
case MINUS:
if (mode == DImode)
return (4 + (REG_OR_SUBREG_REG (XEXP (x, 1)) ? 0 : 8)
+ ((REG_OR_SUBREG_REG (XEXP (x, 0))
|| (GET_CODE (XEXP (x, 0)) == CONST_INT
&& const_ok_for_arm (INTVAL (XEXP (x, 0)))))
? 0 : 8));
if (GET_MODE_CLASS (mode) == MODE_FLOAT)
return (2 + ((REG_OR_SUBREG_REG (XEXP (x, 1))
|| (GET_CODE (XEXP (x, 1)) == CONST_DOUBLE
&& const_double_rtx_ok_for_fpu (XEXP (x, 1))))
? 0 : 8)
+ ((REG_OR_SUBREG_REG (XEXP (x, 0))
|| (GET_CODE (XEXP (x, 0)) == CONST_DOUBLE
&& const_double_rtx_ok_for_fpu (XEXP (x, 0))))
? 0 : 8));
if (((GET_CODE (XEXP (x, 0)) == CONST_INT
&& const_ok_for_arm (INTVAL (XEXP (x, 0)))
&& REG_OR_SUBREG_REG (XEXP (x, 1))))
|| (((subcode = GET_CODE (XEXP (x, 1))) == ASHIFT
|| subcode == ASHIFTRT || subcode == LSHIFTRT
|| subcode == ROTATE || subcode == ROTATERT
|| (subcode == MULT
&& GET_CODE (XEXP (XEXP (x, 1), 1)) == CONST_INT
&& ((INTVAL (XEXP (XEXP (x, 1), 1)) &
(INTVAL (XEXP (XEXP (x, 1), 1)) - 1)) == 0)))
&& REG_OR_SUBREG_REG (XEXP (XEXP (x, 1), 0))
&& (REG_OR_SUBREG_REG (XEXP (XEXP (x, 1), 1))
|| GET_CODE (XEXP (XEXP (x, 1), 1)) == CONST_INT)
&& REG_OR_SUBREG_REG (XEXP (x, 0))))
return 1;
/* Fall through */
case PLUS:
if (GET_MODE_CLASS (mode) == MODE_FLOAT)
return (2 + (REG_OR_SUBREG_REG (XEXP (x, 0)) ? 0 : 8)
+ ((REG_OR_SUBREG_REG (XEXP (x, 1))
|| (GET_CODE (XEXP (x, 1)) == CONST_DOUBLE
&& const_double_rtx_ok_for_fpu (XEXP (x, 1))))
? 0 : 8));
/* Fall through */
case AND: case XOR: case IOR:
extra_cost = 0;
/* Normally the frame registers will be spilt into reg+const during
reload, so it is a bad idea to combine them with other instructions,
since then they might not be moved outside of loops. As a compromise
we allow integration with ops that have a constant as their second
operand. */
if ((REG_OR_SUBREG_REG (XEXP (x, 0))
&& ARM_FRAME_RTX (REG_OR_SUBREG_RTX (XEXP (x, 0)))
&& GET_CODE (XEXP (x, 1)) != CONST_INT)
|| (REG_OR_SUBREG_REG (XEXP (x, 0))
&& ARM_FRAME_RTX (REG_OR_SUBREG_RTX (XEXP (x, 0)))))
extra_cost = 4;
if (mode == DImode)
return (4 + extra_cost + (REG_OR_SUBREG_REG (XEXP (x, 0)) ? 0 : 8)
+ ((REG_OR_SUBREG_REG (XEXP (x, 1))
|| (GET_CODE (XEXP (x, 1)) == CONST_INT
&& const_ok_for_op (INTVAL (XEXP (x, 1)), code, mode)))
? 0 : 8));
if (REG_OR_SUBREG_REG (XEXP (x, 0)))
return (1 + (GET_CODE (XEXP (x, 1)) == CONST_INT ? 0 : extra_cost)
+ ((REG_OR_SUBREG_REG (XEXP (x, 1))
|| (GET_CODE (XEXP (x, 1)) == CONST_INT
&& const_ok_for_op (INTVAL (XEXP (x, 1)), code, mode)))
? 0 : 4));
else if (REG_OR_SUBREG_REG (XEXP (x, 1)))
return (1 + extra_cost
+ ((((subcode = GET_CODE (XEXP (x, 0))) == ASHIFT
|| subcode == LSHIFTRT || subcode == ASHIFTRT
|| subcode == ROTATE || subcode == ROTATERT
|| (subcode == MULT
&& GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
&& ((INTVAL (XEXP (XEXP (x, 0), 1)) &
(INTVAL (XEXP (XEXP (x, 0), 1)) - 1)) == 0))
&& (REG_OR_SUBREG_REG (XEXP (XEXP (x, 0), 0)))
&& ((REG_OR_SUBREG_REG (XEXP (XEXP (x, 0), 1)))
|| GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT)))
? 0 : 4));
return 8;
case MULT:
/* There is no point basing this on the tuning, since it is always the
fast variant if it exists at all */
if (arm_fast_multiply && mode == DImode
&& (GET_CODE (XEXP (x, 0)) == GET_CODE (XEXP (x, 1)))
&& (GET_CODE (XEXP (x, 0)) == ZERO_EXTEND
|| GET_CODE (XEXP (x, 0)) == SIGN_EXTEND))
return 8;
if (GET_MODE_CLASS (mode) == MODE_FLOAT
|| mode == DImode)
return 30;
if (GET_CODE (XEXP (x, 1)) == CONST_INT)
{
unsigned HOST_WIDE_INT i = (INTVAL (XEXP (x, 1))
& (unsigned HOST_WIDE_INT) 0xffffffff);
int add_cost = const_ok_for_arm (i) ? 4 : 8;
int j;
/* Tune as appropriate */
int booth_unit_size = ((tune_flags & FL_FAST_MULT) ? 8 : 2);
for (j = 0; i && j < 32; j += booth_unit_size)
{
i >>= booth_unit_size;
add_cost += 2;
}
return add_cost;
}
return (((tune_flags & FL_FAST_MULT) ? 8 : 30)
+ (REG_OR_SUBREG_REG (XEXP (x, 0)) ? 0 : 4)
+ (REG_OR_SUBREG_REG (XEXP (x, 1)) ? 0 : 4));
case NEG:
if (GET_MODE_CLASS (mode) == MODE_FLOAT)
return 4 + (REG_OR_SUBREG_REG (XEXP (x, 0)) ? 0 : 6);
/* Fall through */
case NOT:
if (mode == DImode)
return 4 + (REG_OR_SUBREG_REG (XEXP (x, 0)) ? 0 : 4);
return 1 + (REG_OR_SUBREG_REG (XEXP (x, 0)) ? 0 : 4);
case IF_THEN_ELSE:
if (GET_CODE (XEXP (x, 1)) == PC || GET_CODE (XEXP (x, 2)) == PC)
return 14;
return 2;
case COMPARE:
return 1;
case ABS:
return 4 + (mode == DImode ? 4 : 0);
case SIGN_EXTEND:
if (GET_MODE (XEXP (x, 0)) == QImode)
return (4 + (mode == DImode ? 4 : 0)
+ (GET_CODE (XEXP (x, 0)) == MEM ? 10 : 0));
/* Fall through */
case ZERO_EXTEND:
switch (GET_MODE (XEXP (x, 0)))
{
case QImode:
return (1 + (mode == DImode ? 4 : 0)
+ (GET_CODE (XEXP (x, 0)) == MEM ? 10 : 0));
case HImode:
return (4 + (mode == DImode ? 4 : 0)
+ (GET_CODE (XEXP (x, 0)) == MEM ? 10 : 0));
case SImode:
return (1 + (GET_CODE (XEXP (x, 0)) == MEM ? 10 : 0));
}
abort ();
default:
return 99;
}
}
int
arm_adjust_cost (insn, link, dep, cost)
rtx insn;
rtx link;
rtx dep;
int cost;
{
rtx i_pat, d_pat;
if ((i_pat = single_set (insn)) != NULL
&& GET_CODE (SET_SRC (i_pat)) == MEM
&& (d_pat = single_set (dep)) != NULL
&& GET_CODE (SET_DEST (d_pat)) == MEM)
{
/* This is a load after a store, there is no conflict if the load reads
from a cached area. Assume that loads from the stack, and from the
constant pool are cached, and that others will miss. This is a
hack. */
/* debug_rtx (insn);
debug_rtx (dep);
debug_rtx (link);
fprintf (stderr, "costs %d\n", cost); */
if (CONSTANT_POOL_ADDRESS_P (XEXP (SET_SRC (i_pat), 0))
|| reg_mentioned_p (stack_pointer_rtx, XEXP (SET_SRC (i_pat), 0))
|| reg_mentioned_p (frame_pointer_rtx, XEXP (SET_SRC (i_pat), 0))
|| reg_mentioned_p (hard_frame_pointer_rtx,
XEXP (SET_SRC (i_pat), 0)))
{
/* fprintf (stderr, "***** Now 1\n"); */
return 1;
}
}
return cost;
}
/* This code has been fixed for cross compilation. */
static int fpa_consts_inited = 0;
char *strings_fpa[8] = {
"0", "1", "2", "3",
"4", "5", "0.5", "10"
};
static REAL_VALUE_TYPE values_fpa[8];
static void
init_fpa_table ()
{
int i;
REAL_VALUE_TYPE r;
for (i = 0; i < 8; i++)
{
r = REAL_VALUE_ATOF (strings_fpa[i], DFmode);
values_fpa[i] = r;
}
fpa_consts_inited = 1;
}
/* Return TRUE if rtx X is a valid immediate FPU constant. */
int
const_double_rtx_ok_for_fpu (x)
rtx x;
{
REAL_VALUE_TYPE r;
int i;
if (!fpa_consts_inited)
init_fpa_table ();
REAL_VALUE_FROM_CONST_DOUBLE (r, x);
if (REAL_VALUE_MINUS_ZERO (r))
return 0;
for (i = 0; i < 8; i++)
if (REAL_VALUES_EQUAL (r, values_fpa[i]))
return 1;
return 0;
}
/* Return TRUE if rtx X is a valid immediate FPU constant. */
int
neg_const_double_rtx_ok_for_fpu (x)
rtx x;
{
REAL_VALUE_TYPE r;
int i;
if (!fpa_consts_inited)
init_fpa_table ();
REAL_VALUE_FROM_CONST_DOUBLE (r, x);
r = REAL_VALUE_NEGATE (r);
if (REAL_VALUE_MINUS_ZERO (r))
return 0;
for (i = 0; i < 8; i++)
if (REAL_VALUES_EQUAL (r, values_fpa[i]))
return 1;
return 0;
}
/* Predicates for `match_operand' and `match_operator'. */
/* s_register_operand is the same as register_operand, but it doesn't accept
(SUBREG (MEM)...).
This function exists because at the time it was put in it led to better
code. SUBREG(MEM) always needs a reload in the places where
s_register_operand is used, and this seemed to lead to excessive
reloading. */
int
s_register_operand (op, mode)
register rtx op;
enum machine_mode mode;
{
if (GET_MODE (op) != mode && mode != VOIDmode)
return 0;
if (GET_CODE (op) == SUBREG)
op = SUBREG_REG (op);
/* We don't consider registers whose class is NO_REGS
to be a register operand. */
return (GET_CODE (op) == REG
&& (REGNO (op) >= FIRST_PSEUDO_REGISTER
|| REGNO_REG_CLASS (REGNO (op)) != NO_REGS));
}
/* Only accept reg, subreg(reg), const_int. */
int
reg_or_int_operand (op, mode)
register rtx op;
enum machine_mode mode;
{
if (GET_CODE (op) == CONST_INT)
return 1;
if (GET_MODE (op) != mode && mode != VOIDmode)
return 0;
if (GET_CODE (op) == SUBREG)
op = SUBREG_REG (op);
/* We don't consider registers whose class is NO_REGS
to be a register operand. */
return (GET_CODE (op) == REG
&& (REGNO (op) >= FIRST_PSEUDO_REGISTER
|| REGNO_REG_CLASS (REGNO (op)) != NO_REGS));
}
/* Return 1 if OP is an item in memory, given that we are in reload. */
int
reload_memory_operand (op, mode)
rtx op;
enum machine_mode mode;
{
int regno = true_regnum (op);
return (! CONSTANT_P (op)
&& (regno == -1
|| (GET_CODE (op) == REG
&& REGNO (op) >= FIRST_PSEUDO_REGISTER)));
}
/* Return TRUE for valid operands for the rhs of an ARM instruction. */
int
arm_rhs_operand (op, mode)
rtx op;
enum machine_mode mode;
{
return (s_register_operand (op, mode)
|| (GET_CODE (op) == CONST_INT && const_ok_for_arm (INTVAL (op))));
}
/* Return TRUE for valid operands for the rhs of an ARM instruction, or a load.
*/
int
arm_rhsm_operand (op, mode)
rtx op;
enum machine_mode mode;
{
return (s_register_operand (op, mode)
|| (GET_CODE (op) == CONST_INT && const_ok_for_arm (INTVAL (op)))
|| memory_operand (op, mode));
}
/* Return TRUE for valid operands for the rhs of an ARM instruction, or if a
constant that is valid when negated. */
int
arm_add_operand (op, mode)
rtx op;
enum machine_mode mode;
{
return (s_register_operand (op, mode)
|| (GET_CODE (op) == CONST_INT
&& (const_ok_for_arm (INTVAL (op))
|| const_ok_for_arm (-INTVAL (op)))));
}
int
arm_not_operand (op, mode)
rtx op;
enum machine_mode mode;
{
return (s_register_operand (op, mode)
|| (GET_CODE (op) == CONST_INT
&& (const_ok_for_arm (INTVAL (op))
|| const_ok_for_arm (~INTVAL (op)))));
}
/* Return TRUE if the operand is a memory reference which contains an
offsettable address. */
int
offsettable_memory_operand (op, mode)
register rtx op;
enum machine_mode mode;
{
if (mode == VOIDmode)
mode = GET_MODE (op);
return (mode == GET_MODE (op)
&& GET_CODE (op) == MEM
&& offsettable_address_p (reload_completed | reload_in_progress,
mode, XEXP (op, 0)));
}
/* Return TRUE if the operand is a memory reference which is, or can be
made word aligned by adjusting the offset. */
int
alignable_memory_operand (op, mode)
register rtx op;
enum machine_mode mode;
{
rtx reg;
if (mode == VOIDmode)
mode = GET_MODE (op);
if (mode != GET_MODE (op) || GET_CODE (op) != MEM)
return 0;
op = XEXP (op, 0);
return ((GET_CODE (reg = op) == REG
|| (GET_CODE (op) == SUBREG
&& GET_CODE (reg = SUBREG_REG (op)) == REG)
|| (GET_CODE (op) == PLUS
&& GET_CODE (XEXP (op, 1)) == CONST_INT
&& (GET_CODE (reg = XEXP (op, 0)) == REG
|| (GET_CODE (XEXP (op, 0)) == SUBREG
&& GET_CODE (reg = SUBREG_REG (XEXP (op, 0))) == REG))))
&& REGNO_POINTER_ALIGN (REGNO (reg)) >= 4);
}
/* Similar to s_register_operand, but does not allow hard integer
registers. */
int
f_register_operand (op, mode)
register rtx op;
enum machine_mode mode;
{
if (GET_MODE (op) != mode && mode != VOIDmode)
return 0;
if (GET_CODE (op) == SUBREG)
op = SUBREG_REG (op);
/* We don't consider registers whose class is NO_REGS
to be a register operand. */
return (GET_CODE (op) == REG
&& (REGNO (op) >= FIRST_PSEUDO_REGISTER
|| REGNO_REG_CLASS (REGNO (op)) == FPU_REGS));
}
/* Return TRUE for valid operands for the rhs of an FPU instruction. */
int
fpu_rhs_operand (op, mode)
rtx op;
enum machine_mode mode;
{
if (s_register_operand (op, mode))
return TRUE;
else if (GET_CODE (op) == CONST_DOUBLE)
return (const_double_rtx_ok_for_fpu (op));
return FALSE;
}
int
fpu_add_operand (op, mode)
rtx op;
enum machine_mode mode;
{
if (s_register_operand (op, mode))
return TRUE;
else if (GET_CODE (op) == CONST_DOUBLE)
return (const_double_rtx_ok_for_fpu (op)
|| neg_const_double_rtx_ok_for_fpu (op));
return FALSE;
}
/* Return nonzero if OP is a constant power of two. */
int
power_of_two_operand (op, mode)
rtx op;
enum machine_mode mode;
{
if (GET_CODE (op) == CONST_INT)
{
HOST_WIDE_INT value = INTVAL(op);
return value != 0 && (value & (value - 1)) == 0;
}
return FALSE;
}
/* Return TRUE for a valid operand of a DImode operation.
Either: REG, CONST_DOUBLE or MEM(DImode_address).
Note that this disallows MEM(REG+REG), but allows
MEM(PRE/POST_INC/DEC(REG)). */
int
di_operand (op, mode)
rtx op;
enum machine_mode mode;
{
if (s_register_operand (op, mode))
return TRUE;
switch (GET_CODE (op))
{
case CONST_DOUBLE:
case CONST_INT:
return TRUE;
case MEM:
return memory_address_p (DImode, XEXP (op, 0));
default:
return FALSE;
}
}
/* Return TRUE for a valid operand of a DFmode operation when -msoft-float.
Either: REG, CONST_DOUBLE or MEM(DImode_address).
Note that this disallows MEM(REG+REG), but allows
MEM(PRE/POST_INC/DEC(REG)). */
int
soft_df_operand (op, mode)
rtx op;
enum machine_mode mode;
{
if (s_register_operand (op, mode))
return TRUE;
switch (GET_CODE (op))
{
case CONST_DOUBLE:
return TRUE;
case MEM:
return memory_address_p (DFmode, XEXP (op, 0));
default:
return FALSE;
}
}
/* Return TRUE for valid index operands. */
int
index_operand (op, mode)
rtx op;
enum machine_mode mode;
{
return (s_register_operand(op, mode)
|| (immediate_operand (op, mode)
&& INTVAL (op) < 4096 && INTVAL (op) > -4096));
}
/* Return TRUE for valid shifts by a constant. This also accepts any
power of two on the (somewhat overly relaxed) assumption that the
shift operator in this case was a mult. */
int
const_shift_operand (op, mode)
rtx op;
enum machine_mode mode;
{
return (power_of_two_operand (op, mode)
|| (immediate_operand (op, mode)
&& (INTVAL (op) < 32 && INTVAL (op) > 0)));
}
/* Return TRUE for arithmetic operators which can be combined with a multiply
(shift). */
int
shiftable_operator (x, mode)
rtx x;
enum machine_mode mode;
{
if (GET_MODE (x) != mode)
return FALSE;
else
{
enum rtx_code code = GET_CODE (x);
return (code == PLUS || code == MINUS
|| code == IOR || code == XOR || code == AND);
}
}
/* Return TRUE for shift operators. */
int
shift_operator (x, mode)
rtx x;
enum machine_mode mode;
{
if (GET_MODE (x) != mode)
return FALSE;
else
{
enum rtx_code code = GET_CODE (x);
if (code == MULT)
return power_of_two_operand (XEXP (x, 1));
return (code == ASHIFT || code == ASHIFTRT || code == LSHIFTRT
|| code == ROTATERT);
}
}
int equality_operator (x, mode)
rtx x;
enum machine_mode mode;
{
return GET_CODE (x) == EQ || GET_CODE (x) == NE;
}
/* Return TRUE for SMIN SMAX UMIN UMAX operators. */
int
minmax_operator (x, mode)
rtx x;
enum machine_mode mode;
{
enum rtx_code code = GET_CODE (x);
if (GET_MODE (x) != mode)
return FALSE;
return code == SMIN || code == SMAX || code == UMIN || code == UMAX;
}
/* return TRUE if x is EQ or NE */
/* Return TRUE if this is the condition code register, if we aren't given
a mode, accept any class CCmode register */
int
cc_register (x, mode)
rtx x;
enum machine_mode mode;
{
if (mode == VOIDmode)
{
mode = GET_MODE (x);
if (GET_MODE_CLASS (mode) != MODE_CC)
return FALSE;
}
if (mode == GET_MODE (x) && GET_CODE (x) == REG && REGNO (x) == 24)
return TRUE;
return FALSE;
}
/* Return TRUE if this is the condition code register, if we aren't given
a mode, accept any class CCmode register which indicates a dominance
expression. */
int
dominant_cc_register (x, mode)
rtx x;
enum machine_mode mode;
{
if (mode == VOIDmode)
{
mode = GET_MODE (x);
if (GET_MODE_CLASS (mode) != MODE_CC)
return FALSE;
}
if (mode != CC_DNEmode && mode != CC_DEQmode
&& mode != CC_DLEmode && mode != CC_DLTmode
&& mode != CC_DGEmode && mode != CC_DGTmode
&& mode != CC_DLEUmode && mode != CC_DLTUmode
&& mode != CC_DGEUmode && mode != CC_DGTUmode)
return FALSE;
if (mode == GET_MODE (x) && GET_CODE (x) == REG && REGNO (x) == 24)
return TRUE;
return FALSE;
}
/* Return TRUE if X references a SYMBOL_REF. */
int
symbol_mentioned_p (x)
rtx x;
{
register char *fmt;
register int i;
if (GET_CODE (x) == SYMBOL_REF)
return 1;
fmt = GET_RTX_FORMAT (GET_CODE (x));
for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
{
if (fmt[i] == 'E')
{
register int j;
for (j = XVECLEN (x, i) - 1; j >= 0; j--)
if (symbol_mentioned_p (XVECEXP (x, i, j)))
return 1;
}
else if (fmt[i] == 'e' && symbol_mentioned_p (XEXP (x, i)))
return 1;
}
return 0;
}
/* Return TRUE if X references a LABEL_REF. */
int
label_mentioned_p (x)
rtx x;
{
register char *fmt;
register int i;
if (GET_CODE (x) == LABEL_REF)
return 1;
fmt = GET_RTX_FORMAT (GET_CODE (x));
for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
{
if (fmt[i] == 'E')
{
register int j;
for (j = XVECLEN (x, i) - 1; j >= 0; j--)
if (label_mentioned_p (XVECEXP (x, i, j)))
return 1;
}
else if (fmt[i] == 'e' && label_mentioned_p (XEXP (x, i)))
return 1;
}
return 0;
}
enum rtx_code
minmax_code (x)
rtx x;
{
enum rtx_code code = GET_CODE (x);
if (code == SMAX)
return GE;
else if (code == SMIN)
return LE;
else if (code == UMIN)
return LEU;
else if (code == UMAX)
return GEU;
abort ();
}
/* Return 1 if memory locations are adjacent */
int
adjacent_mem_locations (a, b)
rtx a, b;
{
int val0 = 0, val1 = 0;
int reg0, reg1;
if ((GET_CODE (XEXP (a, 0)) == REG
|| (GET_CODE (XEXP (a, 0)) == PLUS
&& GET_CODE (XEXP (XEXP (a, 0), 1)) == CONST_INT))
&& (GET_CODE (XEXP (b, 0)) == REG
|| (GET_CODE (XEXP (b, 0)) == PLUS
&& GET_CODE (XEXP (XEXP (b, 0), 1)) == CONST_INT)))
{
if (GET_CODE (XEXP (a, 0)) == PLUS)
{
reg0 = REGNO (XEXP (XEXP (a, 0), 0));
val0 = INTVAL (XEXP (XEXP (a, 0), 1));
}
else
reg0 = REGNO (XEXP (a, 0));
if (GET_CODE (XEXP (b, 0)) == PLUS)
{
reg1 = REGNO (XEXP (XEXP (b, 0), 0));
val1 = INTVAL (XEXP (XEXP (b, 0), 1));
}
else
reg1 = REGNO (XEXP (b, 0));
return (reg0 == reg1) && ((val1 - val0) == 4 || (val0 - val1) == 4);
}
return 0;
}
/* Return 1 if OP is a load multiple operation. It is known to be
parallel and the first section will be tested. */
int
load_multiple_operation (op, mode)
rtx op;
enum machine_mode mode;
{
HOST_WIDE_INT count = XVECLEN (op, 0);
int dest_regno;
rtx src_addr;
HOST_WIDE_INT i = 1, base = 0;
rtx elt;
if (count <= 1
|| GET_CODE (XVECEXP (op, 0, 0)) != SET)
return 0;
/* Check to see if this might be a write-back */
if (GET_CODE (SET_SRC (elt = XVECEXP (op, 0, 0))) == PLUS)
{
i++;
base = 1;
/* Now check it more carefully */
if (GET_CODE (SET_DEST (elt)) != REG
|| GET_CODE (XEXP (SET_SRC (elt), 0)) != REG
|| REGNO (XEXP (SET_SRC (elt), 0)) != REGNO (SET_DEST (elt))
|| GET_CODE (XEXP (SET_SRC (elt), 1)) != CONST_INT
|| INTVAL (XEXP (SET_SRC (elt), 1)) != (count - 2) * 4
|| GET_CODE (XVECEXP (op, 0, count - 1)) != CLOBBER
|| GET_CODE (XEXP (XVECEXP (op, 0, count - 1), 0)) != REG
|| REGNO (XEXP (XVECEXP (op, 0, count - 1), 0))
!= REGNO (SET_DEST (elt)))
return 0;
count--;
}
/* Perform a quick check so we don't blow up below. */
if (count <= i
|| GET_CODE (XVECEXP (op, 0, i - 1)) != SET
|| GET_CODE (SET_DEST (XVECEXP (op, 0, i - 1))) != REG
|| GET_CODE (SET_SRC (XVECEXP (op, 0, i - 1))) != MEM)
return 0;
dest_regno = REGNO (SET_DEST (XVECEXP (op, 0, i - 1)));
src_addr = XEXP (SET_SRC (XVECEXP (op, 0, i - 1)), 0);
for (; i < count; i++)
{
rtx elt = XVECEXP (op, 0, i);
if (GET_CODE (elt) != SET
|| GET_CODE (SET_DEST (elt)) != REG
|| GET_MODE (SET_DEST (elt)) != SImode
|| REGNO (SET_DEST (elt)) != dest_regno + i - base
|| GET_CODE (SET_SRC (elt)) != MEM
|| GET_MODE (SET_SRC (elt)) != SImode
|| GET_CODE (XEXP (SET_SRC (elt), 0)) != PLUS
|| ! rtx_equal_p (XEXP (XEXP (SET_SRC (elt), 0), 0), src_addr)
|| GET_CODE (XEXP (XEXP (SET_SRC (elt), 0), 1)) != CONST_INT
|| INTVAL (XEXP (XEXP (SET_SRC (elt), 0), 1)) != (i - base) * 4)
return 0;
}
return 1;
}
/* Return 1 if OP is a store multiple operation. It is known to be
parallel and the first section will be tested. */
int
store_multiple_operation (op, mode)
rtx op;
enum machine_mode mode;
{
HOST_WIDE_INT count = XVECLEN (op, 0);
int src_regno;
rtx dest_addr;
HOST_WIDE_INT i = 1, base = 0;
rtx elt;
if (count <= 1
|| GET_CODE (XVECEXP (op, 0, 0)) != SET)
return 0;
/* Check to see if this might be a write-back */
if (GET_CODE (SET_SRC (elt = XVECEXP (op, 0, 0))) == PLUS)
{
i++;
base = 1;
/* Now check it more carefully */
if (GET_CODE (SET_DEST (elt)) != REG
|| GET_CODE (XEXP (SET_SRC (elt), 0)) != REG
|| REGNO (XEXP (SET_SRC (elt), 0)) != REGNO (SET_DEST (elt))
|| GET_CODE (XEXP (SET_SRC (elt), 1)) != CONST_INT
|| INTVAL (XEXP (SET_SRC (elt), 1)) != (count - 2) * 4
|| GET_CODE (XVECEXP (op, 0, count - 1)) != CLOBBER
|| GET_CODE (XEXP (XVECEXP (op, 0, count - 1), 0)) != REG
|| REGNO (XEXP (XVECEXP (op, 0, count - 1), 0))
!= REGNO (SET_DEST (elt)))
return 0;
count--;
}
/* Perform a quick check so we don't blow up below. */
if (count <= i
|| GET_CODE (XVECEXP (op, 0, i - 1)) != SET
|| GET_CODE (SET_DEST (XVECEXP (op, 0, i - 1))) != MEM
|| GET_CODE (SET_SRC (XVECEXP (op, 0, i - 1))) != REG)
return 0;
src_regno = REGNO (SET_SRC (XVECEXP (op, 0, i - 1)));
dest_addr = XEXP (SET_DEST (XVECEXP (op, 0, i - 1)), 0);
for (; i < count; i++)
{
elt = XVECEXP (op, 0, i);
if (GET_CODE (elt) != SET
|| GET_CODE (SET_SRC (elt)) != REG
|| GET_MODE (SET_SRC (elt)) != SImode
|| REGNO (SET_SRC (elt)) != src_regno + i - base
|| GET_CODE (SET_DEST (elt)) != MEM
|| GET_MODE (SET_DEST (elt)) != SImode
|| GET_CODE (XEXP (SET_DEST (elt), 0)) != PLUS
|| ! rtx_equal_p (XEXP (XEXP (SET_DEST (elt), 0), 0), dest_addr)
|| GET_CODE (XEXP (XEXP (SET_DEST (elt), 0), 1)) != CONST_INT
|| INTVAL (XEXP (XEXP (SET_DEST (elt), 0), 1)) != (i - base) * 4)
return 0;
}
return 1;
}
int
load_multiple_sequence (operands, nops, regs, base, load_offset)
rtx *operands;
int nops;
int *regs;
int *base;
HOST_WIDE_INT *load_offset;
{
int unsorted_regs[4];
HOST_WIDE_INT unsorted_offsets[4];
int order[4];
int base_reg;
int i;
/* Can only handle 2, 3, or 4 insns at present, though could be easily
extended if required. */
if (nops < 2 || nops > 4)
abort ();
/* Loop over the operands and check that the memory references are
suitable (ie immediate offsets from the same base register). At
the same time, extract the target register, and the memory
offsets. */
for (i = 0; i < nops; i++)
{
rtx reg;
rtx offset;
if (GET_CODE (operands[nops + i]) != MEM)
abort ();
/* Don't reorder volatile memory references; it doesn't seem worth
looking for the case where the order is ok anyway. */
if (MEM_VOLATILE_P (operands[nops + i]))
return 0;
offset = const0_rtx;
if ((GET_CODE (reg = XEXP (operands[nops + i], 0)) == REG
|| (GET_CODE (reg) == SUBREG
&& GET_CODE (reg = SUBREG_REG (reg)) == REG))
|| (GET_CODE (XEXP (operands[nops + i], 0)) == PLUS
&& ((GET_CODE (reg = XEXP (XEXP (operands[nops + i], 0), 0))
== REG)
|| (GET_CODE (reg) == SUBREG
&& GET_CODE (reg = SUBREG_REG (reg)) == REG))
&& (GET_CODE (offset = XEXP (XEXP (operands[nops + i], 0), 1))
== CONST_INT)))
{
if (i == 0)
{
base_reg = REGNO(reg);
unsorted_regs[0] = (GET_CODE (operands[i]) == REG
? REGNO (operands[i])
: REGNO (SUBREG_REG (operands[i])));
order[0] = 0;
}
else
{
if (base_reg != REGNO (reg))
/* Not addressed from the same base register. */
return 0;
unsorted_regs[i] = (GET_CODE (operands[i]) == REG
? REGNO (operands[i])
: REGNO (SUBREG_REG (operands[i])));
if (unsorted_regs[i] < unsorted_regs[order[0]])
order[0] = i;
}
/* If it isn't an integer register, or if it overwrites the
base register but isn't the last insn in the list, then
we can't do this. */
if (unsorted_regs[i] < 0 || unsorted_regs[i] > 14
|| (i != nops - 1 && unsorted_regs[i] == base_reg))
return 0;
unsorted_offsets[i] = INTVAL (offset);
}
else
/* Not a suitable memory address. */
return 0;
}
/* All the useful information has now been extracted from the
operands into unsorted_regs and unsorted_offsets; additionally,
order[0] has been set to the lowest numbered register in the
list. Sort the registers into order, and check that the memory
offsets are ascending and adjacent. */
for (i = 1; i < nops; i++)
{
int j;
order[i] = order[i - 1];
for (j = 0; j < nops; j++)
if (unsorted_regs[j] > unsorted_regs[order[i - 1]]
&& (order[i] == order[i - 1]
|| unsorted_regs[j] < unsorted_regs[order[i]]))
order[i] = j;
/* Have we found a suitable register? if not, one must be used more
than once. */
if (order[i] == order[i - 1])
return 0;
/* Is the memory address adjacent and ascending? */
if (unsorted_offsets[order[i]] != unsorted_offsets[order[i - 1]] + 4)
return 0;
}
if (base)
{
*base = base_reg;
for (i = 0; i < nops; i++)
regs[i] = unsorted_regs[order[i]];
*load_offset = unsorted_offsets[order[0]];
}
if (unsorted_offsets[order[0]] == 0)
return 1; /* ldmia */
if (unsorted_offsets[order[0]] == 4)
return 2; /* ldmib */
if (unsorted_offsets[order[nops - 1]] == 0)
return 3; /* ldmda */
if (unsorted_offsets[order[nops - 1]] == -4)
return 4; /* ldmdb */
/* Can't do it without setting up the offset, only do this if it takes
no more than one insn. */
return (const_ok_for_arm (unsorted_offsets[order[0]])
|| const_ok_for_arm (-unsorted_offsets[order[0]])) ? 5 : 0;
}
char *
emit_ldm_seq (operands, nops)
rtx *operands;
int nops;
{
int regs[4];
int base_reg;
HOST_WIDE_INT offset;
char buf[100];
int i;
switch (load_multiple_sequence (operands, nops, regs, &base_reg, &offset))
{
case 1:
strcpy (buf, "ldm%?ia\t");
break;
case 2:
strcpy (buf, "ldm%?ib\t");
break;
case 3:
strcpy (buf, "ldm%?da\t");
break;
case 4:
strcpy (buf, "ldm%?db\t");
break;
case 5:
if (offset >= 0)
sprintf (buf, "add%%?\t%s%s, %s%s, #%ld", REGISTER_PREFIX,
reg_names[regs[0]], REGISTER_PREFIX, reg_names[base_reg],
(long) offset);
else
sprintf (buf, "sub%%?\t%s%s, %s%s, #%ld", REGISTER_PREFIX,
reg_names[regs[0]], REGISTER_PREFIX, reg_names[base_reg],
(long) -offset);
output_asm_insn (buf, operands);
base_reg = regs[0];
strcpy (buf, "ldm%?ia\t");
break;
default:
abort ();
}
sprintf (buf + strlen (buf), "%s%s, {%s%s", REGISTER_PREFIX,
reg_names[base_reg], REGISTER_PREFIX, reg_names[regs[0]]);
for (i = 1; i < nops; i++)
sprintf (buf + strlen (buf), ", %s%s", REGISTER_PREFIX,
reg_names[regs[i]]);
strcat (buf, "}\t%@ phole ldm");
output_asm_insn (buf, operands);
return "";
}
int
store_multiple_sequence (operands, nops, regs, base, load_offset)
rtx *operands;
int nops;
int *regs;
int *base;
HOST_WIDE_INT *load_offset;
{
int unsorted_regs[4];
HOST_WIDE_INT unsorted_offsets[4];
int order[4];
int base_reg;
int i;
/* Can only handle 2, 3, or 4 insns at present, though could be easily
extended if required. */
if (nops < 2 || nops > 4)
abort ();
/* Loop over the operands and check that the memory references are
suitable (ie immediate offsets from the same base register). At
the same time, extract the target register, and the memory
offsets. */
for (i = 0; i < nops; i++)
{
rtx reg;
rtx offset;
if (GET_CODE (operands[nops + i]) != MEM)
abort ();
/* Don't reorder volatile memory references; it doesn't seem worth
looking for the case where the order is ok anyway. */
if (MEM_VOLATILE_P (operands[nops + i]))
return 0;
offset = const0_rtx;
if ((GET_CODE (reg = XEXP (operands[nops + i], 0)) == REG
|| (GET_CODE (reg) == SUBREG
&& GET_CODE (reg = SUBREG_REG (reg)) == REG))
|| (GET_CODE (XEXP (operands[nops + i], 0)) == PLUS
&& ((GET_CODE (reg = XEXP (XEXP (operands[nops + i], 0), 0))
== REG)
|| (GET_CODE (reg) == SUBREG
&& GET_CODE (reg = SUBREG_REG (reg)) == REG))
&& (GET_CODE (offset = XEXP (XEXP (operands[nops + i], 0), 1))
== CONST_INT)))
{
if (i == 0)
{
base_reg = REGNO(reg);
unsorted_regs[0] = (GET_CODE (operands[i]) == REG
? REGNO (operands[i])
: REGNO (SUBREG_REG (operands[i])));
order[0] = 0;
}
else
{
if (base_reg != REGNO (reg))
/* Not addressed from the same base register. */
return 0;
unsorted_regs[i] = (GET_CODE (operands[i]) == REG
? REGNO (operands[i])
: REGNO (SUBREG_REG (operands[i])));
if (unsorted_regs[i] < unsorted_regs[order[0]])
order[0] = i;
}
/* If it isn't an integer register, then we can't do this. */
if (unsorted_regs[i] < 0 || unsorted_regs[i] > 14)
return 0;
unsorted_offsets[i] = INTVAL (offset);
}
else
/* Not a suitable memory address. */
return 0;
}
/* All the useful information has now been extracted from the
operands into unsorted_regs and unsorted_offsets; additionally,
order[0] has been set to the lowest numbered register in the
list. Sort the registers into order, and check that the memory
offsets are ascending and adjacent. */
for (i = 1; i < nops; i++)
{
int j;
order[i] = order[i - 1];
for (j = 0; j < nops; j++)
if (unsorted_regs[j] > unsorted_regs[order[i - 1]]
&& (order[i] == order[i - 1]
|| unsorted_regs[j] < unsorted_regs[order[i]]))
order[i] = j;
/* Have we found a suitable register? if not, one must be used more
than once. */
if (order[i] == order[i - 1])
return 0;
/* Is the memory address adjacent and ascending? */
if (unsorted_offsets[order[i]] != unsorted_offsets[order[i - 1]] + 4)
return 0;
}
if (base)
{
*base = base_reg;
for (i = 0; i < nops; i++)
regs[i] = unsorted_regs[order[i]];
*load_offset = unsorted_offsets[order[0]];
}
if (unsorted_offsets[order[0]] == 0)
return 1; /* stmia */
if (unsorted_offsets[order[0]] == 4)
return 2; /* stmib */
if (unsorted_offsets[order[nops - 1]] == 0)
return 3; /* stmda */
if (unsorted_offsets[order[nops - 1]] == -4)
return 4; /* stmdb */
return 0;
}
char *
emit_stm_seq (operands, nops)
rtx *operands;
int nops;
{
int regs[4];
int base_reg;
HOST_WIDE_INT offset;
char buf[100];
int i;
switch (store_multiple_sequence (operands, nops, regs, &base_reg, &offset))
{
case 1:
strcpy (buf, "stm%?ia\t");
break;
case 2:
strcpy (buf, "stm%?ib\t");
break;
case 3:
strcpy (buf, "stm%?da\t");
break;
case 4:
strcpy (buf, "stm%?db\t");
break;
default:
abort ();
}
sprintf (buf + strlen (buf), "%s%s, {%s%s", REGISTER_PREFIX,
reg_names[base_reg], REGISTER_PREFIX, reg_names[regs[0]]);
for (i = 1; i < nops; i++)
sprintf (buf + strlen (buf), ", %s%s", REGISTER_PREFIX,
reg_names[regs[i]]);
strcat (buf, "}\t%@ phole stm");
output_asm_insn (buf, operands);
return "";
}
int
multi_register_push (op, mode)
rtx op;
enum machine_mode mode;
{
if (GET_CODE (op) != PARALLEL
|| (GET_CODE (XVECEXP (op, 0, 0)) != SET)
|| (GET_CODE (SET_SRC (XVECEXP (op, 0, 0))) != UNSPEC)
|| (XINT (SET_SRC (XVECEXP (op, 0, 0)), 1) != 2))
return 0;
return 1;
}
/* Routines for use with attributes */
/* Return nonzero if ATTR is a valid attribute for DECL.
ATTRIBUTES are any existing attributes and ARGS are the arguments
supplied with ATTR.
Supported attributes:
naked: don't output any prologue or epilogue code, the user is assumed
to do the right thing. */
int
arm_valid_machine_decl_attribute (decl, attributes, attr, args)
tree decl;
tree attributes;
tree attr;
tree args;
{
if (args != NULL_TREE)
return 0;
if (is_attribute_p ("naked", attr))
return TREE_CODE (decl) == FUNCTION_DECL;
return 0;
}
/* Return non-zero if FUNC is a naked function. */
static int
arm_naked_function_p (func)
tree func;
{
tree a;
if (TREE_CODE (func) != FUNCTION_DECL)
abort ();
a = lookup_attribute ("naked", DECL_MACHINE_ATTRIBUTES (func));
return a != NULL_TREE;
}
/* Routines for use in generating RTL */
rtx
arm_gen_load_multiple (base_regno, count, from, up, write_back)
int base_regno;
int count;
rtx from;
int up;
int write_back;
{
int i = 0, j;
rtx result;
int sign = up ? 1 : -1;
result = gen_rtx (PARALLEL, VOIDmode,
rtvec_alloc (count + (write_back ? 2 : 0)));
if (write_back)
{
XVECEXP (result, 0, 0)
= gen_rtx (SET, GET_MODE (from), from,
plus_constant (from, count * 4 * sign));
i = 1;
count++;
}
for (j = 0; i < count; i++, j++)
{
XVECEXP (result, 0, i)
= gen_rtx (SET, VOIDmode, gen_rtx (REG, SImode, base_regno + j),
gen_rtx (MEM, SImode,
plus_constant (from, j * 4 * sign)));
}
if (write_back)
XVECEXP (result, 0, i) = gen_rtx (CLOBBER, SImode, from);
return result;
}
rtx
arm_gen_store_multiple (base_regno, count, to, up, write_back)
int base_regno;
int count;
rtx to;
int up;
int write_back;
{
int i = 0, j;
rtx result;
int sign = up ? 1 : -1;
result = gen_rtx (PARALLEL, VOIDmode,
rtvec_alloc (count + (write_back ? 2 : 0)));
if (write_back)
{
XVECEXP (result, 0, 0)
= gen_rtx (SET, GET_MODE (to), to,
plus_constant (to, count * 4 * sign));
i = 1;
count++;
}
for (j = 0; i < count; i++, j++)
{
XVECEXP (result, 0, i)
= gen_rtx (SET, VOIDmode,
gen_rtx (MEM, SImode, plus_constant (to, j * 4 * sign)),
gen_rtx (REG, SImode, base_regno + j));
}
if (write_back)
XVECEXP (result, 0, i) = gen_rtx (CLOBBER, SImode, to);
return result;
}
int
arm_gen_movstrqi (operands)
rtx *operands;
{
HOST_WIDE_INT in_words_to_go, out_words_to_go, last_bytes;
int i, r;
rtx src, dst;
rtx st_src, st_dst, end_src, end_dst, fin_src, fin_dst;
rtx part_bytes_reg = NULL;
extern int optimize;
if (GET_CODE (operands[2]) != CONST_INT
|| GET_CODE (operands[3]) != CONST_INT
|| INTVAL (operands[2]) > 64
|| INTVAL (operands[3]) & 3)
return 0;
st_dst = XEXP (operands[0], 0);
st_src = XEXP (operands[1], 0);
fin_dst = dst = copy_to_mode_reg (SImode, st_dst);
fin_src = src = copy_to_mode_reg (SImode, st_src);
in_words_to_go = (INTVAL (operands[2]) + 3) / 4;
out_words_to_go = INTVAL (operands[2]) / 4;
last_bytes = INTVAL (operands[2]) & 3;
if (out_words_to_go != in_words_to_go && ((in_words_to_go - 1) & 3) != 0)
part_bytes_reg = gen_rtx (REG, SImode, (in_words_to_go - 1) & 3);
for (i = 0; in_words_to_go >= 2; i+=4)
{
if (in_words_to_go > 4)
emit_insn (arm_gen_load_multiple (0, 4, src, TRUE, TRUE));
else
emit_insn (arm_gen_load_multiple (0, in_words_to_go, src, TRUE,
FALSE));
if (out_words_to_go)
{
if (out_words_to_go > 4)
emit_insn (arm_gen_store_multiple (0, 4, dst, TRUE, TRUE));
else if (out_words_to_go != 1)
emit_insn (arm_gen_store_multiple (0, out_words_to_go,
dst, TRUE,
(last_bytes == 0
? FALSE : TRUE)));
else
{
emit_move_insn (gen_rtx (MEM, SImode, dst),
gen_rtx (REG, SImode, 0));
if (last_bytes != 0)
emit_insn (gen_addsi3 (dst, dst, GEN_INT (4)));
}
}
in_words_to_go -= in_words_to_go < 4 ? in_words_to_go : 4;
out_words_to_go -= out_words_to_go < 4 ? out_words_to_go : 4;
}
/* OUT_WORDS_TO_GO will be zero here if there are byte stores to do. */
if (out_words_to_go)
{
rtx sreg;
emit_move_insn (sreg = gen_reg_rtx (SImode), gen_rtx (MEM, SImode, src));
emit_move_insn (fin_src = gen_reg_rtx (SImode), plus_constant (src, 4));
emit_move_insn (gen_rtx (MEM, SImode, dst), sreg);
emit_move_insn (fin_dst = gen_reg_rtx (SImode), plus_constant (dst, 4));
in_words_to_go--;
if (in_words_to_go) /* Sanity check */
abort ();
}
if (in_words_to_go)
{
if (in_words_to_go < 0)
abort ();
part_bytes_reg = copy_to_mode_reg (SImode, gen_rtx (MEM, SImode, src));
}
if (BYTES_BIG_ENDIAN && last_bytes)
{
rtx tmp = gen_reg_rtx (SImode);
if (part_bytes_reg == NULL)
abort ();
/* The bytes we want are in the top end of the word */
emit_insn (gen_lshrsi3 (tmp, part_bytes_reg,
GEN_INT (8 * (4 - last_bytes))));
part_bytes_reg = tmp;
while (last_bytes)
{
emit_move_insn (gen_rtx (MEM, QImode,
plus_constant (dst, last_bytes - 1)),
gen_rtx (SUBREG, QImode, part_bytes_reg, 0));
if (--last_bytes)
{
tmp = gen_reg_rtx (SImode);
emit_insn (gen_lshrsi3 (tmp, part_bytes_reg, GEN_INT (8)));
part_bytes_reg = tmp;
}
}
}
else
{
while (last_bytes)
{
if (part_bytes_reg == NULL)
abort ();
emit_move_insn (gen_rtx (MEM, QImode, dst),
gen_rtx (SUBREG, QImode, part_bytes_reg, 0));
if (--last_bytes)