/* Output routines for GCC for ARM.
Copyright (C) 1991, 1993, 1994, 1995, 1996, 1997, 1998, 1999, 2000, 2001,
- 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009
+ 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010
Free Software Foundation, Inc.
Contributed by Pieter `Tiggr' Schoenmakers (rcpieter@win.tue.nl)
and Martin Simmons (@harleqn.co.uk).
static bool arm_cirrus_insn_p (rtx);
static void cirrus_reorg (rtx);
static void arm_init_builtins (void);
-static rtx arm_expand_builtin (tree, rtx, rtx, enum machine_mode, int);
static void arm_init_iwmmxt_builtins (void);
static rtx safe_vector_operand (rtx, enum machine_mode);
static rtx arm_expand_binop_builtin (enum insn_code, tree, rtx);
static void arm_asm_trampoline_template (FILE *);
static void arm_trampoline_init (rtx, tree, rtx);
static rtx arm_trampoline_adjust_address (rtx);
+static rtx arm_pic_static_addr (rtx orig, rtx reg);
\f
/* Table of machine attributes. */
/* The processor for which instructions should be scheduled. */
enum processor_type arm_tune = arm_none;
+/* The current tuning set. */
+const struct tune_params *current_tune;
+
/* The default processor used if not overridden by commandline. */
static enum processor_type arm_default_cpu = arm_none;
the next function. */
static int after_arm_reorg = 0;
-/* The maximum number of insns to be used when loading a constant. */
-static int arm_constant_limit = 3;
-
static enum arm_pcs arm_pcs_default;
/* For an explanation of these variables, see final_prescan_insn below. */
enum processor_type core;
const char *arch;
const unsigned long flags;
- bool (* rtx_costs) (rtx, enum rtx_code, enum rtx_code, int *, bool);
+ const struct tune_params *const tune;
+};
+
+const struct tune_params arm_slowmul_tune =
+{
+ arm_slowmul_rtx_costs,
+ 3
+};
+
+const struct tune_params arm_fastmul_tune =
+{
+ arm_fastmul_rtx_costs,
+ 1
+};
+
+const struct tune_params arm_xscale_tune =
+{
+ arm_xscale_rtx_costs,
+ 2
+};
+
+const struct tune_params arm_9e_tune =
+{
+ arm_9e_rtx_costs,
+ 1
};
/* Not all of these give usefully different compilation alternatives,
{
/* ARM Cores */
#define ARM_CORE(NAME, IDENT, ARCH, FLAGS, COSTS) \
- {NAME, arm_none, #ARCH, FLAGS | FL_FOR_ARCH##ARCH, arm_##COSTS##_rtx_costs},
+ {NAME, arm_none, #ARCH, FLAGS | FL_FOR_ARCH##ARCH, &arm_##COSTS##_tune},
#include "arm-cores.def"
#undef ARM_CORE
{NULL, arm_none, NULL, 0, NULL}
static const struct processors all_architectures[] =
{
/* ARM Architectures */
- /* We don't specify rtx_costs here as it will be figured out
+ /* We don't specify tuning costs here as it will be figured out
from the core. */
{"armv2", arm2, "2", FL_CO_PROC | FL_MODE26 | FL_FOR_ARCH2, NULL},
TLS_LE32
};
+/* The maximum number of insns to be used when loading a constant. */
+inline static int
+arm_constant_limit (bool size_p)
+{
+ return size_p ? 1 : current_tune->constant_limit;
+}
+
/* Emit an insn that's a simple single-set. Both the operands must be known
to be valid. */
inline static rtx
gcc_assert (arm_tune != arm_none);
tune_flags = all_cores[(int)arm_tune].flags;
+ current_tune = all_cores[(int)arm_tune].tune;
if (target_fp16_format_name)
{
if (optimize_size)
{
- arm_constant_limit = 1;
-
/* If optimizing for size, bump the number of instructions that we
are prepared to conditionally execute (even on a StrongARM). */
max_insns_skipped = 6;
}
else
{
- /* For processors with load scheduling, it never costs more than
- 2 cycles to load a constant, and the load scheduler may well
- reduce that to 1. */
- if (arm_ld_sched)
- arm_constant_limit = 1;
-
- /* On XScale the longer latency of a load makes it more difficult
- to achieve a good schedule, so it's faster to synthesize
- constants that can be done in two insns. */
- if (arm_tune_xscale)
- arm_constant_limit = 2;
-
/* StrongARM has early execution of branches, so a sequence
that is worth skipping is shorter. */
if (arm_tune_strongarm)
&& !cond
&& (arm_gen_constant (code, mode, NULL_RTX, val, target, source,
1, 0)
- > arm_constant_limit + (code != SET)))
+ > (arm_constant_limit (optimize_function_for_size_p (cfun))
+ + (code != SET))))
{
if (code == SET)
{
int can_negate = 0;
int final_invert = 0;
int can_negate_initial = 0;
- int can_shift = 0;
int i;
int num_bits_set = 0;
int set_sign_bit_copies = 0;
{
case SET:
can_invert = 1;
- can_shift = 1;
can_negate = 1;
break;
{
rtx pic_ref, address;
rtx insn;
- int subregs = 0;
-
- /* If this function doesn't have a pic register, create one now. */
- require_pic_register ();
if (reg == 0)
{
gcc_assert (can_create_pseudo_p ());
reg = gen_reg_rtx (Pmode);
-
- subregs = 1;
+ address = gen_reg_rtx (Pmode);
}
-
- if (subregs)
- address = gen_reg_rtx (Pmode);
else
address = reg;
- if (TARGET_32BIT)
- emit_insn (gen_pic_load_addr_32bit (address, orig));
- else /* TARGET_THUMB1 */
- emit_insn (gen_pic_load_addr_thumb1 (address, orig));
-
/* VxWorks does not impose a fixed gap between segments; the run-time
gap can be different from the object-file gap. We therefore can't
use GOTOFF unless we are absolutely sure that the symbol is in the
SYMBOL_REF_LOCAL_P (orig)))
&& NEED_GOT_RELOC
&& !TARGET_VXWORKS_RTP)
- pic_ref = gen_rtx_PLUS (Pmode, cfun->machine->pic_reg, address);
+ insn = arm_pic_static_addr (orig, reg);
else
{
+ /* If this function doesn't have a pic register, create one now. */
+ require_pic_register ();
+
+ if (TARGET_32BIT)
+ emit_insn (gen_pic_load_addr_32bit (address, orig));
+ else /* TARGET_THUMB1 */
+ emit_insn (gen_pic_load_addr_thumb1 (address, orig));
+
pic_ref = gen_const_mem (Pmode,
gen_rtx_PLUS (Pmode, cfun->machine->pic_reg,
address));
+ insn = emit_move_insn (reg, pic_ref);
}
- insn = emit_move_insn (reg, pic_ref);
-
/* Put a REG_EQUAL note on this insn, so that it can be optimized
by loop. */
set_unique_reg_note (insn, REG_EQUAL, orig);
emit_use (pic_reg);
}
+/* Generate code to load the address of a static var when flag_pic is set. */
+static rtx
+arm_pic_static_addr (rtx orig, rtx reg)
+{
+ rtx l1, labelno, offset_rtx, insn;
+
+ gcc_assert (flag_pic);
+
+ /* We use an UNSPEC rather than a LABEL_REF because this label
+ never appears in the code stream. */
+ labelno = GEN_INT (pic_labelno++);
+ l1 = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, labelno), UNSPEC_PIC_LABEL);
+ l1 = gen_rtx_CONST (VOIDmode, l1);
+
+ /* On the ARM the PC register contains 'dot + 8' at the time of the
+ addition, on the Thumb it is 'dot + 4'. */
+ offset_rtx = plus_constant (l1, TARGET_ARM ? 8 : 4);
+ offset_rtx = gen_rtx_UNSPEC (Pmode, gen_rtvec (2, orig, offset_rtx),
+ UNSPEC_SYMBOL_OFFSET);
+ offset_rtx = gen_rtx_CONST (Pmode, offset_rtx);
+
+ if (TARGET_32BIT)
+ {
+ emit_insn (gen_pic_load_addr_32bit (reg, offset_rtx));
+ if (TARGET_ARM)
+ insn = emit_insn (gen_pic_add_dot_plus_eight (reg, reg, labelno));
+ else
+ insn = emit_insn (gen_pic_add_dot_plus_four (reg, reg, labelno));
+ }
+ else /* TARGET_THUMB1 */
+ {
+ emit_insn (gen_pic_load_addr_thumb1 (reg, offset_rtx));
+ insn = emit_insn (gen_pic_add_dot_plus_four (reg, reg, labelno));
+ }
+
+ return insn;
+}
/* Return nonzero if X is valid as an ARM state addressing register. */
static int
if (TARGET_ARM)
emit_insn (gen_tls_load_dot_plus_eight (reg, reg, labelno));
else if (TARGET_THUMB2)
- emit_insn (gen_tls_load_dot_plus_four (reg, reg, labelno));
+ emit_insn (gen_tls_load_dot_plus_four (reg, NULL, reg, labelno));
else
{
emit_insn (gen_pic_add_dot_plus_four (reg, reg, labelno));
enum rtx_code subcode;
rtx operand;
enum rtx_code code = GET_CODE (x);
- int extra_cost;
*total = 0;
switch (code)
/* 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,
}
}
+/* Estimates the size cost of thumb1 instructions.
+ For now most of the code is copied from thumb1_rtx_costs. We need more
+ fine grain tuning when we have more related test cases. */
+static inline int
+thumb1_size_rtx_costs (rtx x, enum rtx_code code, enum rtx_code outer)
+{
+ enum machine_mode mode = GET_MODE (x);
+
+ switch (code)
+ {
+ case ASHIFT:
+ case ASHIFTRT:
+ case LSHIFTRT:
+ case ROTATERT:
+ case PLUS:
+ case MINUS:
+ case COMPARE:
+ case NEG:
+ case NOT:
+ return COSTS_N_INSNS (1);
+
+ case MULT:
+ if (GET_CODE (XEXP (x, 1)) == CONST_INT)
+ {
+ /* Thumb1 mul instruction can't operate on const. We must Load it
+ into a register first. */
+ int const_size = thumb1_size_rtx_costs (XEXP (x, 1), CONST_INT, SET);
+ return COSTS_N_INSNS (1) + const_size;
+ }
+ return COSTS_N_INSNS (1);
+
+ case SET:
+ return (COSTS_N_INSNS (1)
+ + 4 * ((GET_CODE (SET_SRC (x)) == MEM)
+ + GET_CODE (SET_DEST (x)) == MEM));
+
+ case CONST_INT:
+ if (outer == SET)
+ {
+ if ((unsigned HOST_WIDE_INT) INTVAL (x) < 256)
+ return 0;
+ if (thumb_shiftable_const (INTVAL (x)))
+ return COSTS_N_INSNS (2);
+ return COSTS_N_INSNS (3);
+ }
+ else if ((outer == PLUS || outer == COMPARE)
+ && INTVAL (x) < 256 && INTVAL (x) > -256)
+ return 0;
+ else if ((outer == IOR || outer == XOR || outer == AND)
+ && INTVAL (x) < 256 && INTVAL (x) >= -256)
+ return COSTS_N_INSNS (1);
+ else if (outer == AND)
+ {
+ int i;
+ /* This duplicates the tests in the andsi3 expander. */
+ for (i = 9; i <= 31; i++)
+ if ((((HOST_WIDE_INT) 1) << i) - 1 == INTVAL (x)
+ || (((HOST_WIDE_INT) 1) << i) - 1 == ~INTVAL (x))
+ return COSTS_N_INSNS (2);
+ }
+ else if (outer == ASHIFT || outer == ASHIFTRT
+ || outer == LSHIFTRT)
+ return 0;
+ return COSTS_N_INSNS (2);
+
+ case CONST:
+ case CONST_DOUBLE:
+ case LABEL_REF:
+ case SYMBOL_REF:
+ return COSTS_N_INSNS (3);
+
+ case UDIV:
+ case UMOD:
+ case DIV:
+ case MOD:
+ return 100;
+
+ case TRUNCATE:
+ return 99;
+
+ case AND:
+ case XOR:
+ case IOR:
+ /* XXX guess. */
+ return 8;
+
+ case MEM:
+ /* XXX another guess. */
+ /* 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)
+ + ((GET_CODE (x) == SYMBOL_REF && CONSTANT_POOL_ADDRESS_P (x))
+ ? 4 : 0));
+
+ case IF_THEN_ELSE:
+ /* XXX a guess. */
+ if (GET_CODE (XEXP (x, 1)) == PC || GET_CODE (XEXP (x, 2)) == PC)
+ return 14;
+ return 2;
+
+ case ZERO_EXTEND:
+ /* XXX still guessing. */
+ 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));
+
+ default:
+ return 99;
+ }
+
+ default:
+ return 99;
+ }
+}
+
/* RTX costs when optimizing for size. */
static bool
arm_size_rtx_costs (rtx x, enum rtx_code code, enum rtx_code outer_code,
enum machine_mode mode = GET_MODE (x);
if (TARGET_THUMB1)
{
- /* XXX TBD. For now, use the standard costs. */
- *total = thumb1_rtx_costs (x, code, outer_code);
+ *total = thumb1_size_rtx_costs (x, code, outer_code);
return true;
}
return arm_size_rtx_costs (x, (enum rtx_code) code,
(enum rtx_code) outer_code, total);
else
- return all_cores[(int)arm_tune].rtx_costs (x, (enum rtx_code) code,
- (enum rtx_code) outer_code,
- total, speed);
+ return current_tune->rtx_costs (x, (enum rtx_code) code,
+ (enum rtx_code) outer_code,
+ total, speed);
}
/* RTX costs for cores with a slow MUL implementation. Thumb-2 is not
so it can be ignored. */
static bool
-arm_xscale_rtx_costs (rtx x, enum rtx_code code, enum rtx_code outer_code, int *total, bool speed)
+arm_xscale_rtx_costs (rtx x, enum rtx_code code, enum rtx_code outer_code,
+ int *total, bool speed)
{
enum machine_mode mode = GET_MODE (x);
}
}
-/* Must not copy a SET whose source operand is PC-relative. */
+/* Must not copy any rtx that uses a pc-relative address. */
+
+static int
+arm_note_pic_base (rtx *x, void *date ATTRIBUTE_UNUSED)
+{
+ if (GET_CODE (*x) == UNSPEC
+ && XINT (*x, 1) == UNSPEC_PIC_BASE)
+ return 1;
+ return 0;
+}
static bool
arm_cannot_copy_insn_p (rtx insn)
{
- rtx pat = PATTERN (insn);
-
- if (GET_CODE (pat) == SET)
- {
- rtx rhs = SET_SRC (pat);
-
- if (GET_CODE (rhs) == UNSPEC
- && XINT (rhs, 1) == UNSPEC_PIC_BASE)
- return TRUE;
-
- if (GET_CODE (rhs) == MEM
- && GET_CODE (XEXP (rhs, 0)) == UNSPEC
- && XINT (XEXP (rhs, 0), 1) == UNSPEC_PIC_BASE)
- return TRUE;
- }
-
- return FALSE;
+ return for_each_rtx (&PATTERN (insn), arm_note_pic_base, NULL);
}
enum rtx_code
for (reg = LAST_IWMMXT_REGNUM; reg >= FIRST_IWMMXT_REGNUM; reg--)
if (df_regs_ever_live_p (reg) && ! call_used_regs[reg])
{
- insn = gen_rtx_PRE_DEC (V2SImode, stack_pointer_rtx);
+ insn = gen_rtx_PRE_DEC (Pmode, stack_pointer_rtx);
insn = gen_rtx_MEM (V2SImode, insn);
insn = emit_set_insn (insn, gen_rtx_REG (V2SImode, reg));
RTX_FRAME_RELATED_P (insn) = 1;
for (reg = LAST_FPA_REGNUM; reg >= FIRST_FPA_REGNUM; reg--)
if (df_regs_ever_live_p (reg) && !call_used_regs[reg])
{
- insn = gen_rtx_PRE_DEC (XFmode, stack_pointer_rtx);
+ insn = gen_rtx_PRE_DEC (Pmode, stack_pointer_rtx);
insn = gen_rtx_MEM (XFmode, insn);
insn = emit_set_insn (insn, gen_rtx_REG (XFmode, reg));
RTX_FRAME_RELATED_P (insn) = 1;
}
}
+/* Given the stack offsets and register mask in OFFSETS, decide
+ how many additional registers to push instead of subtracting
+ a constant from SP. */
+static int
+thumb1_extra_regs_pushed (arm_stack_offsets *offsets)
+{
+ HOST_WIDE_INT amount = offsets->outgoing_args - offsets->saved_regs;
+ unsigned long live_regs_mask = offsets->saved_regs_mask;
+ /* Extract a mask of the ones we can give to the Thumb's push instruction. */
+ unsigned long l_mask = live_regs_mask & 0x40ff;
+ /* Then count how many other high registers will need to be pushed. */
+ unsigned long high_regs_pushed = bit_count (live_regs_mask & 0x0f00);
+ int n_free;
+
+ /* If the stack frame size is 512 exactly, we can save one load
+ instruction, which should make this a win even when optimizing
+ for speed. */
+ if (!optimize_size && amount != 512)
+ return 0;
+
+ /* Can't do this if there are high registers to push, or if we
+ are not going to do a push at all. */
+ if (high_regs_pushed != 0 || l_mask == 0)
+ return 0;
+
+ /* Don't do this if thumb1_expand_prologue wants to emit instructions
+ between the push and the stack frame allocation. */
+ if ((flag_pic && arm_pic_register != INVALID_REGNUM)
+ || (!frame_pointer_needed && CALLER_INTERWORKING_SLOT_SIZE > 0))
+ return 0;
+
+ for (n_free = 0; n_free < 8 && !(live_regs_mask & 1); live_regs_mask >>= 1)
+ n_free++;
+
+ if (n_free == 0)
+ return 0;
+ gcc_assert (amount / 4 * 4 == amount);
+
+ if (amount >= 512 && (amount - n_free * 4) < 512)
+ return (amount - 508) / 4;
+ if (amount <= n_free * 4)
+ return amount / 4;
+ return 0;
+}
+
/* Generate the rest of a function's prologue. */
void
thumb1_expand_prologue (void)
stack_pointer_rtx);
amount = offsets->outgoing_args - offsets->saved_regs;
+ amount -= 4 * thumb1_extra_regs_pushed (offsets);
if (amount)
{
if (amount < 512)
register. */
else if ((l_mask & 0xff) != 0
|| (high_regs_pushed == 0 && l_mask))
- thumb_pushpop (f, l_mask, 1, &cfa_offset, l_mask);
+ {
+ unsigned long mask = l_mask;
+ mask |= (1 << thumb1_extra_regs_pushed (offsets)) - 1;
+ thumb_pushpop (f, mask, 1, &cfa_offset, mask);
+ }
if (high_regs_pushed)
{
offset = INTVAL (XEXP (e1, 1));
asm_fprintf (asm_out_file, "\t.setfp %r, %r, #%wd\n",
HARD_FRAME_POINTER_REGNUM, reg,
- INTVAL (XEXP (e1, 1)));
+ offset);
}
else if (GET_CODE (e1) == REG)
{
fputc (')', fp);
return TRUE;
}
+ else if (GET_CODE (x) == UNSPEC && XINT (x, 1) == UNSPEC_SYMBOL_OFFSET)
+ {
+ output_addr_const (fp, XVECEXP (x, 0, 0));
+ if (GOT_PCREL)
+ fputs ("+.", fp);
+ fputs ("-(", fp);
+ output_addr_const (fp, XVECEXP (x, 0, 1));
+ fputc (')', fp);
+ return TRUE;
+ }
else if (GET_CODE (x) == CONST_VECTOR)
return arm_emit_vector_const (fp, x);
thumb1_output_casesi (rtx *operands)
{
rtx diff_vec = PATTERN (next_real_insn (operands[0]));
- addr_diff_vec_flags flags;
gcc_assert (GET_CODE (diff_vec) == ADDR_DIFF_VEC);
- flags = ADDR_DIFF_VEC_FLAGS (diff_vec);
-
switch (GET_MODE(diff_vec))
{
case QImode: