blob: 1868c53b05f36669e661c54dff71d193fc38af56 [file] [log] [blame]
/* The Blackfin code generation auxiliary output file.
Copyright (C) 2005-2015 Free Software Foundation, Inc.
Contributed by Analog Devices.
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 3, or (at your
option) any later version.
GCC 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 GCC; see the file COPYING3. If not see
<http://www.gnu.org/licenses/>. */
#include "config.h"
#include "system.h"
#include "coretypes.h"
#include "tm.h"
#include "rtl.h"
#include "regs.h"
#include "hard-reg-set.h"
#include "insn-config.h"
#include "insn-codes.h"
#include "conditions.h"
#include "insn-flags.h"
#include "output.h"
#include "insn-attr.h"
#include "hash-set.h"
#include "machmode.h"
#include "vec.h"
#include "double-int.h"
#include "input.h"
#include "alias.h"
#include "symtab.h"
#include "wide-int.h"
#include "inchash.h"
#include "tree.h"
#include "fold-const.h"
#include "varasm.h"
#include "calls.h"
#include "flags.h"
#include "except.h"
#include "function.h"
#include "target.h"
#include "target-def.h"
#include "hashtab.h"
#include "statistics.h"
#include "real.h"
#include "fixed-value.h"
#include "expmed.h"
#include "dojump.h"
#include "explow.h"
#include "emit-rtl.h"
#include "stmt.h"
#include "expr.h"
#include "diagnostic-core.h"
#include "recog.h"
#include "optabs.h"
#include "ggc.h"
#include "predict.h"
#include "dominance.h"
#include "cfg.h"
#include "cfgrtl.h"
#include "cfganal.h"
#include "lcm.h"
#include "cfgbuild.h"
#include "cfgcleanup.h"
#include "basic-block.h"
#include "hash-map.h"
#include "is-a.h"
#include "plugin-api.h"
#include "ipa-ref.h"
#include "cgraph.h"
#include "langhooks.h"
#include "bfin-protos.h"
#include "tm_p.h"
#include "tm-preds.h"
#include "tm-constrs.h"
#include "gt-bfin.h"
#include "timevar.h"
#include "df.h"
#include "sel-sched.h"
#include "hw-doloop.h"
#include "opts.h"
#include "dumpfile.h"
#include "builtins.h"
/* A C structure for machine-specific, per-function data.
This is added to the cfun structure. */
struct GTY(()) machine_function
{
/* Set if we are notified by the doloop pass that a hardware loop
was created. */
int has_hardware_loops;
/* Set if we create a memcpy pattern that uses loop registers. */
int has_loopreg_clobber;
};
/* RTX for condition code flag register and RETS register */
extern GTY(()) rtx bfin_cc_rtx;
extern GTY(()) rtx bfin_rets_rtx;
rtx bfin_cc_rtx, bfin_rets_rtx;
int max_arg_registers = 0;
/* Arrays used when emitting register names. */
const char *short_reg_names[] = SHORT_REGISTER_NAMES;
const char *high_reg_names[] = HIGH_REGISTER_NAMES;
const char *dregs_pair_names[] = DREGS_PAIR_NAMES;
const char *byte_reg_names[] = BYTE_REGISTER_NAMES;
static int arg_regs[] = FUNCTION_ARG_REGISTERS;
static int ret_regs[] = FUNCTION_RETURN_REGISTERS;
int splitting_for_sched, splitting_loops;
static void
bfin_globalize_label (FILE *stream, const char *name)
{
fputs (".global ", stream);
assemble_name (stream, name);
fputc (';',stream);
fputc ('\n',stream);
}
static void
output_file_start (void)
{
FILE *file = asm_out_file;
int i;
fprintf (file, ".file \"%s\";\n", LOCATION_FILE (input_location));
for (i = 0; arg_regs[i] >= 0; i++)
;
max_arg_registers = i; /* how many arg reg used */
}
/* Examine machine-dependent attributes of function type FUNTYPE and return its
type. See the definition of E_FUNKIND. */
static e_funkind
funkind (const_tree funtype)
{
tree attrs = TYPE_ATTRIBUTES (funtype);
if (lookup_attribute ("interrupt_handler", attrs))
return INTERRUPT_HANDLER;
else if (lookup_attribute ("exception_handler", attrs))
return EXCPT_HANDLER;
else if (lookup_attribute ("nmi_handler", attrs))
return NMI_HANDLER;
else
return SUBROUTINE;
}
/* Legitimize PIC addresses. If the address is already position-independent,
we return ORIG. Newly generated position-independent addresses go into a
reg. This is REG if nonzero, otherwise we allocate register(s) as
necessary. PICREG is the register holding the pointer to the PIC offset
table. */
static rtx
legitimize_pic_address (rtx orig, rtx reg, rtx picreg)
{
rtx addr = orig;
rtx new_rtx = orig;
if (GET_CODE (addr) == SYMBOL_REF || GET_CODE (addr) == LABEL_REF)
{
int unspec;
rtx tmp;
if (TARGET_ID_SHARED_LIBRARY)
unspec = UNSPEC_MOVE_PIC;
else if (GET_CODE (addr) == SYMBOL_REF
&& SYMBOL_REF_FUNCTION_P (addr))
unspec = UNSPEC_FUNCDESC_GOT17M4;
else
unspec = UNSPEC_MOVE_FDPIC;
if (reg == 0)
{
gcc_assert (can_create_pseudo_p ());
reg = gen_reg_rtx (Pmode);
}
tmp = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, addr), unspec);
new_rtx = gen_const_mem (Pmode, gen_rtx_PLUS (Pmode, picreg, tmp));
emit_move_insn (reg, new_rtx);
if (picreg == pic_offset_table_rtx)
crtl->uses_pic_offset_table = 1;
return reg;
}
else if (GET_CODE (addr) == CONST || GET_CODE (addr) == PLUS)
{
rtx base;
if (GET_CODE (addr) == CONST)
{
addr = XEXP (addr, 0);
gcc_assert (GET_CODE (addr) == PLUS);
}
if (XEXP (addr, 0) == picreg)
return orig;
if (reg == 0)
{
gcc_assert (can_create_pseudo_p ());
reg = gen_reg_rtx (Pmode);
}
base = legitimize_pic_address (XEXP (addr, 0), reg, picreg);
addr = legitimize_pic_address (XEXP (addr, 1),
base == reg ? NULL_RTX : reg,
picreg);
if (GET_CODE (addr) == CONST_INT)
{
gcc_assert (! reload_in_progress && ! reload_completed);
addr = force_reg (Pmode, addr);
}
if (GET_CODE (addr) == PLUS && CONSTANT_P (XEXP (addr, 1)))
{
base = gen_rtx_PLUS (Pmode, base, XEXP (addr, 0));
addr = XEXP (addr, 1);
}
return gen_rtx_PLUS (Pmode, base, addr);
}
return new_rtx;
}
/* Stack frame layout. */
/* For a given REGNO, determine whether it must be saved in the function
prologue. IS_INTHANDLER specifies whether we're generating a normal
prologue or an interrupt/exception one. */
static bool
must_save_p (bool is_inthandler, unsigned regno)
{
if (D_REGNO_P (regno))
{
bool is_eh_return_reg = false;
if (crtl->calls_eh_return)
{
unsigned j;
for (j = 0; ; j++)
{
unsigned test = EH_RETURN_DATA_REGNO (j);
if (test == INVALID_REGNUM)
break;
if (test == regno)
is_eh_return_reg = true;
}
}
return (is_eh_return_reg
|| (df_regs_ever_live_p (regno)
&& !fixed_regs[regno]
&& (is_inthandler || !call_used_regs[regno])));
}
else if (P_REGNO_P (regno))
{
return ((df_regs_ever_live_p (regno)
&& !fixed_regs[regno]
&& (is_inthandler || !call_used_regs[regno]))
|| (is_inthandler
&& (ENABLE_WA_05000283 || ENABLE_WA_05000315)
&& regno == REG_P5)
|| (!TARGET_FDPIC
&& regno == PIC_OFFSET_TABLE_REGNUM
&& (crtl->uses_pic_offset_table
|| (TARGET_ID_SHARED_LIBRARY && !crtl->is_leaf))));
}
else
return ((is_inthandler || !call_used_regs[regno])
&& (df_regs_ever_live_p (regno)
|| (!leaf_function_p () && call_used_regs[regno])));
}
/* Compute the number of DREGS to save with a push_multiple operation.
This could include registers that aren't modified in the function,
since push_multiple only takes a range of registers.
If IS_INTHANDLER, then everything that is live must be saved, even
if normally call-clobbered.
If CONSECUTIVE, return the number of registers we can save in one
instruction with a push/pop multiple instruction. */
static int
n_dregs_to_save (bool is_inthandler, bool consecutive)
{
int count = 0;
unsigned i;
for (i = REG_R7 + 1; i-- != REG_R0;)
{
if (must_save_p (is_inthandler, i))
count++;
else if (consecutive)
return count;
}
return count;
}
/* Like n_dregs_to_save, but compute number of PREGS to save. */
static int
n_pregs_to_save (bool is_inthandler, bool consecutive)
{
int count = 0;
unsigned i;
for (i = REG_P5 + 1; i-- != REG_P0;)
if (must_save_p (is_inthandler, i))
count++;
else if (consecutive)
return count;
return count;
}
/* Determine if we are going to save the frame pointer in the prologue. */
static bool
must_save_fp_p (void)
{
return df_regs_ever_live_p (REG_FP);
}
/* Determine if we are going to save the RETS register. */
static bool
must_save_rets_p (void)
{
return df_regs_ever_live_p (REG_RETS);
}
static bool
stack_frame_needed_p (void)
{
/* EH return puts a new return address into the frame using an
address relative to the frame pointer. */
if (crtl->calls_eh_return)
return true;
return frame_pointer_needed;
}
/* Emit code to save registers in the prologue. SAVEALL is nonzero if we
must save all registers; this is used for interrupt handlers.
SPREG contains (reg:SI REG_SP). IS_INTHANDLER is true if we're doing
this for an interrupt (or exception) handler. */
static void
expand_prologue_reg_save (rtx spreg, int saveall, bool is_inthandler)
{
rtx predec1 = gen_rtx_PRE_DEC (SImode, spreg);
rtx predec = gen_rtx_MEM (SImode, predec1);
int ndregs = saveall ? 8 : n_dregs_to_save (is_inthandler, false);
int npregs = saveall ? 6 : n_pregs_to_save (is_inthandler, false);
int ndregs_consec = saveall ? 8 : n_dregs_to_save (is_inthandler, true);
int npregs_consec = saveall ? 6 : n_pregs_to_save (is_inthandler, true);
int dregno, pregno;
int total_consec = ndregs_consec + npregs_consec;
int i, d_to_save;
if (saveall || is_inthandler)
{
rtx_insn *insn = emit_move_insn (predec, gen_rtx_REG (SImode, REG_ASTAT));
RTX_FRAME_RELATED_P (insn) = 1;
for (dregno = REG_LT0; dregno <= REG_LB1; dregno++)
if (! crtl->is_leaf
|| cfun->machine->has_hardware_loops
|| cfun->machine->has_loopreg_clobber
|| (ENABLE_WA_05000257
&& (dregno == REG_LC0 || dregno == REG_LC1)))
{
insn = emit_move_insn (predec, gen_rtx_REG (SImode, dregno));
RTX_FRAME_RELATED_P (insn) = 1;
}
}
if (total_consec != 0)
{
rtx_insn *insn;
rtx val = GEN_INT (-total_consec * 4);
rtx pat = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (total_consec + 2));
XVECEXP (pat, 0, 0) = gen_rtx_UNSPEC (VOIDmode, gen_rtvec (1, val),
UNSPEC_PUSH_MULTIPLE);
XVECEXP (pat, 0, total_consec + 1) = gen_rtx_SET (VOIDmode, spreg,
gen_rtx_PLUS (Pmode,
spreg,
val));
RTX_FRAME_RELATED_P (XVECEXP (pat, 0, total_consec + 1)) = 1;
d_to_save = ndregs_consec;
dregno = REG_R7 + 1 - ndregs_consec;
pregno = REG_P5 + 1 - npregs_consec;
for (i = 0; i < total_consec; i++)
{
rtx memref = gen_rtx_MEM (word_mode,
gen_rtx_PLUS (Pmode, spreg,
GEN_INT (- i * 4 - 4)));
rtx subpat;
if (d_to_save > 0)
{
subpat = gen_rtx_SET (VOIDmode, memref, gen_rtx_REG (word_mode,
dregno++));
d_to_save--;
}
else
{
subpat = gen_rtx_SET (VOIDmode, memref, gen_rtx_REG (word_mode,
pregno++));
}
XVECEXP (pat, 0, i + 1) = subpat;
RTX_FRAME_RELATED_P (subpat) = 1;
}
insn = emit_insn (pat);
RTX_FRAME_RELATED_P (insn) = 1;
}
for (dregno = REG_R0; ndregs != ndregs_consec; dregno++)
{
if (must_save_p (is_inthandler, dregno))
{
rtx_insn *insn =
emit_move_insn (predec, gen_rtx_REG (word_mode, dregno));
RTX_FRAME_RELATED_P (insn) = 1;
ndregs--;
}
}
for (pregno = REG_P0; npregs != npregs_consec; pregno++)
{
if (must_save_p (is_inthandler, pregno))
{
rtx_insn *insn =
emit_move_insn (predec, gen_rtx_REG (word_mode, pregno));
RTX_FRAME_RELATED_P (insn) = 1;
npregs--;
}
}
for (i = REG_P7 + 1; i < REG_CC; i++)
if (saveall
|| (is_inthandler
&& (df_regs_ever_live_p (i)
|| (!leaf_function_p () && call_used_regs[i]))))
{
rtx_insn *insn;
if (i == REG_A0 || i == REG_A1)
insn = emit_move_insn (gen_rtx_MEM (PDImode, predec1),
gen_rtx_REG (PDImode, i));
else
insn = emit_move_insn (predec, gen_rtx_REG (SImode, i));
RTX_FRAME_RELATED_P (insn) = 1;
}
}
/* Emit code to restore registers in the epilogue. SAVEALL is nonzero if we
must save all registers; this is used for interrupt handlers.
SPREG contains (reg:SI REG_SP). IS_INTHANDLER is true if we're doing
this for an interrupt (or exception) handler. */
static void
expand_epilogue_reg_restore (rtx spreg, bool saveall, bool is_inthandler)
{
rtx postinc1 = gen_rtx_POST_INC (SImode, spreg);
rtx postinc = gen_rtx_MEM (SImode, postinc1);
int ndregs = saveall ? 8 : n_dregs_to_save (is_inthandler, false);
int npregs = saveall ? 6 : n_pregs_to_save (is_inthandler, false);
int ndregs_consec = saveall ? 8 : n_dregs_to_save (is_inthandler, true);
int npregs_consec = saveall ? 6 : n_pregs_to_save (is_inthandler, true);
int total_consec = ndregs_consec + npregs_consec;
int i, regno;
rtx_insn *insn;
/* A slightly crude technique to stop flow from trying to delete "dead"
insns. */
MEM_VOLATILE_P (postinc) = 1;
for (i = REG_CC - 1; i > REG_P7; i--)
if (saveall
|| (is_inthandler
&& (df_regs_ever_live_p (i)
|| (!leaf_function_p () && call_used_regs[i]))))
{
if (i == REG_A0 || i == REG_A1)
{
rtx mem = gen_rtx_MEM (PDImode, postinc1);
MEM_VOLATILE_P (mem) = 1;
emit_move_insn (gen_rtx_REG (PDImode, i), mem);
}
else
emit_move_insn (gen_rtx_REG (SImode, i), postinc);
}
regno = REG_P5 - npregs_consec;
for (; npregs != npregs_consec; regno--)
{
if (must_save_p (is_inthandler, regno))
{
emit_move_insn (gen_rtx_REG (word_mode, regno), postinc);
npregs--;
}
}
regno = REG_R7 - ndregs_consec;
for (; ndregs != ndregs_consec; regno--)
{
if (must_save_p (is_inthandler, regno))
{
emit_move_insn (gen_rtx_REG (word_mode, regno), postinc);
ndregs--;
}
}
if (total_consec != 0)
{
rtx pat = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (total_consec + 1));
XVECEXP (pat, 0, 0)
= gen_rtx_SET (VOIDmode, spreg,
gen_rtx_PLUS (Pmode, spreg,
GEN_INT (total_consec * 4)));
if (npregs_consec > 0)
regno = REG_P5 + 1;
else
regno = REG_R7 + 1;
for (i = 0; i < total_consec; i++)
{
rtx addr = (i > 0
? gen_rtx_PLUS (Pmode, spreg, GEN_INT (i * 4))
: spreg);
rtx memref = gen_rtx_MEM (word_mode, addr);
regno--;
XVECEXP (pat, 0, i + 1)
= gen_rtx_SET (VOIDmode, gen_rtx_REG (word_mode, regno), memref);
if (npregs_consec > 0)
{
if (--npregs_consec == 0)
regno = REG_R7 + 1;
}
}
insn = emit_insn (pat);
RTX_FRAME_RELATED_P (insn) = 1;
}
if (saveall || is_inthandler)
{
for (regno = REG_LB1; regno >= REG_LT0; regno--)
if (! crtl->is_leaf
|| cfun->machine->has_hardware_loops
|| cfun->machine->has_loopreg_clobber
|| (ENABLE_WA_05000257 && (regno == REG_LC0 || regno == REG_LC1)))
emit_move_insn (gen_rtx_REG (SImode, regno), postinc);
emit_move_insn (gen_rtx_REG (SImode, REG_ASTAT), postinc);
}
}
/* Perform any needed actions needed for a function that is receiving a
variable number of arguments.
CUM is as above.
MODE and TYPE are the mode and type of the current parameter.
PRETEND_SIZE is a variable that should be set to the amount of stack
that must be pushed by the prolog to pretend that our caller pushed
it.
Normally, this macro will push all remaining incoming registers on the
stack and set PRETEND_SIZE to the length of the registers pushed.
Blackfin specific :
- VDSP C compiler manual (our ABI) says that a variable args function
should save the R0, R1 and R2 registers in the stack.
- The caller will always leave space on the stack for the
arguments that are passed in registers, so we dont have
to leave any extra space.
- now, the vastart pointer can access all arguments from the stack. */
static void
setup_incoming_varargs (cumulative_args_t cum,
machine_mode mode ATTRIBUTE_UNUSED,
tree type ATTRIBUTE_UNUSED, int *pretend_size,
int no_rtl)
{
rtx mem;
int i;
if (no_rtl)
return;
/* The move for named arguments will be generated automatically by the
compiler. We need to generate the move rtx for the unnamed arguments
if they are in the first 3 words. We assume at least 1 named argument
exists, so we never generate [ARGP] = R0 here. */
for (i = get_cumulative_args (cum)->words + 1; i < max_arg_registers; i++)
{
mem = gen_rtx_MEM (Pmode,
plus_constant (Pmode, arg_pointer_rtx,
(i * UNITS_PER_WORD)));
emit_move_insn (mem, gen_rtx_REG (Pmode, i));
}
*pretend_size = 0;
}
/* Value should be nonzero if functions must have frame pointers.
Zero means the frame pointer need not be set up (and parms may
be accessed via the stack pointer) in functions that seem suitable. */
static bool
bfin_frame_pointer_required (void)
{
e_funkind fkind = funkind (TREE_TYPE (current_function_decl));
if (fkind != SUBROUTINE)
return true;
/* We turn on -fomit-frame-pointer if -momit-leaf-frame-pointer is used,
so we have to override it for non-leaf functions. */
if (TARGET_OMIT_LEAF_FRAME_POINTER && ! crtl->is_leaf)
return true;
return false;
}
/* Return the number of registers pushed during the prologue. */
static int
n_regs_saved_by_prologue (void)
{
e_funkind fkind = funkind (TREE_TYPE (current_function_decl));
bool is_inthandler = fkind != SUBROUTINE;
tree attrs = TYPE_ATTRIBUTES (TREE_TYPE (current_function_decl));
bool all = (lookup_attribute ("saveall", attrs) != NULL_TREE
|| (is_inthandler && !crtl->is_leaf));
int ndregs = all ? 8 : n_dregs_to_save (is_inthandler, false);
int npregs = all ? 6 : n_pregs_to_save (is_inthandler, false);
int n = ndregs + npregs;
int i;
if (all || stack_frame_needed_p ())
n += 2;
else
{
if (must_save_fp_p ())
n++;
if (must_save_rets_p ())
n++;
}
if (fkind != SUBROUTINE || all)
{
/* Increment once for ASTAT. */
n++;
if (! crtl->is_leaf
|| cfun->machine->has_hardware_loops
|| cfun->machine->has_loopreg_clobber)
{
n += 6;
}
}
if (fkind != SUBROUTINE)
{
/* RETE/X/N. */
if (lookup_attribute ("nesting", attrs))
n++;
}
for (i = REG_P7 + 1; i < REG_CC; i++)
if (all
|| (fkind != SUBROUTINE
&& (df_regs_ever_live_p (i)
|| (!leaf_function_p () && call_used_regs[i]))))
n += i == REG_A0 || i == REG_A1 ? 2 : 1;
return n;
}
/* Given FROM and TO register numbers, say whether this elimination is
allowed. Frame pointer elimination is automatically handled.
All other eliminations are valid. */
static bool
bfin_can_eliminate (const int from ATTRIBUTE_UNUSED, const int to)
{
return (to == STACK_POINTER_REGNUM ? ! frame_pointer_needed : true);
}
/* Return the offset between two registers, one to be eliminated, and the other
its replacement, at the start of a routine. */
HOST_WIDE_INT
bfin_initial_elimination_offset (int from, int to)
{
HOST_WIDE_INT offset = 0;
if (from == ARG_POINTER_REGNUM)
offset = n_regs_saved_by_prologue () * 4;
if (to == STACK_POINTER_REGNUM)
{
if (crtl->outgoing_args_size >= FIXED_STACK_AREA)
offset += crtl->outgoing_args_size;
else if (crtl->outgoing_args_size)
offset += FIXED_STACK_AREA;
offset += get_frame_size ();
}
return offset;
}
/* Emit code to load a constant CONSTANT into register REG; setting
RTX_FRAME_RELATED_P on all insns we generate if RELATED is true.
Make sure that the insns we generate need not be split. */
static void
frame_related_constant_load (rtx reg, HOST_WIDE_INT constant, bool related)
{
rtx_insn *insn;
rtx cst = GEN_INT (constant);
if (constant >= -32768 && constant < 65536)
insn = emit_move_insn (reg, cst);
else
{
/* We don't call split_load_immediate here, since dwarf2out.c can get
confused about some of the more clever sequences it can generate. */
insn = emit_insn (gen_movsi_high (reg, cst));
if (related)
RTX_FRAME_RELATED_P (insn) = 1;
insn = emit_insn (gen_movsi_low (reg, reg, cst));
}
if (related)
RTX_FRAME_RELATED_P (insn) = 1;
}
/* Generate efficient code to add a value to a P register.
Set RTX_FRAME_RELATED_P on the generated insns if FRAME is nonzero.
EPILOGUE_P is zero if this function is called for prologue,
otherwise it's nonzero. And it's less than zero if this is for
sibcall epilogue. */
static void
add_to_reg (rtx reg, HOST_WIDE_INT value, int frame, int epilogue_p)
{
if (value == 0)
return;
/* Choose whether to use a sequence using a temporary register, or
a sequence with multiple adds. We can add a signed 7-bit value
in one instruction. */
if (value > 120 || value < -120)
{
rtx tmpreg;
rtx tmpreg2;
rtx_insn *insn;
tmpreg2 = NULL_RTX;
/* For prologue or normal epilogue, P1 can be safely used
as the temporary register. For sibcall epilogue, we try to find
a call used P register, which will be restored in epilogue.
If we cannot find such a P register, we have to use one I register
to help us. */
if (epilogue_p >= 0)
tmpreg = gen_rtx_REG (SImode, REG_P1);
else
{
int i;
for (i = REG_P0; i <= REG_P5; i++)
if ((df_regs_ever_live_p (i) && ! call_used_regs[i])
|| (!TARGET_FDPIC
&& i == PIC_OFFSET_TABLE_REGNUM
&& (crtl->uses_pic_offset_table
|| (TARGET_ID_SHARED_LIBRARY
&& ! crtl->is_leaf))))
break;
if (i <= REG_P5)
tmpreg = gen_rtx_REG (SImode, i);
else
{
tmpreg = gen_rtx_REG (SImode, REG_P1);
tmpreg2 = gen_rtx_REG (SImode, REG_I0);
emit_move_insn (tmpreg2, tmpreg);
}
}
if (frame)
frame_related_constant_load (tmpreg, value, TRUE);
else
insn = emit_move_insn (tmpreg, GEN_INT (value));
insn = emit_insn (gen_addsi3 (reg, reg, tmpreg));
if (frame)
RTX_FRAME_RELATED_P (insn) = 1;
if (tmpreg2 != NULL_RTX)
emit_move_insn (tmpreg, tmpreg2);
}
else
do
{
int size = value;
rtx_insn *insn;
if (size > 60)
size = 60;
else if (size < -60)
/* We could use -62, but that would leave the stack unaligned, so
it's no good. */
size = -60;
insn = emit_insn (gen_addsi3 (reg, reg, GEN_INT (size)));
if (frame)
RTX_FRAME_RELATED_P (insn) = 1;
value -= size;
}
while (value != 0);
}
/* Generate a LINK insn for a frame sized FRAME_SIZE. If this constant
is too large, generate a sequence of insns that has the same effect.
SPREG contains (reg:SI REG_SP). */
static void
emit_link_insn (rtx spreg, HOST_WIDE_INT frame_size)
{
HOST_WIDE_INT link_size = frame_size;
rtx_insn *insn;
int i;
if (link_size > 262140)
link_size = 262140;
/* Use a LINK insn with as big a constant as possible, then subtract
any remaining size from the SP. */
insn = emit_insn (gen_link (GEN_INT (-8 - link_size)));
RTX_FRAME_RELATED_P (insn) = 1;
for (i = 0; i < XVECLEN (PATTERN (insn), 0); i++)
{
rtx set = XVECEXP (PATTERN (insn), 0, i);
gcc_assert (GET_CODE (set) == SET);
RTX_FRAME_RELATED_P (set) = 1;
}
frame_size -= link_size;
if (frame_size > 0)
{
/* Must use a call-clobbered PREG that isn't the static chain. */
rtx tmpreg = gen_rtx_REG (Pmode, REG_P1);
frame_related_constant_load (tmpreg, -frame_size, TRUE);
insn = emit_insn (gen_addsi3 (spreg, spreg, tmpreg));
RTX_FRAME_RELATED_P (insn) = 1;
}
}
/* Return the number of bytes we must reserve for outgoing arguments
in the current function's stack frame. */
static HOST_WIDE_INT
arg_area_size (void)
{
if (crtl->outgoing_args_size)
{
if (crtl->outgoing_args_size >= FIXED_STACK_AREA)
return crtl->outgoing_args_size;
else
return FIXED_STACK_AREA;
}
return 0;
}
/* Save RETS and FP, and allocate a stack frame. ALL is true if the
function must save all its registers (true only for certain interrupt
handlers). */
static void
do_link (rtx spreg, HOST_WIDE_INT frame_size, bool all)
{
frame_size += arg_area_size ();
if (all
|| stack_frame_needed_p ()
|| (must_save_rets_p () && must_save_fp_p ()))
emit_link_insn (spreg, frame_size);
else
{
if (must_save_rets_p ())
{
rtx pat = gen_movsi (gen_rtx_MEM (Pmode,
gen_rtx_PRE_DEC (Pmode, spreg)),
bfin_rets_rtx);
rtx_insn *insn = emit_insn (pat);
RTX_FRAME_RELATED_P (insn) = 1;
}
if (must_save_fp_p ())
{
rtx pat = gen_movsi (gen_rtx_MEM (Pmode,
gen_rtx_PRE_DEC (Pmode, spreg)),
gen_rtx_REG (Pmode, REG_FP));
rtx_insn *insn = emit_insn (pat);
RTX_FRAME_RELATED_P (insn) = 1;
}
add_to_reg (spreg, -frame_size, 1, 0);
}
}
/* Like do_link, but used for epilogues to deallocate the stack frame.
EPILOGUE_P is zero if this function is called for prologue,
otherwise it's nonzero. And it's less than zero if this is for
sibcall epilogue. */
static void
do_unlink (rtx spreg, HOST_WIDE_INT frame_size, bool all, int epilogue_p)
{
frame_size += arg_area_size ();
if (stack_frame_needed_p ())
emit_insn (gen_unlink ());
else
{
rtx postinc = gen_rtx_MEM (Pmode, gen_rtx_POST_INC (Pmode, spreg));
add_to_reg (spreg, frame_size, 0, epilogue_p);
if (all || must_save_fp_p ())
{
rtx fpreg = gen_rtx_REG (Pmode, REG_FP);
emit_move_insn (fpreg, postinc);
emit_use (fpreg);
}
if (all || must_save_rets_p ())
{
emit_move_insn (bfin_rets_rtx, postinc);
emit_use (bfin_rets_rtx);
}
}
}
/* Generate a prologue suitable for a function of kind FKIND. This is
called for interrupt and exception handler prologues.
SPREG contains (reg:SI REG_SP). */
static void
expand_interrupt_handler_prologue (rtx spreg, e_funkind fkind, bool all)
{
HOST_WIDE_INT frame_size = get_frame_size ();
rtx predec1 = gen_rtx_PRE_DEC (SImode, spreg);
rtx predec = gen_rtx_MEM (SImode, predec1);
rtx_insn *insn;
tree attrs = TYPE_ATTRIBUTES (TREE_TYPE (current_function_decl));
tree kspisusp = lookup_attribute ("kspisusp", attrs);
if (kspisusp)
{
insn = emit_move_insn (spreg, gen_rtx_REG (Pmode, REG_USP));
RTX_FRAME_RELATED_P (insn) = 1;
}
/* We need space on the stack in case we need to save the argument
registers. */
if (fkind == EXCPT_HANDLER)
{
insn = emit_insn (gen_addsi3 (spreg, spreg, GEN_INT (-12)));
RTX_FRAME_RELATED_P (insn) = 1;
}
/* If we're calling other functions, they won't save their call-clobbered
registers, so we must save everything here. */
if (!crtl->is_leaf)
all = true;
expand_prologue_reg_save (spreg, all, true);
if (ENABLE_WA_05000283 || ENABLE_WA_05000315)
{
rtx chipid = GEN_INT (trunc_int_for_mode (0xFFC00014, SImode));
rtx p5reg = gen_rtx_REG (Pmode, REG_P5);
emit_insn (gen_movbi (bfin_cc_rtx, const1_rtx));
emit_insn (gen_movsi_high (p5reg, chipid));
emit_insn (gen_movsi_low (p5reg, p5reg, chipid));
emit_insn (gen_dummy_load (p5reg, bfin_cc_rtx));
}
if (lookup_attribute ("nesting", attrs))
{
rtx srcreg = gen_rtx_REG (Pmode, ret_regs[fkind]);
insn = emit_move_insn (predec, srcreg);
RTX_FRAME_RELATED_P (insn) = 1;
}
do_link (spreg, frame_size, all);
if (fkind == EXCPT_HANDLER)
{
rtx r0reg = gen_rtx_REG (SImode, REG_R0);
rtx r1reg = gen_rtx_REG (SImode, REG_R1);
rtx r2reg = gen_rtx_REG (SImode, REG_R2);
emit_move_insn (r0reg, gen_rtx_REG (SImode, REG_SEQSTAT));
emit_insn (gen_ashrsi3 (r0reg, r0reg, GEN_INT (26)));
emit_insn (gen_ashlsi3 (r0reg, r0reg, GEN_INT (26)));
emit_move_insn (r1reg, spreg);
emit_move_insn (r2reg, gen_rtx_REG (Pmode, REG_FP));
emit_insn (gen_addsi3 (r2reg, r2reg, GEN_INT (8)));
}
}
/* Generate an epilogue suitable for a function of kind FKIND. This is
called for interrupt and exception handler epilogues.
SPREG contains (reg:SI REG_SP). */
static void
expand_interrupt_handler_epilogue (rtx spreg, e_funkind fkind, bool all)
{
tree attrs = TYPE_ATTRIBUTES (TREE_TYPE (current_function_decl));
rtx postinc1 = gen_rtx_POST_INC (SImode, spreg);
rtx postinc = gen_rtx_MEM (SImode, postinc1);
/* A slightly crude technique to stop flow from trying to delete "dead"
insns. */
MEM_VOLATILE_P (postinc) = 1;
do_unlink (spreg, get_frame_size (), all, 1);
if (lookup_attribute ("nesting", attrs))
{
rtx srcreg = gen_rtx_REG (Pmode, ret_regs[fkind]);
emit_move_insn (srcreg, postinc);
}
/* If we're calling other functions, they won't save their call-clobbered
registers, so we must save (and restore) everything here. */
if (!crtl->is_leaf)
all = true;
expand_epilogue_reg_restore (spreg, all, true);
/* Deallocate any space we left on the stack in case we needed to save the
argument registers. */
if (fkind == EXCPT_HANDLER)
emit_insn (gen_addsi3 (spreg, spreg, GEN_INT (12)));
emit_jump_insn (gen_return_internal (gen_rtx_REG (Pmode, ret_regs[fkind])));
}
/* Used while emitting the prologue to generate code to load the correct value
into the PIC register, which is passed in DEST. */
static rtx
bfin_load_pic_reg (rtx dest)
{
struct cgraph_local_info *i = NULL;
rtx addr;
i = cgraph_node::local_info (current_function_decl);
/* Functions local to the translation unit don't need to reload the
pic reg, since the caller always passes a usable one. */
if (i && i->local)
return pic_offset_table_rtx;
if (global_options_set.x_bfin_library_id)
addr = plus_constant (Pmode, pic_offset_table_rtx,
-4 - bfin_library_id * 4);
else
addr = gen_rtx_PLUS (Pmode, pic_offset_table_rtx,
gen_rtx_UNSPEC (Pmode, gen_rtvec (1, const0_rtx),
UNSPEC_LIBRARY_OFFSET));
emit_insn (gen_movsi (dest, gen_rtx_MEM (Pmode, addr)));
return dest;
}
/* Generate RTL for the prologue of the current function. */
void
bfin_expand_prologue (void)
{
HOST_WIDE_INT frame_size = get_frame_size ();
rtx spreg = gen_rtx_REG (Pmode, REG_SP);
e_funkind fkind = funkind (TREE_TYPE (current_function_decl));
rtx pic_reg_loaded = NULL_RTX;
tree attrs = TYPE_ATTRIBUTES (TREE_TYPE (current_function_decl));
bool all = lookup_attribute ("saveall", attrs) != NULL_TREE;
if (fkind != SUBROUTINE)
{
expand_interrupt_handler_prologue (spreg, fkind, all);
return;
}
if (crtl->limit_stack
|| (TARGET_STACK_CHECK_L1
&& !DECL_NO_LIMIT_STACK (current_function_decl)))
{
HOST_WIDE_INT offset
= bfin_initial_elimination_offset (ARG_POINTER_REGNUM,
STACK_POINTER_REGNUM);
rtx lim = crtl->limit_stack ? stack_limit_rtx : NULL_RTX;
rtx tmp = gen_rtx_REG (Pmode, REG_R3);
rtx p2reg = gen_rtx_REG (Pmode, REG_P2);
emit_move_insn (tmp, p2reg);
if (!lim)
{
emit_move_insn (p2reg, gen_int_mode (0xFFB00000, SImode));
emit_move_insn (p2reg, gen_rtx_MEM (Pmode, p2reg));
lim = p2reg;
}
if (GET_CODE (lim) == SYMBOL_REF)
{
if (TARGET_ID_SHARED_LIBRARY)
{
rtx p1reg = gen_rtx_REG (Pmode, REG_P1);
rtx val;
pic_reg_loaded = bfin_load_pic_reg (p2reg);
val = legitimize_pic_address (stack_limit_rtx, p1reg,
pic_reg_loaded);
emit_move_insn (p1reg, val);
frame_related_constant_load (p2reg, offset, FALSE);
emit_insn (gen_addsi3 (p2reg, p2reg, p1reg));
lim = p2reg;
}
else
{
rtx limit = plus_constant (Pmode, lim, offset);
emit_move_insn (p2reg, limit);
lim = p2reg;
}
}
else
{
if (lim != p2reg)
emit_move_insn (p2reg, lim);
add_to_reg (p2reg, offset, 0, 0);
lim = p2reg;
}
emit_insn (gen_compare_lt (bfin_cc_rtx, spreg, lim));
emit_insn (gen_trapifcc ());
emit_move_insn (p2reg, tmp);
}
expand_prologue_reg_save (spreg, all, false);
do_link (spreg, frame_size, all);
if (TARGET_ID_SHARED_LIBRARY
&& !TARGET_SEP_DATA
&& (crtl->uses_pic_offset_table
|| !crtl->is_leaf))
bfin_load_pic_reg (pic_offset_table_rtx);
}
/* Generate RTL for the epilogue of the current function. NEED_RETURN is zero
if this is for a sibcall. EH_RETURN is nonzero if we're expanding an
eh_return pattern. SIBCALL_P is true if this is a sibcall epilogue,
false otherwise. */
void
bfin_expand_epilogue (int need_return, int eh_return, bool sibcall_p)
{
rtx spreg = gen_rtx_REG (Pmode, REG_SP);
e_funkind fkind = funkind (TREE_TYPE (current_function_decl));
int e = sibcall_p ? -1 : 1;
tree attrs = TYPE_ATTRIBUTES (TREE_TYPE (current_function_decl));
bool all = lookup_attribute ("saveall", attrs) != NULL_TREE;
if (fkind != SUBROUTINE)
{
expand_interrupt_handler_epilogue (spreg, fkind, all);
return;
}
do_unlink (spreg, get_frame_size (), all, e);
expand_epilogue_reg_restore (spreg, all, false);
/* Omit the return insn if this is for a sibcall. */
if (! need_return)
return;
if (eh_return)
emit_insn (gen_addsi3 (spreg, spreg, gen_rtx_REG (Pmode, REG_P2)));
emit_jump_insn (gen_return_internal (gen_rtx_REG (Pmode, REG_RETS)));
}
/* Return nonzero if register OLD_REG can be renamed to register NEW_REG. */
int
bfin_hard_regno_rename_ok (unsigned int old_reg ATTRIBUTE_UNUSED,
unsigned int new_reg)
{
/* Interrupt functions can only use registers that have already been
saved by the prologue, even if they would normally be
call-clobbered. */
if (funkind (TREE_TYPE (current_function_decl)) != SUBROUTINE
&& !df_regs_ever_live_p (new_reg))
return 0;
return 1;
}
/* Implement TARGET_EXTRA_LIVE_ON_ENTRY. */
static void
bfin_extra_live_on_entry (bitmap regs)
{
if (TARGET_FDPIC)
bitmap_set_bit (regs, FDPIC_REGNO);
}
/* Return the value of the return address for the frame COUNT steps up
from the current frame, after the prologue.
We punt for everything but the current frame by returning const0_rtx. */
rtx
bfin_return_addr_rtx (int count)
{
if (count != 0)
return const0_rtx;
return get_hard_reg_initial_val (Pmode, REG_RETS);
}
static rtx
bfin_delegitimize_address (rtx orig_x)
{
rtx x = orig_x;
if (GET_CODE (x) != MEM)
return orig_x;
x = XEXP (x, 0);
if (GET_CODE (x) == PLUS
&& GET_CODE (XEXP (x, 1)) == UNSPEC
&& XINT (XEXP (x, 1), 1) == UNSPEC_MOVE_PIC
&& GET_CODE (XEXP (x, 0)) == REG
&& REGNO (XEXP (x, 0)) == PIC_OFFSET_TABLE_REGNUM)
return XVECEXP (XEXP (x, 1), 0, 0);
return orig_x;
}
/* This predicate is used to compute the length of a load/store insn.
OP is a MEM rtx, we return nonzero if its addressing mode requires a
32-bit instruction. */
int
effective_address_32bit_p (rtx op, machine_mode mode)
{
HOST_WIDE_INT offset;
mode = GET_MODE (op);
op = XEXP (op, 0);
if (GET_CODE (op) != PLUS)
{
gcc_assert (REG_P (op) || GET_CODE (op) == POST_INC
|| GET_CODE (op) == PRE_DEC || GET_CODE (op) == POST_DEC);
return 0;
}
if (GET_CODE (XEXP (op, 1)) == UNSPEC)
return 1;
offset = INTVAL (XEXP (op, 1));
/* All byte loads use a 16-bit offset. */
if (GET_MODE_SIZE (mode) == 1)
return 1;
if (GET_MODE_SIZE (mode) == 4)
{
/* Frame pointer relative loads can use a negative offset, all others
are restricted to a small positive one. */
if (XEXP (op, 0) == frame_pointer_rtx)
return offset < -128 || offset > 60;
return offset < 0 || offset > 60;
}
/* Must be HImode now. */
return offset < 0 || offset > 30;
}
/* Returns true if X is a memory reference using an I register. */
bool
bfin_dsp_memref_p (rtx x)
{
if (! MEM_P (x))
return false;
x = XEXP (x, 0);
if (GET_CODE (x) == POST_INC || GET_CODE (x) == PRE_INC
|| GET_CODE (x) == POST_DEC || GET_CODE (x) == PRE_DEC)
x = XEXP (x, 0);
return IREG_P (x);
}
/* Return cost of the memory address ADDR.
All addressing modes are equally cheap on the Blackfin. */
static int
bfin_address_cost (rtx addr ATTRIBUTE_UNUSED,
machine_mode mode ATTRIBUTE_UNUSED,
addr_space_t as ATTRIBUTE_UNUSED,
bool speed ATTRIBUTE_UNUSED)
{
return 1;
}
/* Subroutine of print_operand; used to print a memory reference X to FILE. */
void
print_address_operand (FILE *file, rtx x)
{
switch (GET_CODE (x))
{
case PLUS:
output_address (XEXP (x, 0));
fprintf (file, "+");
output_address (XEXP (x, 1));
break;
case PRE_DEC:
fprintf (file, "--");
output_address (XEXP (x, 0));
break;
case POST_INC:
output_address (XEXP (x, 0));
fprintf (file, "++");
break;
case POST_DEC:
output_address (XEXP (x, 0));
fprintf (file, "--");
break;
default:
gcc_assert (GET_CODE (x) != MEM);
print_operand (file, x, 0);
break;
}
}
/* Adding intp DImode support by Tony
* -- Q: (low word)
* -- R: (high word)
*/
void
print_operand (FILE *file, rtx x, char code)
{
machine_mode mode;
if (code == '!')
{
if (GET_MODE (current_output_insn) == SImode)
fprintf (file, " ||");
else
fprintf (file, ";");
return;
}
mode = GET_MODE (x);
switch (code)
{
case 'j':
switch (GET_CODE (x))
{
case EQ:
fprintf (file, "e");
break;
case NE:
fprintf (file, "ne");
break;
case GT:
fprintf (file, "g");
break;
case LT:
fprintf (file, "l");
break;
case GE:
fprintf (file, "ge");
break;
case LE:
fprintf (file, "le");
break;
case GTU:
fprintf (file, "g");
break;
case LTU:
fprintf (file, "l");
break;
case GEU:
fprintf (file, "ge");
break;
case LEU:
fprintf (file, "le");
break;
default:
output_operand_lossage ("invalid %%j value");
}
break;
case 'J': /* reverse logic */
switch (GET_CODE(x))
{
case EQ:
fprintf (file, "ne");
break;
case NE:
fprintf (file, "e");
break;
case GT:
fprintf (file, "le");
break;
case LT:
fprintf (file, "ge");
break;
case GE:
fprintf (file, "l");
break;
case LE:
fprintf (file, "g");
break;
case GTU:
fprintf (file, "le");
break;
case LTU:
fprintf (file, "ge");
break;
case GEU:
fprintf (file, "l");
break;
case LEU:
fprintf (file, "g");
break;
default:
output_operand_lossage ("invalid %%J value");
}
break;
default:
switch (GET_CODE (x))
{
case REG:
if (code == 'h')
{
if (REGNO (x) < 32)
fprintf (file, "%s", short_reg_names[REGNO (x)]);
else
output_operand_lossage ("invalid operand for code '%c'", code);
}
else if (code == 'd')
{
if (REGNO (x) < 32)
fprintf (file, "%s", high_reg_names[REGNO (x)]);
else
output_operand_lossage ("invalid operand for code '%c'", code);
}
else if (code == 'w')
{
if (REGNO (x) == REG_A0 || REGNO (x) == REG_A1)
fprintf (file, "%s.w", reg_names[REGNO (x)]);
else
output_operand_lossage ("invalid operand for code '%c'", code);
}
else if (code == 'x')
{
if (REGNO (x) == REG_A0 || REGNO (x) == REG_A1)
fprintf (file, "%s.x", reg_names[REGNO (x)]);
else
output_operand_lossage ("invalid operand for code '%c'", code);
}
else if (code == 'v')
{
if (REGNO (x) == REG_A0)
fprintf (file, "AV0");
else if (REGNO (x) == REG_A1)
fprintf (file, "AV1");
else
output_operand_lossage ("invalid operand for code '%c'", code);
}
else if (code == 'D')
{
if (D_REGNO_P (REGNO (x)))
fprintf (file, "%s", dregs_pair_names[REGNO (x)]);
else
output_operand_lossage ("invalid operand for code '%c'", code);
}
else if (code == 'H')
{
if ((mode == DImode || mode == DFmode) && REG_P (x))
fprintf (file, "%s", reg_names[REGNO (x) + 1]);
else
output_operand_lossage ("invalid operand for code '%c'", code);
}
else if (code == 'T')
{
if (D_REGNO_P (REGNO (x)))
fprintf (file, "%s", byte_reg_names[REGNO (x)]);
else
output_operand_lossage ("invalid operand for code '%c'", code);
}
else
fprintf (file, "%s", reg_names[REGNO (x)]);
break;
case MEM:
fputc ('[', file);
x = XEXP (x,0);
print_address_operand (file, x);
fputc (']', file);
break;
case CONST_INT:
if (code == 'M')
{
switch (INTVAL (x))
{
case MACFLAG_NONE:
break;
case MACFLAG_FU:
fputs ("(FU)", file);
break;
case MACFLAG_T:
fputs ("(T)", file);
break;
case MACFLAG_TFU:
fputs ("(TFU)", file);
break;
case MACFLAG_W32:
fputs ("(W32)", file);
break;
case MACFLAG_IS:
fputs ("(IS)", file);
break;
case MACFLAG_IU:
fputs ("(IU)", file);
break;
case MACFLAG_IH:
fputs ("(IH)", file);
break;
case MACFLAG_M:
fputs ("(M)", file);
break;
case MACFLAG_IS_M:
fputs ("(IS,M)", file);
break;
case MACFLAG_ISS2:
fputs ("(ISS2)", file);
break;
case MACFLAG_S2RND:
fputs ("(S2RND)", file);
break;
default:
gcc_unreachable ();
}
break;
}
else if (code == 'b')
{
if (INTVAL (x) == 0)
fputs ("+=", file);
else if (INTVAL (x) == 1)
fputs ("-=", file);
else
gcc_unreachable ();
break;
}
/* Moves to half registers with d or h modifiers always use unsigned
constants. */
else if (code == 'd')
x = GEN_INT ((INTVAL (x) >> 16) & 0xffff);
else if (code == 'h')
x = GEN_INT (INTVAL (x) & 0xffff);
else if (code == 'N')
x = GEN_INT (-INTVAL (x));
else if (code == 'X')
x = GEN_INT (exact_log2 (0xffffffff & INTVAL (x)));
else if (code == 'Y')
x = GEN_INT (exact_log2 (0xffffffff & ~INTVAL (x)));
else if (code == 'Z')
/* Used for LINK insns. */
x = GEN_INT (-8 - INTVAL (x));
/* fall through */
case SYMBOL_REF:
output_addr_const (file, x);
break;
case CONST_DOUBLE:
output_operand_lossage ("invalid const_double operand");
break;
case UNSPEC:
switch (XINT (x, 1))
{
case UNSPEC_MOVE_PIC:
output_addr_const (file, XVECEXP (x, 0, 0));
fprintf (file, "@GOT");
break;
case UNSPEC_MOVE_FDPIC:
output_addr_const (file, XVECEXP (x, 0, 0));
fprintf (file, "@GOT17M4");
break;
case UNSPEC_FUNCDESC_GOT17M4:
output_addr_const (file, XVECEXP (x, 0, 0));
fprintf (file, "@FUNCDESC_GOT17M4");
break;
case UNSPEC_LIBRARY_OFFSET:
fprintf (file, "_current_shared_library_p5_offset_");
break;
default:
gcc_unreachable ();
}
break;
default:
output_addr_const (file, x);
}
}
}
/* Argument support functions. */
/* Initialize a variable CUM of type CUMULATIVE_ARGS
for a call to a function whose data type is FNTYPE.
For a library call, FNTYPE is 0.
VDSP C Compiler manual, our ABI says that
first 3 words of arguments will use R0, R1 and R2.
*/
void
init_cumulative_args (CUMULATIVE_ARGS *cum, tree fntype,
rtx libname ATTRIBUTE_UNUSED)
{
static CUMULATIVE_ARGS zero_cum;
*cum = zero_cum;
/* Set up the number of registers to use for passing arguments. */
cum->nregs = max_arg_registers;
cum->arg_regs = arg_regs;
cum->call_cookie = CALL_NORMAL;
/* Check for a longcall attribute. */
if (fntype && lookup_attribute ("shortcall", TYPE_ATTRIBUTES (fntype)))
cum->call_cookie |= CALL_SHORT;
else if (fntype && lookup_attribute ("longcall", TYPE_ATTRIBUTES (fntype)))
cum->call_cookie |= CALL_LONG;
return;
}
/* Update the data in CUM to advance over an argument
of mode MODE and data type TYPE.
(TYPE is null for libcalls where that information may not be available.) */
static void
bfin_function_arg_advance (cumulative_args_t cum_v, machine_mode mode,
const_tree type, bool named ATTRIBUTE_UNUSED)
{
CUMULATIVE_ARGS *cum = get_cumulative_args (cum_v);
int count, bytes, words;
bytes = (mode == BLKmode) ? int_size_in_bytes (type) : GET_MODE_SIZE (mode);
words = (bytes + UNITS_PER_WORD - 1) / UNITS_PER_WORD;
cum->words += words;
cum->nregs -= words;
if (cum->nregs <= 0)
{
cum->nregs = 0;
cum->arg_regs = NULL;
}
else
{
for (count = 1; count <= words; count++)
cum->arg_regs++;
}
return;
}
/* Define where to put the arguments to a function.
Value is zero to push the argument on the stack,
or a hard register in which to store the argument.
MODE is the argument's machine mode.
TYPE is the data type of the argument (as a tree).
This is null for libcalls where that information may
not be available.
CUM is a variable of type CUMULATIVE_ARGS which gives info about
the preceding args and about the function being called.
NAMED is nonzero if this argument is a named parameter
(otherwise it is an extra parameter matching an ellipsis). */
static rtx
bfin_function_arg (cumulative_args_t cum_v, machine_mode mode,
const_tree type, bool named ATTRIBUTE_UNUSED)
{
CUMULATIVE_ARGS *cum = get_cumulative_args (cum_v);
int bytes
= (mode == BLKmode) ? int_size_in_bytes (type) : GET_MODE_SIZE (mode);
if (mode == VOIDmode)
/* Compute operand 2 of the call insn. */
return GEN_INT (cum->call_cookie);
if (bytes == -1)
return NULL_RTX;
if (cum->nregs)
return gen_rtx_REG (mode, *(cum->arg_regs));
return NULL_RTX;
}
/* For an arg passed partly in registers and partly in memory,
this is the number of bytes passed in registers.
For args passed entirely in registers or entirely in memory, zero.
Refer VDSP C Compiler manual, our ABI.
First 3 words are in registers. So, if an argument is larger
than the registers available, it will span the register and
stack. */
static int
bfin_arg_partial_bytes (cumulative_args_t cum, machine_mode mode,
tree type ATTRIBUTE_UNUSED,
bool named ATTRIBUTE_UNUSED)
{
int bytes
= (mode == BLKmode) ? int_size_in_bytes (type) : GET_MODE_SIZE (mode);
int bytes_left = get_cumulative_args (cum)->nregs * UNITS_PER_WORD;
if (bytes == -1)
return 0;
if (bytes_left == 0)
return 0;
if (bytes > bytes_left)
return bytes_left;
return 0;
}
/* Variable sized types are passed by reference. */
static bool
bfin_pass_by_reference (cumulative_args_t cum ATTRIBUTE_UNUSED,
machine_mode mode ATTRIBUTE_UNUSED,
const_tree type, bool named ATTRIBUTE_UNUSED)
{
return type && TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST;
}
/* Decide whether a type should be returned in memory (true)
or in a register (false). This is called by the macro
TARGET_RETURN_IN_MEMORY. */
static bool
bfin_return_in_memory (const_tree type, const_tree fntype ATTRIBUTE_UNUSED)
{
int size = int_size_in_bytes (type);
return size > 2 * UNITS_PER_WORD || size == -1;
}
/* Register in which address to store a structure value
is passed to a function. */
static rtx
bfin_struct_value_rtx (tree fntype ATTRIBUTE_UNUSED,
int incoming ATTRIBUTE_UNUSED)
{
return gen_rtx_REG (Pmode, REG_P0);
}
/* Return true when register may be used to pass function parameters. */
bool
function_arg_regno_p (int n)
{
int i;
for (i = 0; arg_regs[i] != -1; i++)
if (n == arg_regs[i])
return true;
return false;
}
/* Returns 1 if OP contains a symbol reference */
int
symbolic_reference_mentioned_p (rtx op)
{
register const char *fmt;
register int i;
if (GET_CODE (op) == SYMBOL_REF || GET_CODE (op) == LABEL_REF)
return 1;
fmt = GET_RTX_FORMAT (GET_CODE (op));
for (i = GET_RTX_LENGTH (GET_CODE (op)) - 1; i >= 0; i--)
{
if (fmt[i] == 'E')
{
register int j;
for (j = XVECLEN (op, i) - 1; j >= 0; j--)
if (symbolic_reference_mentioned_p (XVECEXP (op, i, j)))
return 1;
}
else if (fmt[i] == 'e' && symbolic_reference_mentioned_p (XEXP (op, i)))
return 1;
}
return 0;
}
/* Decide whether we can make a sibling call to a function. DECL is the
declaration of the function being targeted by the call and EXP is the
CALL_EXPR representing the call. */
static bool
bfin_function_ok_for_sibcall (tree decl ATTRIBUTE_UNUSED,
tree exp ATTRIBUTE_UNUSED)
{
struct cgraph_local_info *this_func, *called_func;
e_funkind fkind = funkind (TREE_TYPE (current_function_decl));
if (fkind != SUBROUTINE)
return false;
if (!TARGET_ID_SHARED_LIBRARY || TARGET_SEP_DATA)
return true;
/* When compiling for ID shared libraries, can't sibcall a local function
from a non-local function, because the local function thinks it does
not need to reload P5 in the prologue, but the sibcall wil pop P5 in the
sibcall epilogue, and we end up with the wrong value in P5. */
if (!decl)
/* Not enough information. */
return false;
this_func = cgraph_node::local_info (current_function_decl);
called_func = cgraph_node::local_info (decl);
if (!called_func)
return false;
return !called_func->local || this_func->local;
}
/* Write a template for a trampoline to F. */
static void
bfin_asm_trampoline_template (FILE *f)
{
if (TARGET_FDPIC)
{
fprintf (f, "\t.dd\t0x00000000\n"); /* 0 */
fprintf (f, "\t.dd\t0x00000000\n"); /* 0 */
fprintf (f, "\t.dd\t0x0000e109\n"); /* p1.l = fn low */
fprintf (f, "\t.dd\t0x0000e149\n"); /* p1.h = fn high */
fprintf (f, "\t.dd\t0x0000e10a\n"); /* p2.l = sc low */
fprintf (f, "\t.dd\t0x0000e14a\n"); /* p2.h = sc high */
fprintf (f, "\t.dw\t0xac4b\n"); /* p3 = [p1 + 4] */
fprintf (f, "\t.dw\t0x9149\n"); /* p1 = [p1] */
fprintf (f, "\t.dw\t0x0051\n"); /* jump (p1)*/
}
else
{
fprintf (f, "\t.dd\t0x0000e109\n"); /* p1.l = fn low */
fprintf (f, "\t.dd\t0x0000e149\n"); /* p1.h = fn high */
fprintf (f, "\t.dd\t0x0000e10a\n"); /* p2.l = sc low */
fprintf (f, "\t.dd\t0x0000e14a\n"); /* p2.h = sc high */
fprintf (f, "\t.dw\t0x0051\n"); /* jump (p1)*/
}
}
/* Emit RTL insns to initialize the variable parts of a trampoline at
M_TRAMP. FNDECL is the target function. CHAIN_VALUE is an RTX for
the static chain value for the function. */
static void
bfin_trampoline_init (rtx m_tramp, tree fndecl, rtx chain_value)
{
rtx t1 = copy_to_reg (XEXP (DECL_RTL (fndecl), 0));
rtx t2 = copy_to_reg (chain_value);
rtx mem;
int i = 0;
emit_block_move (m_tramp, assemble_trampoline_template (),
GEN_INT (TRAMPOLINE_SIZE), BLOCK_OP_NORMAL);
if (TARGET_FDPIC)
{
rtx a = force_reg (Pmode, plus_constant (Pmode, XEXP (m_tramp, 0), 8));
mem = adjust_address (m_tramp, Pmode, 0);
emit_move_insn (mem, a);
i = 8;
}
mem = adjust_address (m_tramp, HImode, i + 2);
emit_move_insn (mem, gen_lowpart (HImode, t1));
emit_insn (gen_ashrsi3 (t1, t1, GEN_INT (16)));
mem = adjust_address (m_tramp, HImode, i + 6);
emit_move_insn (mem, gen_lowpart (HImode, t1));
mem = adjust_address (m_tramp, HImode, i + 10);
emit_move_insn (mem, gen_lowpart (HImode, t2));
emit_insn (gen_ashrsi3 (t2, t2, GEN_INT (16)));
mem = adjust_address (m_tramp, HImode, i + 14);
emit_move_insn (mem, gen_lowpart (HImode, t2));
}
/* Emit insns to move operands[1] into operands[0]. */
void
emit_pic_move (rtx *operands, machine_mode mode ATTRIBUTE_UNUSED)
{
rtx temp = reload_in_progress ? operands[0] : gen_reg_rtx (Pmode);
gcc_assert (!TARGET_FDPIC || !(reload_in_progress || reload_completed));
if (GET_CODE (operands[0]) == MEM && SYMBOLIC_CONST (operands[1]))
operands[1] = force_reg (SImode, operands[1]);
else
operands[1] = legitimize_pic_address (operands[1], temp,
TARGET_FDPIC ? OUR_FDPIC_REG
: pic_offset_table_rtx);
}
/* Expand a move operation in mode MODE. The operands are in OPERANDS.
Returns true if no further code must be generated, false if the caller
should generate an insn to move OPERANDS[1] to OPERANDS[0]. */
bool
expand_move (rtx *operands, machine_mode mode)
{
rtx op = operands[1];
if ((TARGET_ID_SHARED_LIBRARY || TARGET_FDPIC)
&& SYMBOLIC_CONST (op))
emit_pic_move (operands, mode);
else if (mode == SImode && GET_CODE (op) == CONST
&& GET_CODE (XEXP (op, 0)) == PLUS
&& GET_CODE (XEXP (XEXP (op, 0), 0)) == SYMBOL_REF
&& !targetm.legitimate_constant_p (mode, op))
{
rtx dest = operands[0];
rtx op0, op1;
gcc_assert (!reload_in_progress && !reload_completed);
op = XEXP (op, 0);
op0 = force_reg (mode, XEXP (op, 0));
op1 = XEXP (op, 1);
if (!insn_data[CODE_FOR_addsi3].operand[2].predicate (op1, mode))
op1 = force_reg (mode, op1);
if (GET_CODE (dest) == MEM)
dest = gen_reg_rtx (mode);
emit_insn (gen_addsi3 (dest, op0, op1));
if (dest == operands[0])
return true;
operands[1] = dest;
}
/* Don't generate memory->memory or constant->memory moves, go through a
register */
else if ((reload_in_progress | reload_completed) == 0
&& GET_CODE (operands[0]) == MEM
&& GET_CODE (operands[1]) != REG)
operands[1] = force_reg (mode, operands[1]);
return false;
}
/* Split one or more DImode RTL references into pairs of SImode
references. The RTL can be REG, offsettable MEM, integer constant, or
CONST_DOUBLE. "operands" is a pointer to an array of DImode RTL to
split and "num" is its length. lo_half and hi_half are output arrays
that parallel "operands". */
void
split_di (rtx operands[], int num, rtx lo_half[], rtx hi_half[])
{
while (num--)
{
rtx op = operands[num];
/* simplify_subreg refuse to split volatile memory addresses,
but we still have to handle it. */
if (GET_CODE (op) == MEM)
{
lo_half[num] = adjust_address (op, SImode, 0);
hi_half[num] = adjust_address (op, SImode, 4);
}
else
{
lo_half[num] = simplify_gen_subreg (SImode, op,
GET_MODE (op) == VOIDmode
? DImode : GET_MODE (op), 0);
hi_half[num] = simplify_gen_subreg (SImode, op,
GET_MODE (op) == VOIDmode
? DImode : GET_MODE (op), 4);
}
}
}
bool
bfin_longcall_p (rtx op, int call_cookie)
{
gcc_assert (GET_CODE (op) == SYMBOL_REF);
if (SYMBOL_REF_WEAK (op))
return 1;
if (call_cookie & CALL_SHORT)
return 0;
if (call_cookie & CALL_LONG)
return 1;
if (TARGET_LONG_CALLS)
return 1;
return 0;
}
/* Expand a call instruction. FNADDR is the call target, RETVAL the return value.
COOKIE is a CONST_INT holding the call_cookie prepared init_cumulative_args.
SIBCALL is nonzero if this is a sibling call. */
void
bfin_expand_call (rtx retval, rtx fnaddr, rtx callarg1, rtx cookie, int sibcall)
{
rtx use = NULL, call;
rtx callee = XEXP (fnaddr, 0);
int nelts = 3;
rtx pat;
rtx picreg = get_hard_reg_initial_val (SImode, FDPIC_REGNO);
rtx retsreg = gen_rtx_REG (Pmode, REG_RETS);
int n;
/* In an untyped call, we can get NULL for operand 2. */
if (cookie == NULL_RTX)
cookie = const0_rtx;
/* Static functions and indirect calls don't need the pic register. */
if (!TARGET_FDPIC && flag_pic
&& GET_CODE (callee) == SYMBOL_REF
&& !SYMBOL_REF_LOCAL_P (callee))
use_reg (&use, pic_offset_table_rtx);
if (TARGET_FDPIC)
{
int caller_in_sram, callee_in_sram;
/* 0 is not in sram, 1 is in L1 sram, 2 is in L2 sram. */
caller_in_sram = callee_in_sram = 0;
if (lookup_attribute ("l1_text",
DECL_ATTRIBUTES (cfun->decl)) != NULL_TREE)
caller_in_sram = 1;
else if (lookup_attribute ("l2",
DECL_ATTRIBUTES (cfun->decl)) != NULL_TREE)
caller_in_sram = 2;
if (GET_CODE (callee) == SYMBOL_REF
&& SYMBOL_REF_DECL (callee) && DECL_P (SYMBOL_REF_DECL (callee)))
{
if (lookup_attribute
("l1_text",
DECL_ATTRIBUTES (SYMBOL_REF_DECL (callee))) != NULL_TREE)
callee_in_sram = 1;
else if (lookup_attribute
("l2",
DECL_ATTRIBUTES (SYMBOL_REF_DECL (callee))) != NULL_TREE)
callee_in_sram = 2;
}
if (GET_CODE (callee) != SYMBOL_REF
|| bfin_longcall_p (callee, INTVAL (cookie))
|| (GET_CODE (callee) == SYMBOL_REF
&& !SYMBOL_REF_LOCAL_P (callee)
&& TARGET_INLINE_PLT)
|| caller_in_sram != callee_in_sram
|| (caller_in_sram && callee_in_sram
&& (GET_CODE (callee) != SYMBOL_REF
|| !SYMBOL_REF_LOCAL_P (callee))))
{
rtx addr = callee;
if (! address_operand (addr, Pmode))
addr = force_reg (Pmode, addr);
fnaddr = gen_reg_rtx (SImode);
emit_insn (gen_load_funcdescsi (fnaddr, addr));
fnaddr = gen_rtx_MEM (Pmode, fnaddr);
picreg = gen_reg_rtx (SImode);
emit_insn (gen_load_funcdescsi (picreg,
plus_constant (Pmode, addr, 4)));
}
nelts++;
}
else if ((!register_no_elim_operand (callee, Pmode)
&& GET_CODE (callee) != SYMBOL_REF)
|| (GET_CODE (callee) == SYMBOL_REF
&& ((TARGET_ID_SHARED_LIBRARY && !TARGET_LEAF_ID_SHARED_LIBRARY)
|| bfin_longcall_p (callee, INTVAL (cookie)))))
{
callee = copy_to_mode_reg (Pmode, callee);
fnaddr = gen_rtx_MEM (Pmode, callee);
}
call = gen_rtx_CALL (VOIDmode, fnaddr, callarg1);
if (retval)
call = gen_rtx_SET (VOIDmode, retval, call);
pat = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (nelts));
n = 0;
XVECEXP (pat, 0, n++) = call;
if (TARGET_FDPIC)
XVECEXP (pat, 0, n++) = gen_rtx_USE (VOIDmode, picreg);
XVECEXP (pat, 0, n++) = gen_rtx_USE (VOIDmode, cookie);
if (sibcall)
XVECEXP (pat, 0, n++) = ret_rtx;
else
XVECEXP (pat, 0, n++) = gen_rtx_CLOBBER (VOIDmode, retsreg);
call = emit_call_insn (pat);
if (use)
CALL_INSN_FUNCTION_USAGE (call) = use;
}
/* Return 1 if hard register REGNO can hold a value of machine-mode MODE. */
int
hard_regno_mode_ok (int regno, machine_mode mode)
{
/* Allow only dregs to store value of mode HI or QI */
enum reg_class rclass = REGNO_REG_CLASS (regno);
if (mode == CCmode)
return 0;
if (mode == V2HImode)
return D_REGNO_P (regno);
if (rclass == CCREGS)
return mode == BImode;
if (mode == PDImode || mode == V2PDImode)
return regno == REG_A0 || regno == REG_A1;
/* Allow all normal 32-bit regs, except REG_M3, in case regclass ever comes
up with a bad register class (such as ALL_REGS) for DImode. */
if (mode == DImode)
return regno < REG_M3;
if (mode == SImode
&& TEST_HARD_REG_BIT (reg_class_contents[PROLOGUE_REGS], regno))
return 1;
return TEST_HARD_REG_BIT (reg_class_contents[MOST_REGS], regno);
}
/* Implements target hook vector_mode_supported_p. */
static bool
bfin_vector_mode_supported_p (machine_mode mode)
{
return mode == V2HImode;
}
/* Worker function for TARGET_REGISTER_MOVE_COST. */
static int
bfin_register_move_cost (machine_mode mode,
reg_class_t class1, reg_class_t class2)
{
/* These need secondary reloads, so they're more expensive. */
if ((class1 == CCREGS && !reg_class_subset_p (class2, DREGS))
|| (class2 == CCREGS && !reg_class_subset_p (class1, DREGS)))
return 4;
/* If optimizing for size, always prefer reg-reg over reg-memory moves. */
if (optimize_size)
return 2;
if (GET_MODE_CLASS (mode) == MODE_INT)
{
/* Discourage trying to use the accumulators. */
if (TEST_HARD_REG_BIT (reg_class_contents[class1], REG_A0)
|| TEST_HARD_REG_BIT (reg_class_contents[class1], REG_A1)
|| TEST_HARD_REG_BIT (reg_class_contents[class2], REG_A0)
|| TEST_HARD_REG_BIT (reg_class_contents[class2], REG_A1))
return 20;
}
return 2;
}
/* Worker function for TARGET_MEMORY_MOVE_COST.
??? In theory L1 memory has single-cycle latency. We should add a switch
that tells the compiler whether we expect to use only L1 memory for the
program; it'll make the costs more accurate. */
static int
bfin_memory_move_cost (machine_mode mode ATTRIBUTE_UNUSED,
reg_class_t rclass,
bool in ATTRIBUTE_UNUSED)
{
/* Make memory accesses slightly more expensive than any register-register
move. Also, penalize non-DP registers, since they need secondary
reloads to load and store. */
if (! reg_class_subset_p (rclass, DPREGS))
return 10;
return 8;
}
/* Inform reload about cases where moving X with a mode MODE to a register in
RCLASS requires an extra scratch register. Return the class needed for the
scratch register. */
static reg_class_t
bfin_secondary_reload (bool in_p, rtx x, reg_class_t rclass_i,
machine_mode mode, secondary_reload_info *sri)
{
/* If we have HImode or QImode, we can only use DREGS as secondary registers;
in most other cases we can also use PREGS. */
enum reg_class default_class = GET_MODE_SIZE (mode) >= 4 ? DPREGS : DREGS;
enum reg_class x_class = NO_REGS;
enum rtx_code code = GET_CODE (x);
enum reg_class rclass = (enum reg_class) rclass_i;
if (code == SUBREG)
x = SUBREG_REG (x), code = GET_CODE (x);
if (REG_P (x))
{
int regno = REGNO (x);
if (regno >= FIRST_PSEUDO_REGISTER)
regno = reg_renumber[regno];
if (regno == -1)
code = MEM;
else
x_class = REGNO_REG_CLASS (regno);
}
/* We can be asked to reload (plus (FP) (large_constant)) into a DREG.
This happens as a side effect of register elimination, and we need
a scratch register to do it. */
if (fp_plus_const_operand (x, mode))
{
rtx op2 = XEXP (x, 1);
int large_constant_p = ! satisfies_constraint_Ks7 (op2);
if (rclass == PREGS || rclass == PREGS_CLOBBERED)
return NO_REGS;
/* If destination is a DREG, we can do this without a scratch register
if the constant is valid for an add instruction. */
if ((rclass == DREGS || rclass == DPREGS)
&& ! large_constant_p)
return NO_REGS;
/* Reloading to anything other than a DREG? Use a PREG scratch
register. */
sri->icode = CODE_FOR_reload_insi;
return NO_REGS;
}
/* Data can usually be moved freely between registers of most classes.
AREGS are an exception; they can only move to or from another register
in AREGS or one in DREGS. They can also be assigned the constant 0. */
if (x_class == AREGS || x_class == EVEN_AREGS || x_class == ODD_AREGS)
return (rclass == DREGS || rclass == AREGS || rclass == EVEN_AREGS
|| rclass == ODD_AREGS
? NO_REGS : DREGS);
if (rclass == AREGS || rclass == EVEN_AREGS || rclass == ODD_AREGS)
{
if (code == MEM)
{
sri->icode = in_p ? CODE_FOR_reload_inpdi : CODE_FOR_reload_outpdi;
return NO_REGS;
}
if (x != const0_rtx && x_class != DREGS)
{
return DREGS;
}
else
return NO_REGS;
}
/* CCREGS can only be moved from/to DREGS. */
if (rclass == CCREGS && x_class != DREGS)
return DREGS;
if (x_class == CCREGS && rclass != DREGS)
return DREGS;
/* All registers other than AREGS can load arbitrary constants. The only
case that remains is MEM. */
if (code == MEM)
if (! reg_class_subset_p (rclass, default_class))
return default_class;
return NO_REGS;
}
/* Implement TARGET_CLASS_LIKELY_SPILLED_P. */
static bool
bfin_class_likely_spilled_p (reg_class_t rclass)
{
switch (rclass)
{
case PREGS_CLOBBERED:
case PROLOGUE_REGS:
case P0REGS:
case D0REGS:
case D1REGS:
case D2REGS:
case CCREGS:
return true;
default:
break;
}
return false;
}
static struct machine_function *
bfin_init_machine_status (void)
{
return ggc_cleared_alloc<machine_function> ();
}
/* Implement the TARGET_OPTION_OVERRIDE hook. */
static void
bfin_option_override (void)
{
/* If processor type is not specified, enable all workarounds. */
if (bfin_cpu_type == BFIN_CPU_UNKNOWN)
{
int i;
for (i = 0; bfin_cpus[i].name != NULL; i++)
bfin_workarounds |= bfin_cpus[i].workarounds;
bfin_si_revision = 0xffff;
}
if (bfin_csync_anomaly == 1)
bfin_workarounds |= WA_SPECULATIVE_SYNCS;
else if (bfin_csync_anomaly == 0)
bfin_workarounds &= ~WA_SPECULATIVE_SYNCS;
if (bfin_specld_anomaly == 1)
bfin_workarounds |= WA_SPECULATIVE_LOADS;
else if (bfin_specld_anomaly == 0)
bfin_workarounds &= ~WA_SPECULATIVE_LOADS;
if (TARGET_OMIT_LEAF_FRAME_POINTER)
flag_omit_frame_pointer = 1;
#ifdef SUBTARGET_FDPIC_NOT_SUPPORTED
if (TARGET_FDPIC)
error ("-mfdpic is not supported, please use a bfin-linux-uclibc target");
#endif
/* Library identification */
if (global_options_set.x_bfin_library_id && ! TARGET_ID_SHARED_LIBRARY)
error ("-mshared-library-id= specified without -mid-shared-library");
if (stack_limit_rtx && TARGET_FDPIC)
{
warning (0, "-fstack-limit- options are ignored with -mfdpic; use -mstack-check-l1");
stack_limit_rtx = NULL_RTX;
}
if (stack_limit_rtx && TARGET_STACK_CHECK_L1)
error ("can%'t use multiple stack checking methods together");
if (TARGET_ID_SHARED_LIBRARY && TARGET_FDPIC)
error ("ID shared libraries and FD-PIC mode can%'t be used together");
/* Don't allow the user to specify -mid-shared-library and -msep-data
together, as it makes little sense from a user's point of view... */
if (TARGET_SEP_DATA && TARGET_ID_SHARED_LIBRARY)
error ("cannot specify both -msep-data and -mid-shared-library");
/* ... internally, however, it's nearly the same. */
if (TARGET_SEP_DATA)
target_flags |= MASK_ID_SHARED_LIBRARY | MASK_LEAF_ID_SHARED_LIBRARY;
if (TARGET_ID_SHARED_LIBRARY && flag_pic == 0)
flag_pic = 1;
/* There is no single unaligned SI op for PIC code. Sometimes we
need to use ".4byte" and sometimes we need to use ".picptr".
See bfin_assemble_integer for details. */
if (TARGET_FDPIC)
targetm.asm_out.unaligned_op.si = 0;
/* Silently turn off flag_pic if not doing FDPIC or ID shared libraries,
since we don't support it and it'll just break. */
if (flag_pic && !TARGET_FDPIC && !TARGET_ID_SHARED_LIBRARY)
flag_pic = 0;
if (TARGET_MULTICORE && bfin_cpu_type != BFIN_CPU_BF561)
error ("-mmulticore can only be used with BF561");
if (TARGET_COREA && !TARGET_MULTICORE)
error ("-mcorea should be used with -mmulticore");
if (TARGET_COREB && !TARGET_MULTICORE)
error ("-mcoreb should be used with -mmulticore");
if (TARGET_COREA && TARGET_COREB)
error ("-mcorea and -mcoreb can%'t be used together");
flag_schedule_insns = 0;
init_machine_status = bfin_init_machine_status;
}
/* Return the destination address of BRANCH.
We need to use this instead of get_attr_length, because the
cbranch_with_nops pattern conservatively sets its length to 6, and
we still prefer to use shorter sequences. */
static int
branch_dest (rtx_insn *branch)
{
rtx dest;
int dest_uid;
rtx pat = PATTERN (branch);
if (GET_CODE (pat) == PARALLEL)
pat = XVECEXP (pat, 0, 0);
dest = SET_SRC (pat);
if (GET_CODE (dest) == IF_THEN_ELSE)
dest = XEXP (dest, 1);
dest = XEXP (dest, 0);
dest_uid = INSN_UID (dest);
return INSN_ADDRESSES (dest_uid);
}
/* Return nonzero if INSN is annotated with a REG_BR_PROB note that indicates
it's a branch that's predicted taken. */
static int
cbranch_predicted_taken_p (rtx insn)
{
rtx x = find_reg_note (insn, REG_BR_PROB, 0);
if (x)
{
int pred_val = XINT (x, 0);
return pred_val >= REG_BR_PROB_BASE / 2;
}
return 0;
}
/* Templates for use by asm_conditional_branch. */
static const char *ccbranch_templates[][3] = {
{ "if !cc jump %3;", "if cc jump 4 (bp); jump.s %3;", "if cc jump 6 (bp); jump.l %3;" },
{ "if cc jump %3;", "if !cc jump 4 (bp); jump.s %3;", "if !cc jump 6 (bp); jump.l %3;" },
{ "if !cc jump %3 (bp);", "if cc jump 4; jump.s %3;", "if cc jump 6; jump.l %3;" },
{ "if cc jump %3 (bp);", "if !cc jump 4; jump.s %3;", "if !cc jump 6; jump.l %3;" },
};
/* Output INSN, which is a conditional branch instruction with operands
OPERANDS.
We deal with the various forms of conditional branches that can be generated
by bfin_reorg to prevent the hardware from doing speculative loads, by
- emitting a sufficient number of nops, if N_NOPS is nonzero, or
- always emitting the branch as predicted taken, if PREDICT_TAKEN is true.
Either of these is only necessary if the branch is short, otherwise the
template we use ends in an unconditional jump which flushes the pipeline
anyway. */
void
asm_conditional_branch (rtx_insn *insn, rtx *operands, int n_nops, int predict_taken)
{
int offset = branch_dest (insn) - INSN_ADDRESSES (INSN_UID (insn));
/* Note : offset for instructions like if cc jmp; jump.[sl] offset
is to be taken from start of if cc rather than jump.
Range for jump.s is (-4094, 4096) instead of (-4096, 4094)
*/
int len = (offset >= -1024 && offset <= 1022 ? 0
: offset >= -4094 && offset <= 4096 ? 1
: 2);
int bp = predict_taken && len == 0 ? 1 : cbranch_predicted_taken_p (insn);
int idx = (bp << 1) | (GET_CODE (operands[0]) == EQ ? BRF : BRT);
output_asm_insn (ccbranch_templates[idx][len], operands);
gcc_assert (n_nops == 0 || !bp);
if (len == 0)
while (n_nops-- > 0)
output_asm_insn ("nop;", NULL);
}
/* Emit rtl for a comparison operation CMP in mode MODE. Operands have been
stored in bfin_compare_op0 and bfin_compare_op1 already. */
rtx
bfin_gen_compare (rtx cmp, machine_mode mode ATTRIBUTE_UNUSED)
{
enum rtx_code code1, code2;
rtx op0 = XEXP (cmp, 0), op1 = XEXP (cmp, 1);
rtx tem = bfin_cc_rtx;
enum rtx_code code = GET_CODE (cmp);
/* If we have a BImode input, then we already have a compare result, and
do not need to emit another comparison. */
if (GET_MODE (op0) == BImode)
{
gcc_assert ((code == NE || code == EQ) && op1 == const0_rtx);
tem = op0, code2 = code;
}
else
{
switch (code) {
/* bfin has these conditions */
case EQ:
case LT:
case LE:
case LEU:
case LTU:
code1 = code;
code2 = NE;
break;
default:
code1 = reverse_condition (code);
code2 = EQ;
break;
}
emit_insn (gen_rtx_SET (VOIDmode, tem,
gen_rtx_fmt_ee (code1, BImode, op0, op1)));
}
return gen_rtx_fmt_ee (code2, BImode, tem, CONST0_RTX (BImode));
}
/* Return nonzero iff C has exactly one bit set if it is interpreted
as a 32-bit constant. */
int
log2constp (unsigned HOST_WIDE_INT c)
{
c &= 0xFFFFFFFF;
return c != 0 && (c & (c-1)) == 0;
}
/* Returns the number of consecutive least significant zeros in the binary
representation of *V.
We modify *V to contain the original value arithmetically shifted right by
the number of zeroes. */
static int
shiftr_zero (HOST_WIDE_INT *v)
{
unsigned HOST_WIDE_INT tmp = *v;
unsigned HOST_WIDE_INT sgn;
int n = 0;
if (tmp == 0)
return 0;
sgn = tmp & ((unsigned HOST_WIDE_INT) 1 << (HOST_BITS_PER_WIDE_INT - 1));
while ((tmp & 0x1) == 0 && n <= 32)
{
tmp = (tmp >> 1) | sgn;
n++;
}
*v = tmp;
return n;
}
/* After reload, split the load of an immediate constant. OPERANDS are the
operands of the movsi_insn pattern which we are splitting. We return
nonzero if we emitted a sequence to load the constant, zero if we emitted
nothing because we want to use the splitter's default sequence. */
int
split_load_immediate (rtx operands[])
{
HOST_WIDE_INT val = INTVAL (operands[1]);
HOST_WIDE_INT tmp;
HOST_WIDE_INT shifted = val;
HOST_WIDE_INT shifted_compl = ~val;
int num_zero = shiftr_zero (&shifted);
int num_compl_zero = shiftr_zero (&shifted_compl);
unsigned int regno = REGNO (operands[0]);
/* This case takes care of single-bit set/clear constants, which we could
also implement with BITSET/BITCLR. */
if (num_zero
&& shifted >= -32768 && shifted < 65536
&& (D_REGNO_P (regno)
|| (regno >= REG_P0 && regno <= REG_P7 && num_zero <= 2)))
{
emit_insn (gen_movsi (operands[0], gen_int_mode (shifted, SImode)));
emit_insn (gen_ashlsi3 (operands[0], operands[0], GEN_INT (num_zero)));
return 1;
}
tmp = val & 0xFFFF;
tmp |= -(tmp & 0x8000);
/* If high word has one bit set or clear, try to use a bit operation. */
if (D_REGNO_P (regno))
{
if (log2constp (val & 0xFFFF0000))
{
emit_insn (gen_movsi (operands[0], GEN_INT (val & 0xFFFF)));
emit_insn (gen_iorsi3 (operands[0], operands[0],
gen_int_mode (val & 0xFFFF0000, SImode)));
return 1;
}
else if (log2constp (val | 0xFFFF) && (val & 0x8000) != 0)
{
emit_insn (gen_movsi (operands[0], GEN_INT (tmp)));
emit_insn (gen_andsi3 (operands[0], operands[0],
gen_int_mode (val | 0xFFFF, SImode)));
}
}
if (D_REGNO_P (regno))
{
if (tmp >= -64 && tmp <= 63)
{
emit_insn (gen_movsi (operands[0], GEN_INT (tmp)));
emit_insn (gen_movstricthi_high (operands[0],
gen_int_mode (val & -65536,
SImode)));
return 1;
}
if ((val & 0xFFFF0000) == 0)
{
emit_insn (gen_movsi (operands[0], const0_rtx));
emit_insn (gen_movsi_low (operands[0], operands[0], operands[1]));
return 1;
}
if ((val & 0xFFFF0000) == 0xFFFF0000)
{
emit_insn (gen_movsi (operands[0], constm1_rtx));
emit_insn (gen_movsi_low (operands[0], operands[0], operands[1]));
return 1;
}
}
/* Need DREGs for the remaining case. */
if (regno > REG_R7)
return 0;
if (optimize_size
&& num_compl_zero && shifted_compl >= -64 && shifted_compl <= 63)
{
/* If optimizing for size, generate a sequence that has more instructions
but is shorter. */
emit_insn (gen_movsi (operands[0], gen_int_mode (shifted_compl, SImode)));
emit_insn (gen_ashlsi3 (operands[0], operands[0],
GEN_INT (num_compl_zero)));
emit_insn (gen_one_cmplsi2 (operands[0], operands[0]));
return 1;
}
return 0;
}
/* Return true if the legitimate memory address for a memory operand of mode
MODE. Return false if not. */
static bool
bfin_valid_add (machine_mode mode, HOST_WIDE_INT value)
{
unsigned HOST_WIDE_INT v = value > 0 ? value : -value;
int sz = GET_MODE_SIZE (mode);
int shift = sz == 1 ? 0 : sz == 2 ? 1 : 2;
/* The usual offsettable_memref machinery doesn't work so well for this
port, so we deal with the problem here. */
if (value > 0 && sz == 8)
v += 4;
return (v & ~(0x7fff << shift)) == 0;
}
static bool
bfin_valid_reg_p (unsigned int regno, int strict, machine_mode mode,
enum rtx_code outer_code)
{
if (strict)
return REGNO_OK_FOR_BASE_STRICT_P (regno, mode, outer_code, SCRATCH);
else
return REGNO_OK_FOR_BASE_NONSTRICT_P (regno, mode, outer_code, SCRATCH);
}
/* Recognize an RTL expression that is a valid memory address for an
instruction. The MODE argument is the machine mode for the MEM expression
that wants to use this address.
Blackfin addressing modes are as follows:
[preg]
[preg + imm16]
B [ Preg + uimm15 ]
W [ Preg + uimm16m2 ]
[ Preg + uimm17m4 ]
[preg++]
[preg--]
[--sp]
*/
static bool
bfin_legitimate_address_p (machine_mode mode, rtx x, bool strict)
{
switch (GET_CODE (x)) {
case REG:
if (bfin_valid_reg_p (REGNO (x), strict, mode, MEM))
return true;
break;
case PLUS:
if (REG_P (XEXP (x, 0))
&& bfin_valid_reg_p (REGNO (XEXP (x, 0)), strict, mode, PLUS)
&& ((GET_CODE (XEXP (x, 1)) == UNSPEC && mode == SImode)
|| (GET_CODE (XEXP (x, 1)) == CONST_INT
&& bfin_valid_add (mode, INTVAL (XEXP (x, 1))))))
return true;
break;
case POST_INC:
case POST_DEC:
if (LEGITIMATE_MODE_FOR_AUTOINC_P (mode)
&& REG_P (XEXP (x, 0))
&& bfin_valid_reg_p (REGNO (XEXP (x, 0)), strict, mode, POST_INC))
return true;
case PRE_DEC:
if (LEGITIMATE_MODE_FOR_AUTOINC_P (mode)
&& XEXP (x, 0) == stack_pointer_rtx
&& REG_P (XEXP (x, 0))
&& bfin_valid_reg_p (REGNO (XEXP (x, 0)), strict, mode, PRE_DEC))
return true;
break;
default:
break;
}
return false;
}
/* Decide whether we can force certain constants to memory. If we
decide we can't, the caller should be able to cope with it in
another way. */
static bool
bfin_cannot_force_const_mem (machine_mode mode ATTRIBUTE_UNUSED,
rtx x ATTRIBUTE_UNUSED)
{
/* We have only one class of non-legitimate constants, and our movsi
expander knows how to handle them. Dropping these constants into the
data section would only shift the problem - we'd still get relocs
outside the object, in the data section rather than the text section. */
return true;
}
/* Ensure that for any constant of the form symbol + offset, the offset
remains within the object. Any other constants are ok.
This ensures that flat binaries never have to deal with relocations
crossing section boundaries. */
static bool
bfin_legitimate_constant_p (machine_mode mode ATTRIBUTE_UNUSED, rtx x)
{
rtx sym;
HOST_WIDE_INT offset;
if (GET_CODE (x) != CONST)
return true;
x = XEXP (x, 0);
gcc_assert (GET_CODE (x) == PLUS);
sym = XEXP (x, 0);
x = XEXP (x, 1);
if (GET_CODE (sym) != SYMBOL_REF
|| GET_CODE (x) != CONST_INT)
return true;
offset = INTVAL (x);
if (SYMBOL_REF_DECL (sym) == 0)
return true;
if (offset < 0
|| offset >= int_size_in_bytes (TREE_TYPE (SYMBOL_REF_DECL (sym))))
return false;
return true;
}
static bool
bfin_rtx_costs (rtx x, int code_i, int outer_code_i, int opno, int *total,
bool speed)
{
enum rtx_code code = (enum rtx_code) code_i;
enum rtx_code outer_code = (enum rtx_code) outer_code_i;
int cost2 = COSTS_N_INSNS (1);
rtx op0, op1;
switch (code)
{
case CONST_INT:
if (outer_code == SET || outer_code == PLUS)
*total = satisfies_constraint_Ks7 (x) ? 0 : cost2;
else if (outer_code == AND)
*total = log2constp (~INTVAL (x)) ? 0 : cost2;
else if (outer_code == LE || outer_code == LT || outer_code == EQ)
*total = (INTVAL (x) >= -4 && INTVAL (x) <= 3) ? 0 : cost2;
else if (outer_code == LEU || outer_code == LTU)
*total = (INTVAL (x) >= 0 && INTVAL (x) <= 7) ? 0 : cost2;
else if (outer_code == MULT)
*total = (INTVAL (x) == 2 || INTVAL (x) == 4) ? 0 : cost2;
else if (outer_code == ASHIFT && (INTVAL (x) == 1 || INTVAL (x) == 2))
*total = 0;
else if (outer_code == ASHIFT || outer_code == ASHIFTRT
|| outer_code == LSHIFTRT)
*total = (INTVAL (x) >= 0 && INTVAL (x) <= 31) ? 0 : cost2;
else if (outer_code == IOR || outer_code == XOR)
*total = (INTVAL (x) & (INTVAL (x) - 1)) == 0 ? 0 : cost2;
else
*total = cost2;
return true;
case CONST:
case LABEL_REF:
case SYMBOL_REF:
case CONST_DOUBLE:
*total = COSTS_N_INSNS (2);
return true;
case PLUS:
op0 = XEXP (x, 0);
op1 = XEXP (x, 1);
if (GET_MODE (x) == SImode)
{
if (GET_CODE (op0) == MULT
&& GET_CODE (XEXP (op0, 1)) == CONST_INT)
{
HOST_WIDE_INT val = INTVAL (XEXP (op0, 1));
if (val == 2 || val == 4)
{
*total = cost2;
*total += rtx_cost (XEXP (op0, 0), outer_code, opno, speed);
*total += rtx_cost (op1, outer_code, opno, speed);
return true;
}
}
*total = cost2;
if (GET_CODE (op0) != REG
&& (GET_CODE (op0) != SUBREG || GET_CODE (SUBREG_REG (op0)) != REG))
*total += set_src_cost (op0, speed);
#if 0 /* We'd like to do this for accuracy, but it biases the loop optimizer
towards creating too many induction variables. */
if (!reg_or_7bit_operand (op1, SImode))
*total += set_src_cost (op1, speed);
#endif
}
else if (GET_MODE (x) == DImode)
{
*total = 6 * cost2;
if (GET_CODE (op1) != CONST_INT
|| !satisfies_constraint_Ks7 (op1))
*total += rtx_cost (op1, PLUS, 1, speed);
if (GET_CODE (op0) != REG
&& (GET_CODE (op0) != SUBREG || GET_CODE (SUBREG_REG (op0)) != REG))
*total += rtx_cost (op0, PLUS, 0, speed);
}
return true;
case MINUS:
if (GET_MODE (x) == DImode)
*total = 6 * cost2;
else
*total = cost2;
return true;
case ASHIFT:
case ASHIFTRT:
case LSHIFTRT:
if (GET_MODE (x) == DImode)
*total = 6 * cost2;
else
*total = cost2;
op0 = XEXP (x, 0);
op1 = XEXP (x, 1);
if (GET_CODE (op0) != REG
&& (GET_CODE (op0) != SUBREG || GET_CODE (SUBREG_REG (op0)) != REG))
*total += rtx_cost (op0, code, 0, speed);
return true;
case IOR:
case AND:
case XOR:
op0 = XEXP (x, 0);
op1 = XEXP (x, 1);
/* Handle special cases of IOR: rotates, ALIGN insns, movstricthi_high. */
if (code == IOR)
{
if ((GET_CODE (op0) == LSHIFTRT && GET_CODE (op1) == ASHIFT)
|| (GET_CODE (op0) == ASHIFT && GET_CODE (op1) == ZERO_EXTEND)
|| (GET_CODE (op0) == ASHIFT && GET_CODE (op1) == LSHIFTRT)
|| (GET_CODE (op0) == AND && GET_CODE (op1) == CONST_INT))
{
*total = cost2;
return true;
}
}
if (GET_CODE (op0) != REG
&& (GET_CODE (op0) != SUBREG || GET_CODE (SUBREG_REG (op0)) != REG))
*total += rtx_cost (op0, code, 0, speed);
if (GET_MODE (x) == DImode)
{
*total = 2 * cost2;
return true;
}
*total = cost2;
if (GET_MODE (x) != SImode)
return true;
if (code == AND)
{
if (! rhs_andsi3_operand (XEXP (x, 1), SImode))
*total += rtx_cost (XEXP (x, 1), code, 1, speed);
}
else
{
if (! regorlog2_operand (XEXP (x, 1), SImode))
*total += rtx_cost (XEXP (x, 1), code, 1, speed);
}
return true;
case ZERO_EXTRACT:
case SIGN_EXTRACT:
if (outer_code == SET
&& XEXP (x, 1) == const1_rtx
&& GET_CODE (XEXP (x, 2)) == CONST_INT)
{
*total = 2 * cost2;
return true;
}
/* fall through */
case SIGN_EXTEND:
case ZERO_EXTEND:
*total = cost2;
return true;
case MULT:
{
op0 = XEXP (x, 0);
op1 = XEXP (x, 1);
if (GET_CODE (op0) == GET_CODE (op1)
&& (GET_CODE (op0) == ZERO_EXTEND
|| GET_CODE (op0) == SIGN_EXTEND))
{
*total = COSTS_N_INSNS (1);
op0 = XEXP (op0, 0);
op1 = XEXP (op1, 0);
}
else if (!speed)
*total = COSTS_N_INSNS (1);
else
*total = COSTS_N_INSNS (3);
if (GET_CODE (op0) != REG
&& (GET_CODE (op0) != SUBREG || GET_CODE (SUBREG_REG (op0)) != REG))
*total += rtx_cost (op0, MULT, 0, speed);
if (GET_CODE (op1) != REG
&& (GET_CODE (op1) != SUBREG || GET_CODE (SUBREG_REG (op1)) != REG))
*total += rtx_cost (op1, MULT, 1, speed);
}
return true;
case UDIV:
case UMOD:
*total = COSTS_N_INSNS (32);
return true;
case VEC_CONCAT:
case VEC_SELECT:
if (outer_code == SET)
*total = cost2;
return true;
default:
return false;
}
}
/* Used for communication between {push,pop}_multiple_operation (which
we use not only as a predicate) and the corresponding output functions. */
static int first_preg_to_save, first_dreg_to_save;
static int n_regs_to_save;
int
analyze_push_multiple_operation (rtx op)
{
int lastdreg = 8, lastpreg = 6;
int i, group;
first_preg_to_save = lastpreg;
first_dreg_to_save = lastdreg;
for (i = 1, group = 0; i < XVECLEN (op, 0) - 1; i++)
{
rtx t = XVECEXP (op, 0, i);
rtx src, dest;
int regno;
if (GET_CODE (t) != SET)
return 0;
src = SET_SRC (t);
dest = SET_DEST (t);
if (GET_CODE (dest) != MEM || ! REG_P (src))
return 0;
dest = XEXP (dest, 0);
if (GET_CODE (dest) != PLUS
|| ! REG_P (XEXP (dest, 0))
|| REGNO (XEXP (dest, 0)) != REG_SP
|| GET_CODE (XEXP (dest, 1)) != CONST_INT
|| INTVAL (XEXP (dest, 1)) != -i * 4)
return 0;
regno = REGNO (src);
if (group == 0)
{
if (D_REGNO_P (regno))
{
group = 1;
first_dreg_to_save = lastdreg = regno - REG_R0;
}
else if (regno >= REG_P0 && regno <= REG_P7)