[pf3gnuchains/gcc-fork.git] / gcc / config / ia64 / ia64.c

/* Definitions of target machine for GNU compiler.
   Copyright (C) 1999, 2000, 2001, 2002, 2003, 2004
   Free Software Foundation, Inc.
   Contributed by James E. Wilson <wilson@cygnus.com> and
                  David Mosberger <davidm@hpl.hp.com>.

This file is part of GCC.

GCC is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2, or (at your option)
any later 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 COPYING.  If not, write to
the Free Software Foundation, 59 Temple Place - Suite 330,
Boston, MA 02111-1307, USA.  */

#include "config.h"
#include "system.h"
#include "coretypes.h"
#include "tm.h"
#include "rtl.h"
#include "tree.h"
#include "regs.h"
#include "hard-reg-set.h"
#include "real.h"
#include "insn-config.h"
#include "conditions.h"
#include "output.h"
#include "insn-attr.h"
#include "flags.h"
#include "recog.h"
#include "expr.h"
#include "optabs.h"
#include "except.h"
#include "function.h"
#include "ggc.h"
#include "basic-block.h"
#include "toplev.h"
#include "sched-int.h"
#include "timevar.h"
#include "target.h"
#include "target-def.h"
#include "tm_p.h"
#include "hashtab.h"
#include "langhooks.h"
#include "cfglayout.h"
#include "tree-gimple.h"

/* This is used for communication between ASM_OUTPUT_LABEL and
   ASM_OUTPUT_LABELREF.  */
int ia64_asm_output_label = 0;

/* Define the information needed to generate branch and scc insns.  This is
   stored from the compare operation.  */
struct rtx_def * ia64_compare_op0;
struct rtx_def * ia64_compare_op1;

/* Register names for ia64_expand_prologue.  */
static const char * const ia64_reg_numbers[96] =
{ "r32", "r33", "r34", "r35", "r36", "r37", "r38", "r39",
  "r40", "r41", "r42", "r43", "r44", "r45", "r46", "r47",
  "r48", "r49", "r50", "r51", "r52", "r53", "r54", "r55",
  "r56", "r57", "r58", "r59", "r60", "r61", "r62", "r63",
  "r64", "r65", "r66", "r67", "r68", "r69", "r70", "r71",
  "r72", "r73", "r74", "r75", "r76", "r77", "r78", "r79",
  "r80", "r81", "r82", "r83", "r84", "r85", "r86", "r87",
  "r88", "r89", "r90", "r91", "r92", "r93", "r94", "r95",
  "r96", "r97", "r98", "r99", "r100","r101","r102","r103",
  "r104","r105","r106","r107","r108","r109","r110","r111",
  "r112","r113","r114","r115","r116","r117","r118","r119",
  "r120","r121","r122","r123","r124","r125","r126","r127"};

/* ??? These strings could be shared with REGISTER_NAMES.  */
static const char * const ia64_input_reg_names[8] =
{ "in0",  "in1",  "in2",  "in3",  "in4",  "in5",  "in6",  "in7" };

/* ??? These strings could be shared with REGISTER_NAMES.  */
static const char * const ia64_local_reg_names[80] =
{ "loc0", "loc1", "loc2", "loc3", "loc4", "loc5", "loc6", "loc7",
  "loc8", "loc9", "loc10","loc11","loc12","loc13","loc14","loc15",
  "loc16","loc17","loc18","loc19","loc20","loc21","loc22","loc23",
  "loc24","loc25","loc26","loc27","loc28","loc29","loc30","loc31",
  "loc32","loc33","loc34","loc35","loc36","loc37","loc38","loc39",
  "loc40","loc41","loc42","loc43","loc44","loc45","loc46","loc47",
  "loc48","loc49","loc50","loc51","loc52","loc53","loc54","loc55",
  "loc56","loc57","loc58","loc59","loc60","loc61","loc62","loc63",
  "loc64","loc65","loc66","loc67","loc68","loc69","loc70","loc71",
  "loc72","loc73","loc74","loc75","loc76","loc77","loc78","loc79" };

/* ??? These strings could be shared with REGISTER_NAMES.  */
static const char * const ia64_output_reg_names[8] =
{ "out0", "out1", "out2", "out3", "out4", "out5", "out6", "out7" };

/* String used with the -mfixed-range= option.  */
const char *ia64_fixed_range_string;

/* Determines whether we use adds, addl, or movl to generate our
   TLS immediate offsets.  */
int ia64_tls_size = 22;

/* String used with the -mtls-size= option.  */
const char *ia64_tls_size_string;

/* Which cpu are we scheduling for.  */
enum processor_type ia64_tune;

/* String used with the -tune= option.  */
const char *ia64_tune_string;

/* Determines whether we run our final scheduling pass or not.  We always
   avoid the normal second scheduling pass.  */
static int ia64_flag_schedule_insns2;

/* Determines whether we run variable tracking in machine dependent
   reorganization.  */
static int ia64_flag_var_tracking;

/* Variables which are this size or smaller are put in the sdata/sbss
   sections.  */

unsigned int ia64_section_threshold;

/* The following variable is used by the DFA insn scheduler.  The value is
   TRUE if we do insn bundling instead of insn scheduling.  */
int bundling_p = 0;

/* Structure to be filled in by ia64_compute_frame_size with register
   save masks and offsets for the current function.  */

struct ia64_frame_info
{
  HOST_WIDE_INT total_size;     /* size of the stack frame, not including
                                   the caller's scratch area.  */
  HOST_WIDE_INT spill_cfa_off;  /* top of the reg spill area from the cfa.  */
  HOST_WIDE_INT spill_size;     /* size of the gr/br/fr spill area.  */
  HOST_WIDE_INT extra_spill_size;  /* size of spill area for others.  */
  HARD_REG_SET mask;            /* mask of saved registers.  */
  unsigned int gr_used_mask;    /* mask of registers in use as gr spill
                                   registers or long-term scratches.  */
  int n_spilled;                /* number of spilled registers.  */
  int reg_fp;                   /* register for fp.  */
  int reg_save_b0;              /* save register for b0.  */
  int reg_save_pr;              /* save register for prs.  */
  int reg_save_ar_pfs;          /* save register for ar.pfs.  */
  int reg_save_ar_unat;         /* save register for ar.unat.  */
  int reg_save_ar_lc;           /* save register for ar.lc.  */
  int reg_save_gp;              /* save register for gp.  */
  int n_input_regs;             /* number of input registers used.  */
  int n_local_regs;             /* number of local registers used.  */
  int n_output_regs;            /* number of output registers used.  */
  int n_rotate_regs;            /* number of rotating registers used.  */

  char need_regstk;             /* true if a .regstk directive needed.  */
  char initialized;             /* true if the data is finalized.  */
};

/* Current frame information calculated by ia64_compute_frame_size.  */
static struct ia64_frame_info current_frame_info;
\f
static int ia64_first_cycle_multipass_dfa_lookahead (void);
static void ia64_dependencies_evaluation_hook (rtx, rtx);
static void ia64_init_dfa_pre_cycle_insn (void);
static rtx ia64_dfa_pre_cycle_insn (void);
static int ia64_first_cycle_multipass_dfa_lookahead_guard (rtx);
static int ia64_dfa_new_cycle (FILE *, int, rtx, int, int, int *);
static rtx gen_tls_get_addr (void);
static rtx gen_thread_pointer (void);
static rtx ia64_expand_tls_address (enum tls_model, rtx, rtx);
static int find_gr_spill (int);
static int next_scratch_gr_reg (void);
static void mark_reg_gr_used_mask (rtx, void *);
static void ia64_compute_frame_size (HOST_WIDE_INT);
static void setup_spill_pointers (int, rtx, HOST_WIDE_INT);
static void finish_spill_pointers (void);
static rtx spill_restore_mem (rtx, HOST_WIDE_INT);
static void do_spill (rtx (*)(rtx, rtx, rtx), rtx, HOST_WIDE_INT, rtx);
static void do_restore (rtx (*)(rtx, rtx, rtx), rtx, HOST_WIDE_INT);
static rtx gen_movdi_x (rtx, rtx, rtx);
static rtx gen_fr_spill_x (rtx, rtx, rtx);
static rtx gen_fr_restore_x (rtx, rtx, rtx);

static enum machine_mode hfa_element_mode (tree, int);
static void ia64_setup_incoming_varargs (CUMULATIVE_ARGS *, enum machine_mode,
                                         tree, int *, int);
static bool ia64_pass_by_reference (CUMULATIVE_ARGS *, enum machine_mode,
                                    tree, bool);
static bool ia64_function_ok_for_sibcall (tree, tree);
static bool ia64_return_in_memory (tree, tree);
static bool ia64_rtx_costs (rtx, int, int, int *);
static void fix_range (const char *);
static struct machine_function * ia64_init_machine_status (void);
static void emit_insn_group_barriers (FILE *);
static void emit_all_insn_group_barriers (FILE *);
static void final_emit_insn_group_barriers (FILE *);
static void emit_predicate_relation_info (void);
static void ia64_reorg (void);
static bool ia64_in_small_data_p (tree);
static void process_epilogue (void);
static int process_set (FILE *, rtx);

static rtx ia64_expand_fetch_and_op (optab, enum machine_mode, tree, rtx);
static rtx ia64_expand_op_and_fetch (optab, enum machine_mode, tree, rtx);
static rtx ia64_expand_compare_and_swap (enum machine_mode, enum machine_mode,
                                         int, tree, rtx);
static rtx ia64_expand_lock_test_and_set (enum machine_mode, tree, rtx);
static rtx ia64_expand_lock_release (enum machine_mode, tree, rtx);
static bool ia64_assemble_integer (rtx, unsigned int, int);
static void ia64_output_function_prologue (FILE *, HOST_WIDE_INT);
static void ia64_output_function_epilogue (FILE *, HOST_WIDE_INT);
static void ia64_output_function_end_prologue (FILE *);

static int ia64_issue_rate (void);
static int ia64_adjust_cost (rtx, rtx, rtx, int);
static void ia64_sched_init (FILE *, int, int);
static void ia64_sched_finish (FILE *, int);
static int ia64_dfa_sched_reorder (FILE *, int, rtx *, int *, int, int);
static int ia64_sched_reorder (FILE *, int, rtx *, int *, int);
static int ia64_sched_reorder2 (FILE *, int, rtx *, int *, int);
static int ia64_variable_issue (FILE *, int, rtx, int);

static struct bundle_state *get_free_bundle_state (void);
static void free_bundle_state (struct bundle_state *);
static void initiate_bundle_states (void);
static void finish_bundle_states (void);
static unsigned bundle_state_hash (const void *);
static int bundle_state_eq_p (const void *, const void *);
static int insert_bundle_state (struct bundle_state *);
static void initiate_bundle_state_table (void);
static void finish_bundle_state_table (void);
static int try_issue_nops (struct bundle_state *, int);
static int try_issue_insn (struct bundle_state *, rtx);
static void issue_nops_and_insn (struct bundle_state *, int, rtx, int, int);
static int get_max_pos (state_t);
static int get_template (state_t, int);

static rtx get_next_important_insn (rtx, rtx);
static void bundling (FILE *, int, rtx, rtx);

static void ia64_output_mi_thunk (FILE *, tree, HOST_WIDE_INT,
                                  HOST_WIDE_INT, tree);
static void ia64_file_start (void);

static void ia64_select_rtx_section (enum machine_mode, rtx,
                                     unsigned HOST_WIDE_INT);
static void ia64_rwreloc_select_section (tree, int, unsigned HOST_WIDE_INT)
     ATTRIBUTE_UNUSED;
static void ia64_rwreloc_unique_section (tree, int)
     ATTRIBUTE_UNUSED;
static void ia64_rwreloc_select_rtx_section (enum machine_mode, rtx,
                                             unsigned HOST_WIDE_INT)
     ATTRIBUTE_UNUSED;
static unsigned int ia64_rwreloc_section_type_flags (tree, const char *, int)
     ATTRIBUTE_UNUSED;

static void ia64_hpux_add_extern_decl (tree decl)
     ATTRIBUTE_UNUSED;
static void ia64_hpux_file_end (void)
     ATTRIBUTE_UNUSED;
static void ia64_init_libfuncs (void)
     ATTRIBUTE_UNUSED;
static void ia64_hpux_init_libfuncs (void)
     ATTRIBUTE_UNUSED;
static void ia64_sysv4_init_libfuncs (void)
     ATTRIBUTE_UNUSED;
static void ia64_vms_init_libfuncs (void)
     ATTRIBUTE_UNUSED;

static tree ia64_handle_model_attribute (tree *, tree, tree, int, bool *);
static void ia64_encode_section_info (tree, rtx, int);
static rtx ia64_struct_value_rtx (tree, int);
static tree ia64_gimplify_va_arg (tree, tree, tree *, tree *);
static bool ia64_scalar_mode_supported_p (enum machine_mode mode);

\f
/* Table of valid machine attributes.  */
static const struct attribute_spec ia64_attribute_table[] =
{
  /* { name, min_len, max_len, decl_req, type_req, fn_type_req, handler } */
  { "syscall_linkage", 0, 0, false, true,  true,  NULL },
  { "model",           1, 1, true, false, false, ia64_handle_model_attribute },
  { NULL,              0, 0, false, false, false, NULL }
};

/* Initialize the GCC target structure.  */
#undef TARGET_ATTRIBUTE_TABLE
#define TARGET_ATTRIBUTE_TABLE ia64_attribute_table

#undef TARGET_INIT_BUILTINS
#define TARGET_INIT_BUILTINS ia64_init_builtins

#undef TARGET_EXPAND_BUILTIN
#define TARGET_EXPAND_BUILTIN ia64_expand_builtin

#undef TARGET_ASM_BYTE_OP
#define TARGET_ASM_BYTE_OP "\tdata1\t"
#undef TARGET_ASM_ALIGNED_HI_OP
#define TARGET_ASM_ALIGNED_HI_OP "\tdata2\t"
#undef TARGET_ASM_ALIGNED_SI_OP
#define TARGET_ASM_ALIGNED_SI_OP "\tdata4\t"
#undef TARGET_ASM_ALIGNED_DI_OP
#define TARGET_ASM_ALIGNED_DI_OP "\tdata8\t"
#undef TARGET_ASM_UNALIGNED_HI_OP
#define TARGET_ASM_UNALIGNED_HI_OP "\tdata2.ua\t"
#undef TARGET_ASM_UNALIGNED_SI_OP
#define TARGET_ASM_UNALIGNED_SI_OP "\tdata4.ua\t"
#undef TARGET_ASM_UNALIGNED_DI_OP
#define TARGET_ASM_UNALIGNED_DI_OP "\tdata8.ua\t"
#undef TARGET_ASM_INTEGER
#define TARGET_ASM_INTEGER ia64_assemble_integer

#undef TARGET_ASM_FUNCTION_PROLOGUE
#define TARGET_ASM_FUNCTION_PROLOGUE ia64_output_function_prologue
#undef TARGET_ASM_FUNCTION_END_PROLOGUE
#define TARGET_ASM_FUNCTION_END_PROLOGUE ia64_output_function_end_prologue
#undef TARGET_ASM_FUNCTION_EPILOGUE
#define TARGET_ASM_FUNCTION_EPILOGUE ia64_output_function_epilogue

#undef TARGET_IN_SMALL_DATA_P
#define TARGET_IN_SMALL_DATA_P  ia64_in_small_data_p

#undef TARGET_SCHED_ADJUST_COST
#define TARGET_SCHED_ADJUST_COST ia64_adjust_cost
#undef TARGET_SCHED_ISSUE_RATE
#define TARGET_SCHED_ISSUE_RATE ia64_issue_rate
#undef TARGET_SCHED_VARIABLE_ISSUE
#define TARGET_SCHED_VARIABLE_ISSUE ia64_variable_issue
#undef TARGET_SCHED_INIT
#define TARGET_SCHED_INIT ia64_sched_init
#undef TARGET_SCHED_FINISH
#define TARGET_SCHED_FINISH ia64_sched_finish
#undef TARGET_SCHED_REORDER
#define TARGET_SCHED_REORDER ia64_sched_reorder
#undef TARGET_SCHED_REORDER2
#define TARGET_SCHED_REORDER2 ia64_sched_reorder2

#undef TARGET_SCHED_DEPENDENCIES_EVALUATION_HOOK
#define TARGET_SCHED_DEPENDENCIES_EVALUATION_HOOK ia64_dependencies_evaluation_hook

#undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD
#define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD ia64_first_cycle_multipass_dfa_lookahead

#undef TARGET_SCHED_INIT_DFA_PRE_CYCLE_INSN
#define TARGET_SCHED_INIT_DFA_PRE_CYCLE_INSN ia64_init_dfa_pre_cycle_insn
#undef TARGET_SCHED_DFA_PRE_CYCLE_INSN
#define TARGET_SCHED_DFA_PRE_CYCLE_INSN ia64_dfa_pre_cycle_insn

#undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD_GUARD
#define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD_GUARD\
  ia64_first_cycle_multipass_dfa_lookahead_guard

#undef TARGET_SCHED_DFA_NEW_CYCLE
#define TARGET_SCHED_DFA_NEW_CYCLE ia64_dfa_new_cycle

#undef TARGET_FUNCTION_OK_FOR_SIBCALL
#define TARGET_FUNCTION_OK_FOR_SIBCALL ia64_function_ok_for_sibcall
#undef TARGET_PASS_BY_REFERENCE
#define TARGET_PASS_BY_REFERENCE ia64_pass_by_reference

#undef TARGET_ASM_OUTPUT_MI_THUNK
#define TARGET_ASM_OUTPUT_MI_THUNK ia64_output_mi_thunk
#undef TARGET_ASM_CAN_OUTPUT_MI_THUNK
#define TARGET_ASM_CAN_OUTPUT_MI_THUNK hook_bool_tree_hwi_hwi_tree_true

#undef TARGET_ASM_FILE_START
#define TARGET_ASM_FILE_START ia64_file_start

#undef TARGET_RTX_COSTS
#define TARGET_RTX_COSTS ia64_rtx_costs
#undef TARGET_ADDRESS_COST
#define TARGET_ADDRESS_COST hook_int_rtx_0

#undef TARGET_MACHINE_DEPENDENT_REORG
#define TARGET_MACHINE_DEPENDENT_REORG ia64_reorg

#undef TARGET_ENCODE_SECTION_INFO
#define TARGET_ENCODE_SECTION_INFO ia64_encode_section_info

/* ??? ABI doesn't allow us to define this.  */
#if 0
#undef TARGET_PROMOTE_FUNCTION_ARGS
#define TARGET_PROMOTE_FUNCTION_ARGS hook_bool_tree_true
#endif

/* ??? ABI doesn't allow us to define this.  */
#if 0
#undef TARGET_PROMOTE_FUNCTION_RETURN
#define TARGET_PROMOTE_FUNCTION_RETURN hook_bool_tree_true
#endif

/* ??? Investigate.  */
#if 0
#undef TARGET_PROMOTE_PROTOTYPES
#define TARGET_PROMOTE_PROTOTYPES hook_bool_tree_true
#endif

#undef TARGET_STRUCT_VALUE_RTX
#define TARGET_STRUCT_VALUE_RTX ia64_struct_value_rtx
#undef TARGET_RETURN_IN_MEMORY
#define TARGET_RETURN_IN_MEMORY ia64_return_in_memory
#undef TARGET_SETUP_INCOMING_VARARGS
#define TARGET_SETUP_INCOMING_VARARGS ia64_setup_incoming_varargs
#undef TARGET_STRICT_ARGUMENT_NAMING
#define TARGET_STRICT_ARGUMENT_NAMING hook_bool_CUMULATIVE_ARGS_true
#undef TARGET_MUST_PASS_IN_STACK
#define TARGET_MUST_PASS_IN_STACK must_pass_in_stack_var_size

#undef TARGET_GIMPLIFY_VA_ARG_EXPR
#define TARGET_GIMPLIFY_VA_ARG_EXPR ia64_gimplify_va_arg

#undef TARGET_UNWIND_EMIT
#define TARGET_UNWIND_EMIT process_for_unwind_directive

#undef TARGET_SCALAR_MODE_SUPPORTED_P
#define TARGET_SCALAR_MODE_SUPPORTED_P ia64_scalar_mode_supported_p

struct gcc_target targetm = TARGET_INITIALIZER;
\f
typedef enum
  {
    ADDR_AREA_NORMAL,   /* normal address area */
    ADDR_AREA_SMALL     /* addressable by "addl" (-2MB < addr < 2MB) */
  }
ia64_addr_area;

static GTY(()) tree small_ident1;
static GTY(()) tree small_ident2;

static void
init_idents (void)
{
  if (small_ident1 == 0)
    {
      small_ident1 = get_identifier ("small");
      small_ident2 = get_identifier ("__small__");
    }
}

/* Retrieve the address area that has been chosen for the given decl.  */

static ia64_addr_area
ia64_get_addr_area (tree decl)
{
  tree model_attr;

  model_attr = lookup_attribute ("model", DECL_ATTRIBUTES (decl));
  if (model_attr)
    {
      tree id;

      init_idents ();
      id = TREE_VALUE (TREE_VALUE (model_attr));
      if (id == small_ident1 || id == small_ident2)
        return ADDR_AREA_SMALL;
    }
  return ADDR_AREA_NORMAL;
}

static tree
ia64_handle_model_attribute (tree *node, tree name, tree args, int flags ATTRIBUTE_UNUSED, bool *no_add_attrs)
{
  ia64_addr_area addr_area = ADDR_AREA_NORMAL;
  ia64_addr_area area;
  tree arg, decl = *node;

  init_idents ();
  arg = TREE_VALUE (args);
  if (arg == small_ident1 || arg == small_ident2)
    {
      addr_area = ADDR_AREA_SMALL;
    }
  else
    {
      warning ("invalid argument of `%s' attribute",
               IDENTIFIER_POINTER (name));
      *no_add_attrs = true;
    }

  switch (TREE_CODE (decl))
    {
    case VAR_DECL:
      if ((DECL_CONTEXT (decl) && TREE_CODE (DECL_CONTEXT (decl))
           == FUNCTION_DECL)
          && !TREE_STATIC (decl))
        {
          error ("%Jan address area attribute cannot be specified for "
                 "local variables", decl, decl);
          *no_add_attrs = true;
        }
      area = ia64_get_addr_area (decl);
      if (area != ADDR_AREA_NORMAL && addr_area != area)
        {
          error ("%Jaddress area of '%s' conflicts with previous "
                 "declaration", decl, decl);
          *no_add_attrs = true;
        }
      break;

    case FUNCTION_DECL:
      error ("%Jaddress area attribute cannot be specified for functions",
             decl, decl);
      *no_add_attrs = true;
      break;

    default:
      warning ("`%s' attribute ignored", IDENTIFIER_POINTER (name));
      *no_add_attrs = true;
      break;
    }

  return NULL_TREE;
}

static void
ia64_encode_addr_area (tree decl, rtx symbol)
{
  int flags;

  flags = SYMBOL_REF_FLAGS (symbol);
  switch (ia64_get_addr_area (decl))
    {
    case ADDR_AREA_NORMAL: break;
    case ADDR_AREA_SMALL: flags |= SYMBOL_FLAG_SMALL_ADDR; break;
    default: abort ();
    }
  SYMBOL_REF_FLAGS (symbol) = flags;
}

static void
ia64_encode_section_info (tree decl, rtx rtl, int first)
{
  default_encode_section_info (decl, rtl, first);

  /* Careful not to prod global register variables.  */
  if (TREE_CODE (decl) == VAR_DECL
      && GET_CODE (DECL_RTL (decl)) == MEM
      && GET_CODE (XEXP (DECL_RTL (decl), 0)) == SYMBOL_REF
      && (TREE_STATIC (decl) || DECL_EXTERNAL (decl)))
    ia64_encode_addr_area (decl, XEXP (rtl, 0));
}
\f
/* Return 1 if the operands of a move are ok.  */

int
ia64_move_ok (rtx dst, rtx src)
{
  /* If we're under init_recog_no_volatile, we'll not be able to use
     memory_operand.  So check the code directly and don't worry about
     the validity of the underlying address, which should have been
     checked elsewhere anyway.  */
  if (GET_CODE (dst) != MEM)
    return 1;
  if (GET_CODE (src) == MEM)
    return 0;
  if (register_operand (src, VOIDmode))
    return 1;

  /* Otherwise, this must be a constant, and that either 0 or 0.0 or 1.0.  */
  if (INTEGRAL_MODE_P (GET_MODE (dst)))
    return src == const0_rtx;
  else
    return GET_CODE (src) == CONST_DOUBLE && CONST_DOUBLE_OK_FOR_G (src);
}

int
addp4_optimize_ok (rtx op1, rtx op2)
{
  return (basereg_operand (op1, GET_MODE(op1)) !=
          basereg_operand (op2, GET_MODE(op2)));
}

/* Check if OP is a mask suitable for use with SHIFT in a dep.z instruction.
   Return the length of the field, or <= 0 on failure.  */

int
ia64_depz_field_mask (rtx rop, rtx rshift)
{
  unsigned HOST_WIDE_INT op = INTVAL (rop);
  unsigned HOST_WIDE_INT shift = INTVAL (rshift);

  /* Get rid of the zero bits we're shifting in.  */
  op >>= shift;

  /* We must now have a solid block of 1's at bit 0.  */
  return exact_log2 (op + 1);
}

/* Expand a symbolic constant load.  */

void
ia64_expand_load_address (rtx dest, rtx src)
{
  if (GET_CODE (src) == SYMBOL_REF && SYMBOL_REF_TLS_MODEL (src))
    abort ();
  if (GET_CODE (dest) != REG)
    abort ();

  /* ILP32 mode still loads 64-bits of data from the GOT.  This avoids
     having to pointer-extend the value afterward.  Other forms of address
     computation below are also more natural to compute as 64-bit quantities.
     If we've been given an SImode destination register, change it.  */
  if (GET_MODE (dest) != Pmode)
    dest = gen_rtx_REG (Pmode, REGNO (dest));

  if (GET_CODE (src) == SYMBOL_REF && SYMBOL_REF_SMALL_ADDR_P (src))
    {
      emit_insn (gen_rtx_SET (VOIDmode, dest, src));
      return;
    }
  else if (TARGET_AUTO_PIC)
    {
      emit_insn (gen_load_gprel64 (dest, src));
      return;
    }
  else if (GET_CODE (src) == SYMBOL_REF && SYMBOL_REF_FUNCTION_P (src))
    {
      emit_insn (gen_load_fptr (dest, src));
      return;
    }
  else if (sdata_symbolic_operand (src, VOIDmode))
    {
      emit_insn (gen_load_gprel (dest, src));
      return;
    }

  if (GET_CODE (src) == CONST
      && GET_CODE (XEXP (src, 0)) == PLUS
      && GET_CODE (XEXP (XEXP (src, 0), 1)) == CONST_INT
      && (INTVAL (XEXP (XEXP (src, 0), 1)) & 0x3fff) != 0)
    {
      rtx sym = XEXP (XEXP (src, 0), 0);
      HOST_WIDE_INT ofs, hi, lo;

      /* Split the offset into a sign extended 14-bit low part
         and a complementary high part.  */
      ofs = INTVAL (XEXP (XEXP (src, 0), 1));
      lo = ((ofs & 0x3fff) ^ 0x2000) - 0x2000;
      hi = ofs - lo;

      ia64_expand_load_address (dest, plus_constant (sym, hi));
      emit_insn (gen_adddi3 (dest, dest, GEN_INT (lo)));
    }
  else
    {
      rtx tmp;

      tmp = gen_rtx_HIGH (Pmode, src);
      tmp = gen_rtx_PLUS (Pmode, tmp, pic_offset_table_rtx);
      emit_insn (gen_rtx_SET (VOIDmode, dest, tmp));

      tmp = gen_rtx_LO_SUM (GET_MODE (dest), dest, src);
      emit_insn (gen_rtx_SET (VOIDmode, dest, tmp));
    }
}

static GTY(()) rtx gen_tls_tga;
static rtx
gen_tls_get_addr (void)
{
  if (!gen_tls_tga)
    gen_tls_tga = init_one_libfunc ("__tls_get_addr");
  return gen_tls_tga;
}

static GTY(()) rtx thread_pointer_rtx;
static rtx
gen_thread_pointer (void)
{
  if (!thread_pointer_rtx)
    thread_pointer_rtx = gen_rtx_REG (Pmode, 13);
  return thread_pointer_rtx;
}

static rtx
ia64_expand_tls_address (enum tls_model tls_kind, rtx op0, rtx op1)
{
  rtx tga_op1, tga_op2, tga_ret, tga_eqv, tmp, insns;
  rtx orig_op0 = op0;

  switch (tls_kind)
    {
    case TLS_MODEL_GLOBAL_DYNAMIC:
      start_sequence ();

      tga_op1 = gen_reg_rtx (Pmode);
      emit_insn (gen_load_ltoff_dtpmod (tga_op1, op1));
      tga_op1 = gen_const_mem (Pmode, tga_op1);

      tga_op2 = gen_reg_rtx (Pmode);
      emit_insn (gen_load_ltoff_dtprel (tga_op2, op1));
      tga_op2 = gen_const_mem (Pmode, tga_op2);

      tga_ret = emit_library_call_value (gen_tls_get_addr (), NULL_RTX,
                                         LCT_CONST, Pmode, 2, tga_op1,
                                         Pmode, tga_op2, Pmode);

      insns = get_insns ();
      end_sequence ();

      if (GET_MODE (op0) != Pmode)
        op0 = tga_ret;
      emit_libcall_block (insns, op0, tga_ret, op1);
      break;

    case TLS_MODEL_LOCAL_DYNAMIC:
      /* ??? This isn't the completely proper way to do local-dynamic
         If the call to __tls_get_addr is used only by a single symbol,
         then we should (somehow) move the dtprel to the second arg
         to avoid the extra add.  */
      start_sequence ();

      tga_op1 = gen_reg_rtx (Pmode);
      emit_insn (gen_load_ltoff_dtpmod (tga_op1, op1));
      tga_op1 = gen_const_mem (Pmode, tga_op1);

      tga_op2 = const0_rtx;

      tga_ret = emit_library_call_value (gen_tls_get_addr (), NULL_RTX,
                                         LCT_CONST, Pmode, 2, tga_op1,
                                         Pmode, tga_op2, Pmode);

      insns = get_insns ();
      end_sequence ();

      tga_eqv = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, const0_rtx),
                                UNSPEC_LD_BASE);
      tmp = gen_reg_rtx (Pmode);
      emit_libcall_block (insns, tmp, tga_ret, tga_eqv);

      if (!register_operand (op0, Pmode))
        op0 = gen_reg_rtx (Pmode);
      if (TARGET_TLS64)
        {
          emit_insn (gen_load_dtprel (op0, op1));
          emit_insn (gen_adddi3 (op0, tmp, op0));
        }
      else
        emit_insn (gen_add_dtprel (op0, tmp, op1));
      break;

    case TLS_MODEL_INITIAL_EXEC:
      tmp = gen_reg_rtx (Pmode);
      emit_insn (gen_load_ltoff_tprel (tmp, op1));
      tmp = gen_const_mem (Pmode, tmp);
      tmp = force_reg (Pmode, tmp);

      if (!register_operand (op0, Pmode))
        op0 = gen_reg_rtx (Pmode);
      emit_insn (gen_adddi3 (op0, tmp, gen_thread_pointer ()));
      break;

    case TLS_MODEL_LOCAL_EXEC:
      if (!register_operand (op0, Pmode))
        op0 = gen_reg_rtx (Pmode);
      if (TARGET_TLS64)
        {
          emit_insn (gen_load_tprel (op0, op1));
          emit_insn (gen_adddi3 (op0, gen_thread_pointer (), op0));
        }
      else
        emit_insn (gen_add_tprel (op0, gen_thread_pointer (), op1));
      break;

    default:
      abort ();
    }

  if (orig_op0 == op0)
    return NULL_RTX;
  if (GET_MODE (orig_op0) == Pmode)
    return op0;
  return gen_lowpart (GET_MODE (orig_op0), op0);
}

rtx
ia64_expand_move (rtx op0, rtx op1)
{
  enum machine_mode mode = GET_MODE (op0);

  if (!reload_in_progress && !reload_completed && !ia64_move_ok (op0, op1))
    op1 = force_reg (mode, op1);

  if ((mode == Pmode || mode == ptr_mode) && symbolic_operand (op1, VOIDmode))
    {
      enum tls_model tls_kind;
      if (GET_CODE (op1) == SYMBOL_REF
          && (tls_kind = SYMBOL_REF_TLS_MODEL (op1)))
        return ia64_expand_tls_address (tls_kind, op0, op1);

      if (!TARGET_NO_PIC && reload_completed)
        {
          ia64_expand_load_address (op0, op1);
          return NULL_RTX;
        }
    }

  return op1;
}

/* Split a move from OP1 to OP0 conditional on COND.  */

void
ia64_emit_cond_move (rtx op0, rtx op1, rtx cond)
{
  rtx insn, first = get_last_insn ();

  emit_move_insn (op0, op1);

  for (insn = get_last_insn (); insn != first; insn = PREV_INSN (insn))
    if (INSN_P (insn))
      PATTERN (insn) = gen_rtx_COND_EXEC (VOIDmode, copy_rtx (cond),
                                          PATTERN (insn));
}

/* Split a post-reload TImode or TFmode reference into two DImode
   components.  This is made extra difficult by the fact that we do
   not get any scratch registers to work with, because reload cannot
   be prevented from giving us a scratch that overlaps the register
   pair involved.  So instead, when addressing memory, we tweak the
   pointer register up and back down with POST_INCs.  Or up and not
   back down when we can get away with it.

   REVERSED is true when the loads must be done in reversed order
   (high word first) for correctness.  DEAD is true when the pointer
   dies with the second insn we generate and therefore the second
   address must not carry a postmodify.

   May return an insn which is to be emitted after the moves.  */

static rtx
ia64_split_tmode (rtx out[2], rtx in, bool reversed, bool dead)
{
  rtx fixup = 0;

  switch (GET_CODE (in))
    {
    case REG:
      out[reversed] = gen_rtx_REG (DImode, REGNO (in));
      out[!reversed] = gen_rtx_REG (DImode, REGNO (in) + 1);
      break;

    case CONST_INT:
    case CONST_DOUBLE:
      /* Cannot occur reversed.  */
      if (reversed) abort ();
      
      if (GET_MODE (in) != TFmode)
        split_double (in, &out[0], &out[1]);
      else
        /* split_double does not understand how to split a TFmode
           quantity into a pair of DImode constants.  */
        {
          REAL_VALUE_TYPE r;
          unsigned HOST_WIDE_INT p[2];
          long l[4];  /* TFmode is 128 bits */

          REAL_VALUE_FROM_CONST_DOUBLE (r, in);
          real_to_target (l, &r, TFmode);

          if (FLOAT_WORDS_BIG_ENDIAN)
            {
              p[0] = (((unsigned HOST_WIDE_INT) l[0]) << 32) + l[1];
              p[1] = (((unsigned HOST_WIDE_INT) l[2]) << 32) + l[3];
            }
          else
            {
              p[0] = (((unsigned HOST_WIDE_INT) l[3]) << 32) + l[2];
              p[1] = (((unsigned HOST_WIDE_INT) l[1]) << 32) + l[0];
            }
          out[0] = GEN_INT (p[0]);
          out[1] = GEN_INT (p[1]);
        }
      break;

    case MEM:
      {
        rtx base = XEXP (in, 0);
        rtx offset;

        switch (GET_CODE (base))
          {
          case REG:
            if (!reversed)
              {
                out[0] = adjust_automodify_address
                  (in, DImode, gen_rtx_POST_INC (Pmode, base), 0);
                out[1] = adjust_automodify_address
                  (in, DImode, dead ? 0 : gen_rtx_POST_DEC (Pmode, base), 8);
              }
            else
              {
                /* Reversal requires a pre-increment, which can only
                   be done as a separate insn.  */
                emit_insn (gen_adddi3 (base, base, GEN_INT (8)));
                out[0] = adjust_automodify_address
                  (in, DImode, gen_rtx_POST_DEC (Pmode, base), 8);
                out[1] = adjust_address (in, DImode, 0);
              }
            break;

          case POST_INC:
            if (reversed || dead) abort ();
            /* Just do the increment in two steps.  */
            out[0] = adjust_automodify_address (in, DImode, 0, 0);
            out[1] = adjust_automodify_address (in, DImode, 0, 8);
            break;

          case POST_DEC:
            if (reversed || dead) abort ();
            /* Add 8, subtract 24.  */
            base = XEXP (base, 0);
            out[0] = adjust_automodify_address
              (in, DImode, gen_rtx_POST_INC (Pmode, base), 0);
            out[1] = adjust_automodify_address
              (in, DImode,
               gen_rtx_POST_MODIFY (Pmode, base, plus_constant (base, -24)),
               8);
            break;

          case POST_MODIFY:
            if (reversed || dead) abort ();
            /* Extract and adjust the modification.  This case is
               trickier than the others, because we might have an
               index register, or we might have a combined offset that
               doesn't fit a signed 9-bit displacement field.  We can
               assume the incoming expression is already legitimate.  */
            offset = XEXP (base, 1);
            base = XEXP (base, 0);

            out[0] = adjust_automodify_address
              (in, DImode, gen_rtx_POST_INC (Pmode, base), 0);

            if (GET_CODE (XEXP (offset, 1)) == REG)
              {
                /* Can't adjust the postmodify to match.  Emit the
                   original, then a separate addition insn.  */
                out[1] = adjust_automodify_address (in, DImode, 0, 8);
                fixup = gen_adddi3 (base, base, GEN_INT (-8));
              }
            else if (GET_CODE (XEXP (offset, 1)) != CONST_INT)
              abort ();
            else if (INTVAL (XEXP (offset, 1)) < -256 + 8)
              {
                /* Again the postmodify cannot be made to match, but
                   in this case it's more efficient to get rid of the
                   postmodify entirely and fix up with an add insn.  */
                out[1] = adjust_automodify_address (in, DImode, base, 8);
                fixup = gen_adddi3 (base, base,
                                    GEN_INT (INTVAL (XEXP (offset, 1)) - 8));
              }
            else
              {
                /* Combined offset still fits in the displacement field.
                   (We cannot overflow it at the high end.)  */
                out[1] = adjust_automodify_address
                  (in, DImode,
                   gen_rtx_POST_MODIFY (Pmode, base,
                     gen_rtx_PLUS (Pmode, base,
                                   GEN_INT (INTVAL (XEXP (offset, 1)) - 8))),
                   8);
              }
            break;

          default:
            abort ();
          }
        break;
      }

    default:
      abort ();
    }

  return fixup;
}

/* Split a TImode or TFmode move instruction after reload.
   This is used by *movtf_internal and *movti_internal.  */
void
ia64_split_tmode_move (rtx operands[])
{
  rtx in[2], out[2], insn;
  rtx fixup[2];
  bool dead = false;
  bool reversed = false;

  /* It is possible for reload to decide to overwrite a pointer with
     the value it points to.  In that case we have to do the loads in
     the appropriate order so that the pointer is not destroyed too
     early.  Also we must not generate a postmodify for that second
     load, or rws_access_regno will abort.  */
  if (GET_CODE (operands[1]) == MEM
      && reg_overlap_mentioned_p (operands[0], operands[1]))
    {
      rtx base = XEXP (operands[1], 0);
      while (GET_CODE (base) != REG)
        base = XEXP (base, 0);

      if (REGNO (base) == REGNO (operands[0]))
        reversed = true;
      dead = true;
    }
  /* Another reason to do the moves in reversed order is if the first
     element of the target register pair is also the second element of
     the source register pair.  */
  if (GET_CODE (operands[0]) == REG && GET_CODE (operands[1]) == REG
      && REGNO (operands[0]) == REGNO (operands[1]) + 1)
    reversed = true;

  fixup[0] = ia64_split_tmode (in, operands[1], reversed, dead);
  fixup[1] = ia64_split_tmode (out, operands[0], reversed, dead);

#define MAYBE_ADD_REG_INC_NOTE(INSN, EXP)                               \
  if (GET_CODE (EXP) == MEM                                             \
      && (GET_CODE (XEXP (EXP, 0)) == POST_MODIFY                       \
          || GET_CODE (XEXP (EXP, 0)) == POST_INC                       \
          || GET_CODE (XEXP (EXP, 0)) == POST_DEC))                     \
    REG_NOTES (INSN) = gen_rtx_EXPR_LIST (REG_INC,                      \
                                          XEXP (XEXP (EXP, 0), 0),      \
                                          REG_NOTES (INSN))

  insn = emit_insn (gen_rtx_SET (VOIDmode, out[0], in[0]));
  MAYBE_ADD_REG_INC_NOTE (insn, in[0]);
  MAYBE_ADD_REG_INC_NOTE (insn, out[0]);

  insn = emit_insn (gen_rtx_SET (VOIDmode, out[1], in[1]));
  MAYBE_ADD_REG_INC_NOTE (insn, in[1]);
  MAYBE_ADD_REG_INC_NOTE (insn, out[1]);

  if (fixup[0])
    emit_insn (fixup[0]);
  if (fixup[1])
    emit_insn (fixup[1]);

#undef MAYBE_ADD_REG_INC_NOTE
}

/* ??? Fixing GR->FR XFmode moves during reload is hard.  You need to go
   through memory plus an extra GR scratch register.  Except that you can
   either get the first from SECONDARY_MEMORY_NEEDED or the second from
   SECONDARY_RELOAD_CLASS, but not both.

   We got into problems in the first place by allowing a construct like
   (subreg:XF (reg:TI)), which we got from a union containing a long double.
   This solution attempts to prevent this situation from occurring.  When
   we see something like the above, we spill the inner register to memory.  */

rtx
spill_xfmode_operand (rtx in, int force)
{
  if (GET_CODE (in) == SUBREG
      && GET_MODE (SUBREG_REG (in)) == TImode
      && GET_CODE (SUBREG_REG (in)) == REG)
    {
      rtx memt = assign_stack_temp (TImode, 16, 0);
      emit_move_insn (memt, SUBREG_REG (in));
      return adjust_address (memt, XFmode, 0);
    }
  else if (force && GET_CODE (in) == REG)
    {
      rtx memx = assign_stack_temp (XFmode, 16, 0);
      emit_move_insn (memx, in);
      return memx;
    }
  else
    return in;
}

/* Emit comparison instruction if necessary, returning the expression
   that holds the compare result in the proper mode.  */

static GTY(()) rtx cmptf_libfunc;

rtx
ia64_expand_compare (enum rtx_code code, enum machine_mode mode)
{
  rtx op0 = ia64_compare_op0, op1 = ia64_compare_op1;
  rtx 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)
    {
      if ((code == NE || code == EQ) && op1 == const0_rtx)
        cmp = op0;
      else
        abort ();
    }
  /* HPUX TFmode compare requires a library call to _U_Qfcmp, which takes a
     magic number as its third argument, that indicates what to do.
     The return value is an integer to be compared against zero.  */
  else if (GET_MODE (op0) == TFmode)
    {
      enum qfcmp_magic {
        QCMP_INV = 1,   /* Raise FP_INVALID on SNaN as a side effect.  */
        QCMP_UNORD = 2,
        QCMP_EQ = 4,
        QCMP_LT = 8,
        QCMP_GT = 16
      } magic;
      enum rtx_code ncode;
      rtx ret, insns;
      if (!cmptf_libfunc || GET_MODE (op1) != TFmode)
        abort ();
      switch (code)
        {
          /* 1 = equal, 0 = not equal.  Equality operators do
             not raise FP_INVALID when given an SNaN operand.  */
        case EQ:        magic = QCMP_EQ;                  ncode = NE; break;
        case NE:        magic = QCMP_EQ;                  ncode = EQ; break;
          /* isunordered() from C99.  */
        case UNORDERED: magic = QCMP_UNORD;               ncode = NE; break;
        case ORDERED:   magic = QCMP_UNORD;               ncode = EQ; break;
          /* Relational operators raise FP_INVALID when given
             an SNaN operand.  */
        case LT:        magic = QCMP_LT        |QCMP_INV; ncode = NE; break;
        case LE:        magic = QCMP_LT|QCMP_EQ|QCMP_INV; ncode = NE; break;
        case GT:        magic = QCMP_GT        |QCMP_INV; ncode = NE; break;
        case GE:        magic = QCMP_GT|QCMP_EQ|QCMP_INV; ncode = NE; break;
          /* FUTURE: Implement UNEQ, UNLT, UNLE, UNGT, UNGE, LTGT.
             Expanders for buneq etc. weuld have to be added to ia64.md
             for this to be useful.  */
        default: abort ();
        }

      start_sequence ();

      ret = emit_library_call_value (cmptf_libfunc, 0, LCT_CONST, DImode, 3,
                                     op0, TFmode, op1, TFmode,
                                     GEN_INT (magic), DImode);
      cmp = gen_reg_rtx (BImode);
      emit_insn (gen_rtx_SET (VOIDmode, cmp,
                              gen_rtx_fmt_ee (ncode, BImode,
                                              ret, const0_rtx)));

      insns = get_insns ();
      end_sequence ();

      emit_libcall_block (insns, cmp, cmp,
                          gen_rtx_fmt_ee (code, BImode, op0, op1));
      code = NE;
    }
  else
    {
      cmp = gen_reg_rtx (BImode);
      emit_insn (gen_rtx_SET (VOIDmode, cmp,
                              gen_rtx_fmt_ee (code, BImode, op0, op1)));
      code = NE;
    }

  return gen_rtx_fmt_ee (code, mode, cmp, const0_rtx);
}

/* Emit the appropriate sequence for a call.  */

void
ia64_expand_call (rtx retval, rtx addr, rtx nextarg ATTRIBUTE_UNUSED,
                  int sibcall_p)
{
  rtx insn, b0;

  addr = XEXP (addr, 0);
  addr = convert_memory_address (DImode, addr);
  b0 = gen_rtx_REG (DImode, R_BR (0));

  /* ??? Should do this for functions known to bind local too.  */
  if (TARGET_NO_PIC || TARGET_AUTO_PIC)
    {
      if (sibcall_p)
        insn = gen_sibcall_nogp (addr);
      else if (! retval)
        insn = gen_call_nogp (addr, b0);
      else
        insn = gen_call_value_nogp (retval, addr, b0);
      insn = emit_call_insn (insn);
    }
  else
    {
      if (sibcall_p)
        insn = gen_sibcall_gp (addr);
      else if (! retval)
        insn = gen_call_gp (addr, b0);
      else
        insn = gen_call_value_gp (retval, addr, b0);
      insn = emit_call_insn (insn);

      use_reg (&CALL_INSN_FUNCTION_USAGE (insn), pic_offset_table_rtx);
    }

  if (sibcall_p)
    use_reg (&CALL_INSN_FUNCTION_USAGE (insn), b0);
}

void
ia64_reload_gp (void)
{
  rtx tmp;

  if (current_frame_info.reg_save_gp)
    tmp = gen_rtx_REG (DImode, current_frame_info.reg_save_gp);
  else
    {
      HOST_WIDE_INT offset;

      offset = (current_frame_info.spill_cfa_off
                + current_frame_info.spill_size);
      if (frame_pointer_needed)
        {
          tmp = hard_frame_pointer_rtx;
          offset = -offset;
        }
      else
        {
          tmp = stack_pointer_rtx;
          offset = current_frame_info.total_size - offset;
        }

      if (CONST_OK_FOR_I (offset))
        emit_insn (gen_adddi3 (pic_offset_table_rtx,
                               tmp, GEN_INT (offset)));
      else
        {
          emit_move_insn (pic_offset_table_rtx, GEN_INT (offset));
          emit_insn (gen_adddi3 (pic_offset_table_rtx,
                                 pic_offset_table_rtx, tmp));
        }

      tmp = gen_rtx_MEM (DImode, pic_offset_table_rtx);
    }

  emit_move_insn (pic_offset_table_rtx, tmp);
}

void
ia64_split_call (rtx retval, rtx addr, rtx retaddr, rtx scratch_r,
                 rtx scratch_b, int noreturn_p, int sibcall_p)
{
  rtx insn;
  bool is_desc = false;

  /* If we find we're calling through a register, then we're actually
     calling through a descriptor, so load up the values.  */
  if (REG_P (addr) && GR_REGNO_P (REGNO (addr)))
    {
      rtx tmp;
      bool addr_dead_p;

      /* ??? We are currently constrained to *not* use peep2, because
         we can legitimately change the global lifetime of the GP
         (in the form of killing where previously live).  This is
         because a call through a descriptor doesn't use the previous
         value of the GP, while a direct call does, and we do not
         commit to either form until the split here.

         That said, this means that we lack precise life info for
         whether ADDR is dead after this call.  This is not terribly
         important, since we can fix things up essentially for free
         with the POST_DEC below, but it's nice to not use it when we
         can immediately tell it's not necessary.  */
      addr_dead_p = ((noreturn_p || sibcall_p
                      || TEST_HARD_REG_BIT (regs_invalidated_by_call,
                                            REGNO (addr)))
                     && !FUNCTION_ARG_REGNO_P (REGNO (addr)));

      /* Load the code address into scratch_b.  */
      tmp = gen_rtx_POST_INC (Pmode, addr);
      tmp = gen_rtx_MEM (Pmode, tmp);
      emit_move_insn (scratch_r, tmp);
      emit_move_insn (scratch_b, scratch_r);

      /* Load the GP address.  If ADDR is not dead here, then we must
         revert the change made above via the POST_INCREMENT.  */
      if (!addr_dead_p)
        tmp = gen_rtx_POST_DEC (Pmode, addr);
      else
        tmp = addr;
      tmp = gen_rtx_MEM (Pmode, tmp);
      emit_move_insn (pic_offset_table_rtx, tmp);

      is_desc = true;
      addr = scratch_b;
    }

  if (sibcall_p)
    insn = gen_sibcall_nogp (addr);
  else if (retval)
    insn = gen_call_value_nogp (retval, addr, retaddr);
  else
    insn = gen_call_nogp (addr, retaddr);
  emit_call_insn (insn);

  if ((!TARGET_CONST_GP || is_desc) && !noreturn_p && !sibcall_p)
    ia64_reload_gp ();
}
\f
/* Begin the assembly file.  */

static void
ia64_file_start (void)
{
  default_file_start ();
  emit_safe_across_calls ();
}

void
emit_safe_across_calls (void)
{
  unsigned int rs, re;
  int out_state;

  rs = 1;
  out_state = 0;
  while (1)
    {
      while (rs < 64 && call_used_regs[PR_REG (rs)])
        rs++;
      if (rs >= 64)
        break;
      for (re = rs + 1; re < 64 && ! call_used_regs[PR_REG (re)]; re++)
        continue;
      if (out_state == 0)
        {
          fputs ("\t.pred.safe_across_calls ", asm_out_file);
          out_state = 1;
        }
      else
        fputc (',', asm_out_file);
      if (re == rs + 1)
        fprintf (asm_out_file, "p%u", rs);
      else
        fprintf (asm_out_file, "p%u-p%u", rs, re - 1);
      rs = re + 1;
    }
  if (out_state)
    fputc ('\n', asm_out_file);
}

/* Helper function for ia64_compute_frame_size: find an appropriate general
   register to spill some special register to.  SPECIAL_SPILL_MASK contains
   bits in GR0 to GR31 that have already been allocated by this routine.
   TRY_LOCALS is true if we should attempt to locate a local regnum.  */

static int
find_gr_spill (int try_locals)
{
  int regno;

  /* If this is a leaf function, first try an otherwise unused
     call-clobbered register.  */
  if (current_function_is_leaf)
    {
      for (regno = GR_REG (1); regno <= GR_REG (31); regno++)
        if (! regs_ever_live[regno]
            && call_used_regs[regno]
            && ! fixed_regs[regno]
            && ! global_regs[regno]
            && ((current_frame_info.gr_used_mask >> regno) & 1) == 0)
          {
            current_frame_info.gr_used_mask |= 1 << regno;
            return regno;
          }
    }

  if (try_locals)
    {
      regno = current_frame_info.n_local_regs;
      /* If there is a frame pointer, then we can't use loc79, because
         that is HARD_FRAME_POINTER_REGNUM.  In particular, see the
         reg_name switching code in ia64_expand_prologue.  */
      if (regno < (80 - frame_pointer_needed))
        {
          current_frame_info.n_local_regs = regno + 1;
          return LOC_REG (0) + regno;
        }
    }

  /* Failed to find a general register to spill to.  Must use stack.  */
  return 0;
}

/* In order to make for nice schedules, we try to allocate every temporary
   to a different register.  We must of course stay away from call-saved,
   fixed, and global registers.  We must also stay away from registers
   allocated in current_frame_info.gr_used_mask, since those include regs
   used all through the prologue.

   Any register allocated here must be used immediately.  The idea is to
   aid scheduling, not to solve data flow problems.  */

static int last_scratch_gr_reg;

static int
next_scratch_gr_reg (void)
{
  int i, regno;

  for (i = 0; i < 32; ++i)
    {
      regno = (last_scratch_gr_reg + i + 1) & 31;
      if (call_used_regs[regno]
          && ! fixed_regs[regno]
          && ! global_regs[regno]
          && ((current_frame_info.gr_used_mask >> regno) & 1) == 0)
        {
          last_scratch_gr_reg = regno;
          return regno;
        }
    }

  /* There must be _something_ available.  */
  abort ();
}

/* Helper function for ia64_compute_frame_size, called through
   diddle_return_value.  Mark REG in current_frame_info.gr_used_mask.  */

static void
mark_reg_gr_used_mask (rtx reg, void *data ATTRIBUTE_UNUSED)
{
  unsigned int regno = REGNO (reg);
  if (regno < 32)
    {
      unsigned int i, n = HARD_REGNO_NREGS (regno, GET_MODE (reg));
      for (i = 0; i < n; ++i)
        current_frame_info.gr_used_mask |= 1 << (regno + i);
    }
}

/* Returns the number of bytes offset between the frame pointer and the stack
   pointer for the current function.  SIZE is the number of bytes of space
   needed for local variables.  */

static void
ia64_compute_frame_size (HOST_WIDE_INT size)
{
  HOST_WIDE_INT total_size;
  HOST_WIDE_INT spill_size = 0;
  HOST_WIDE_INT extra_spill_size = 0;
  HOST_WIDE_INT pretend_args_size;
  HARD_REG_SET mask;
  int n_spilled = 0;
  int spilled_gr_p = 0;
  int spilled_fr_p = 0;
  unsigned int regno;
  int i;

  if (current_frame_info.initialized)
    return;

  memset (&current_frame_info, 0, sizeof current_frame_info);
  CLEAR_HARD_REG_SET (mask);

  /* Don't allocate scratches to the return register.  */
  diddle_return_value (mark_reg_gr_used_mask, NULL);

  /* Don't allocate scratches to the EH scratch registers.  */
  if (cfun->machine->ia64_eh_epilogue_sp)
    mark_reg_gr_used_mask (cfun->machine->ia64_eh_epilogue_sp, NULL);
  if (cfun->machine->ia64_eh_epilogue_bsp)
    mark_reg_gr_used_mask (cfun->machine->ia64_eh_epilogue_bsp, NULL);

  /* Find the size of the register stack frame.  We have only 80 local
     registers, because we reserve 8 for the inputs and 8 for the
     outputs.  */

  /* Skip HARD_FRAME_POINTER_REGNUM (loc79) when frame_pointer_needed,
     since we'll be adjusting that down later.  */
  regno = LOC_REG (78) + ! frame_pointer_needed;
  for (; regno >= LOC_REG (0); regno--)
    if (regs_ever_live[regno])
      break;
  current_frame_info.n_local_regs = regno - LOC_REG (0) + 1;

  /* For functions marked with the syscall_linkage attribute, we must mark
     all eight input registers as in use, so that locals aren't visible to
     the caller.  */

  if (cfun->machine->n_varargs > 0
      || lookup_attribute ("syscall_linkage",
                           TYPE_ATTRIBUTES (TREE_TYPE (current_function_decl))))
    current_frame_info.n_input_regs = 8;
  else
    {
      for (regno = IN_REG (7); regno >= IN_REG (0); regno--)
        if (regs_ever_live[regno])
          break;
      current_frame_info.n_input_regs = regno - IN_REG (0) + 1;
    }

  for (regno = OUT_REG (7); regno >= OUT_REG (0); regno--)
    if (regs_ever_live[regno])
      break;
  i = regno - OUT_REG (0) + 1;

  /* When -p profiling, we need one output register for the mcount argument.
     Likewise for -a profiling for the bb_init_func argument.  For -ax
     profiling, we need two output registers for the two bb_init_trace_func
     arguments.  */
  if (current_function_profile)
    i = MAX (i, 1);
  current_frame_info.n_output_regs = i;

  /* ??? No rotating register support yet.  */
  current_frame_info.n_rotate_regs = 0;

  /* Discover which registers need spilling, and how much room that
     will take.  Begin with floating point and general registers,
     which will always wind up on the stack.  */

  for (regno = FR_REG (2); regno <= FR_REG (127); regno++)
    if (regs_ever_live[regno] && ! call_used_regs[regno])
      {
        SET_HARD_REG_BIT (mask, regno);
        spill_size += 16;
        n_spilled += 1;
        spilled_fr_p = 1;
      }

  for (regno = GR_REG (1); regno <= GR_REG (31); regno++)
    if (regs_ever_live[regno] && ! call_used_regs[regno])
      {
        SET_HARD_REG_BIT (mask, regno);
        spill_size += 8;
        n_spilled += 1;
        spilled_gr_p = 1;
      }

  for (regno = BR_REG (1); regno <= BR_REG (7); regno++)
    if (regs_ever_live[regno] && ! call_used_regs[regno])
      {
        SET_HARD_REG_BIT (mask, regno);
        spill_size += 8;
        n_spilled += 1;
      }

  /* Now come all special registers that might get saved in other
     general registers.  */

  if (frame_pointer_needed)
    {
      current_frame_info.reg_fp = find_gr_spill (1);
      /* If we did not get a register, then we take LOC79.  This is guaranteed
         to be free, even if regs_ever_live is already set, because this is
         HARD_FRAME_POINTER_REGNUM.  This requires incrementing n_local_regs,
         as we don't count loc79 above.  */
      if (current_frame_info.reg_fp == 0)
        {
          current_frame_info.reg_fp = LOC_REG (79);
          current_frame_info.n_local_regs++;
        }
    }

  if (! current_function_is_leaf)
    {
      /* Emit a save of BR0 if we call other functions.  Do this even
         if this function doesn't return, as EH depends on this to be
         able to unwind the stack.  */
      SET_HARD_REG_BIT (mask, BR_REG (0));

      current_frame_info.reg_save_b0 = find_gr_spill (1);
      if (current_frame_info.reg_save_b0 == 0)
        {
          spill_size += 8;
          n_spilled += 1;
        }

      /* Similarly for ar.pfs.  */
      SET_HARD_REG_BIT (mask, AR_PFS_REGNUM);
      current_frame_info.reg_save_ar_pfs = find_gr_spill (1);
      if (current_frame_info.reg_save_ar_pfs == 0)
        {
          extra_spill_size += 8;
          n_spilled += 1;
        }

      /* Similarly for gp.  Note that if we're calling setjmp, the stacked
         registers are clobbered, so we fall back to the stack.  */
      current_frame_info.reg_save_gp
        = (current_function_calls_setjmp ? 0 : find_gr_spill (1));
      if (current_frame_info.reg_save_gp == 0)
        {
          SET_HARD_REG_BIT (mask, GR_REG (1));
          spill_size += 8;
          n_spilled += 1;
        }
    }
  else
    {
      if (regs_ever_live[BR_REG (0)] && ! call_used_regs[BR_REG (0)])
        {
          SET_HARD_REG_BIT (mask, BR_REG (0));
          spill_size += 8;
          n_spilled += 1;
        }

      if (regs_ever_live[AR_PFS_REGNUM])
        {
          SET_HARD_REG_BIT (mask, AR_PFS_REGNUM);
          current_frame_info.reg_save_ar_pfs = find_gr_spill (1);
          if (current_frame_info.reg_save_ar_pfs == 0)
            {
              extra_spill_size += 8;
              n_spilled += 1;
            }
        }
    }

  /* Unwind descriptor hackery: things are most efficient if we allocate
     consecutive GR save registers for RP, PFS, FP in that order. However,
     it is absolutely critical that FP get the only hard register that's
     guaranteed to be free, so we allocated it first.  If all three did
     happen to be allocated hard regs, and are consecutive, rearrange them
     into the preferred order now.  */
  if (current_frame_info.reg_fp != 0
      && current_frame_info.reg_save_b0 == current_frame_info.reg_fp + 1
      && current_frame_info.reg_save_ar_pfs == current_frame_info.reg_fp + 2)
    {
      current_frame_info.reg_save_b0 = current_frame_info.reg_fp;
      current_frame_info.reg_save_ar_pfs = current_frame_info.reg_fp + 1;
      current_frame_info.reg_fp = current_frame_info.reg_fp + 2;
    }

  /* See if we need to store the predicate register block.  */
  for (regno = PR_REG (0); regno <= PR_REG (63); regno++)
    if (regs_ever_live[regno] && ! call_used_regs[regno])
      break;
  if (regno <= PR_REG (63))
    {
      SET_HARD_REG_BIT (mask, PR_REG (0));
      current_frame_info.reg_save_pr = find_gr_spill (1);
      if (current_frame_info.reg_save_pr == 0)
        {
          extra_spill_size += 8;
          n_spilled += 1;
        }

      /* ??? Mark them all as used so that register renaming and such
         are free to use them.  */
      for (regno = PR_REG (0); regno <= PR_REG (63); regno++)
        regs_ever_live[regno] = 1;
    }

  /* If we're forced to use st8.spill, we're forced to save and restore
     ar.unat as well.  The check for existing liveness allows inline asm
     to touch ar.unat.  */
  if (spilled_gr_p || cfun->machine->n_varargs
      || regs_ever_live[AR_UNAT_REGNUM])
    {
      regs_ever_live[AR_UNAT_REGNUM] = 1;
      SET_HARD_REG_BIT (mask, AR_UNAT_REGNUM);
      current_frame_info.reg_save_ar_unat = find_gr_spill (spill_size == 0);
      if (current_frame_info.reg_save_ar_unat == 0)
        {
          extra_spill_size += 8;
          n_spilled += 1;
        }
    }

  if (regs_ever_live[AR_LC_REGNUM])
    {
      SET_HARD_REG_BIT (mask, AR_LC_REGNUM);
      current_frame_info.reg_save_ar_lc = find_gr_spill (spill_size == 0);
      if (current_frame_info.reg_save_ar_lc == 0)
        {
          extra_spill_size += 8;
          n_spilled += 1;
        }
    }

  /* If we have an odd number of words of pretend arguments written to
     the stack, then the FR save area will be unaligned.  We round the
     size of this area up to keep things 16 byte aligned.  */
  if (spilled_fr_p)
    pretend_args_size = IA64_STACK_ALIGN (current_function_pretend_args_size);
  else
    pretend_args_size = current_function_pretend_args_size;

  total_size = (spill_size + extra_spill_size + size + pretend_args_size
                + current_function_outgoing_args_size);
  total_size = IA64_STACK_ALIGN (total_size);

  /* We always use the 16-byte scratch area provided by the caller, but
     if we are a leaf function, there's no one to which we need to provide
     a scratch area.  */
  if (current_function_is_leaf)
    total_size = MAX (0, total_size - 16);

  current_frame_info.total_size = total_size;
  current_frame_info.spill_cfa_off = pretend_args_size - 16;
  current_frame_info.spill_size = spill_size;
  current_frame_info.extra_spill_size = extra_spill_size;
  COPY_HARD_REG_SET (current_frame_info.mask, mask);
  current_frame_info.n_spilled = n_spilled;
  current_frame_info.initialized = reload_completed;
}

/* Compute the initial difference between the specified pair of registers.  */

HOST_WIDE_INT
ia64_initial_elimination_offset (int from, int to)
{
  HOST_WIDE_INT offset;

  ia64_compute_frame_size (get_frame_size ());
  switch (from)
    {
    case FRAME_POINTER_REGNUM:
      if (to == HARD_FRAME_POINTER_REGNUM)
        {
          if (current_function_is_leaf)
            offset = -current_frame_info.total_size;
          else
            offset = -(current_frame_info.total_size
                       - current_function_outgoing_args_size - 16);
        }
      else if (to == STACK_POINTER_REGNUM)
        {
          if (current_function_is_leaf)
            offset = 0;
          else
            offset = 16 + current_function_outgoing_args_size;
        }
      else
        abort ();
      break;

    case ARG_POINTER_REGNUM:
      /* Arguments start above the 16 byte save area, unless stdarg
         in which case we store through the 16 byte save area.  */
      if (to == HARD_FRAME_POINTER_REGNUM)
        offset = 16 - current_function_pretend_args_size;
      else if (to == STACK_POINTER_REGNUM)
        offset = (current_frame_info.total_size
                  + 16 - current_function_pretend_args_size);
      else
        abort ();
      break;

    default:
      abort ();
    }

  return offset;
}

/* If there are more than a trivial number of register spills, we use
   two interleaved iterators so that we can get two memory references
   per insn group.

   In order to simplify things in the prologue and epilogue expanders,
   we use helper functions to fix up the memory references after the
   fact with the appropriate offsets to a POST_MODIFY memory mode.
   The following data structure tracks the state of the two iterators
   while insns are being emitted.  */

struct spill_fill_data
{
  rtx init_after;               /* point at which to emit initializations */
  rtx init_reg[2];              /* initial base register */
  rtx iter_reg[2];              /* the iterator registers */
  rtx *prev_addr[2];            /* address of last memory use */
  rtx prev_insn[2];             /* the insn corresponding to prev_addr */
  HOST_WIDE_INT prev_off[2];    /* last offset */
  int n_iter;                   /* number of iterators in use */
  int next_iter;                /* next iterator to use */
  unsigned int save_gr_used_mask;
};

static struct spill_fill_data spill_fill_data;

static void
setup_spill_pointers (int n_spills, rtx init_reg, HOST_WIDE_INT cfa_off)
{
  int i;

  spill_fill_data.init_after = get_last_insn ();
  spill_fill_data.init_reg[0] = init_reg;
  spill_fill_data.init_reg[1] = init_reg;
  spill_fill_data.prev_addr[0] = NULL;
  spill_fill_data.prev_addr[1] = NULL;
  spill_fill_data.prev_insn[0] = NULL;
  spill_fill_data.prev_insn[1] = NULL;
  spill_fill_data.prev_off[0] = cfa_off;
  spill_fill_data.prev_off[1] = cfa_off;
  spill_fill_data.next_iter = 0;
  spill_fill_data.save_gr_used_mask = current_frame_info.gr_used_mask;

  spill_fill_data.n_iter = 1 + (n_spills > 2);
  for (i = 0; i < spill_fill_data.n_iter; ++i)
    {
      int regno = next_scratch_gr_reg ();
      spill_fill_data.iter_reg[i] = gen_rtx_REG (DImode, regno);
      current_frame_info.gr_used_mask |= 1 << regno;
    }
}

static void
finish_spill_pointers (void)
{
  current_frame_info.gr_used_mask = spill_fill_data.save_gr_used_mask;
}

static rtx
spill_restore_mem (rtx reg, HOST_WIDE_INT cfa_off)
{
  int iter = spill_fill_data.next_iter;
  HOST_WIDE_INT disp = spill_fill_data.prev_off[iter] - cfa_off;
  rtx disp_rtx = GEN_INT (disp);
  rtx mem;

  if (spill_fill_data.prev_addr[iter])
    {
      if (CONST_OK_FOR_N (disp))
        {
          *spill_fill_data.prev_addr[iter]
            = gen_rtx_POST_MODIFY (DImode, spill_fill_data.iter_reg[iter],
                                   gen_rtx_PLUS (DImode,
                                                 spill_fill_data.iter_reg[iter],
                                                 disp_rtx));
          REG_NOTES (spill_fill_data.prev_insn[iter])
            = gen_rtx_EXPR_LIST (REG_INC, spill_fill_data.iter_reg[iter],
                                 REG_NOTES (spill_fill_data.prev_insn[iter]));
        }
      else
        {
          /* ??? Could use register post_modify for loads.  */
          if (! CONST_OK_FOR_I (disp))
            {
              rtx tmp = gen_rtx_REG (DImode, next_scratch_gr_reg ());
              emit_move_insn (tmp, disp_rtx);
              disp_rtx = tmp;
            }
          emit_insn (gen_adddi3 (spill_fill_data.iter_reg[iter],
                                 spill_fill_data.iter_reg[iter], disp_rtx));
        }
    }
  /* Micro-optimization: if we've created a frame pointer, it's at
     CFA 0, which may allow the real iterator to be initialized lower,
     slightly increasing parallelism.  Also, if there are few saves
     it may eliminate the iterator entirely.  */
  else if (disp == 0
           && spill_fill_data.init_reg[iter] == stack_pointer_rtx
           && frame_pointer_needed)
    {
      mem = gen_rtx_MEM (GET_MODE (reg), hard_frame_pointer_rtx);
      set_mem_alias_set (mem, get_varargs_alias_set ());
      return mem;
    }
  else
    {
      rtx seq, insn;

      if (disp == 0)
        seq = gen_movdi (spill_fill_data.iter_reg[iter],
                         spill_fill_data.init_reg[iter]);
      else
        {
          start_sequence ();

          if (! CONST_OK_FOR_I (disp))
            {
              rtx tmp = gen_rtx_REG (DImode, next_scratch_gr_reg ());
              emit_move_insn (tmp, disp_rtx);
              disp_rtx = tmp;
            }

          emit_insn (gen_adddi3 (spill_fill_data.iter_reg[iter],
                                 spill_fill_data.init_reg[iter],
                                 disp_rtx));

          seq = get_insns ();
          end_sequence ();
        }

      /* Careful for being the first insn in a sequence.  */
      if (spill_fill_data.init_after)
        insn = emit_insn_after (seq, spill_fill_data.init_after);
      else
        {
          rtx first = get_insns ();
          if (first)
            insn = emit_insn_before (seq, first);
          else
            insn = emit_insn (seq);
        }
      spill_fill_data.init_after = insn;

      /* If DISP is 0, we may or may not have a further adjustment
         afterward.  If we do, then the load/store insn may be modified
         to be a post-modify.  If we don't, then this copy may be
         eliminated by copyprop_hardreg_forward, which makes this
         insn garbage, which runs afoul of the sanity check in
         propagate_one_insn.  So mark this insn as legal to delete.  */
      if (disp == 0)
        REG_NOTES(insn) = gen_rtx_EXPR_LIST (REG_MAYBE_DEAD, const0_rtx,
                                             REG_NOTES (insn));
    }

  mem = gen_rtx_MEM (GET_MODE (reg), spill_fill_data.iter_reg[iter]);

  /* ??? Not all of the spills are for varargs, but some of them are.
     The rest of the spills belong in an alias set of their own.  But
     it doesn't actually hurt to include them here.  */
  set_mem_alias_set (mem, get_varargs_alias_set ());

  spill_fill_data.prev_addr[iter] = &XEXP (mem, 0);
  spill_fill_data.prev_off[iter] = cfa_off;

  if (++iter >= spill_fill_data.n_iter)
    iter = 0;
  spill_fill_data.next_iter = iter;

  return mem;
}

static void
do_spill (rtx (*move_fn) (rtx, rtx, rtx), rtx reg, HOST_WIDE_INT cfa_off,
          rtx frame_reg)
{
  int iter = spill_fill_data.next_iter;
  rtx mem, insn;

  mem = spill_restore_mem (reg, cfa_off);
  insn = emit_insn ((*move_fn) (mem, reg, GEN_INT (cfa_off)));
  spill_fill_data.prev_insn[iter] = insn;

  if (frame_reg)
    {
      rtx base;
      HOST_WIDE_INT off;

      RTX_FRAME_RELATED_P (insn) = 1;

      /* Don't even pretend that the unwind code can intuit its way
         through a pair of interleaved post_modify iterators.  Just
         provide the correct answer.  */

      if (frame_pointer_needed)
        {
          base = hard_frame_pointer_rtx;
          off = - cfa_off;
        }
      else
        {
          base = stack_pointer_rtx;
          off = current_frame_info.total_size - cfa_off;
        }

      REG_NOTES (insn)
        = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR,
                gen_rtx_SET (VOIDmode,
                             gen_rtx_MEM (GET_MODE (reg),
                                          plus_constant (base, off)),
                             frame_reg),
                REG_NOTES (insn));
    }
}

static void
do_restore (rtx (*move_fn) (rtx, rtx, rtx), rtx reg, HOST_WIDE_INT cfa_off)
{
  int iter = spill_fill_data.next_iter;
  rtx insn;

  insn = emit_insn ((*move_fn) (reg, spill_restore_mem (reg, cfa_off),
                                GEN_INT (cfa_off)));
  spill_fill_data.prev_insn[iter] = insn;
}

/* Wrapper functions that discards the CONST_INT spill offset.  These
   exist so that we can give gr_spill/gr_fill the offset they need and
   use a consistent function interface.  */

static rtx
gen_movdi_x (rtx dest, rtx src, rtx offset ATTRIBUTE_UNUSED)
{
  return gen_movdi (dest, src);
}

static rtx
gen_fr_spill_x (rtx dest, rtx src, rtx offset ATTRIBUTE_UNUSED)
{
  return gen_fr_spill (dest, src);
}

static rtx
gen_fr_restore_x (rtx dest, rtx src, rtx offset ATTRIBUTE_UNUSED)
{
  return gen_fr_restore (dest, src);
}

/* Called after register allocation to add any instructions needed for the
   prologue.  Using a prologue insn is favored compared to putting all of the
   instructions in output_function_prologue(), since it allows the scheduler
   to intermix instructions with the saves of the caller saved registers.  In
   some cases, it might be necessary to emit a barrier instruction as the last
   insn to prevent such scheduling.

   Also any insns generated here should have RTX_FRAME_RELATED_P(insn) = 1
   so that the debug info generation code can handle them properly.

   The register save area is layed out like so:
   cfa+16
        [ varargs spill area ]
        [ fr register spill area ]
        [ br register spill area ]
        [ ar register spill area ]
        [ pr register spill area ]
        [ gr register spill area ] */

/* ??? Get inefficient code when the frame size is larger than can fit in an
   adds instruction.  */

void
ia64_expand_prologue (void)
{
  rtx insn, ar_pfs_save_reg, ar_unat_save_reg;
  int i, epilogue_p, regno, alt_regno, cfa_off, n_varargs;
  rtx reg, alt_reg;

  ia64_compute_frame_size (get_frame_size ());
  last_scratch_gr_reg = 15;

  /* If there is no epilogue, then we don't need some prologue insns.
     We need to avoid emitting the dead prologue insns, because flow
     will complain about them.  */
  if (optimize)
    {
      edge e;
      edge_iterator ei;

      FOR_EACH_EDGE (e, ei, EXIT_BLOCK_PTR->preds)
        if ((e->flags & EDGE_FAKE) == 0
            && (e->flags & EDGE_FALLTHRU) != 0)
          break;
      epilogue_p = (e != NULL);
    }
  else
    epilogue_p = 1;

  /* Set the local, input, and output register names.  We need to do this
     for GNU libc, which creates crti.S/crtn.S by splitting initfini.c in
     half.  If we use in/loc/out register names, then we get assembler errors
     in crtn.S because there is no alloc insn or regstk directive in there.  */
  if (! TARGET_REG_NAMES)
    {
      int inputs = current_frame_info.n_input_regs;
      int locals = current_frame_info.n_local_regs;
      int outputs = current_frame_info.n_output_regs;

      for (i = 0; i < inputs; i++)
        reg_names[IN_REG (i)] = ia64_reg_numbers[i];
      for (i = 0; i < locals; i++)
        reg_names[LOC_REG (i)] = ia64_reg_numbers[inputs + i];
      for (i = 0; i < outputs; i++)
        reg_names[OUT_REG (i)] = ia64_reg_numbers[inputs + locals + i];
    }

  /* Set the frame pointer register name.  The regnum is logically loc79,
     but of course we'll not have allocated that many locals.  Rather than
     worrying about renumbering the existing rtxs, we adjust the name.  */
  /* ??? This code means that we can never use one local register when
     there is a frame pointer.  loc79 gets wasted in this case, as it is
     renamed to a register that will never be used.  See also the try_locals
     code in find_gr_spill.  */
  if (current_frame_info.reg_fp)
    {
      const char *tmp = reg_names[HARD_FRAME_POINTER_REGNUM];
      reg_names[HARD_FRAME_POINTER_REGNUM]
        = reg_names[current_frame_info.reg_fp];
      reg_names[current_frame_info.reg_fp] = tmp;
    }

  /* We don't need an alloc instruction if we've used no outputs or locals.  */
  if (current_frame_info.n_local_regs == 0
      && current_frame_info.n_output_regs == 0
      && current_frame_info.n_input_regs <= current_function_args_info.int_regs
      && !TEST_HARD_REG_BIT (current_frame_info.mask, AR_PFS_REGNUM))
    {
      /* If there is no alloc, but there are input registers used, then we
         need a .regstk directive.  */
      current_frame_info.need_regstk = (TARGET_REG_NAMES != 0);
      ar_pfs_save_reg = NULL_RTX;
    }
  else
    {
      current_frame_info.need_regstk = 0;

      if (current_frame_info.reg_save_ar_pfs)
        regno = current_frame_info.reg_save_ar_pfs;
      else
        regno = next_scratch_gr_reg ();
      ar_pfs_save_reg = gen_rtx_REG (DImode, regno);

      insn = emit_insn (gen_alloc (ar_pfs_save_reg,
                                   GEN_INT (current_frame_info.n_input_regs),
                                   GEN_INT (current_frame_info.n_local_regs),
                                   GEN_INT (current_frame_info.n_output_regs),
                                   GEN_INT (current_frame_info.n_rotate_regs)));
      RTX_FRAME_RELATED_P (insn) = (current_frame_info.reg_save_ar_pfs != 0);
    }

  /* Set up frame pointer, stack pointer, and spill iterators.  */

  n_varargs = cfun->machine->n_varargs;
  setup_spill_pointers (current_frame_info.n_spilled + n_varargs,
                        stack_pointer_rtx, 0);

  if (frame_pointer_needed)
    {
      insn = emit_move_insn (hard_frame_pointer_rtx, stack_pointer_rtx);
      RTX_FRAME_RELATED_P (insn) = 1;
    }

  if (current_frame_info.total_size != 0)
    {
      rtx frame_size_rtx = GEN_INT (- current_frame_info.total_size);
      rtx offset;

      if (CONST_OK_FOR_I (- current_frame_info.total_size))
        offset = frame_size_rtx;
      else
        {
          regno = next_scratch_gr_reg ();
          offset = gen_rtx_REG (DImode, regno);
          emit_move_insn (offset, frame_size_rtx);
        }

      insn = emit_insn (gen_adddi3 (stack_pointer_rtx,
                                    stack_pointer_rtx, offset));

      if (! frame_pointer_needed)
        {
          RTX_FRAME_RELATED_P (insn) = 1;
          if (GET_CODE (offset) != CONST_INT)
            {
              REG_NOTES (insn)
                = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR,
                        gen_rtx_SET (VOIDmode,
                                     stack_pointer_rtx,
                                     gen_rtx_PLUS (DImode,
                                                   stack_pointer_rtx,
                                                   frame_size_rtx)),
                        REG_NOTES (insn));
            }
        }

      /* ??? At this point we must generate a magic insn that appears to
         modify the stack pointer, the frame pointer, and all spill
         iterators.  This would allow the most scheduling freedom.  For
         now, just hard stop.  */
      emit_insn (gen_blockage ());
    }

  /* Must copy out ar.unat before doing any integer spills.  */
  if (TEST_HARD_REG_BIT (current_frame_info.mask, AR_UNAT_REGNUM))
    {
      if (current_frame_info.reg_save_ar_unat)
        ar_unat_save_reg
          = gen_rtx_REG (DImode, current_frame_info.reg_save_ar_unat);
      else
        {
          alt_regno = next_scratch_gr_reg ();
          ar_unat_save_reg = gen_rtx_REG (DImode, alt_regno);
          current_frame_info.gr_used_mask |= 1 << alt_regno;
        }

      reg = gen_rtx_REG (DImode, AR_UNAT_REGNUM);
      insn = emit_move_insn (ar_unat_save_reg, reg);
      RTX_FRAME_RELATED_P (insn) = (current_frame_info.reg_save_ar_unat != 0);

      /* Even if we're not going to generate an epilogue, we still
         need to save the register so that EH works.  */
      if (! epilogue_p && current_frame_info.reg_save_ar_unat)
        emit_insn (gen_prologue_use (ar_unat_save_reg));
    }
  else
    ar_unat_save_reg = NULL_RTX;

  /* Spill all varargs registers.  Do this before spilling any GR registers,
     since we want the UNAT bits for the GR registers to override the UNAT
     bits from varargs, which we don't care about.  */

  cfa_off = -16;
  for (regno = GR_ARG_FIRST + 7; n_varargs > 0; --n_varargs, --regno)
    {
      reg = gen_rtx_REG (DImode, regno);
      do_spill (gen_gr_spill, reg, cfa_off += 8, NULL_RTX);
    }

  /* Locate the bottom of the register save area.  */
  cfa_off = (current_frame_info.spill_cfa_off
             + current_frame_info.spill_size
             + current_frame_info.extra_spill_size);

  /* Save the predicate register block either in a register or in memory.  */
  if (TEST_HARD_REG_BIT (current_frame_info.mask, PR_REG (0)))
    {
      reg = gen_rtx_REG (DImode, PR_REG (0));
      if (current_frame_info.reg_save_pr != 0)
        {
          alt_reg = gen_rtx_REG (DImode, current_frame_info.reg_save_pr);
          insn = emit_move_insn (alt_reg, reg);

          /* ??? Denote pr spill/fill by a DImode move that modifies all
             64 hard registers.  */
          RTX_FRAME_RELATED_P (insn) = 1;
          REG_NOTES (insn)
            = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR,
                        gen_rtx_SET (VOIDmode, alt_reg, reg),
                        REG_NOTES (insn));

          /* Even if we're not going to generate an epilogue, we still
             need to save the register so that EH works.  */
          if (! epilogue_p)
            emit_insn (gen_prologue_use (alt_reg));
        }
      else
        {
          alt_regno = next_scratch_gr_reg ();
          alt_reg = gen_rtx_REG (DImode, alt_regno);
          insn = emit_move_insn (alt_reg, reg);
          do_spill (gen_movdi_x, alt_reg, cfa_off, reg);
          cfa_off -= 8;
        }
    }

  /* Handle AR regs in numerical order.  All of them get special handling.  */
  if (TEST_HARD_REG_BIT (current_frame_info.mask, AR_UNAT_REGNUM)
      && current_frame_info.reg_save_ar_unat == 0)
    {
      reg = gen_rtx_REG (DImode, AR_UNAT_REGNUM);
      do_spill (gen_movdi_x, ar_unat_save_reg, cfa_off, reg);
      cfa_off -= 8;
    }

  /* The alloc insn already copied ar.pfs into a general register.  The
     only thing we have to do now is copy that register to a stack slot
     if we'd not allocated a local register for the job.  */
  if (TEST_HARD_REG_BIT (current_frame_info.mask, AR_PFS_REGNUM)
      && current_frame_info.reg_save_ar_pfs == 0)
    {
      reg = gen_rtx_REG (DImode, AR_PFS_REGNUM);
      do_spill (gen_movdi_x, ar_pfs_save_reg, cfa_off, reg);
      cfa_off -= 8;
    }

  if (TEST_HARD_REG_BIT (current_frame_info.mask, AR_LC_REGNUM))
    {
      reg = gen_rtx_REG (DImode, AR_LC_REGNUM);
      if (current_frame_info.reg_save_ar_lc != 0)
        {
          alt_reg = gen_rtx_REG (DImode, current_frame_info.reg_save_ar_lc);
          insn = emit_move_insn (alt_reg, reg);
          RTX_FRAME_RELATED_P (insn) = 1;

          /* Even if we're not going to generate an epilogue, we still
             need to save the register so that EH works.  */
          if (! epilogue_p)
            emit_insn (gen_prologue_use (alt_reg));
        }
      else
        {
          alt_regno = next_scratch_gr_reg ();
          alt_reg = gen_rtx_REG (DImode, alt_regno);
          emit_move_insn (alt_reg, reg);
          do_spill (gen_movdi_x, alt_reg, cfa_off, reg);
          cfa_off -= 8;
        }
    }

  if (current_frame_info.reg_save_gp)
    {
      insn = emit_move_insn (gen_rtx_REG (DImode,
                                          current_frame_info.reg_save_gp),
                             pic_offset_table_rtx);
      /* We don't know for sure yet if this is actually needed, since
         we've not split the PIC call patterns.  If all of the calls
         are indirect, and not followed by any uses of the gp, then
         this save is dead.  Allow it to go away.  */
      REG_NOTES (insn)
        = gen_rtx_EXPR_LIST (REG_MAYBE_DEAD, const0_rtx, REG_NOTES (insn));
    }

  /* We should now be at the base of the gr/br/fr spill area.  */
  if (cfa_off != (current_frame_info.spill_cfa_off
                  + current_frame_info.spill_size))
    abort ();

  /* Spill all general registers.  */
  for (regno = GR_REG (1); regno <= GR_REG (31); ++regno)
    if (TEST_HARD_REG_BIT (current_frame_info.mask, regno))
      {
        reg = gen_rtx_REG (DImode, regno);
        do_spill (gen_gr_spill, reg, cfa_off, reg);
        cfa_off -= 8;
      }

  /* Handle BR0 specially -- it may be getting stored permanently in
     some GR register.  */
  if (TEST_HARD_REG_BIT (current_frame_info.mask, BR_REG (0)))
    {
      reg = gen_rtx_REG (DImode, BR_REG (0));
      if (current_frame_info.reg_save_b0 != 0)
        {
          alt_reg = gen_rtx_REG (DImode, current_frame_info.reg_save_b0);
          insn = emit_move_insn (alt_reg, reg);
          RTX_FRAME_RELATED_P (insn) = 1;

          /* Even if we're not going to generate an epilogue, we still
             need to save the register so that EH works.  */
          if (! epilogue_p)
            emit_insn (gen_prologue_use (alt_reg));
        }
      else
        {
          alt_regno = next_scratch_gr_reg ();
          alt_reg = gen_rtx_REG (DImode, alt_regno);
          emit_move_insn (alt_reg, reg);
          do_spill (gen_movdi_x, alt_reg, cfa_off, reg);
          cfa_off -= 8;
        }
    }

  /* Spill the rest of the BR registers.  */
  for (regno = BR_REG (1); regno <= BR_REG (7); ++regno)
    if (TEST_HARD_REG_BIT (current_frame_info.mask, regno))
      {
        alt_regno = next_scratch_gr_reg ();
        alt_reg = gen_rtx_REG (DImode, alt_regno);
        reg = gen_rtx_REG (DImode, regno);
        emit_move_insn (alt_reg, reg);
        do_spill (gen_movdi_x, alt_reg, cfa_off, reg);
        cfa_off -= 8;
      }

  /* Align the frame and spill all FR registers.  */
  for (regno = FR_REG (2); regno <= FR_REG (127); ++regno)
    if (TEST_HARD_REG_BIT (current_frame_info.mask, regno))
      {
        if (cfa_off & 15)
          abort ();
        reg = gen_rtx_REG (XFmode, regno);
        do_spill (gen_fr_spill_x, reg, cfa_off, reg);
        cfa_off -= 16;
      }

  if (cfa_off != current_frame_info.spill_cfa_off)
    abort ();

  finish_spill_pointers ();
}

/* Called after register allocation to add any instructions needed for the
   epilogue.  Using an epilogue insn is favored compared to putting all of the
   instructions in output_function_prologue(), since it allows the scheduler
   to intermix instructions with the saves of the caller saved registers.  In
   some cases, it might be necessary to emit a barrier instruction as the last
   insn to prevent such scheduling.  */

void
ia64_expand_epilogue (int sibcall_p)
{
  rtx insn, reg, alt_reg, ar_unat_save_reg;
  int regno, alt_regno, cfa_off;

  ia64_compute_frame_size (get_frame_size ());

  /* If there is a frame pointer, then we use it instead of the stack
     pointer, so that the stack pointer does not need to be valid when
     the epilogue starts.  See EXIT_IGNORE_STACK.  */
  if (frame_pointer_needed)
    setup_spill_pointers (current_frame_info.n_spilled,
                          hard_frame_pointer_rtx, 0);
  else
    setup_spill_pointers (current_frame_info.n_spilled, stack_pointer_rtx,
                          current_frame_info.total_size);

  if (current_frame_info.total_size != 0)
    {
      /* ??? At this point we must generate a magic insn that appears to
         modify the spill iterators and the frame pointer.  This would
         allow the most scheduling freedom.  For now, just hard stop.  */
      emit_insn (gen_blockage ());
    }

  /* Locate the bottom of the register save area.  */
  cfa_off = (current_frame_info.spill_cfa_off
             + current_frame_info.spill_size
             + current_frame_info.extra_spill_size);

  /* Restore the predicate registers.  */
  if (TEST_HARD_REG_BIT (current_frame_info.mask, PR_REG (0)))
    {
      if (current_frame_info.reg_save_pr != 0)
        alt_reg = gen_rtx_REG (DImode, current_frame_info.reg_save_pr);
      else
        {
          alt_regno = next_scratch_gr_reg ();
          alt_reg = gen_rtx_REG (DImode, alt_regno);
          do_restore (gen_movdi_x, alt_reg, cfa_off);
          cfa_off -= 8;
        }
      reg = gen_rtx_REG (DImode, PR_REG (0));
      emit_move_insn (reg, alt_reg);
    }

  /* Restore the application registers.  */

  /* Load the saved unat from the stack, but do not restore it until
     after the GRs have been restored.  */
  if (TEST_HARD_REG_BIT (current_frame_info.mask, AR_UNAT_REGNUM))
    {
      if (current_frame_info.reg_save_ar_unat != 0)
        ar_unat_save_reg
          = gen_rtx_REG (DImode, current_frame_info.reg_save_ar_unat);
      else
        {
          alt_regno = next_scratch_gr_reg ();
          ar_unat_save_reg = gen_rtx_REG (DImode, alt_regno);
          current_frame_info.gr_used_mask |= 1 << alt_regno;
          do_restore (gen_movdi_x, ar_unat_save_reg, cfa_off);
          cfa_off -= 8;
        }
    }
  else
    ar_unat_save_reg = NULL_RTX;

  if (current_frame_info.reg_save_ar_pfs != 0)
    {
      alt_reg = gen_rtx_REG (DImode, current_frame_info.reg_save_ar_pfs);
      reg = gen_rtx_REG (DImode, AR_PFS_REGNUM);
      emit_move_insn (reg, alt_reg);
    }
  else if (TEST_HARD_REG_BIT (current_frame_info.mask, AR_PFS_REGNUM))
    {
      alt_regno = next_scratch_gr_reg ();
      alt_reg = gen_rtx_REG (DImode, alt_regno);
      do_restore (gen_movdi_x, alt_reg, cfa_off);
      cfa_off -= 8;
      reg = gen_rtx_REG (DImode, AR_PFS_REGNUM);
      emit_move_insn (reg, alt_reg);
    }

  if (TEST_HARD_REG_BIT (current_frame_info.mask, AR_LC_REGNUM))
    {
      if (current_frame_info.reg_save_ar_lc != 0)
        alt_reg = gen_rtx_REG (DImode, current_frame_info.reg_save_ar_lc);
      else
        {
          alt_regno = next_scratch_gr_reg ();
          alt_reg = gen_rtx_REG (DImode, alt_regno);
          do_restore (gen_movdi_x, alt_reg, cfa_off);
          cfa_off -= 8;
        }
      reg = gen_rtx_REG (DImode, AR_LC_REGNUM);
      emit_move_insn (reg, alt_reg);
    }

  /* We should now be at the base of the gr/br/fr spill area.  */
  if (cfa_off != (current_frame_info.spill_cfa_off
                  + current_frame_info.spill_size))
    abort ();

  /* The GP may be stored on the stack in the prologue, but it's
     never restored in the epilogue.  Skip the stack slot.  */
  if (TEST_HARD_REG_BIT (current_frame_info.mask, GR_REG (1)))
    cfa_off -= 8;

  /* Restore all general registers.  */
  for (regno = GR_REG (2); regno <= GR_REG (31); ++regno)
    if (TEST_HARD_REG_BIT (current_frame_info.mask, regno))
      {
        reg = gen_rtx_REG (DImode, regno);
        do_restore (gen_gr_restore, reg, cfa_off);
        cfa_off -= 8;
      }

  /* Restore the branch registers.  Handle B0 specially, as it may
     have gotten stored in some GR register.  */
  if (TEST_HARD_REG_BIT (current_frame_info.mask, BR_REG (0)))
    {
      if (current_frame_info.reg_save_b0 != 0)
        alt_reg = gen_rtx_REG (DImode, current_frame_info.reg_save_b0);
      else
        {
          alt_regno = next_scratch_gr_reg ();
          alt_reg = gen_rtx_REG (DImode, alt_regno);
          do_restore (gen_movdi_x, alt_reg, cfa_off);
          cfa_off -= 8;
        }
      reg = gen_rtx_REG (DImode, BR_REG (0));
      emit_move_insn (reg, alt_reg);
    }

  for (regno = BR_REG (1); regno <= BR_REG (7); ++regno)
    if (TEST_HARD_REG_BIT (current_frame_info.mask, regno))
      {
        alt_regno = next_scratch_gr_reg ();
        alt_reg = gen_rtx_REG (DImode, alt_regno);
        do_restore (gen_movdi_x, alt_reg, cfa_off);
        cfa_off -= 8;
        reg = gen_rtx_REG (DImode, regno);
        emit_move_insn (reg, alt_reg);
      }

  /* Restore floating point registers.  */
  for (regno = FR_REG (2); regno <= FR_REG (127); ++regno)
    if (TEST_HARD_REG_BIT (current_frame_info.mask, regno))
      {
        if (cfa_off & 15)
          abort ();
        reg = gen_rtx_REG (XFmode, regno);
        do_restore (gen_fr_restore_x, reg, cfa_off);
        cfa_off -= 16;
      }

  /* Restore ar.unat for real.  */
  if (TEST_HARD_REG_BIT (current_frame_info.mask, AR_UNAT_REGNUM))
    {
      reg = gen_rtx_REG (DImode, AR_UNAT_REGNUM);
      emit_move_insn (reg, ar_unat_save_reg);
    }

  if (cfa_off != current_frame_info.spill_cfa_off)
    abort ();

  finish_spill_pointers ();

  if (current_frame_info.total_size || cfun->machine->ia64_eh_epilogue_sp)
    {
      /* ??? At this point we must generate a magic insn that appears to
         modify the spill iterators, the stack pointer, and the frame
         pointer.  This would allow the most scheduling freedom.  For now,
         just hard stop.  */
      emit_insn (gen_blockage ());
    }

  if (cfun->machine->ia64_eh_epilogue_sp)
    emit_move_insn (stack_pointer_rtx, cfun->machine->ia64_eh_epilogue_sp);
  else if (frame_pointer_needed)
    {
      insn = emit_move_insn (stack_pointer_rtx, hard_frame_pointer_rtx);
      RTX_FRAME_RELATED_P (insn) = 1;
    }
  else if (current_frame_info.total_size)
    {
      rtx offset, frame_size_rtx;

      frame_size_rtx = GEN_INT (current_frame_info.total_size);
      if (CONST_OK_FOR_I (current_frame_info.total_size))
        offset = frame_size_rtx;
      else
        {
          regno = next_scratch_gr_reg ();
          offset = gen_rtx_REG (DImode, regno);
          emit_move_insn (offset, frame_size_rtx);
        }

      insn = emit_insn (gen_adddi3 (stack_pointer_rtx, stack_pointer_rtx,
                                    offset));

      RTX_FRAME_RELATED_P (insn) = 1;
      if (GET_CODE (offset) != CONST_INT)
        {
          REG_NOTES (insn)
            = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR,
                        gen_rtx_SET (VOIDmode,
                                     stack_pointer_rtx,
                                     gen_rtx_PLUS (DImode,
                                                   stack_pointer_rtx,
                                                   frame_size_rtx)),
                        REG_NOTES (insn));
        }
    }

  if (cfun->machine->ia64_eh_epilogue_bsp)
    emit_insn (gen_set_bsp (cfun->machine->ia64_eh_epilogue_bsp));

  if (! sibcall_p)
    emit_jump_insn (gen_return_internal (gen_rtx_REG (DImode, BR_REG (0))));
  else
    {
      int fp = GR_REG (2);
      /* We need a throw away register here, r0 and r1 are reserved, so r2 is the
         first available call clobbered register.  If there was a frame_pointer
         register, we may have swapped the names of r2 and HARD_FRAME_POINTER_REGNUM,
         so we have to make sure we're using the string "r2" when emitting
         the register name for the assembler.  */
      if (current_frame_info.reg_fp && current_frame_info.reg_fp == GR_REG (2))
        fp = HARD_FRAME_POINTER_REGNUM;

      /* We must emit an alloc to force the input registers to become output
         registers.  Otherwise, if the callee tries to pass its parameters
         through to another call without an intervening alloc, then these
         values get lost.  */
      /* ??? We don't need to preserve all input registers.  We only need to
         preserve those input registers used as arguments to the sibling call.
         It is unclear how to compute that number here.  */
      if (current_frame_info.n_input_regs != 0)
        emit_insn (gen_alloc (gen_rtx_REG (DImode, fp),
                              const0_rtx, const0_rtx,
                              GEN_INT (current_frame_info.n_input_regs),
                              const0_rtx));
    }
}

/* Return 1 if br.ret can do all the work required to return from a
   function.  */

int
ia64_direct_return (void)
{
  if (reload_completed && ! frame_pointer_needed)
    {
      ia64_compute_frame_size (get_frame_size ());

      return (current_frame_info.total_size == 0
              && current_frame_info.n_spilled == 0
              && current_frame_info.reg_save_b0 == 0
              && current_frame_info.reg_save_pr == 0
              && current_frame_info.reg_save_ar_pfs == 0
              && current_frame_info.reg_save_ar_unat == 0
              && current_frame_info.reg_save_ar_lc == 0);
    }
  return 0;
}

/* Return the magic cookie that we use to hold the return address
   during early compilation.  */

rtx
ia64_return_addr_rtx (HOST_WIDE_INT count, rtx frame ATTRIBUTE_UNUSED)
{
  if (count != 0)
    return NULL;
  return gen_rtx_UNSPEC (Pmode, gen_rtvec (1, const0_rtx), UNSPEC_RET_ADDR);
}

/* Split this value after reload, now that we know where the return
   address is saved.  */

void
ia64_split_return_addr_rtx (rtx dest)
{
  rtx src;

  if (TEST_HARD_REG_BIT (current_frame_info.mask, BR_REG (0)))
    {
      if (current_frame_info.reg_save_b0 != 0)
        src = gen_rtx_REG (DImode, current_frame_info.reg_save_b0);
      else
        {
          HOST_WIDE_INT off;
          unsigned int regno;

          /* Compute offset from CFA for BR0.  */
          /* ??? Must be kept in sync with ia64_expand_prologue.  */
          off = (current_frame_info.spill_cfa_off
                 + current_frame_info.spill_size);
          for (regno = GR_REG (1); regno <= GR_REG (31); ++regno)
            if (TEST_HARD_REG_BIT (current_frame_info.mask, regno))
              off -= 8;

          /* Convert CFA offset to a register based offset.  */
          if (frame_pointer_needed)
            src = hard_frame_pointer_rtx;
          else
            {
              src = stack_pointer_rtx;
              off += current_frame_info.total_size;
            }

          /* Load address into scratch register.  */
          if (CONST_OK_FOR_I (off))
            emit_insn (gen_adddi3 (dest, src, GEN_INT (off)));
          else
            {
              emit_move_insn (dest, GEN_INT (off));
              emit_insn (gen_adddi3 (dest, src, dest));
            }

          src = gen_rtx_MEM (Pmode, dest);
        }
    }
  else
    src = gen_rtx_REG (DImode, BR_REG (0));

  emit_move_insn (dest, src);
}

int
ia64_hard_regno_rename_ok (int from, int to)
{
  /* Don't clobber any of the registers we reserved for the prologue.  */
  if (to == current_frame_info.reg_fp
      || to == current_frame_info.reg_save_b0
      || to == current_frame_info.reg_save_pr
      || to == current_frame_info.reg_save_ar_pfs
      || to == current_frame_info.reg_save_ar_unat
      || to == current_frame_info.reg_save_ar_lc)
    return 0;

  if (from == current_frame_info.reg_fp
      || from == current_frame_info.reg_save_b0
      || from == current_frame_info.reg_save_pr
      || from == current_frame_info.reg_save_ar_pfs
      || from == current_frame_info.reg_save_ar_unat
      || from == current_frame_info.reg_save_ar_lc)
    return 0;

  /* Don't use output registers outside the register frame.  */
  if (OUT_REGNO_P (to) && to >= OUT_REG (current_frame_info.n_output_regs))
    return 0;

  /* Retain even/oddness on predicate register pairs.  */
  if (PR_REGNO_P (from) && PR_REGNO_P (to))
    return (from & 1) == (to & 1);

  return 1;
}

/* Target hook for assembling integer objects.  Handle word-sized
   aligned objects and detect the cases when @fptr is needed.  */

static bool
ia64_assemble_integer (rtx x, unsigned int size, int aligned_p)
{
  if (size == POINTER_SIZE / BITS_PER_UNIT
      && aligned_p
      && !(TARGET_NO_PIC || TARGET_AUTO_PIC)
      && GET_CODE (x) == SYMBOL_REF
      && SYMBOL_REF_FUNCTION_P (x))
    {
      if (POINTER_SIZE == 32)
        fputs ("\tdata4\t@fptr(", asm_out_file);
      else
        fputs ("\tdata8\t@fptr(", asm_out_file);
      output_addr_const (asm_out_file, x);
      fputs (")\n", asm_out_file);
      return true;
    }
  return default_assemble_integer (x, size, aligned_p);
}

/* Emit the function prologue.  */

static void
ia64_output_function_prologue (FILE *file, HOST_WIDE_INT size ATTRIBUTE_UNUSED)
{
  int mask, grsave, grsave_prev;

  if (current_frame_info.need_regstk)
    fprintf (file, "\t.regstk %d, %d, %d, %d\n",
             current_frame_info.n_input_regs,
             current_frame_info.n_local_regs,
             current_frame_info.n_output_regs,
             current_frame_info.n_rotate_regs);

  if (!flag_unwind_tables && (!flag_exceptions || USING_SJLJ_EXCEPTIONS))
    return;

  /* Emit the .prologue directive.  */

  mask = 0;
  grsave = grsave_prev = 0;
  if (current_frame_info.reg_save_b0 != 0)
    {
      mask |= 8;
      grsave = grsave_prev = current_frame_info.reg_save_b0;
    }
  if (current_frame_info.reg_save_ar_pfs != 0
      && (grsave_prev == 0
          || current_frame_info.reg_save_ar_pfs == grsave_prev + 1))
    {
      mask |= 4;
      if (grsave_prev == 0)
        grsave = current_frame_info.reg_save_ar_pfs;
      grsave_prev = current_frame_info.reg_save_ar_pfs;
    }
  if (current_frame_info.reg_fp != 0
      && (grsave_prev == 0
          || current_frame_info.reg_fp == grsave_prev + 1))
    {
      mask |= 2;
      if (grsave_prev == 0)
        grsave = HARD_FRAME_POINTER_REGNUM;
      grsave_prev = current_frame_info.reg_fp;
    }
  if (current_frame_info.reg_save_pr != 0
      && (grsave_prev == 0
          || current_frame_info.reg_save_pr == grsave_prev + 1))
    {
      mask |= 1;
      if (grsave_prev == 0)
        grsave = current_frame_info.reg_save_pr;
    }

  if (mask && TARGET_GNU_AS)
    fprintf (file, "\t.prologue %d, %d\n", mask,
             ia64_dbx_register_number (grsave));
  else
    fputs ("\t.prologue\n", file);

  /* Emit a .spill directive, if necessary, to relocate the base of
     the register spill area.  */
  if (current_frame_info.spill_cfa_off != -16)
    fprintf (file, "\t.spill %ld\n",
             (long) (current_frame_info.spill_cfa_off
                     + current_frame_info.spill_size));
}

/* Emit the .body directive at the scheduled end of the prologue.  */

static void
ia64_output_function_end_prologue (FILE *file)
{
  if (!flag_unwind_tables && (!flag_exceptions || USING_SJLJ_EXCEPTIONS))
    return;

  fputs ("\t.body\n", file);
}

/* Emit the function epilogue.  */

static void
ia64_output_function_epilogue (FILE *file ATTRIBUTE_UNUSED,
                               HOST_WIDE_INT size ATTRIBUTE_UNUSED)
{
  int i;

  if (current_frame_info.reg_fp)
    {
      const char *tmp = reg_names[HARD_FRAME_POINTER_REGNUM];
      reg_names[HARD_FRAME_POINTER_REGNUM]
        = reg_names[current_frame_info.reg_fp];
      reg_names[current_frame_info.reg_fp] = tmp;
    }
  if (! TARGET_REG_NAMES)
    {
      for (i = 0; i < current_frame_info.n_input_regs; i++)
        reg_names[IN_REG (i)] = ia64_input_reg_names[i];
      for (i = 0; i < current_frame_info.n_local_regs; i++)
        reg_names[LOC_REG (i)] = ia64_local_reg_names[i];
      for (i = 0; i < current_frame_info.n_output_regs; i++)
        reg_names[OUT_REG (i)] = ia64_output_reg_names[i];
    }

  current_frame_info.initialized = 0;
}

int
ia64_dbx_register_number (int regno)
{
  /* In ia64_expand_prologue we quite literally renamed the frame pointer
     from its home at loc79 to something inside the register frame.  We
     must perform the same renumbering here for the debug info.  */
  if (current_frame_info.reg_fp)
    {
      if (regno == HARD_FRAME_POINTER_REGNUM)
        regno = current_frame_info.reg_fp;
      else if (regno == current_frame_info.reg_fp)
        regno = HARD_FRAME_POINTER_REGNUM;
    }

  if (IN_REGNO_P (regno))
    return 32 + regno - IN_REG (0);
  else if (LOC_REGNO_P (regno))
    return 32 + current_frame_info.n_input_regs + regno - LOC_REG (0);
  else if (OUT_REGNO_P (regno))
    return (32 + current_frame_info.n_input_regs
            + current_frame_info.n_local_regs + regno - OUT_REG (0));
  else
    return regno;
}

void
ia64_initialize_trampoline (rtx addr, rtx fnaddr, rtx static_chain)
{
  rtx addr_reg, eight = GEN_INT (8);

  /* The Intel assembler requires that the global __ia64_trampoline symbol
     be declared explicitly */
  if (!TARGET_GNU_AS)
    {
      static bool declared_ia64_trampoline = false;

      if (!declared_ia64_trampoline)
        {
          declared_ia64_trampoline = true;
          (*targetm.asm_out.globalize_label) (asm_out_file,
                                              "__ia64_trampoline");
        }
    }

  /* Make sure addresses are Pmode even if we are in ILP32 mode. */
  addr = convert_memory_address (Pmode, addr);
  fnaddr = convert_memory_address (Pmode, fnaddr);
  static_chain = convert_memory_address (Pmode, static_chain);

  /* Load up our iterator.  */
  addr_reg = gen_reg_rtx (Pmode);
  emit_move_insn (addr_reg, addr);

  /* The first two words are the fake descriptor:
     __ia64_trampoline, ADDR+16.  */
  emit_move_insn (gen_rtx_MEM (Pmode, addr_reg),
                  gen_rtx_SYMBOL_REF (Pmode, "__ia64_trampoline"));
  emit_insn (gen_adddi3 (addr_reg, addr_reg, eight));

  emit_move_insn (gen_rtx_MEM (Pmode, addr_reg),
                  copy_to_reg (plus_constant (addr, 16)));
  emit_insn (gen_adddi3 (addr_reg, addr_reg, eight));

  /* The third word is the target descriptor.  */
  emit_move_insn (gen_rtx_MEM (Pmode, addr_reg), fnaddr);
  emit_insn (gen_adddi3 (addr_reg, addr_reg, eight));

  /* The fourth word is the static chain.  */
  emit_move_insn (gen_rtx_MEM (Pmode, addr_reg), static_chain);
}
\f
/* Do any needed setup for a variadic function.  CUM has not been updated
   for the last named argument which has type TYPE and mode MODE.

   We generate the actual spill instructions during prologue generation.  */

static void
ia64_setup_incoming_varargs (CUMULATIVE_ARGS *cum, enum machine_mode mode,
                             tree type, int * pretend_size,
                             int second_time ATTRIBUTE_UNUSED)
{
  CUMULATIVE_ARGS next_cum = *cum;

  /* Skip the current argument.  */
  ia64_function_arg_advance (&next_cum, mode, type, 1);

  if (next_cum.words < MAX_ARGUMENT_SLOTS)
    {
      int n = MAX_ARGUMENT_SLOTS - next_cum.words;
      *pretend_size = n * UNITS_PER_WORD;
      cfun->machine->n_varargs = n;
    }
}

/* Check whether TYPE is a homogeneous floating point aggregate.  If
   it is, return the mode of the floating point type that appears
   in all leafs.  If it is not, return VOIDmode.

   An aggregate is a homogeneous floating point aggregate is if all
   fields/elements in it have the same floating point type (e.g,
   SFmode).  128-bit quad-precision floats are excluded.  */

static enum machine_mode
hfa_element_mode (tree type, int nested)
{
  enum machine_mode element_mode = VOIDmode;
  enum machine_mode mode;
  enum tree_code code = TREE_CODE (type);
  int know_element_mode = 0;
  tree t;

  switch (code)
    {
    case VOID_TYPE:     case INTEGER_TYPE:      case ENUMERAL_TYPE:
    case BOOLEAN_TYPE:  case CHAR_TYPE:         case POINTER_TYPE:
    case OFFSET_TYPE:   case REFERENCE_TYPE:    case METHOD_TYPE:
    case FILE_TYPE:     case SET_TYPE:          case LANG_TYPE:
    case FUNCTION_TYPE:
      return VOIDmode;

      /* Fortran complex types are supposed to be HFAs, so we need to handle
         gcc's COMPLEX_TYPEs as HFAs.  We need to exclude the integral complex
         types though.  */
    case COMPLEX_TYPE:
      if (GET_MODE_CLASS (TYPE_MODE (type)) == MODE_COMPLEX_FLOAT
          && TYPE_MODE (type) != TCmode)
        return GET_MODE_INNER (TYPE_MODE (type));
      else
        return VOIDmode;

    case REAL_TYPE:
      /* We want to return VOIDmode for raw REAL_TYPEs, but the actual
         mode if this is contained within an aggregate.  */
      if (nested && TYPE_MODE (type) != TFmode)
        return TYPE_MODE (type);
      else
        return VOIDmode;

    case ARRAY_TYPE:
      return hfa_element_mode (TREE_TYPE (type), 1);

    case RECORD_TYPE:
    case UNION_TYPE:
    case QUAL_UNION_TYPE:
      for (t = TYPE_FIELDS (type); t; t = TREE_CHAIN (t))
        {
          if (TREE_CODE (t) != FIELD_DECL)
            continue;

          mode = hfa_element_mode (TREE_TYPE (t), 1);
          if (know_element_mode)
            {
              if (mode != element_mode)
                return VOIDmode;
            }
          else if (GET_MODE_CLASS (mode) != MODE_FLOAT)
            return VOIDmode;
          else
            {
              know_element_mode = 1;
              element_mode = mode;
            }
        }
      return element_mode;

    default:
      /* If we reach here, we probably have some front-end specific type
         that the backend doesn't know about.  This can happen via the
         aggregate_value_p call in init_function_start.  All we can do is
         ignore unknown tree types.  */
      return VOIDmode;
    }

  return VOIDmode;
}

/* Return the number of words required to hold a quantity of TYPE and MODE
   when passed as an argument.  */
static int
ia64_function_arg_words (tree type, enum machine_mode mode)
{
  int words;

  if (mode == BLKmode)
    words = int_size_in_bytes (type);
  else
    words = GET_MODE_SIZE (mode);

  return (words + UNITS_PER_WORD - 1) / UNITS_PER_WORD;  /* round up */
}

/* Return the number of registers that should be skipped so the current
   argument (described by TYPE and WORDS) will be properly aligned.

   Integer and float arguments larger than 8 bytes start at the next
   even boundary.  Aggregates larger than 8 bytes start at the next
   even boundary if the aggregate has 16 byte alignment.  Note that
   in the 32-bit ABI, TImode and TFmode have only 8-byte alignment
   but are still to be aligned in registers.

   ??? The ABI does not specify how to handle aggregates with
   alignment from 9 to 15 bytes, or greater than 16.  We handle them
   all as if they had 16 byte alignment.  Such aggregates can occur
   only if gcc extensions are used.  */
static int
ia64_function_arg_offset (CUMULATIVE_ARGS *cum, tree type, int words)
{
  if ((cum->words & 1) == 0)
    return 0;

  if (type
      && TREE_CODE (type) != INTEGER_TYPE
      && TREE_CODE (type) != REAL_TYPE)
    return TYPE_ALIGN (type) > 8 * BITS_PER_UNIT;
  else
    return words > 1;
}

/* Return rtx for register where argument is passed, or zero if it is passed
   on the stack.  */
/* ??? 128-bit quad-precision floats are always passed in general
   registers.  */

rtx
ia64_function_arg (CUMULATIVE_ARGS *cum, enum machine_mode mode, tree type,
                   int named, int incoming)
{
  int basereg = (incoming ? GR_ARG_FIRST : AR_ARG_FIRST);
  int words = ia64_function_arg_words (type, mode);
  int offset = ia64_function_arg_offset (cum, type, words);
  enum machine_mode hfa_mode = VOIDmode;

  /* If all argument slots are used, then it must go on the stack.  */
  if (cum->words + offset >= MAX_ARGUMENT_SLOTS)
    return 0;

  /* Check for and handle homogeneous FP aggregates.  */
  if (type)
    hfa_mode = hfa_element_mode (type, 0);

  /* Unnamed prototyped hfas are passed as usual.  Named prototyped hfas
     and unprototyped hfas are passed specially.  */
  if (hfa_mode != VOIDmode && (! cum->prototype || named))
    {
      rtx loc[16];
      int i = 0;
      int fp_regs = cum->fp_regs;
      int int_regs = cum->words + offset;
      int hfa_size = GET_MODE_SIZE (hfa_mode);
      int byte_size;
      int args_byte_size;

      /* If prototyped, pass it in FR regs then GR regs.
         If not prototyped, pass it in both FR and GR regs.

         If this is an SFmode aggregate, then it is possible to run out of
         FR regs while GR regs are still left.  In that case, we pass the
         remaining part in the GR regs.  */

      /* Fill the FP regs.  We do this always.  We stop if we reach the end
         of the argument, the last FP register, or the last argument slot.  */

      byte_size = ((mode == BLKmode)
                   ? int_size_in_bytes (type) : GET_MODE_SIZE (mode));
      args_byte_size = int_regs * UNITS_PER_WORD;
      offset = 0;
      for (; (offset < byte_size && fp_regs < MAX_ARGUMENT_SLOTS
              && args_byte_size < (MAX_ARGUMENT_SLOTS * UNITS_PER_WORD)); i++)
        {
          loc[i] = gen_rtx_EXPR_LIST (VOIDmode,
                                      gen_rtx_REG (hfa_mode, (FR_ARG_FIRST
                                                              + fp_regs)),
                                      GEN_INT (offset));
          offset += hfa_size;
          args_byte_size += hfa_size;
          fp_regs++;
        }

      /* If no prototype, then the whole thing must go in GR regs.  */
      if (! cum->prototype)
        offset = 0;
      /* If this is an SFmode aggregate, then we might have some left over
         that needs to go in GR regs.  */
      else if (byte_size != offset)
        int_regs += offset / UNITS_PER_WORD;

      /* Fill in the GR regs.  We must use DImode here, not the hfa mode.  */

      for (; offset < byte_size && int_regs < MAX_ARGUMENT_SLOTS; i++)
        {
          enum machine_mode gr_mode = DImode;
          unsigned int gr_size;

          /* If we have an odd 4 byte hunk because we ran out of FR regs,
             then this goes in a GR reg left adjusted/little endian, right
             adjusted/big endian.  */
          /* ??? Currently this is handled wrong, because 4-byte hunks are
             always right adjusted/little endian.  */
          if (offset & 0x4)
            gr_mode = SImode;
          /* If we have an even 4 byte hunk because the aggregate is a
             multiple of 4 bytes in size, then this goes in a GR reg right
             adjusted/little endian.  */
          else if (byte_size - offset == 4)
            gr_mode = SImode;

          loc[i] = gen_rtx_EXPR_LIST (VOIDmode,
                                      gen_rtx_REG (gr_mode, (basereg
                                                             + int_regs)),
                                      GEN_INT (offset));

          gr_size = GET_MODE_SIZE (gr_mode);
          offset += gr_size;
          if (gr_size == UNITS_PER_WORD
              || (gr_size < UNITS_PER_WORD && offset % UNITS_PER_WORD == 0))
            int_regs++;
          else if (gr_size > UNITS_PER_WORD)
            int_regs += gr_size / UNITS_PER_WORD;
        }
      return gen_rtx_PARALLEL (mode, gen_rtvec_v (i, loc));
    }

  /* Integral and aggregates go in general registers.  If we have run out of
     FR registers, then FP values must also go in general registers.  This can
     happen when we have a SFmode HFA.  */
  else if (mode == TFmode || mode == TCmode
           || (! FLOAT_MODE_P (mode) || cum->fp_regs == MAX_ARGUMENT_SLOTS))
    {
      int byte_size = ((mode == BLKmode)
                       ? int_size_in_bytes (type) : GET_MODE_SIZE (mode));
      if (BYTES_BIG_ENDIAN
        && (mode == BLKmode || (type && AGGREGATE_TYPE_P (type)))
        && byte_size < UNITS_PER_WORD
        && byte_size > 0)
        {
          rtx gr_reg = gen_rtx_EXPR_LIST (VOIDmode,
                                          gen_rtx_REG (DImode,
                                                       (basereg + cum->words
                                                        + offset)),
                                          const0_rtx);
          return gen_rtx_PARALLEL (mode, gen_rtvec (1, gr_reg));
        }
      else
        return gen_rtx_REG (mode, basereg + cum->words + offset);

    }

  /* If there is a prototype, then FP values go in a FR register when
     named, and in a GR register when unnamed.  */
  else if (cum->prototype)
    {
      if (named)
        return gen_rtx_REG (mode, FR_ARG_FIRST + cum->fp_regs);
      /* In big-endian mode, an anonymous SFmode value must be represented
         as (parallel:SF [(expr_list (reg:DI n) (const_int 0))]) to force
         the value into the high half of the general register.  */
      else if (BYTES_BIG_ENDIAN && mode == SFmode)
        return gen_rtx_PARALLEL (mode,
                 gen_rtvec (1,
                   gen_rtx_EXPR_LIST (VOIDmode,
                     gen_rtx_REG (DImode, basereg + cum->words + offset),
                                      const0_rtx)));
      else
        return gen_rtx_REG (mode, basereg + cum->words + offset);
    }
  /* If there is no prototype, then FP values go in both FR and GR
     registers.  */
  else
    {
      /* See comment above.  */
      enum machine_mode inner_mode =
        (BYTES_BIG_ENDIAN && mode == SFmode) ? DImode : mode;

      rtx fp_reg = gen_rtx_EXPR_LIST (VOIDmode,
                                      gen_rtx_REG (mode, (FR_ARG_FIRST
                                                          + cum->fp_regs)),
                                      const0_rtx);
      rtx gr_reg = gen_rtx_EXPR_LIST (VOIDmode,
                                      gen_rtx_REG (inner_mode,
                                                   (basereg + cum->words
                                                    + offset)),
                                      const0_rtx);

      return gen_rtx_PARALLEL (mode, gen_rtvec (2, fp_reg, gr_reg));
    }
}

/* Return number of words, at the beginning of the argument, that must be
   put in registers.  0 is the argument is entirely in registers or entirely
   in memory.  */

int
ia64_function_arg_partial_nregs (CUMULATIVE_ARGS *cum, enum machine_mode mode,
                                 tree type, int named ATTRIBUTE_UNUSED)
{
  int words = ia64_function_arg_words (type, mode);
  int offset = ia64_function_arg_offset (cum, type, words);

  /* If all argument slots are used, then it must go on the stack.  */
  if (cum->words + offset >= MAX_ARGUMENT_SLOTS)
    return 0;

  /* It doesn't matter whether the argument goes in FR or GR regs.  If
     it fits within the 8 argument slots, then it goes entirely in
     registers.  If it extends past the last argument slot, then the rest
     goes on the stack.  */

  if (words + cum->words + offset <= MAX_ARGUMENT_SLOTS)
    return 0;

  return MAX_ARGUMENT_SLOTS - cum->words - offset;
}

/* Update CUM to point after this argument.  This is patterned after
   ia64_function_arg.  */

void
ia64_function_arg_advance (CUMULATIVE_ARGS *cum, enum machine_mode mode,
                           tree type, int named)
{
  int words = ia64_function_arg_words (type, mode);
  int offset = ia64_function_arg_offset (cum, type, words);
  enum machine_mode hfa_mode = VOIDmode;

  /* If all arg slots are already full, then there is nothing to do.  */
  if (cum->words >= MAX_ARGUMENT_SLOTS)
    return;

  cum->words += words + offset;

  /* Check for and handle homogeneous FP aggregates.  */
  if (type)
    hfa_mode = hfa_element_mode (type, 0);

  /* Unnamed prototyped hfas are passed as usual.  Named prototyped hfas
     and unprototyped hfas are passed specially.  */
  if (hfa_mode != VOIDmode && (! cum->prototype || named))
    {
      int fp_regs = cum->fp_regs;
      /* This is the original value of cum->words + offset.  */
      int int_regs = cum->words - words;
      int hfa_size = GET_MODE_SIZE (hfa_mode);
      int byte_size;
      int args_byte_size;

      /* If prototyped, pass it in FR regs then GR regs.
         If not prototyped, pass it in both FR and GR regs.

         If this is an SFmode aggregate, then it is possible to run out of
         FR regs while GR regs are still left.  In that case, we pass the
         remaining part in the GR regs.  */

      /* Fill the FP regs.  We do this always.  We stop if we reach the end
         of the argument, the last FP register, or the last argument slot.  */

      byte_size = ((mode == BLKmode)
                   ? int_size_in_bytes (type) : GET_MODE_SIZE (mode));
      args_byte_size = int_regs * UNITS_PER_WORD;
      offset = 0;
      for (; (offset < byte_size && fp_regs < MAX_ARGUMENT_SLOTS
              && args_byte_size < (MAX_ARGUMENT_SLOTS * UNITS_PER_WORD));)
        {
          offset += hfa_size;
          args_byte_size += hfa_size;
          fp_regs++;
        }

      cum->fp_regs = fp_regs;
    }

  /* Integral and aggregates go in general registers.  So do TFmode FP values.
     If we have run out of FR registers, then other FP values must also go in
     general registers.  This can happen when we have a SFmode HFA.  */
  else if (mode == TFmode || mode == TCmode
           || (! FLOAT_MODE_P (mode) || cum->fp_regs == MAX_ARGUMENT_SLOTS))
    cum->int_regs = cum->words;

  /* If there is a prototype, then FP values go in a FR register when
     named, and in a GR register when unnamed.  */
  else if (cum->prototype)
    {
      if (! named)
        cum->int_regs = cum->words;
      else
        /* ??? Complex types should not reach here.  */
        cum->fp_regs += (GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT ? 2 : 1);
    }
  /* If there is no prototype, then FP values go in both FR and GR
     registers.  */
  else
    {
      /* ??? Complex types should not reach here.  */
      cum->fp_regs += (GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT ? 2 : 1);
      cum->int_regs = cum->words;
    }
}

/* Arguments with alignment larger than 8 bytes start at the next even
   boundary.  On ILP32 HPUX, TFmode arguments start on next even boundary
   even though their normal alignment is 8 bytes.  See ia64_function_arg.  */

int
ia64_function_arg_boundary (enum machine_mode mode, tree type)
{

  if (mode == TFmode && TARGET_HPUX && TARGET_ILP32)
    return PARM_BOUNDARY * 2;

  if (type)
    {
      if (TYPE_ALIGN (type) > PARM_BOUNDARY)
        return PARM_BOUNDARY * 2;
      else
        return PARM_BOUNDARY;
    }

  if (GET_MODE_BITSIZE (mode) > PARM_BOUNDARY)
    return PARM_BOUNDARY * 2;
  else
    return PARM_BOUNDARY;
}

/* Variable sized types are passed by reference.  */
/* ??? At present this is a GCC extension to the IA-64 ABI.  */

static bool
ia64_pass_by_reference (CUMULATIVE_ARGS *cum ATTRIBUTE_UNUSED,
                        enum machine_mode mode ATTRIBUTE_UNUSED,
                        tree type, bool named ATTRIBUTE_UNUSED)
{
  return type && TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST;
}

/* True if it is OK to do sibling call optimization for the specified
   call expression EXP.  DECL will be the called function, or NULL if
   this is an indirect call.  */
static bool
ia64_function_ok_for_sibcall (tree decl, tree exp ATTRIBUTE_UNUSED)
{
  /* We can't perform a sibcall if the current function has the syscall_linkage
     attribute.  */
  if (lookup_attribute ("syscall_linkage",
                        TYPE_ATTRIBUTES (TREE_TYPE (current_function_decl))))
    return false;

  /* We must always return with our current GP.  This means we can
     only sibcall to functions defined in the current module.  */
  return decl && (*targetm.binds_local_p) (decl);
}
\f

/* Implement va_arg.  */

static tree
ia64_gimplify_va_arg (tree valist, tree type, tree *pre_p, tree *post_p)
{
  /* Variable sized types are passed by reference.  */
  if (pass_by_reference (NULL, TYPE_MODE (type), type, false))
    {
      tree ptrtype = build_pointer_type (type);
      tree addr = std_gimplify_va_arg_expr (valist, ptrtype, pre_p, post_p);
      return build_va_arg_indirect_ref (addr);
    }

  /* Aggregate arguments with alignment larger than 8 bytes start at
     the next even boundary.  Integer and floating point arguments
     do so if they are larger than 8 bytes, whether or not they are
     also aligned larger than 8 bytes.  */
  if ((TREE_CODE (type) == REAL_TYPE || TREE_CODE (type) == INTEGER_TYPE)
      ? int_size_in_bytes (type) > 8 : TYPE_ALIGN (type) > 8 * BITS_PER_UNIT)
    {
      tree t = build (PLUS_EXPR, TREE_TYPE (valist), valist,
                      build_int_cst (NULL_TREE, 2 * UNITS_PER_WORD - 1));
      t = build (BIT_AND_EXPR, TREE_TYPE (t), t,
                 build_int_cst (NULL_TREE, -2 * UNITS_PER_WORD));
      t = build (MODIFY_EXPR, TREE_TYPE (valist), valist, t);
      gimplify_and_add (t, pre_p);
    }

  return std_gimplify_va_arg_expr (valist, type, pre_p, post_p);
}
\f
/* Return 1 if function return value returned in memory.  Return 0 if it is
   in a register.  */

static bool
ia64_return_in_memory (tree valtype, tree fntype ATTRIBUTE_UNUSED)
{
  enum machine_mode mode;
  enum machine_mode hfa_mode;
  HOST_WIDE_INT byte_size;

  mode = TYPE_MODE (valtype);
  byte_size = GET_MODE_SIZE (mode);
  if (mode == BLKmode)
    {
      byte_size = int_size_in_bytes (valtype);
      if (byte_size < 0)
        return true;
    }

  /* Hfa's with up to 8 elements are returned in the FP argument registers.  */

  hfa_mode = hfa_element_mode (valtype, 0);
  if (hfa_mode != VOIDmode)
    {
      int hfa_size = GET_MODE_SIZE (hfa_mode);

      if (byte_size / hfa_size > MAX_ARGUMENT_SLOTS)
        return true;
      else
        return false;
    }
  else if (byte_size > UNITS_PER_WORD * MAX_INT_RETURN_SLOTS)
    return true;
  else
    return false;
}

/* Return rtx for register that holds the function return value.  */

rtx
ia64_function_value (tree valtype, tree func ATTRIBUTE_UNUSED)
{
  enum machine_mode mode;
  enum machine_mode hfa_mode;

  mode = TYPE_MODE (valtype);
  hfa_mode = hfa_element_mode (valtype, 0);

  if (hfa_mode != VOIDmode)
    {
      rtx loc[8];
      int i;
      int hfa_size;
      int byte_size;
      int offset;

      hfa_size = GET_MODE_SIZE (hfa_mode);
      byte_size = ((mode == BLKmode)
                   ? int_size_in_bytes (valtype) : GET_MODE_SIZE (mode));
      offset = 0;
      for (i = 0; offset < byte_size; i++)
        {
          loc[i] = gen_rtx_EXPR_LIST (VOIDmode,
                                      gen_rtx_REG (hfa_mode, FR_ARG_FIRST + i),
                                      GEN_INT (offset));
          offset += hfa_size;
        }
      return gen_rtx_PARALLEL (mode, gen_rtvec_v (i, loc));
    }
  else if (FLOAT_TYPE_P (valtype) && mode != TFmode && mode != TCmode)
    return gen_rtx_REG (mode, FR_ARG_FIRST);
  else
    {
      if (BYTES_BIG_ENDIAN
          && (mode == BLKmode || (valtype && AGGREGATE_TYPE_P (valtype))))
        {
          rtx loc[8];
          int offset;
          int bytesize;
          int i;

          offset = 0;
          bytesize = int_size_in_bytes (valtype);
          for (i = 0; offset < bytesize; i++)
            {
              loc[i] = gen_rtx_EXPR_LIST (VOIDmode,
                                          gen_rtx_REG (DImode,
                                                       GR_RET_FIRST + i),
                                          GEN_INT (offset));
              offset += UNITS_PER_WORD;
            }
          return gen_rtx_PARALLEL (mode, gen_rtvec_v (i, loc));
        }
      else
        return gen_rtx_REG (mode, GR_RET_FIRST);
    }
}

/* This is called from dwarf2out.c via ASM_OUTPUT_DWARF_DTPREL.
   We need to emit DTP-relative relocations.  */

void
ia64_output_dwarf_dtprel (FILE *file, int size, rtx x)
{
  if (size != 8)
    abort ();
  fputs ("\tdata8.ua\t@dtprel(", file);
  output_addr_const (file, x);
  fputs (")", file);
}

/* Print a memory address as an operand to reference that memory location.  */

/* ??? Do we need this?  It gets used only for 'a' operands.  We could perhaps
   also call this from ia64_print_operand for memory addresses.  */

void
ia64_print_operand_address (FILE * stream ATTRIBUTE_UNUSED,
                            rtx address ATTRIBUTE_UNUSED)
{
}

/* Print an operand to an assembler instruction.
   C    Swap and print a comparison operator.
   D    Print an FP comparison operator.
   E    Print 32 - constant, for SImode shifts as extract.
   e    Print 64 - constant, for DImode rotates.
   F    A floating point constant 0.0 emitted as f0, or 1.0 emitted as f1, or
        a floating point register emitted normally.
   I    Invert a predicate register by adding 1.
   J    Select the proper predicate register for a condition.
   j    Select the inverse predicate register for a condition.
   O    Append .acq for volatile load.
   P    Postincrement of a MEM.
   Q    Append .rel for volatile store.
   S    Shift amount for shladd instruction.
   T    Print an 8-bit sign extended number (K) as a 32-bit unsigned number
        for Intel assembler.
   U    Print an 8-bit sign extended number (K) as a 64-bit unsigned number
        for Intel assembler.
   r    Print register name, or constant 0 as r0.  HP compatibility for
        Linux kernel.  */
void
ia64_print_operand (FILE * file, rtx x, int code)
{
  const char *str;

  switch (code)
    {
    case 0:
      /* Handled below.  */
      break;

    case 'C':
      {
        enum rtx_code c = swap_condition (GET_CODE (x));
        fputs (GET_RTX_NAME (c), file);
        return;
      }

    case 'D':
      switch (GET_CODE (x))
        {
        case NE:
          str = "neq";
          break;
        case UNORDERED:
          str = "unord";
          break;
        case ORDERED:
          str = "ord";
          break;
        default:
          str = GET_RTX_NAME (GET_CODE (x));
          break;
        }
      fputs (str, file);
      return;

    case 'E':
      fprintf (file, HOST_WIDE_INT_PRINT_DEC, 32 - INTVAL (x));
      return;

    case 'e':
      fprintf (file, HOST_WIDE_INT_PRINT_DEC, 64 - INTVAL (x));
      return;

    case 'F':
      if (x == CONST0_RTX (GET_MODE (x)))
        str = reg_names [FR_REG (0)];
      else if (x == CONST1_RTX (GET_MODE (x)))
        str = reg_names [FR_REG (1)];
      else if (GET_CODE (x) == REG)
        str = reg_names [REGNO (x)];
      else
        abort ();
      fputs (str, file);
      return;

    case 'I':
      fputs (reg_names [REGNO (x) + 1], file);
      return;

    case 'J':
    case 'j':
      {
        unsigned int regno = REGNO (XEXP (x, 0));
        if (GET_CODE (x) == EQ)
          regno += 1;
        if (code == 'j')
          regno ^= 1;
        fputs (reg_names [regno], file);
      }
      return;

    case 'O':
      if (MEM_VOLATILE_P (x))
        fputs(".acq", file);
      return;

    case 'P':
      {
        HOST_WIDE_INT value;

        switch (GET_CODE (XEXP (x, 0)))
          {
          default:
            return;

          case POST_MODIFY:
            x = XEXP (XEXP (XEXP (x, 0), 1), 1);
            if (GET_CODE (x) == CONST_INT)
              value = INTVAL (x);
            else if (GET_CODE (x) == REG)
              {
                fprintf (file, ", %s", reg_names[REGNO (x)]);
                return;
              }
            else
              abort ();
            break;

          case POST_INC:
            value = GET_MODE_SIZE (GET_MODE (x));
            break;

          case POST_DEC:
            value = - (HOST_WIDE_INT) GET_MODE_SIZE (GET_MODE (x));
            break;
          }

        fprintf (file, ", " HOST_WIDE_INT_PRINT_DEC, value);
        return;
      }

    case 'Q':
      if (MEM_VOLATILE_P (x))
        fputs(".rel", file);
      return;

    case 'S':
      fprintf (file, "%d", exact_log2 (INTVAL (x)));
      return;

    case 'T':
      if (! TARGET_GNU_AS && GET_CODE (x) == CONST_INT)
        {
          fprintf (file, "0x%x", (int) INTVAL (x) & 0xffffffff);
          return;
        }
      break;

    case 'U':
      if (! TARGET_GNU_AS && GET_CODE (x) == CONST_INT)
        {
          const char *prefix = "0x";
          if (INTVAL (x) & 0x80000000)
            {
              fprintf (file, "0xffffffff");
              prefix = "";
            }
          fprintf (file, "%s%x", prefix, (int) INTVAL (x) & 0xffffffff);
          return;
        }
      break;

    case 'r':
      /* If this operand is the constant zero, write it as register zero.
         Any register, zero, or CONST_INT value is OK here.  */
      if (GET_CODE (x) == REG)
        fputs (reg_names[REGNO (x)], file);
      else if (x == CONST0_RTX (GET_MODE (x)))
        fputs ("r0", file);
      else if (GET_CODE (x) == CONST_INT)
        output_addr_const (file, x);
      else
        output_operand_lossage ("invalid %%r value");
      return;

    case '+':
      {
        const char *which;

        /* For conditional branches, returns or calls, substitute
           sptk, dptk, dpnt, or spnt for %s.  */
        x = find_reg_note (current_output_insn, REG_BR_PROB, 0);
        if (x)
          {
            int pred_val = INTVAL (XEXP (x, 0));

            /* Guess top and bottom 10% statically predicted.  */
            if (pred_val < REG_BR_PROB_BASE / 50)
              which = ".spnt";
            else if (pred_val < REG_BR_PROB_BASE / 2)
              which = ".dpnt";
            else if (pred_val < REG_BR_PROB_BASE / 100 * 98)
              which = ".dptk";
            else
              which = ".sptk";
          }
        else if (GET_CODE (current_output_insn) == CALL_INSN)
          which = ".sptk";
        else
          which = ".dptk";

        fputs (which, file);
        return;
      }

    case ',':
      x = current_insn_predicate;
      if (x)
        {
          unsigned int regno = REGNO (XEXP (x, 0));
          if (GET_CODE (x) == EQ)
            regno += 1;
          fprintf (file, "(%s) ", reg_names [regno]);
        }
      return;

    default:
      output_operand_lossage ("ia64_print_operand: unknown code");
      return;
    }

  switch (GET_CODE (x))
    {
      /* This happens for the spill/restore instructions.  */
    case POST_INC:
    case POST_DEC:
    case POST_MODIFY:
      x = XEXP (x, 0);
      /* ... fall through ...  */

    case REG:
      fputs (reg_names [REGNO (x)], file);
      break;

    case MEM:
      {
        rtx addr = XEXP (x, 0);
        if (GET_RTX_CLASS (GET_CODE (addr)) == RTX_AUTOINC)
          addr = XEXP (addr, 0);
        fprintf (file, "[%s]", reg_names [REGNO (addr)]);
        break;
      }

    default:
      output_addr_const (file, x);
      break;
    }

  return;
}
\f
/* Compute a (partial) cost for rtx X.  Return true if the complete
   cost has been computed, and false if subexpressions should be
   scanned.  In either case, *TOTAL contains the cost result.  */
/* ??? This is incomplete.  */

static bool
ia64_rtx_costs (rtx x, int code, int outer_code, int *total)
{
  switch (code)
    {
    case CONST_INT:
      switch (outer_code)
        {
        case SET:
          *total = CONST_OK_FOR_J (INTVAL (x)) ? 0 : COSTS_N_INSNS (1);
          return true;
        case PLUS:
          if (CONST_OK_FOR_I (INTVAL (x)))
            *total = 0;
          else if (CONST_OK_FOR_J (INTVAL (x)))
            *total = 1;
          else
            *total = COSTS_N_INSNS (1);
          return true;
        default:
          if (CONST_OK_FOR_K (INTVAL (x)) || CONST_OK_FOR_L (INTVAL (x)))
            *total = 0;
          else
            *total = COSTS_N_INSNS (1);
          return true;
        }

    case CONST_DOUBLE:
      *total = COSTS_N_INSNS (1);
      return true;

    case CONST:
    case SYMBOL_REF:
    case LABEL_REF:
      *total = COSTS_N_INSNS (3);
      return true;

    case MULT:
      /* For multiplies wider than HImode, we have to go to the FPU,
         which normally involves copies.  Plus there's the latency
         of the multiply itself, and the latency of the instructions to
         transfer integer regs to FP regs.  */
      /* ??? Check for FP mode.  */
      if (GET_MODE_SIZE (GET_MODE (x)) > 2)
        *total = COSTS_N_INSNS (10);
      else
        *total = COSTS_N_INSNS (2);
      return true;

    case PLUS:
    case MINUS:
    case ASHIFT:
    case ASHIFTRT:
    case LSHIFTRT:
      *total = COSTS_N_INSNS (1);
      return true;

    case DIV:
    case UDIV:
    case MOD:
    case UMOD:
      /* We make divide expensive, so that divide-by-constant will be
         optimized to a multiply.  */
      *total = COSTS_N_INSNS (60);
      return true;

    default:
      return false;
    }
}

/* Calculate the cost of moving data from a register in class FROM to
   one in class TO, using MODE.  */

int
ia64_register_move_cost (enum machine_mode mode, enum reg_class from,
                         enum reg_class to)
{
  /* ADDL_REGS is the same as GR_REGS for movement purposes.  */
  if (to == ADDL_REGS)
    to = GR_REGS;
  if (from == ADDL_REGS)
    from = GR_REGS;

  /* All costs are symmetric, so reduce cases by putting the
     lower number class as the destination.  */
  if (from < to)
    {
      enum reg_class tmp = to;
      to = from, from = tmp;
    }

  /* Moving from FR<->GR in XFmode must be more expensive than 2,
     so that we get secondary memory reloads.  Between FR_REGS,
     we have to make this at least as expensive as MEMORY_MOVE_COST
     to avoid spectacularly poor register class preferencing.  */
  if (mode == XFmode)
    {
      if (to != GR_REGS || from != GR_REGS)
        return MEMORY_MOVE_COST (mode, to, 0);
      else
        return 3;
    }

  switch (to)
    {
    case PR_REGS:
      /* Moving between PR registers takes two insns.  */
      if (from == PR_REGS)
        return 3;
      /* Moving between PR and anything but GR is impossible.  */
      if (from != GR_REGS)
        return MEMORY_MOVE_COST (mode, to, 0);
      break;

    case BR_REGS:
      /* Moving between BR and anything but GR is impossible.  */
      if (from != GR_REGS && from != GR_AND_BR_REGS)
        return MEMORY_MOVE_COST (mode, to, 0);
      break;

    case AR_I_REGS:
    case AR_M_REGS:
      /* Moving between AR and anything but GR is impossible.  */
      if (from != GR_REGS)
        return MEMORY_MOVE_COST (mode, to, 0);
      break;

    case GR_REGS:
    case FR_REGS:
    case GR_AND_FR_REGS:
    case GR_AND_BR_REGS:
    case ALL_REGS:
      break;

    default:
      abort ();
    }

  return 2;
}

/* This function returns the register class required for a secondary
   register when copying between one of the registers in CLASS, and X,
   using MODE.  A return value of NO_REGS means that no secondary register
   is required.  */

enum reg_class
ia64_secondary_reload_class (enum reg_class class,
                             enum machine_mode mode ATTRIBUTE_UNUSED, rtx x)
{
  int regno = -1;

  if (GET_CODE (x) == REG || GET_CODE (x) == SUBREG)
    regno = true_regnum (x);

  switch (class)
    {
    case BR_REGS:
    case AR_M_REGS:
    case AR_I_REGS:
      /* ??? BR<->BR register copies can happen due to a bad gcse/cse/global
         interaction.  We end up with two pseudos with overlapping lifetimes
         both of which are equiv to the same constant, and both which need
         to be in BR_REGS.  This seems to be a cse bug.  cse_basic_block_end
         changes depending on the path length, which means the qty_first_reg
         check in make_regs_eqv can give different answers at different times.
         At some point I'll probably need a reload_indi pattern to handle
         this.

         We can also get GR_AND_FR_REGS to BR_REGS/AR_REGS copies, where we
         wound up with a FP register from GR_AND_FR_REGS.  Extend that to all
         non-general registers for good measure.  */
      if (regno >= 0 && ! GENERAL_REGNO_P (regno))
        return GR_REGS;

      /* This is needed if a pseudo used as a call_operand gets spilled to a
         stack slot.  */
      if (GET_CODE (x) == MEM)
        return GR_REGS;
      break;

    case FR_REGS:
      /* Need to go through general registers to get to other class regs.  */
      if (regno >= 0 && ! (FR_REGNO_P (regno) || GENERAL_REGNO_P (regno)))
        return GR_REGS;

      /* This can happen when a paradoxical subreg is an operand to the
         muldi3 pattern.  */
      /* ??? This shouldn't be necessary after instruction scheduling is
         enabled, because paradoxical subregs are not accepted by
         register_operand when INSN_SCHEDULING is defined.  Or alternatively,
         stop the paradoxical subreg stupidity in the *_operand functions
         in recog.c.  */
      if (GET_CODE (x) == MEM
          && (GET_MODE (x) == SImode || GET_MODE (x) == HImode
              || GET_MODE (x) == QImode))
        return GR_REGS;

      /* This can happen because of the ior/and/etc patterns that accept FP
         registers as operands.  If the third operand is a constant, then it
         needs to be reloaded into a FP register.  */
      if (GET_CODE (x) == CONST_INT)
        return GR_REGS;

      /* This can happen because of register elimination in a muldi3 insn.
         E.g. `26107 * (unsigned long)&u'.  */
      if (GET_CODE (x) == PLUS)
        return GR_REGS;
      break;