Google Git
Sign in
gnu / gcc / 926d9947847a0683cf34de6fc231209747576088 / . / gcc / config / sh / sh.c
blob: 9db77530c1f6c3624d04514e17bdb01939277ad4 [file] [log] [blame]
/* Output routines for GCC for Renesas / SuperH SH.
Copyright (C) 1993-2015 Free Software Foundation, Inc.
Contributed by Steve Chamberlain (sac@cygnus.com).
Improved by Jim Wilson (wilson@cygnus.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 3, or (at your option)
any later version.
GCC is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with GCC; see the file COPYING3. If not see
<http://www.gnu.org/licenses/>. */
#include <sstream>
#include "config.h"
#define INCLUDE_VECTOR
#include "system.h"
#include "coretypes.h"
#include "tm.h"
#include "insn-config.h"
#include "rtl.h"
#include "hash-set.h"
#include "machmode.h"
#include "vec.h"
#include "double-int.h"
#include "input.h"
#include "alias.h"
#include "symtab.h"
#include "wide-int.h"
#include "inchash.h"
#include "tree.h"
#include "fold-const.h"
#include "stringpool.h"
#include "stor-layout.h"
#include "calls.h"
#include "varasm.h"
#include "flags.h"
#include "hashtab.h"
#include "hard-reg-set.h"
#include "function.h"
#include "statistics.h"
#include "real.h"
#include "fixed-value.h"
#include "expmed.h"
#include "dojump.h"
#include "explow.h"
#include "emit-rtl.h"
#include "stmt.h"
#include "expr.h"
#include "insn-codes.h"
#include "optabs.h"
#include "reload.h"
#include "regs.h"
#include "output.h"
#include "insn-attr.h"
#include "diagnostic-core.h"
#include "recog.h"
#include "dwarf2.h"
#include "tm_p.h"
#include "target.h"
#include "target-def.h"
#include "langhooks.h"
#include "predict.h"
#include "dominance.h"
#include "cfg.h"
#include "cfgrtl.h"
#include "cfganal.h"
#include "lcm.h"
#include "cfgbuild.h"
#include "cfgcleanup.h"
#include "basic-block.h"
#include "df.h"
#include "intl.h"
#include "sched-int.h"
#include "params.h"
#include "ggc.h"
#include "hash-table.h"
#include "tree-ssa-alias.h"
#include "internal-fn.h"
#include "gimple-fold.h"
#include "tree-eh.h"
#include "gimple-expr.h"
#include "is-a.h"
#include "gimple.h"
#include "gimplify.h"
#include "cfgloop.h"
#include "alloc-pool.h"
#include "tm-constrs.h"
#include "opts.h"
#include "tree-pass.h"
#include "pass_manager.h"
#include "context.h"
#include "builtins.h"
#include "rtl-iter.h"
int code_for_indirect_jump_scratch = CODE_FOR_indirect_jump_scratch;
/* These are some macros to abstract register modes. */
#define CONST_OK_FOR_I10(VALUE) (((HOST_WIDE_INT)(VALUE)) >= -512 \
&& ((HOST_WIDE_INT)(VALUE)) <= 511)
#define CONST_OK_FOR_ADD(size) \
(TARGET_SHMEDIA ? CONST_OK_FOR_I10 (size) : CONST_OK_FOR_I08 (size))
#define GEN_MOV (*(TARGET_SHMEDIA64 ? gen_movdi : gen_movsi))
#define GEN_ADD3 (*(TARGET_SHMEDIA64 ? gen_adddi3 : gen_addsi3))
#define GEN_SUB3 (*(TARGET_SHMEDIA64 ? gen_subdi3 : gen_subsi3))
/* Used to simplify the logic below. Find the attributes wherever
they may be. */
#define SH_ATTRIBUTES(decl) \
(TYPE_P (decl)) ? TYPE_ATTRIBUTES (decl) \
: DECL_ATTRIBUTES (decl) \
? (DECL_ATTRIBUTES (decl)) \
: TYPE_ATTRIBUTES (TREE_TYPE (decl))
/* Set to 1 by expand_prologue() when the function is an interrupt handler. */
int current_function_interrupt;
tree sh_deferred_function_attributes;
tree *sh_deferred_function_attributes_tail = &sh_deferred_function_attributes;
/* Global variables for machine-dependent things. */
/* Which cpu are we scheduling for. */
enum processor_type sh_cpu;
/* Definitions used in ready queue reordering for first scheduling pass. */
/* Reg weights arrays for modes SFmode and SImode, indexed by insn LUID. */
static short *regmode_weight[2];
/* Total SFmode and SImode weights of scheduled insns. */
static int curr_regmode_pressure[2];
/* Number of r0 life regions. */
static int r0_life_regions;
/* If true, skip cycles for Q -> R movement. */
static int skip_cycles = 0;
/* Cached value of can_issue_more. This is cached in sh_variable_issue hook
and returned from sh_reorder2. */
static short cached_can_issue_more;
/* Unique number for UNSPEC_BBR pattern. */
static unsigned int unspec_bbr_uid = 1;
/* Provides the class number of the smallest class containing
reg number. */
enum reg_class regno_reg_class[FIRST_PSEUDO_REGISTER] =
{
R0_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS,
GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS,
GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS,
GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS,
GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS,
GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS,
GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS,
GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS,
GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS,
GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS,
GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS,
GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS,
GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS,
GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS,
GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS,
GENERAL_REGS, GENERAL_REGS, GENERAL_REGS, GENERAL_REGS,
FP0_REGS,FP_REGS, FP_REGS, FP_REGS,
FP_REGS, FP_REGS, FP_REGS, FP_REGS,
FP_REGS, FP_REGS, FP_REGS, FP_REGS,
FP_REGS, FP_REGS, FP_REGS, FP_REGS,
FP_REGS, FP_REGS, FP_REGS, FP_REGS,
FP_REGS, FP_REGS, FP_REGS, FP_REGS,
FP_REGS, FP_REGS, FP_REGS, FP_REGS,
FP_REGS, FP_REGS, FP_REGS, FP_REGS,
FP_REGS, FP_REGS, FP_REGS, FP_REGS,
FP_REGS, FP_REGS, FP_REGS, FP_REGS,
FP_REGS, FP_REGS, FP_REGS, FP_REGS,
FP_REGS, FP_REGS, FP_REGS, FP_REGS,
FP_REGS, FP_REGS, FP_REGS, FP_REGS,
FP_REGS, FP_REGS, FP_REGS, FP_REGS,
FP_REGS, FP_REGS, FP_REGS, FP_REGS,
FP_REGS, FP_REGS, FP_REGS, FP_REGS,
TARGET_REGS, TARGET_REGS, TARGET_REGS, TARGET_REGS,
TARGET_REGS, TARGET_REGS, TARGET_REGS, TARGET_REGS,
DF_REGS, DF_REGS, DF_REGS, DF_REGS,
DF_REGS, DF_REGS, DF_REGS, DF_REGS,
NO_REGS, GENERAL_REGS, PR_REGS, T_REGS,
MAC_REGS, MAC_REGS, FPUL_REGS, FPSCR_REGS,
GENERAL_REGS, GENERAL_REGS,
};
char sh_register_names[FIRST_PSEUDO_REGISTER] \
[MAX_REGISTER_NAME_LENGTH + 1] = SH_REGISTER_NAMES_INITIALIZER;
char sh_additional_register_names[ADDREGNAMES_SIZE] \
[MAX_ADDITIONAL_REGISTER_NAME_LENGTH + 1]
= SH_ADDITIONAL_REGISTER_NAMES_INITIALIZER;
int assembler_dialect;
static bool shmedia_space_reserved_for_target_registers;
static void split_branches (rtx_insn *);
static int branch_dest (rtx);
static void print_slot (rtx_sequence *);
static rtx_code_label *add_constant (rtx, machine_mode, rtx);
static void dump_table (rtx_insn *, rtx_insn *);
static bool broken_move (rtx_insn *);
static bool mova_p (rtx_insn *);
static rtx_insn *find_barrier (int, rtx_insn *, rtx_insn *);
static bool noncall_uses_reg (rtx, rtx_insn *, rtx *);
static rtx_insn *gen_block_redirect (rtx_insn *, int, int);
static void sh_reorg (void);
static void sh_option_override (void);
static void sh_override_options_after_change (void);
static void output_stack_adjust (int, rtx, int, HARD_REG_SET *, bool);
static rtx_insn *frame_insn (rtx);
static rtx push (int);
static void pop (int);
static void push_regs (HARD_REG_SET *, int);
static int calc_live_regs (HARD_REG_SET *);
static HOST_WIDE_INT rounded_frame_size (int);
static bool sh_frame_pointer_required (void);
static void sh_emit_mode_set (int, int, int, HARD_REG_SET);
static int sh_mode_needed (int, rtx_insn *);
static int sh_mode_after (int, int, rtx_insn *);
static int sh_mode_entry (int);
static int sh_mode_exit (int);
static int sh_mode_priority (int entity, int n);
static rtx mark_constant_pool_use (rtx);
static tree sh_handle_interrupt_handler_attribute (tree *, tree, tree,
int, bool *);
static tree sh_handle_resbank_handler_attribute (tree *, tree,
tree, int, bool *);
static tree sh2a_handle_function_vector_handler_attribute (tree *, tree,
tree, int, bool *);
static tree sh_handle_sp_switch_attribute (tree *, tree, tree, int, bool *);
static tree sh_handle_trap_exit_attribute (tree *, tree, tree, int, bool *);
static tree sh_handle_renesas_attribute (tree *, tree, tree, int, bool *);
static void sh_print_operand (FILE *, rtx, int);
static void sh_print_operand_address (FILE *, rtx);
static bool sh_print_operand_punct_valid_p (unsigned char code);
static bool sh_asm_output_addr_const_extra (FILE *file, rtx x);
static void sh_output_function_epilogue (FILE *, HOST_WIDE_INT);
static void sh_insert_attributes (tree, tree *);
static const char *sh_check_pch_target_flags (int);
static int sh_register_move_cost (machine_mode, reg_class_t, reg_class_t);
static int sh_adjust_cost (rtx_insn *, rtx, rtx_insn *, int);
static int sh_issue_rate (void);
static int sh_dfa_new_cycle (FILE *, int, rtx_insn *, int, int, int *sort_p);
static short find_set_regmode_weight (rtx, machine_mode);
static short find_insn_regmode_weight (rtx, machine_mode);
static void find_regmode_weight (basic_block, machine_mode);
static int find_r0_life_regions (basic_block);
static void sh_md_init_global (FILE *, int, int);
static void sh_md_finish_global (FILE *, int);
static int rank_for_reorder (const void *, const void *);
static void swap_reorder (rtx_insn **, int);
static void ready_reorder (rtx_insn **, int);
static bool high_pressure (machine_mode);
static int sh_reorder (FILE *, int, rtx_insn **, int *, int);
static int sh_reorder2 (FILE *, int, rtx_insn **, int *, int);
static void sh_md_init (FILE *, int, int);
static int sh_variable_issue (FILE *, int, rtx_insn *, int);
static bool sh_function_ok_for_sibcall (tree, tree);
static bool sh_cannot_modify_jumps_p (void);
static bool sh_can_follow_jump (const rtx_insn *, const rtx_insn *);
static reg_class_t sh_target_reg_class (void);
static bool sh_optimize_target_register_callee_saved (bool);
static bool sh_ms_bitfield_layout_p (const_tree);
static void sh_init_builtins (void);
static tree sh_builtin_decl (unsigned, bool);
static rtx sh_expand_builtin (tree, rtx, rtx, machine_mode, int);
static void sh_output_mi_thunk (FILE *, tree, HOST_WIDE_INT,
HOST_WIDE_INT, tree);
static void sh_file_start (void);
static bool flow_dependent_p (rtx, rtx);
static void flow_dependent_p_1 (rtx, const_rtx, void *);
static int shiftcosts (rtx);
static int and_xor_ior_costs (rtx, int);
static int addsubcosts (rtx);
static int multcosts (rtx);
static bool unspec_caller_rtx_p (rtx);
static bool sh_cannot_copy_insn_p (rtx_insn *);
static bool sh_rtx_costs (rtx, int, int, int, int *, bool);
static int sh_address_cost (rtx, machine_mode, addr_space_t, bool);
static int sh_pr_n_sets (void);
static rtx sh_allocate_initial_value (rtx);
static reg_class_t sh_preferred_reload_class (rtx, reg_class_t);
static reg_class_t sh_secondary_reload (bool, rtx, reg_class_t,
machine_mode,
struct secondary_reload_info *);
static bool sh_legitimate_address_p (machine_mode, rtx, bool);
static rtx sh_legitimize_address (rtx, rtx, machine_mode);
static rtx sh_delegitimize_address (rtx);
static bool sh_cannot_substitute_mem_equiv_p (rtx);
static bool sh_legitimize_address_displacement (rtx *, rtx *, machine_mode);
static int shmedia_target_regs_stack_space (HARD_REG_SET *);
static int shmedia_reserve_space_for_target_registers_p (int, HARD_REG_SET *);
static int shmedia_target_regs_stack_adjust (HARD_REG_SET *);
static int scavenge_reg (HARD_REG_SET *s);
struct save_schedule_s;
static struct save_entry_s *sh5_schedule_saves (HARD_REG_SET *,
struct save_schedule_s *, int);
static rtx sh_struct_value_rtx (tree, int);
static rtx sh_function_value (const_tree, const_tree, bool);
static bool sh_function_value_regno_p (const unsigned int);
static rtx sh_libcall_value (machine_mode, const_rtx);
static bool sh_return_in_memory (const_tree, const_tree);
static rtx sh_builtin_saveregs (void);
static void sh_setup_incoming_varargs (cumulative_args_t, machine_mode,
tree, int *, int);
static bool sh_strict_argument_naming (cumulative_args_t);
static bool sh_pretend_outgoing_varargs_named (cumulative_args_t);
static void sh_atomic_assign_expand_fenv (tree *, tree *, tree *);
static tree sh_build_builtin_va_list (void);
static void sh_va_start (tree, rtx);
static tree sh_gimplify_va_arg_expr (tree, tree, gimple_seq *, gimple_seq *);
static bool sh_promote_prototypes (const_tree);
static machine_mode sh_promote_function_mode (const_tree type,
machine_mode,
int *punsignedp,
const_tree funtype,
int for_return);
static bool sh_pass_by_reference (cumulative_args_t, machine_mode,
const_tree, bool);
static bool sh_callee_copies (cumulative_args_t, machine_mode,
const_tree, bool);
static int sh_arg_partial_bytes (cumulative_args_t, machine_mode,
tree, bool);
static void sh_function_arg_advance (cumulative_args_t, machine_mode,
const_tree, bool);
static rtx sh_function_arg (cumulative_args_t, machine_mode,
const_tree, bool);
static bool sh_scalar_mode_supported_p (machine_mode);
static int sh_dwarf_calling_convention (const_tree);
static void sh_encode_section_info (tree, rtx, int);
static bool sh2a_function_vector_p (tree);
static void sh_trampoline_init (rtx, tree, rtx);
static rtx sh_trampoline_adjust_address (rtx);
static void sh_conditional_register_usage (void);
static bool sh_legitimate_constant_p (machine_mode, rtx);
static int mov_insn_size (machine_mode, bool);
static int mov_insn_alignment_mask (machine_mode, bool);
static bool sh_use_by_pieces_infrastructure_p (unsigned HOST_WIDE_INT,
unsigned int,
enum by_pieces_operation,
bool);
static bool sequence_insn_p (rtx_insn *);
static void sh_canonicalize_comparison (int *, rtx *, rtx *, bool);
static void sh_canonicalize_comparison (enum rtx_code&, rtx&, rtx&,
machine_mode, bool);
static bool sh_legitimate_combined_insn (rtx_insn* insn);
static bool sh_fixed_condition_code_regs (unsigned int* p1, unsigned int* p2);
static void sh_init_sync_libfuncs (void) ATTRIBUTE_UNUSED;
static const struct attribute_spec sh_attribute_table[] =
{
/* { name, min_len, max_len, decl_req, type_req, fn_type_req, handler,
affects_type_identity } */
{ "interrupt_handler", 0, 0, true, false, false,
sh_handle_interrupt_handler_attribute, false },
{ "sp_switch", 1, 1, true, false, false,
sh_handle_sp_switch_attribute, false },
{ "trap_exit", 1, 1, true, false, false,
sh_handle_trap_exit_attribute, false },
{ "renesas", 0, 0, false, true, false,
sh_handle_renesas_attribute, false },
{ "trapa_handler", 0, 0, true, false, false,
sh_handle_interrupt_handler_attribute, false },
{ "nosave_low_regs", 0, 0, true, false, false,
sh_handle_interrupt_handler_attribute, false },
{ "resbank", 0, 0, true, false, false,
sh_handle_resbank_handler_attribute, false },
{ "function_vector", 1, 1, true, false, false,
sh2a_handle_function_vector_handler_attribute, false },
{ NULL, 0, 0, false, false, false, NULL, false }
};
/* Initialize the GCC target structure. */
#undef TARGET_ATTRIBUTE_TABLE
#define TARGET_ATTRIBUTE_TABLE sh_attribute_table
/* The next two are used for debug info when compiling with -gdwarf. */
#undef TARGET_ASM_UNALIGNED_HI_OP
#define TARGET_ASM_UNALIGNED_HI_OP "\t.uaword\t"
#undef TARGET_ASM_UNALIGNED_SI_OP
#define TARGET_ASM_UNALIGNED_SI_OP "\t.ualong\t"
/* These are NULLed out on non-SH5 in TARGET_OPTION_OVERRIDE. */
#undef TARGET_ASM_UNALIGNED_DI_OP
#define TARGET_ASM_UNALIGNED_DI_OP "\t.uaquad\t"
#undef TARGET_ASM_ALIGNED_DI_OP
#define TARGET_ASM_ALIGNED_DI_OP "\t.quad\t"
#undef TARGET_OPTION_OVERRIDE
#define TARGET_OPTION_OVERRIDE sh_option_override
#undef TARGET_OVERRIDE_OPTIONS_AFTER_CHANGE
#define TARGET_OVERRIDE_OPTIONS_AFTER_CHANGE \
sh_override_options_after_change
#undef TARGET_PRINT_OPERAND
#define TARGET_PRINT_OPERAND sh_print_operand
#undef TARGET_PRINT_OPERAND_ADDRESS
#define TARGET_PRINT_OPERAND_ADDRESS sh_print_operand_address
#undef TARGET_PRINT_OPERAND_PUNCT_VALID_P
#define TARGET_PRINT_OPERAND_PUNCT_VALID_P sh_print_operand_punct_valid_p
#undef TARGET_ASM_OUTPUT_ADDR_CONST_EXTRA
#define TARGET_ASM_OUTPUT_ADDR_CONST_EXTRA sh_asm_output_addr_const_extra
#undef TARGET_ASM_FUNCTION_EPILOGUE
#define TARGET_ASM_FUNCTION_EPILOGUE sh_output_function_epilogue
#undef TARGET_ASM_OUTPUT_MI_THUNK
#define TARGET_ASM_OUTPUT_MI_THUNK sh_output_mi_thunk
#undef TARGET_ASM_CAN_OUTPUT_MI_THUNK
#define TARGET_ASM_CAN_OUTPUT_MI_THUNK \
hook_bool_const_tree_hwi_hwi_const_tree_true
#undef TARGET_ASM_FILE_START
#define TARGET_ASM_FILE_START sh_file_start
#undef TARGET_ASM_FILE_START_FILE_DIRECTIVE
#define TARGET_ASM_FILE_START_FILE_DIRECTIVE true
#undef TARGET_REGISTER_MOVE_COST
#define TARGET_REGISTER_MOVE_COST sh_register_move_cost
#undef TARGET_INSERT_ATTRIBUTES
#define TARGET_INSERT_ATTRIBUTES sh_insert_attributes
#undef TARGET_SCHED_ADJUST_COST
#define TARGET_SCHED_ADJUST_COST sh_adjust_cost
#undef TARGET_SCHED_ISSUE_RATE
#define TARGET_SCHED_ISSUE_RATE sh_issue_rate
/* The next 5 hooks have been implemented for reenabling sched1. With the
help of these macros we are limiting the movement of insns in sched1 to
reduce the register pressure. The overall idea is to keep count of SImode
and SFmode regs required by already scheduled insns. When these counts
cross some threshold values; give priority to insns that free registers.
The insn that frees registers is most likely to be the insn with lowest
LUID (original insn order); but such an insn might be there in the stalled
queue (Q) instead of the ready queue (R). To solve this, we skip cycles
up to a max of 8 cycles so that such insns may move from Q -> R.
The description of the hooks are as below:
TARGET_SCHED_INIT_GLOBAL: Added a new target hook in the generic
scheduler; it is called inside the sched_init function just after
find_insn_reg_weights function call. It is used to calculate the SImode
and SFmode weights of insns of basic blocks; much similar to what
find_insn_reg_weights does.
TARGET_SCHED_FINISH_GLOBAL: Corresponding cleanup hook.
TARGET_SCHED_DFA_NEW_CYCLE: Skip cycles if high register pressure is
indicated by TARGET_SCHED_REORDER2; doing this may move insns from
(Q)->(R).
TARGET_SCHED_REORDER: If the register pressure for SImode or SFmode is
high; reorder the ready queue so that the insn with lowest LUID will be
issued next.
TARGET_SCHED_REORDER2: If the register pressure is high, indicate to
TARGET_SCHED_DFA_NEW_CYCLE to skip cycles.
TARGET_SCHED_VARIABLE_ISSUE: Cache the value of can_issue_more so that it
can be returned from TARGET_SCHED_REORDER2.
TARGET_SCHED_INIT: Reset the register pressure counting variables. */
#undef TARGET_SCHED_DFA_NEW_CYCLE
#define TARGET_SCHED_DFA_NEW_CYCLE sh_dfa_new_cycle
#undef TARGET_SCHED_INIT_GLOBAL
#define TARGET_SCHED_INIT_GLOBAL sh_md_init_global
#undef TARGET_SCHED_FINISH_GLOBAL
#define TARGET_SCHED_FINISH_GLOBAL sh_md_finish_global
#undef TARGET_SCHED_VARIABLE_ISSUE
#define TARGET_SCHED_VARIABLE_ISSUE sh_variable_issue
#undef TARGET_SCHED_REORDER
#define TARGET_SCHED_REORDER sh_reorder
#undef TARGET_SCHED_REORDER2
#define TARGET_SCHED_REORDER2 sh_reorder2
#undef TARGET_SCHED_INIT
#define TARGET_SCHED_INIT sh_md_init
#undef TARGET_DELEGITIMIZE_ADDRESS
#define TARGET_DELEGITIMIZE_ADDRESS sh_delegitimize_address
#undef TARGET_LEGITIMIZE_ADDRESS
#define TARGET_LEGITIMIZE_ADDRESS sh_legitimize_address
#undef TARGET_CANNOT_MODIFY_JUMPS_P
#define TARGET_CANNOT_MODIFY_JUMPS_P sh_cannot_modify_jumps_p
#undef TARGET_CAN_FOLLOW_JUMP
#define TARGET_CAN_FOLLOW_JUMP sh_can_follow_jump
#undef TARGET_BRANCH_TARGET_REGISTER_CLASS
#define TARGET_BRANCH_TARGET_REGISTER_CLASS sh_target_reg_class
#undef TARGET_BRANCH_TARGET_REGISTER_CALLEE_SAVED
#define TARGET_BRANCH_TARGET_REGISTER_CALLEE_SAVED \
sh_optimize_target_register_callee_saved
#undef TARGET_MS_BITFIELD_LAYOUT_P
#define TARGET_MS_BITFIELD_LAYOUT_P sh_ms_bitfield_layout_p
#undef TARGET_INIT_BUILTINS
#define TARGET_INIT_BUILTINS sh_init_builtins
#undef TARGET_BUILTIN_DECL
#define TARGET_BUILTIN_DECL sh_builtin_decl
#undef TARGET_EXPAND_BUILTIN
#define TARGET_EXPAND_BUILTIN sh_expand_builtin
#undef TARGET_FUNCTION_OK_FOR_SIBCALL
#define TARGET_FUNCTION_OK_FOR_SIBCALL sh_function_ok_for_sibcall
#undef TARGET_CANNOT_COPY_INSN_P
#define TARGET_CANNOT_COPY_INSN_P sh_cannot_copy_insn_p
#undef TARGET_RTX_COSTS
#define TARGET_RTX_COSTS sh_rtx_costs
#undef TARGET_ADDRESS_COST
#define TARGET_ADDRESS_COST sh_address_cost
#undef TARGET_ALLOCATE_INITIAL_VALUE
#define TARGET_ALLOCATE_INITIAL_VALUE sh_allocate_initial_value
#undef TARGET_MACHINE_DEPENDENT_REORG
#define TARGET_MACHINE_DEPENDENT_REORG sh_reorg
#undef TARGET_DWARF_REGISTER_SPAN
#define TARGET_DWARF_REGISTER_SPAN sh_dwarf_register_span
#ifdef HAVE_AS_TLS
#undef TARGET_HAVE_TLS
#define TARGET_HAVE_TLS true
#endif
#undef TARGET_PROMOTE_PROTOTYPES
#define TARGET_PROMOTE_PROTOTYPES sh_promote_prototypes
#undef TARGET_PROMOTE_FUNCTION_MODE
#define TARGET_PROMOTE_FUNCTION_MODE sh_promote_function_mode
#undef TARGET_FUNCTION_VALUE
#define TARGET_FUNCTION_VALUE sh_function_value
#undef TARGET_FUNCTION_VALUE_REGNO_P
#define TARGET_FUNCTION_VALUE_REGNO_P sh_function_value_regno_p
#undef TARGET_LIBCALL_VALUE
#define TARGET_LIBCALL_VALUE sh_libcall_value
#undef TARGET_STRUCT_VALUE_RTX
#define TARGET_STRUCT_VALUE_RTX sh_struct_value_rtx
#undef TARGET_RETURN_IN_MEMORY
#define TARGET_RETURN_IN_MEMORY sh_return_in_memory
#undef TARGET_EXPAND_BUILTIN_SAVEREGS
#define TARGET_EXPAND_BUILTIN_SAVEREGS sh_builtin_saveregs
#undef TARGET_SETUP_INCOMING_VARARGS
#define TARGET_SETUP_INCOMING_VARARGS sh_setup_incoming_varargs
#undef TARGET_STRICT_ARGUMENT_NAMING
#define TARGET_STRICT_ARGUMENT_NAMING sh_strict_argument_naming
#undef TARGET_PRETEND_OUTGOING_VARARGS_NAMED
#define TARGET_PRETEND_OUTGOING_VARARGS_NAMED sh_pretend_outgoing_varargs_named
#undef TARGET_MUST_PASS_IN_STACK
#define TARGET_MUST_PASS_IN_STACK must_pass_in_stack_var_size
#undef TARGET_PASS_BY_REFERENCE
#define TARGET_PASS_BY_REFERENCE sh_pass_by_reference
#undef TARGET_CALLEE_COPIES
#define TARGET_CALLEE_COPIES sh_callee_copies
#undef TARGET_ARG_PARTIAL_BYTES
#define TARGET_ARG_PARTIAL_BYTES sh_arg_partial_bytes
#undef TARGET_FUNCTION_ARG
#define TARGET_FUNCTION_ARG sh_function_arg
#undef TARGET_FUNCTION_ARG_ADVANCE
#define TARGET_FUNCTION_ARG_ADVANCE sh_function_arg_advance
#undef TARGET_ATOMIC_ASSIGN_EXPAND_FENV
#define TARGET_ATOMIC_ASSIGN_EXPAND_FENV sh_atomic_assign_expand_fenv
#undef TARGET_BUILD_BUILTIN_VA_LIST
#define TARGET_BUILD_BUILTIN_VA_LIST sh_build_builtin_va_list
#undef TARGET_EXPAND_BUILTIN_VA_START
#define TARGET_EXPAND_BUILTIN_VA_START sh_va_start
#undef TARGET_GIMPLIFY_VA_ARG_EXPR
#define TARGET_GIMPLIFY_VA_ARG_EXPR sh_gimplify_va_arg_expr
#undef TARGET_SCALAR_MODE_SUPPORTED_P
#define TARGET_SCALAR_MODE_SUPPORTED_P sh_scalar_mode_supported_p
#undef TARGET_VECTOR_MODE_SUPPORTED_P
#define TARGET_VECTOR_MODE_SUPPORTED_P sh_vector_mode_supported_p
#undef TARGET_CHECK_PCH_TARGET_FLAGS
#define TARGET_CHECK_PCH_TARGET_FLAGS sh_check_pch_target_flags
#undef TARGET_DWARF_CALLING_CONVENTION
#define TARGET_DWARF_CALLING_CONVENTION sh_dwarf_calling_convention
#undef TARGET_FRAME_POINTER_REQUIRED
#define TARGET_FRAME_POINTER_REQUIRED sh_frame_pointer_required
#undef TARGET_MODE_EMIT
#define TARGET_MODE_EMIT sh_emit_mode_set
#undef TARGET_MODE_NEEDED
#define TARGET_MODE_NEEDED sh_mode_needed
#undef TARGET_MODE_AFTER
#define TARGET_MODE_AFTER sh_mode_after
#undef TARGET_MODE_ENTRY
#define TARGET_MODE_ENTRY sh_mode_entry
#undef TARGET_MODE_EXIT
#define TARGET_MODE_EXIT sh_mode_exit
#undef TARGET_MODE_PRIORITY
#define TARGET_MODE_PRIORITY sh_mode_priority
/* Return regmode weight for insn. */
#define INSN_REGMODE_WEIGHT(INSN, MODE)\
regmode_weight[((MODE) == SImode) ? 0 : 1][INSN_UID (INSN)]
/* Return current register pressure for regmode. */
#define CURR_REGMODE_PRESSURE(MODE)\
curr_regmode_pressure[((MODE) == SImode) ? 0 : 1]
#undef TARGET_ENCODE_SECTION_INFO
#define TARGET_ENCODE_SECTION_INFO sh_encode_section_info
#undef TARGET_LRA_P
#define TARGET_LRA_P sh_lra_p
#undef TARGET_SECONDARY_RELOAD
#define TARGET_SECONDARY_RELOAD sh_secondary_reload
#undef TARGET_PREFERRED_RELOAD_CLASS
#define TARGET_PREFERRED_RELOAD_CLASS sh_preferred_reload_class
#undef TARGET_CONDITIONAL_REGISTER_USAGE
#define TARGET_CONDITIONAL_REGISTER_USAGE sh_conditional_register_usage
#undef TARGET_LEGITIMATE_ADDRESS_P
#define TARGET_LEGITIMATE_ADDRESS_P sh_legitimate_address_p
#undef TARGET_CANNOT_SUBSTITUTE_MEM_EQUIV_P
#define TARGET_CANNOT_SUBSTITUTE_MEM_EQUIV_P sh_cannot_substitute_mem_equiv_p
#undef TARGET_LEGITIMIZE_ADDRESS_DISPLACEMENT
#define TARGET_LEGITIMIZE_ADDRESS_DISPLACEMENT \
sh_legitimize_address_displacement
#undef TARGET_TRAMPOLINE_INIT
#define TARGET_TRAMPOLINE_INIT sh_trampoline_init
#undef TARGET_TRAMPOLINE_ADJUST_ADDRESS
#define TARGET_TRAMPOLINE_ADJUST_ADDRESS sh_trampoline_adjust_address
#undef TARGET_LEGITIMATE_CONSTANT_P
#define TARGET_LEGITIMATE_CONSTANT_P sh_legitimate_constant_p
#undef TARGET_CANONICALIZE_COMPARISON
#define TARGET_CANONICALIZE_COMPARISON sh_canonicalize_comparison
#undef TARGET_LEGITIMATE_COMBINED_INSN
#define TARGET_LEGITIMATE_COMBINED_INSN sh_legitimate_combined_insn
#undef TARGET_FIXED_CONDITION_CODE_REGS
#define TARGET_FIXED_CONDITION_CODE_REGS sh_fixed_condition_code_regs
#undef TARGET_USE_BY_PIECES_INFRASTRUCTURE_P
#define TARGET_USE_BY_PIECES_INFRASTRUCTURE_P \
sh_use_by_pieces_infrastructure_p
/* Machine-specific symbol_ref flags. */
#define SYMBOL_FLAG_FUNCVEC_FUNCTION (SYMBOL_FLAG_MACH_DEP << 0)
/* The tas.b instruction sets the 7th bit in the byte, i.e. 0x80. This value
is used by optabs.c atomic op expansion code as well as in sync.md. */
#undef TARGET_ATOMIC_TEST_AND_SET_TRUEVAL
#define TARGET_ATOMIC_TEST_AND_SET_TRUEVAL 0x80
struct gcc_target targetm = TARGET_INITIALIZER;
/* Information on the currently selected atomic model.
This is initialized in sh_option_override. */
static sh_atomic_model selected_atomic_model_;
const sh_atomic_model&
selected_atomic_model (void)
{
return selected_atomic_model_;
}
static sh_atomic_model
parse_validate_atomic_model_option (const char* str)
{
const char* model_names[sh_atomic_model::num_models];
model_names[sh_atomic_model::none] = "none";
model_names[sh_atomic_model::soft_gusa] = "soft-gusa";
model_names[sh_atomic_model::hard_llcs] = "hard-llcs";
model_names[sh_atomic_model::soft_tcb] = "soft-tcb";
model_names[sh_atomic_model::soft_imask] = "soft-imask";
const char* model_cdef_names[sh_atomic_model::num_models];
model_cdef_names[sh_atomic_model::none] = "NONE";
model_cdef_names[sh_atomic_model::soft_gusa] = "SOFT_GUSA";
model_cdef_names[sh_atomic_model::hard_llcs] = "HARD_LLCS";
model_cdef_names[sh_atomic_model::soft_tcb] = "SOFT_TCB";
model_cdef_names[sh_atomic_model::soft_imask] = "SOFT_IMASK";
sh_atomic_model ret;
ret.type = sh_atomic_model::none;
ret.name = model_names[sh_atomic_model::none];
ret.cdef_name = model_cdef_names[sh_atomic_model::none];
ret.strict = false;
ret.tcb_gbr_offset = -1;
/* Handle empty string as 'none'. */
if (str == NULL || *str == '\0')
return ret;
#define err_ret(...) do { error (__VA_ARGS__); return ret; } while (0)
std::vector<std::string> tokens;
for (std::stringstream ss (str); ss.good (); )
{
tokens.push_back (std::string ());
std::getline (ss, tokens.back (), ',');
}
if (tokens.empty ())
err_ret ("invalid atomic model option");
/* The first token must be the atomic model name. */
{
for (size_t i = 0; i < sh_atomic_model::num_models; ++i)
if (tokens.front () == model_names[i])
{
ret.type = (sh_atomic_model::enum_type)i;
ret.name = model_names[i];
ret.cdef_name = model_cdef_names[i];
goto got_mode_name;
}
err_ret ("invalid atomic model name \"%s\"", tokens.front ().c_str ());
got_mode_name:;
}
/* Go through the remaining tokens. */
for (size_t i = 1; i < tokens.size (); ++i)
{
if (tokens[i] == "strict")
ret.strict = true;
else if (tokens[i].find ("gbr-offset=") == 0)
{
std::string offset_str = tokens[i].substr (strlen ("gbr-offset="));
ret.tcb_gbr_offset = integral_argument (offset_str.c_str ());
if (offset_str.empty () || ret.tcb_gbr_offset == -1)
err_ret ("could not parse gbr-offset value \"%s\" in atomic model "
"option", offset_str.c_str ());
}
else
err_ret ("unknown parameter \"%s\" in atomic model option",
tokens[i].c_str ());
}
/* Check that the selection makes sense. */
if (TARGET_SHMEDIA && ret.type != sh_atomic_model::none)
err_ret ("atomic operations are not supported on SHmedia");
if (ret.type == sh_atomic_model::soft_gusa && !TARGET_SH3)
err_ret ("atomic model %s is only available on SH3 and SH4 targets",
ret.name);
if (ret.type == sh_atomic_model::hard_llcs && !TARGET_SH4A)
err_ret ("atomic model %s is only available on SH4A targets", ret.name);
if (ret.type == sh_atomic_model::soft_tcb && ret.tcb_gbr_offset == -1)
err_ret ("atomic model %s requires gbr-offset parameter", ret.name);
if (ret.type == sh_atomic_model::soft_tcb
&& (ret.tcb_gbr_offset < 0 || ret.tcb_gbr_offset > 1020
|| (ret.tcb_gbr_offset & 3) != 0))
err_ret ("invalid gbr-offset value \"%d\" for atomic model %s; it must be "
"a multiple of 4 in the range 0-1020", ret.tcb_gbr_offset,
ret.name);
if (ret.type == sh_atomic_model::soft_imask && TARGET_USERMODE)
err_ret ("cannot use atomic model %s in user mode", ret.name);
return ret;
#undef err_ret
}
/* Register SH specific RTL passes. */
extern opt_pass* make_pass_sh_treg_combine (gcc::context* ctx, bool split_insns,
const char* name);
extern opt_pass* make_pass_sh_optimize_sett_clrt (gcc::context* ctx,
const char* name);
static void
register_sh_passes (void)
{
if (!TARGET_SH1)
return;
/* Running the sh_treg_combine pass after ce1 generates better code when
comparisons are combined and reg-reg moves are introduced, because
reg-reg moves will be eliminated afterwards. However, there are quite
some cases where combine will be unable to fold comparison related insns,
thus for now don't do it.
register_pass (make_pass_sh_treg_combine (g, false, "sh_treg_combine1"),
PASS_POS_INSERT_AFTER, "ce1", 1);
*/
/* Run sh_treg_combine pass after combine but before register allocation. */
register_pass (make_pass_sh_treg_combine (g, true, "sh_treg_combine2"),
PASS_POS_INSERT_AFTER, "split1", 1);
/* Run sh_treg_combine pass after register allocation and basic block
reordering as this sometimes creates new opportunities. */
register_pass (make_pass_sh_treg_combine (g, true, "sh_treg_combine3"),
PASS_POS_INSERT_AFTER, "split4", 1);
/* Optimize sett and clrt insns, by e.g. removing them if the T bit value
is known after a conditional branch.
This must be done after basic blocks and branch conditions have
stabilized and won't be changed by further passes. */
register_pass (make_pass_sh_optimize_sett_clrt (g, "sh_optimize_sett_clrt"),
PASS_POS_INSERT_BEFORE, "sched2", 1);
}
/* Implement TARGET_OPTION_OVERRIDE macro. Validate and override
various options, and do some machine dependent initialization. */
static void
sh_option_override (void)
{
int regno;
SUBTARGET_OVERRIDE_OPTIONS;
if (optimize > 1 && !optimize_size)
target_flags |= MASK_SAVE_ALL_TARGET_REGS;
/* Set default values of TARGET_CBRANCHDI4 and TARGET_CMPEQDI_T. */
TARGET_CBRANCHDI4 = 1;
TARGET_CMPEQDI_T = 0;
sh_cpu = PROCESSOR_SH1;
assembler_dialect = 0;
if (TARGET_SH2)
sh_cpu = PROCESSOR_SH2;
if (TARGET_SH2E)
sh_cpu = PROCESSOR_SH2E;
if (TARGET_SH2A)
sh_cpu = PROCESSOR_SH2A;
if (TARGET_SH3)
sh_cpu = PROCESSOR_SH3;
if (TARGET_SH3E)
sh_cpu = PROCESSOR_SH3E;
if (TARGET_SH4)
{
assembler_dialect = 1;
sh_cpu = PROCESSOR_SH4;
}
if (TARGET_SH4A)
{
assembler_dialect = 1;
sh_cpu = PROCESSOR_SH4A;
}
if (TARGET_SH5)
{
sh_cpu = PROCESSOR_SH5;
target_flags |= MASK_ALIGN_DOUBLE;
if (TARGET_SHMEDIA_FPU)
target_flags |= MASK_FMOVD;
if (TARGET_SHMEDIA)
{
/* There are no delay slots on SHmedia. */
flag_delayed_branch = 0;
/* Relaxation isn't yet supported for SHmedia */
target_flags &= ~MASK_RELAX;
/* After reload, if conversion does little good but can cause
ICEs:
- find_if_block doesn't do anything for SH because we don't
have conditional execution patterns. (We use conditional
move patterns, which are handled differently, and only
before reload).
- find_cond_trap doesn't do anything for the SH because we
don't have conditional traps.
- find_if_case_1 uses redirect_edge_and_branch_force in
the only path that does an optimization, and this causes
an ICE when branch targets are in registers.
- find_if_case_2 doesn't do anything for the SHmedia after
reload except when it can redirect a tablejump - and
that's rather rare. */
flag_if_conversion2 = 0;
if (! strcmp (sh_div_str, "call"))
sh_div_strategy = SH_DIV_CALL;
else if (! strcmp (sh_div_str, "call2"))
sh_div_strategy = SH_DIV_CALL2;
if (! strcmp (sh_div_str, "fp") && TARGET_FPU_ANY)
sh_div_strategy = SH_DIV_FP;
else if (! strcmp (sh_div_str, "inv"))
sh_div_strategy = SH_DIV_INV;
else if (! strcmp (sh_div_str, "inv:minlat"))
sh_div_strategy = SH_DIV_INV_MINLAT;
else if (! strcmp (sh_div_str, "inv20u"))
sh_div_strategy = SH_DIV_INV20U;
else if (! strcmp (sh_div_str, "inv20l"))
sh_div_strategy = SH_DIV_INV20L;
else if (! strcmp (sh_div_str, "inv:call2"))
sh_div_strategy = SH_DIV_INV_CALL2;
else if (! strcmp (sh_div_str, "inv:call"))
sh_div_strategy = SH_DIV_INV_CALL;
else if (! strcmp (sh_div_str, "inv:fp"))
{
if (TARGET_FPU_ANY)
sh_div_strategy = SH_DIV_INV_FP;
else
sh_div_strategy = SH_DIV_INV;
}
TARGET_CBRANCHDI4 = 0;
/* Assembler CFI isn't yet fully supported for SHmedia. */
flag_dwarf2_cfi_asm = 0;
}
}
else
{
/* Only the sh64-elf assembler fully supports .quad properly. */
targetm.asm_out.aligned_op.di = NULL;
targetm.asm_out.unaligned_op.di = NULL;
}
/* User/priviledged mode is supported only on SH3*, SH4* and SH5*.
Disable it for everything else. */
if (! (TARGET_SH3 || TARGET_SH5) && TARGET_USERMODE)
TARGET_USERMODE = false;
if (TARGET_SH1)
{
if (! strcmp (sh_div_str, "call-div1"))
sh_div_strategy = SH_DIV_CALL_DIV1;
else if (! strcmp (sh_div_str, "call-fp")
&& (TARGET_FPU_DOUBLE || TARGET_FPU_SINGLE_ONLY
|| (TARGET_SHCOMPACT && TARGET_FPU_ANY)))
sh_div_strategy = SH_DIV_CALL_FP;
else if (! strcmp (sh_div_str, "call-table") && TARGET_DYNSHIFT)
sh_div_strategy = SH_DIV_CALL_TABLE;
else
/* Pick one that makes most sense for the target in general.
It is not much good to use different functions depending
on -Os, since then we'll end up with two different functions
when some of the code is compiled for size, and some for
speed. */
/* SH4 tends to emphasize speed. */
if (TARGET_HARD_SH4)
sh_div_strategy = SH_DIV_CALL_TABLE;
/* These have their own way of doing things. */
else if (TARGET_SH2A)
sh_div_strategy = SH_DIV_INTRINSIC;
/* ??? Should we use the integer SHmedia function instead? */
else if (TARGET_SHCOMPACT && TARGET_FPU_ANY)
sh_div_strategy = SH_DIV_CALL_FP;
/* SH1 .. SH3 cores often go into small-footprint systems, so
default to the smallest implementation available. */
else
sh_div_strategy = SH_DIV_CALL_DIV1;
}
if (!TARGET_SH1)
TARGET_PRETEND_CMOVE = 0;
if (sh_divsi3_libfunc[0])
; /* User supplied - leave it alone. */
else if (TARGET_DIVIDE_CALL_FP)
sh_divsi3_libfunc = "__sdivsi3_i4";
else if (TARGET_DIVIDE_CALL_TABLE)
sh_divsi3_libfunc = "__sdivsi3_i4i";
else if (TARGET_SH5)
sh_divsi3_libfunc = "__sdivsi3_1";
else
sh_divsi3_libfunc = "__sdivsi3";
if (sh_branch_cost == -1)
{
/* The SH1 does not have delay slots, hence we get a pipeline stall
at every branch. The SH4 is superscalar, so the single delay slot
is not sufficient to keep both pipelines filled.
In any case, set the default branch cost to '2', as it results in
slightly overall smaller code and also enables some if conversions
that are required for matching special T bit related insns. */
sh_branch_cost = 2;
}
/* Set -mzdcbranch for SH4 / SH4A if not otherwise specified by the user. */
if (! global_options_set.x_TARGET_ZDCBRANCH && TARGET_HARD_SH4)
TARGET_ZDCBRANCH = 1;
for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
if (! VALID_REGISTER_P (regno))
sh_register_names[regno][0] = '\0';
for (regno = 0; regno < ADDREGNAMES_SIZE; regno++)
if (! VALID_REGISTER_P (ADDREGNAMES_REGNO (regno)))
sh_additional_register_names[regno][0] = '\0';
if ((flag_pic && ! TARGET_PREFERGOT)
|| (TARGET_SHMEDIA && !TARGET_PT_FIXED))
flag_no_function_cse = 1;
if (targetm.small_register_classes_for_mode_p (VOIDmode))
{
/* Never run scheduling before reload, since that can
break global alloc, and generates slower code anyway due
to the pressure on R0. */
/* Enable sched1 for SH4 if the user explicitly requests.
When sched1 is enabled, the ready queue will be reordered by
the target hooks if pressure is high. We can not do this for
PIC, SH3 and lower as they give spill failures for R0. */
if (!TARGET_HARD_SH4 || flag_pic)
flag_schedule_insns = 0;
/* ??? Current exception handling places basic block boundaries
after call_insns. It causes the high pressure on R0 and gives
spill failures for R0 in reload. See PR 22553 and the thread
on gcc-patches
<http://gcc.gnu.org/ml/gcc-patches/2005-10/msg00816.html>. */
else if (flag_exceptions)
{
if (flag_schedule_insns && global_options_set.x_flag_schedule_insns)
warning (0, "ignoring -fschedule-insns because of exception "
"handling bug");
flag_schedule_insns = 0;
}
else if (flag_schedule_insns
&& !global_options_set.x_flag_schedule_insns)
flag_schedule_insns = 0;
}
/* Unwind info is not correct around the CFG unless either a frame
pointer is present or M_A_O_A is set. Fixing this requires rewriting
unwind info generation to be aware of the CFG and propagating states
around edges. */
if ((flag_unwind_tables || flag_asynchronous_unwind_tables
|| flag_exceptions || flag_non_call_exceptions)
&& flag_omit_frame_pointer && !TARGET_ACCUMULATE_OUTGOING_ARGS)
{
warning (0, "unwind tables currently require either a frame pointer "
"or -maccumulate-outgoing-args for correctness");
TARGET_ACCUMULATE_OUTGOING_ARGS = 1;
}
/* The linker relaxation code breaks when a function contains
alignments that are larger than that at the start of a
compilation unit. */
if (TARGET_RELAX)
{
int min_align = align_loops > align_jumps ? align_loops : align_jumps;
/* Also take possible .long constants / mova tables into account. */
if (min_align < 4)
min_align = 4;
if (align_functions < min_align)
align_functions = min_align;
}
if (flag_unsafe_math_optimizations)
{
/* Enable fsca insn for SH4A if not otherwise specified by the user. */
if (global_options_set.x_TARGET_FSCA == 0 && TARGET_SH4A_FP)
TARGET_FSCA = 1;
/* Enable fsrra insn for SH4A if not otherwise specified by the user. */
if (global_options_set.x_TARGET_FSRRA == 0 && TARGET_SH4A_FP)
TARGET_FSRRA = 1;
}
/* Allow fsrra insn only if -funsafe-math-optimizations and
-ffinite-math-only is enabled. */
TARGET_FSRRA = TARGET_FSRRA
&& flag_unsafe_math_optimizations
&& flag_finite_math_only;
/* If the -mieee option was not explicitly set by the user, turn it on
unless -ffinite-math-only was specified. See also PR 33135. */
if (! global_options_set.x_TARGET_IEEE)
TARGET_IEEE = ! flag_finite_math_only;
if (sh_fixed_range_str)
sh_fix_range (sh_fixed_range_str);
/* This target defaults to strict volatile bitfields. */
if (flag_strict_volatile_bitfields < 0 && abi_version_at_least(2))
flag_strict_volatile_bitfields = 1;
sh_override_options_after_change ();
/* Parse atomic model option and make sure it is valid for the current
target CPU. */
selected_atomic_model_
= parse_validate_atomic_model_option (sh_atomic_model_str);
register_sh_passes ();
}
/* Implement targetm.override_options_after_change. */
static void
sh_override_options_after_change (void)
{
/* Adjust loop, jump and function alignment values (in bytes), if those
were not specified by the user using -falign-loops, -falign-jumps
and -falign-functions options.
32 bit alignment is better for speed, because instructions can be
fetched as a pair from a longword boundary. For size use 16 bit
alignment to get more compact code.
Aligning all jumps increases the code size, even if it might
result in slightly faster code. Thus, it is set to the smallest
alignment possible if not specified by the user. */
if (align_loops == 0)
{
if (TARGET_SH5)
align_loops = 8;
else
align_loops = optimize_size ? 2 : 4;
}
if (align_jumps == 0)
{
if (TARGET_SHMEDIA)
align_jumps = 1 << CACHE_LOG;
else
align_jumps = 2;
}
else if (align_jumps < (TARGET_SHMEDIA ? 4 : 2))
align_jumps = TARGET_SHMEDIA ? 4 : 2;
if (align_functions == 0)
{
if (TARGET_SHMEDIA)
align_functions = optimize_size
? FUNCTION_BOUNDARY/8 : (1 << CACHE_LOG);
else
align_functions = optimize_size ? 2 : 4;
}
/* The linker relaxation code breaks when a function contains
alignments that are larger than that at the start of a
compilation unit. */
if (TARGET_RELAX)
{
int min_align = align_loops > align_jumps ? align_loops : align_jumps;
/* Also take possible .long constants / mova tables into account. */
if (min_align < 4)
min_align = 4;
if (align_functions < min_align)
align_functions = min_align;
}
}
/* Print the operand address in x to the stream. */
static void
sh_print_operand_address (FILE *stream, rtx x)
{
switch (GET_CODE (x))
{
case REG:
case SUBREG:
fprintf (stream, "@%s", reg_names[true_regnum (x)]);
break;
case PLUS:
{
rtx base = XEXP (x, 0);
rtx index = XEXP (x, 1);
switch (GET_CODE (index))
{
case CONST_INT:
fprintf (stream, "@(%d,%s)", (int) INTVAL (index),
reg_names[true_regnum (base)]);
break;
case REG:
case SUBREG:
{
int base_num = true_regnum (base);
int index_num = true_regnum (index);
fprintf (stream, "@(r0,%s)",
reg_names[MAX (base_num, index_num)]);
break;
}
default:
gcc_unreachable ();
}
}
break;
case PRE_DEC:
fprintf (stream, "@-%s", reg_names[true_regnum (XEXP (x, 0))]);
break;
case POST_INC:
fprintf (stream, "@%s+", reg_names[true_regnum (XEXP (x, 0))]);
break;
default:
x = mark_constant_pool_use (x);
output_addr_const (stream, x);
break;
}
}
/* Print operand x (an rtx) in assembler syntax to file stream
according to modifier code.
'.' print a .s if insn needs delay slot
',' print LOCAL_LABEL_PREFIX
'@' print trap, rte or rts depending upon pragma interruptness
'#' output a nop if there is nothing to put in the delay slot
''' print likelihood suffix (/u for unlikely).
'>' print branch target if -fverbose-asm
'O' print a constant without the #
'R' print the LSW of a dp value - changes if in little endian
'S' print the MSW of a dp value - changes if in little endian
'T' print the next word of a dp value - same as 'R' in big endian mode.
'M' SHMEDIA: print an `x' if `m' will print `base,index'.
otherwise: print .b / .w / .l / .s / .d suffix if operand is a MEM.
'N' print 'r63' if the operand is (const_int 0).
'd' print a V2SF reg as dN instead of fpN.
'm' print a pair `base,offset' or `base,index', for LD and ST.
'U' Likewise for {LD,ST}{HI,LO}.
'V' print the position of a single bit set.
'W' print the position of a single bit cleared.
't' print a memory address which is a register.
'u' prints the lowest 16 bits of CONST_INT, as an unsigned value.
'o' output an operator. */
static void
sh_print_operand (FILE *stream, rtx x, int code)
{
int regno;
machine_mode mode;
switch (code)
{
tree trapa_attr;
case '.':
if (final_sequence
&& ! INSN_ANNULLED_BRANCH_P (final_sequence->insn (0))
&& get_attr_length (final_sequence->insn (1)))
fprintf (stream, ASSEMBLER_DIALECT ? "/s" : ".s");
break;
case ',':
fprintf (stream, "%s", LOCAL_LABEL_PREFIX);
break;
case '@':
trapa_attr = lookup_attribute ("trap_exit",
DECL_ATTRIBUTES (current_function_decl));
if (trapa_attr)
fprintf (stream, "trapa #%ld",
(long) TREE_INT_CST_LOW (TREE_VALUE (TREE_VALUE (trapa_attr))));
else if (sh_cfun_interrupt_handler_p ())
{
if (sh_cfun_resbank_handler_p ())
fprintf (stream, "resbank\n");
fprintf (stream, "rte");
}
else
fprintf (stream, "rts");
break;
case '#':
/* Output a nop if there's nothing in the delay slot. */
if (dbr_sequence_length () == 0)
fprintf (stream, "\n\tnop");
break;
case '\'':
{
rtx note = find_reg_note (current_output_insn, REG_BR_PROB, 0);
if (note && XINT (note, 0) * 2 < REG_BR_PROB_BASE)
fputs ("/u", stream);
break;
}
case '>':
if (flag_verbose_asm && JUMP_LABEL (current_output_insn))
{
fputs ("\t! target: ", stream);
output_addr_const (stream, JUMP_LABEL (current_output_insn));
}
break;
case 'O':
x = mark_constant_pool_use (x);
output_addr_const (stream, x);
break;
/* N.B.: %R / %S / %T adjust memory addresses by four.
For SHMEDIA, that means they can be used to access the first and
second 32 bit part of a 64 bit (or larger) value that
might be held in floating point registers or memory.
While they can be used to access 64 bit parts of a larger value
held in general purpose registers, that won't work with memory -
neither for fp registers, since the frxx names are used. */
case 'R':
if (REG_P (x) || GET_CODE (x) == SUBREG)
{
regno = true_regnum (x);
regno += FP_REGISTER_P (regno) ? 1 : SH_REG_LSW_OFFSET;
fputs (reg_names[regno], (stream));
}
else if (MEM_P (x))
{
x = adjust_address (x, SImode, 4 * SH_REG_LSW_OFFSET);
sh_print_operand_address (stream, XEXP (x, 0));
}
else
{
rtx sub = NULL_RTX;
mode = GET_MODE (x);
if (mode == VOIDmode)
mode = DImode;
if (GET_MODE_SIZE (mode) >= 8)
sub = simplify_subreg (SImode, x, mode, 4 * SH_REG_LSW_OFFSET);
if (sub)
sh_print_operand (stream, sub, 0);
else
output_operand_lossage ("invalid operand to %%R");
}
break;
case 'S':
if (REG_P (x) || GET_CODE (x) == SUBREG)
{
regno = true_regnum (x);
regno += FP_REGISTER_P (regno) ? 0 : SH_REG_MSW_OFFSET;
fputs (reg_names[regno], (stream));
}
else if (MEM_P (x))
{
x = adjust_address (x, SImode, 4 * SH_REG_MSW_OFFSET);
sh_print_operand_address (stream, XEXP (x, 0));
}
else
{
rtx sub = NULL_RTX;
mode = GET_MODE (x);
if (mode == VOIDmode)
mode = DImode;
if (GET_MODE_SIZE (mode) >= 8)
sub = simplify_subreg (SImode, x, mode, 4 * SH_REG_MSW_OFFSET);
if (sub)
sh_print_operand (stream, sub, 0);
else
output_operand_lossage ("invalid operand to %%S");
}
break;
case 'T':
/* Next word of a double. */
switch (GET_CODE (x))
{
case REG:
fputs (reg_names[REGNO (x) + 1], (stream));
break;
case MEM:
if (GET_CODE (XEXP (x, 0)) != PRE_DEC
&& GET_CODE (XEXP (x, 0)) != POST_INC)
x = adjust_address (x, SImode, 4);
sh_print_operand_address (stream, XEXP (x, 0));
break;
default:
break;
}
break;
case 't':
gcc_assert (MEM_P (x));
x = XEXP (x, 0);
switch (GET_CODE (x))
{
case REG:
case SUBREG:
sh_print_operand (stream, x, 0);
break;
default:
break;
}
break;
case 'o':
switch (GET_CODE (x))
{
case PLUS: fputs ("add", stream); break;
case MINUS: fputs ("sub", stream); break;
case MULT: fputs ("mul", stream); break;
case DIV: fputs ("div", stream); break;
case EQ: fputs ("eq", stream); break;
case NE: fputs ("ne", stream); break;
case GT: case LT: fputs ("gt", stream); break;
case GE: case LE: fputs ("ge", stream); break;
case GTU: case LTU: fputs ("gtu", stream); break;
case GEU: case LEU: fputs ("geu", stream); break;
default:
break;
}
break;
case 'M':
if (TARGET_SHMEDIA)
{
if (MEM_P (x)
&& GET_CODE (XEXP (x, 0)) == PLUS
&& (REG_P (XEXP (XEXP (x, 0), 1))
|| GET_CODE (XEXP (XEXP (x, 0), 1)) == SUBREG))
fputc ('x', stream);
}
else
{
if (MEM_P (x))
{
switch (GET_MODE (x))
{
case QImode: fputs (".b", stream); break;
case HImode: fputs (".w", stream); break;
case SImode: fputs (".l", stream); break;
case SFmode: fputs (".s", stream); break;
case DFmode: fputs (".d", stream); break;
default: gcc_unreachable ();
}
}
}
break;
case 'm':
gcc_assert (MEM_P (x));
x = XEXP (x, 0);
/* Fall through. */
case 'U':
switch (GET_CODE (x))
{
case REG:
case SUBREG:
sh_print_operand (stream, x, 0);
fputs (", 0", stream);
break;
case PLUS:
sh_print_operand (stream, XEXP (x, 0), 0);
fputs (", ", stream);
sh_print_operand (stream, XEXP (x, 1), 0);
break;
default:
gcc_unreachable ();
}
break;
case 'V':
{
int num = exact_log2 (INTVAL (x));
gcc_assert (num >= 0);
fprintf (stream, "#%d", num);
}
break;
case 'W':
{
int num = exact_log2 (~INTVAL (x));
gcc_assert (num >= 0);
fprintf (stream, "#%d", num);
}
break;
case 'd':
gcc_assert (REG_P (x) && GET_MODE (x) == V2SFmode);
fprintf ((stream), "d%s", reg_names[REGNO (x)] + 1);
break;
case 'N':
if (x == CONST0_RTX (GET_MODE (x)))
{
fprintf ((stream), "r63");
break;
}
goto default_output;
case 'u':
if (CONST_INT_P (x))
{
fprintf ((stream), "%u", (unsigned) INTVAL (x) & (0x10000 - 1));
break;
}
/* Fall through. */
default_output:
default:
regno = 0;
mode = GET_MODE (x);
switch (GET_CODE (x))
{
case TRUNCATE:
{
rtx inner = XEXP (x, 0);
int offset = 0;
machine_mode inner_mode;
/* We might see SUBREGs with vector mode registers inside. */
if (GET_CODE (inner) == SUBREG
&& (GET_MODE_SIZE (GET_MODE (inner))
== GET_MODE_SIZE (GET_MODE (SUBREG_REG (inner))))
&& subreg_lowpart_p (inner))
inner = SUBREG_REG (inner);
if (CONST_INT_P (inner))
{
x = GEN_INT (trunc_int_for_mode (INTVAL (inner), GET_MODE (x)));
goto default_output;
}
inner_mode = GET_MODE (inner);
if (GET_CODE (inner) == SUBREG
&& (GET_MODE_SIZE (GET_MODE (inner))
< GET_MODE_SIZE (GET_MODE (SUBREG_REG (inner))))
&& REG_P (SUBREG_REG (inner)))
{
offset = subreg_regno_offset (REGNO (SUBREG_REG (inner)),
GET_MODE (SUBREG_REG (inner)),
SUBREG_BYTE (inner),
GET_MODE (inner));
inner = SUBREG_REG (inner);
}
if (!REG_P (inner) || GET_MODE_SIZE (inner_mode) > 8)
abort ();
/* Floating point register pairs are always big endian;
general purpose registers are 64 bit wide. */
regno = REGNO (inner);
regno = (HARD_REGNO_NREGS (regno, inner_mode)
- HARD_REGNO_NREGS (regno, mode))
+ offset;
x = inner;
goto reg;
}
case SIGN_EXTEND:
x = XEXP (x, 0);
goto reg;
/* FIXME: We need this on SHmedia32 because reload generates
some sign-extended HI or QI loads into DImode registers
but, because Pmode is SImode, the address ends up with a
subreg:SI of the DImode register. Maybe reload should be
fixed so as to apply alter_subreg to such loads? */
case IF_THEN_ELSE:
gcc_assert (trapping_target_operand (x, VOIDmode));
x = XEXP (XEXP (x, 2), 0);
goto default_output;
case SUBREG:
gcc_assert (SUBREG_BYTE (x) == 0
&& REG_P (SUBREG_REG (x)));
x = SUBREG_REG (x);
/* Fall through. */
reg:
case REG:
regno += REGNO (x);
if (FP_REGISTER_P (regno)
&& mode == V16SFmode)
fprintf ((stream), "mtrx%s", reg_names[regno] + 2);
else if (FP_REGISTER_P (REGNO (x))
&& mode == V4SFmode)
fprintf ((stream), "fv%s", reg_names[regno] + 2);
else if (REG_P (x)
&& mode == V2SFmode)
fprintf ((stream), "fp%s", reg_names[regno] + 2);
else if (FP_REGISTER_P (REGNO (x))
&& GET_MODE_SIZE (mode) > 4)
fprintf ((stream), "d%s", reg_names[regno] + 1);
else
fputs (reg_names[regno], (stream));
break;
case MEM:
output_address (XEXP (x, 0));
break;
default:
if (TARGET_SH1)
fputc ('#', stream);
output_addr_const (stream, x);
break;
}
break;
}
}
static bool
sh_print_operand_punct_valid_p (unsigned char code)
{
return (code == '.' || code == '#' || code == '@' || code == ','
|| code == '$' || code == '\'' || code == '>');
}
/* Implement TARGET_ASM_OUTPUT_ADDR_CONST_EXTRA. */
static bool
sh_asm_output_addr_const_extra (FILE *file, rtx x)
{
if (GET_CODE (x) == UNSPEC)
{
switch (XINT (x, 1))
{
case UNSPEC_DATALABEL:
fputs ("datalabel ", file);
output_addr_const (file, XVECEXP (x, 0, 0));
break;
case UNSPEC_PIC:
/* GLOBAL_OFFSET_TABLE or local symbols, no suffix. */
output_addr_const (file, XVECEXP (x, 0, 0));
break;
case UNSPEC_GOT:
output_addr_const (file, XVECEXP (x, 0, 0));
fputs ("@GOT", file);
break;
case UNSPEC_GOTOFF:
output_addr_const (file, XVECEXP (x, 0, 0));
fputs ("@GOTOFF", file);
break;
case UNSPEC_PLT:
output_addr_const (file, XVECEXP (x, 0, 0));
fputs ("@PLT", file);
break;
case UNSPEC_GOTPLT:
output_addr_const (file, XVECEXP (x, 0, 0));
fputs ("@GOTPLT", file);
break;
case UNSPEC_DTPOFF:
output_addr_const (file, XVECEXP (x, 0, 0));
fputs ("@DTPOFF", file);
break;
case UNSPEC_GOTTPOFF:
output_addr_const (file, XVECEXP (x, 0, 0));
fputs ("@GOTTPOFF", file);
break;
case UNSPEC_TPOFF:
output_addr_const (file, XVECEXP (x, 0, 0));
fputs ("@TPOFF", file);
break;
case UNSPEC_CALLER:
{
char name[32];
/* LPCS stands for Label for PIC Call Site. */
targetm.asm_out.generate_internal_label (name, "LPCS",
INTVAL (XVECEXP (x, 0, 0)));
assemble_name (file, name);
}
break;
case UNSPEC_EXTRACT_S16:
case UNSPEC_EXTRACT_U16:
{
rtx val, shift;
val = XVECEXP (x, 0, 0);
shift = XVECEXP (x, 0, 1);
fputc ('(', file);
if (shift != const0_rtx)
fputc ('(', file);
if (GET_CODE (val) == CONST
|| GET_RTX_CLASS (GET_CODE (val)) != RTX_OBJ)
{
fputc ('(', file);
output_addr_const (file, val);
fputc (')', file);
}
else
output_addr_const (file, val);
if (shift != const0_rtx)
{
fputs (" >> ", file);
output_addr_const (file, shift);
fputc (')', file);
}
fputs (" & 65535)", file);
}
break;
case UNSPEC_SYMOFF:
output_addr_const (file, XVECEXP (x, 0, 0));
fputc ('-', file);
if (GET_CODE (XVECEXP (x, 0, 1)) == CONST)
{
fputc ('(', file);
output_addr_const (file, XVECEXP (x, 0, 1));
fputc (')', file);
}
else
output_addr_const (file, XVECEXP (x, 0, 1));
break;
case UNSPEC_PCREL_SYMOFF:
output_addr_const (file, XVECEXP (x, 0, 0));
fputs ("-(", file);
output_addr_const (file, XVECEXP (x, 0, 1));
fputs ("-.)", file);
break;
default:
return false;
}
return true;
}
else
return false;
}
/* Encode symbol attributes of a SYMBOL_REF into its
SYMBOL_REF_FLAGS. */
static void
sh_encode_section_info (tree decl, rtx rtl, int first)
{
default_encode_section_info (decl, rtl, first);
if (TREE_CODE (decl) == FUNCTION_DECL
&& sh2a_function_vector_p (decl) && TARGET_SH2A)
SYMBOL_REF_FLAGS (XEXP (rtl, 0)) |= SYMBOL_FLAG_FUNCVEC_FUNCTION;
}
/* Prepare operands for a move define_expand; specifically, one of the
operands must be in a register. */
void
prepare_move_operands (rtx operands[], machine_mode mode)
{
if ((mode == SImode || mode == DImode)
&& flag_pic
&& ! ((mode == Pmode || mode == ptr_mode)
&& tls_symbolic_operand (operands[1], Pmode) != TLS_MODEL_NONE))
{
rtx temp;
if (SYMBOLIC_CONST_P (operands[1]))
{
if (MEM_P (operands[0]))
operands[1] = force_reg (Pmode, operands[1]);
else if (TARGET_SHMEDIA
&& GET_CODE (operands[1]) == LABEL_REF
&& target_reg_operand (operands[0], mode))
/* It's ok. */;
else
{
temp = (!can_create_pseudo_p ()
? operands[0]
: gen_reg_rtx (Pmode));
operands[1] = legitimize_pic_address (operands[1], mode, temp);
}
}
else if (GET_CODE (operands[1]) == CONST
&& GET_CODE (XEXP (operands[1], 0)) == PLUS
&& SYMBOLIC_CONST_P (XEXP (XEXP (operands[1], 0), 0)))
{
temp = !can_create_pseudo_p () ? operands[0] : gen_reg_rtx (Pmode);
temp = legitimize_pic_address (XEXP (XEXP (operands[1], 0), 0),
mode, temp);
operands[1] = expand_binop (mode, add_optab, temp,
XEXP (XEXP (operands[1], 0), 1),
(!can_create_pseudo_p ()
? temp
: gen_reg_rtx (Pmode)),
0, OPTAB_LIB_WIDEN);
}
}
if (! reload_in_progress && ! reload_completed)
{
/* Copy the source to a register if both operands aren't registers. */
if (! register_operand (operands[0], mode)
&& ! sh_register_operand (operands[1], mode))
operands[1] = copy_to_mode_reg (mode, operands[1]);
if (MEM_P (operands[0]) && ! memory_operand (operands[0], mode))
{
/* This is like change_address_1 (operands[0], mode, 0, 1) ,
except that we can't use that function because it is static. */
rtx new_rtx = change_address (operands[0], mode, 0);
MEM_COPY_ATTRIBUTES (new_rtx, operands[0]);
operands[0] = new_rtx;
}
/* This case can happen while generating code to move the result
of a library call to the target. Reject `st r0,@(rX,rY)' because
reload will fail to find a spill register for rX, since r0 is already
being used for the source. */
else if (TARGET_SH1
&& refers_to_regno_p (R0_REG, operands[1])
&& MEM_P (operands[0])
&& GET_CODE (XEXP (operands[0], 0)) == PLUS
&& REG_P (XEXP (XEXP (operands[0], 0), 1)))
operands[1] = copy_to_mode_reg (mode, operands[1]);
/* When the displacement addressing is used, RA will assign r0 to
the pseudo register operand for the QI/HImode load/store.
This tends to make a long live range for R0 and might cause
anomalous register spills in some case with LRA. See PR
target/55212.
We split possible load/store to two move insns via r0 so as to
shorten R0 live range. It will make some codes worse but will
win on avarage for LRA. */
else if (sh_lra_p ()
&& TARGET_SH1 && ! TARGET_SH2A
&& (mode == QImode || mode == HImode)
&& ((REG_P (operands[0]) && MEM_P (operands[1]))
|| (REG_P (operands[1]) && MEM_P (operands[0]))))
{
bool load_p = REG_P (operands[0]);
rtx reg = operands[load_p ? 0 : 1];
rtx adr = XEXP (operands[load_p ? 1 : 0], 0);
if (REGNO (reg) >= FIRST_PSEUDO_REGISTER
&& GET_CODE (adr) == PLUS
&& REG_P (XEXP (adr, 0))
&& (REGNO (XEXP (adr, 0)) >= FIRST_PSEUDO_REGISTER)
&& CONST_INT_P (XEXP (adr, 1))
&& INTVAL (XEXP (adr, 1)) != 0
&& sh_legitimate_index_p (mode, XEXP (adr, 1), false, true))
{
rtx r0_rtx = gen_rtx_REG (mode, R0_REG);
emit_move_insn (r0_rtx, operands[1]);
operands[1] = r0_rtx;
}
}
}
if (mode == Pmode || mode == ptr_mode)
{
rtx op0, op1, opc;
enum tls_model tls_kind;
op0 = operands[0];
op1 = operands[1];
if (GET_CODE (op1) == CONST
&& GET_CODE (XEXP (op1, 0)) == PLUS
&& (tls_symbolic_operand (XEXP (XEXP (op1, 0), 0), Pmode)
!= TLS_MODEL_NONE))
{
opc = XEXP (XEXP (op1, 0), 1);
op1 = XEXP (XEXP (op1, 0), 0);
}
else
opc = NULL_RTX;
if (! reload_in_progress && ! reload_completed
&& (tls_kind = tls_symbolic_operand (op1, Pmode)) != TLS_MODEL_NONE)
{
rtx tga_op1, tga_ret, tmp, tmp2;
if (! flag_pic
&& (tls_kind == TLS_MODEL_GLOBAL_DYNAMIC
|| tls_kind == TLS_MODEL_LOCAL_DYNAMIC
|| tls_kind == TLS_MODEL_INITIAL_EXEC))
{
static int got_labelno;
/* Don't schedule insns for getting GOT address when
the first scheduling is enabled, to avoid spill
failures for R0. */
if (flag_schedule_insns)
emit_insn (gen_blockage ());
emit_insn (gen_GOTaddr2picreg (GEN_INT (++got_labelno)));
emit_use (gen_rtx_REG (SImode, PIC_REG));
if (flag_schedule_insns)
emit_insn (gen_blockage ());
}
switch (tls_kind)
{
case TLS_MODEL_GLOBAL_DYNAMIC:
tga_ret = gen_rtx_REG (Pmode, R0_REG);
emit_call_insn (gen_tls_global_dynamic (tga_ret, op1));
tmp = gen_reg_rtx (Pmode);
emit_move_insn (tmp, tga_ret);
op1 = tmp;
break;
case TLS_MODEL_LOCAL_DYNAMIC:
tga_ret = gen_rtx_REG (Pmode, R0_REG);
emit_call_insn (gen_tls_local_dynamic (tga_ret, op1));
tmp = gen_reg_rtx (Pmode);
emit_move_insn (tmp, tga_ret);
if (register_operand (op0, Pmode))
tmp2 = op0;
else
tmp2 = gen_reg_rtx (Pmode);
emit_insn (gen_symDTPOFF2reg (tmp2, op1, tmp));
op1 = tmp2;
break;
case TLS_MODEL_INITIAL_EXEC:
tga_op1 = !can_create_pseudo_p () ? op0 : gen_reg_rtx (Pmode);
tmp = gen_sym2GOTTPOFF (op1);
emit_insn (gen_tls_initial_exec (tga_op1, tmp));
op1 = tga_op1;
break;
case TLS_MODEL_LOCAL_EXEC:
tmp2 = gen_reg_rtx (Pmode);
emit_insn (gen_store_gbr (tmp2));
tmp = gen_reg_rtx (Pmode);
emit_insn (gen_symTPOFF2reg (tmp, op1));
if (register_operand (op0, Pmode))
op1 = op0;
else
op1 = gen_reg_rtx (Pmode);
emit_insn (gen_addsi3 (op1, tmp, tmp2));
break;
default:
gcc_unreachable ();
}
if (opc)
emit_insn (gen_addsi3 (op1, op1, force_reg (SImode, opc)));
operands[1] = op1;
}
}
}
/* Implement the canonicalize_comparison target hook for the combine
pass. For the target hook this function is invoked via
sh_canonicalize_comparison. This function is also re-used to
canonicalize comparisons in cbranch pattern expanders. */
static void
sh_canonicalize_comparison (enum rtx_code& cmp, rtx& op0, rtx& op1,
machine_mode mode,
bool op0_preserve_value)
{
/* When invoked from within the combine pass the mode is not specified,
so try to get it from one of the operands. */
if (mode == VOIDmode)
mode = GET_MODE (op0);
if (mode == VOIDmode)
mode = GET_MODE (op1);
// We need to have a mode to do something useful here.
if (mode == VOIDmode)
return;
// Currently, we don't deal with floats here.
if (GET_MODE_CLASS (mode) == MODE_FLOAT)
return;
// Make sure that the constant operand is the second operand.
if (CONST_INT_P (op0) && !CONST_INT_P (op1))
{
if (op0_preserve_value)
return;
std::swap (op0, op1);
cmp = swap_condition (cmp);
}
if (CONST_INT_P (op1))
{
/* Try to adjust the constant operand in such a way that available
comparison insns can be utilized better and the constant can be
loaded with a 'mov #imm,Rm' insn. This avoids a load from the
constant pool. */
const HOST_WIDE_INT val = INTVAL (op1);
/* x > -1 --> x >= 0
x > 0xFFFFFF7F --> x >= 0xFFFFFF80
x <= -1 --> x < 0
x <= 0xFFFFFF7F --> x < 0xFFFFFF80 */
if ((val == -1 || val == -0x81) && (cmp == GT || cmp == LE))
{
cmp = cmp == GT ? GE : LT;
op1 = gen_int_mode (val + 1, mode);
}
/* x >= 1 --> x > 0
x >= 0x80 --> x > 0x7F
x < 1 --> x <= 0
x < 0x80 --> x <= 0x7F */
else if ((val == 1 || val == 0x80) && (cmp == GE || cmp == LT))
{
cmp = cmp == GE ? GT : LE;
op1 = gen_int_mode (val - 1, mode);
}
/* unsigned x >= 1 --> x != 0
unsigned x < 1 --> x == 0 */
else if (val == 1 && (cmp == GEU || cmp == LTU))
{
cmp = cmp == GEU ? NE : EQ;
op1 = CONST0_RTX (mode);
}
/* unsigned x >= 0x80 --> unsigned x > 0x7F
unsigned x < 0x80 --> unsigned x < 0x7F */
else if (val == 0x80 && (cmp == GEU || cmp == LTU))
{
cmp = cmp == GEU ? GTU : LEU;
op1 = gen_int_mode (val - 1, mode);
}
/* unsigned x > 0 --> x != 0
unsigned x <= 0 --> x == 0 */
else if (val == 0 && (cmp == GTU || cmp == LEU))
cmp = cmp == GTU ? NE : EQ;
/* unsigned x > 0x7FFFFFFF --> signed x < 0
unsigned x <= 0x7FFFFFFF --> signed x >= 0 */
else if (mode == SImode && (cmp == GTU || cmp == LEU)
&& val == 0x7FFFFFFF)
{
cmp = cmp == GTU ? LT : GE;
op1 = const0_rtx;
}
/* unsigned x >= 0x80000000 --> signed x < 0
unsigned x < 0x80000000 --> signed x >= 0 */
else if (mode == SImode && (cmp == GEU || cmp == LTU)
&& (unsigned HOST_WIDE_INT)val
== ((unsigned HOST_WIDE_INT)0x7FFFFFFF + 1))
{
cmp = cmp == GEU ? LT : GE;
op1 = const0_rtx;
}
}
}
/* This function implements the canonicalize_comparison target hook.
This wrapper around the internally used sh_canonicalize_comparison
function is needed to do the enum rtx_code <-> int conversion.
Target hooks cannot use enum rtx_code in its definition. */
static void
sh_canonicalize_comparison (int *code, rtx *op0, rtx *op1,
bool op0_preserve_value)
{
enum rtx_code tmp_code = (enum rtx_code)*code;
sh_canonicalize_comparison (tmp_code, *op0, *op1,
VOIDmode, op0_preserve_value);
*code = (int)tmp_code;
}
/* This function implements the legitimate_combined_insn target hook,
which the combine pass uses to early reject combined insns, before
it tries to recog the insn and determine its cost. */
static bool
sh_legitimate_combined_insn (rtx_insn* insn)
{
/* Reject combinations of memory loads and zero extensions, as these
interfere with other combine patterns such as zero extracts and bit
tests. The SH2A movu.{b|w} insns are formed later in the
'sh_optimize_extu_exts' pass after combine/split1. */
rtx p = PATTERN (insn);
if (GET_CODE (p) == SET
&& REG_P (XEXP (p, 0)) && GET_MODE (XEXP (p, 0)) == SImode
&& GET_CODE (XEXP (p, 1)) == ZERO_EXTEND
&& MEM_P (XEXP (XEXP (p, 1), 0)))
return false;
return true;
}
bool
sh_fixed_condition_code_regs (unsigned int* p1, unsigned int* p2)
{
*p1 = T_REG;
*p2 = INVALID_REGNUM;
return true;
}
enum rtx_code
prepare_cbranch_operands (rtx *operands, machine_mode mode,
enum rtx_code comparison)
{
/* The scratch reg is only available when this is invoked from within
the cbranchdi4_i splitter, through expand_cbranchdi4. */
rtx scratch = NULL_RTX;
if (comparison == LAST_AND_UNUSED_RTX_CODE)
comparison = GET_CODE (operands[0]);
else
scratch = operands[4];
sh_canonicalize_comparison (comparison, operands[1], operands[2],
mode, false);
/* Notice that this function is also invoked after reload by
the cbranchdi4_i pattern, through expand_cbranchdi4. */
rtx op1 = operands[1];
if (can_create_pseudo_p ())
operands[1] = force_reg (mode, op1);
/* When we are handling DImode comparisons, we want to keep constants so
that we can optimize the component comparisons; however, memory loads
are better issued as a whole so that they can be scheduled well.
SImode equality comparisons allow I08 constants, but only when they
compare r0. Hence, if operands[1] has to be loaded from somewhere else
into a register, that register might as well be r0, and we allow the
constant. If it is already in a register, this is likely to be
allocated to a different hard register, thus we load the constant into
a register unless it is zero. */
if (!REG_P (operands[2])
&& (!CONST_INT_P (operands[2])
|| (mode == SImode && operands[2] != CONST0_RTX (SImode)
&& ((comparison != EQ && comparison != NE)
|| (REG_P (op1) && REGNO (op1) != R0_REG)
|| !satisfies_constraint_I08 (operands[2])))))
{
if (scratch && GET_MODE (scratch) == mode)
{
emit_move_insn (scratch, operands[2]);
operands[2] = scratch;
}
else if (can_create_pseudo_p ())
operands[2] = force_reg (mode, operands[2]);
}
return comparison;
}
void
expand_cbranchsi4 (rtx *operands, enum rtx_code comparison, int probability)
{
rtx (*branch_expander) (rtx) = gen_branch_true;
comparison = prepare_cbranch_operands (operands, SImode, comparison);
switch (comparison)
{
case NE: case LT: case LE: case LTU: case LEU:
comparison = reverse_condition (comparison);
branch_expander = gen_branch_false;
default: ;
}
emit_insn (gen_rtx_SET (VOIDmode, get_t_reg_rtx (),
gen_rtx_fmt_ee (comparison, SImode,
operands[1], operands[2])));
rtx_insn *jump = emit_jump_insn (branch_expander (operands[3]));
if (probability >= 0)
add_int_reg_note (jump, REG_BR_PROB, probability);
}
/* ??? How should we distribute probabilities when more than one branch
is generated. So far we only have some ad-hoc observations:
- If the operands are random, they are likely to differ in both parts.
- If comparing items in a hash chain, the operands are random or equal;
operation should be EQ or NE.
- If items are searched in an ordered tree from the root, we can expect
the highpart to be unequal about half of the time; operation should be
an inequality comparison, operands non-constant, and overall probability
about 50%. Likewise for quicksort.
- Range checks will be often made against constants. Even if we assume for
simplicity an even distribution of the non-constant operand over a
sub-range here, the same probability could be generated with differently
wide sub-ranges - as long as the ratio of the part of the subrange that
is before the threshold to the part that comes after the threshold stays
the same. Thus, we can't really tell anything here;
assuming random distribution is at least simple.
*/
bool
expand_cbranchdi4 (rtx *operands, enum rtx_code comparison)
{
enum rtx_code msw_taken, msw_skip, lsw_taken;
rtx_code_label *skip_label = NULL;
rtx op1h, op1l, op2h, op2l;
int num_branches;
int prob, rev_prob;
int msw_taken_prob = -1, msw_skip_prob = -1, lsw_taken_prob = -1;
rtx scratch = operands[4];
comparison = prepare_cbranch_operands (operands, DImode, comparison);
op1h = gen_highpart_mode (SImode, DImode, operands[1]);
op2h = gen_highpart_mode (SImode, DImode, operands[2]);
op1l = gen_lowpart (SImode, operands[1]);
op2l = gen_lowpart (SImode, operands[2]);
msw_taken = msw_skip = lsw_taken = LAST_AND_UNUSED_RTX_CODE;
prob = split_branch_probability;
rev_prob = REG_BR_PROB_BASE - prob;
switch (comparison)
{
/* ??? Should we use the cmpeqdi_t pattern for equality comparisons?
That costs 1 cycle more when the first branch can be predicted taken,
but saves us mispredicts because only one branch needs prediction.
It also enables generating the cmpeqdi_t-1 pattern. */
case EQ:
if (TARGET_CMPEQDI_T)
{
emit_insn (gen_cmpeqdi_t (operands[1], operands[2]));
emit_jump_insn (gen_branch_true (operands[3]));
return true;
}
msw_skip = NE;
lsw_taken = EQ;
if (prob >= 0)
{
// If we had more precision, we'd use rev_prob - (rev_prob >> 32) .
msw_skip_prob = rev_prob;
if (REG_BR_PROB_BASE <= 65535)
lsw_taken_prob = prob ? REG_BR_PROB_BASE : 0;
else
{
lsw_taken_prob
= (prob
? (REG_BR_PROB_BASE
- ((gcov_type) REG_BR_PROB_BASE * rev_prob
/ ((gcov_type) prob << 32)))
: 0);
}
}
break;
case NE:
if (TARGET_CMPEQDI_T)
{
emit_insn (gen_cmpeqdi_t (operands[1], operands[2]));
emit_jump_insn (gen_branch_false (operands[3]));
return true;
}
msw_taken = NE;
msw_taken_prob = prob;
lsw_taken = NE;
lsw_taken_prob = 0;
break;
case GTU: case GT:
msw_taken = comparison;
if (CONST_INT_P (op2l) && INTVAL (op2l) == -1)
break;
if (comparison != GTU || op2h != CONST0_RTX (SImode))
msw_skip = swap_condition (msw_taken);
lsw_taken = GTU;
break;
case GEU: case GE:
if (op2l == CONST0_RTX (SImode))
msw_taken = comparison;
else
{
msw_taken = comparison == GE ? GT : GTU;
msw_skip = swap_condition (msw_taken);
lsw_taken = GEU;
}
break;
case LTU: case LT:
msw_taken = comparison;
if (op2l == CONST0_RTX (SImode))
break;
msw_skip = swap_condition (msw_taken);
lsw_taken = LTU;
break;
case LEU: case LE:
if (CONST_INT_P (op2l) && INTVAL (op2l) == -1)
msw_taken = comparison;
else
{
lsw_taken = LEU;
if (comparison == LE)
msw_taken = LT;
else if (op2h != CONST0_RTX (SImode))
msw_taken = LTU;
else
{
msw_skip = swap_condition (LTU);
break;
}
msw_skip = swap_condition (msw_taken);
}
break;
default: return false;
}
num_branches = ((msw_taken != LAST_AND_UNUSED_RTX_CODE)
+ (msw_skip != LAST_AND_UNUSED_RTX_CODE)
+ (lsw_taken != LAST_AND_UNUSED_RTX_CODE));
if (comparison != EQ && comparison != NE && num_branches > 1)
{
if (!CONSTANT_P (operands[2])
&& prob >= (int) (REG_BR_PROB_BASE * 3 / 8U)
&& prob <= (int) (REG_BR_PROB_BASE * 5 / 8U))
{
msw_taken_prob = prob / 2U;
msw_skip_prob
= REG_BR_PROB_BASE * rev_prob / (REG_BR_PROB_BASE + rev_prob);
lsw_taken_prob = prob;
}
else
{
msw_taken_prob = prob;
msw_skip_prob = REG_BR_PROB_BASE;
/* ??? If we have a constant op2h, should we use that when
calculating lsw_taken_prob? */
lsw_taken_prob = prob;
}
}
operands[1] = op1h;
operands[2] = op2h;
operands[4] = NULL_RTX;
if (reload_completed
&& ! arith_reg_or_0_operand (op2h, SImode)
&& (true_regnum (op1h) || (comparison != EQ && comparison != NE))
&& (msw_taken != LAST_AND_UNUSED_RTX_CODE
|| msw_skip != LAST_AND_UNUSED_RTX_CODE))
{
emit_move_insn (scratch, operands[2]);
operands[2] = scratch;
}
if (msw_taken != LAST_AND_UNUSED_RTX_CODE)
expand_cbranchsi4 (operands, msw_taken, msw_taken_prob);
if (msw_skip != LAST_AND_UNUSED_RTX_CODE)
{
rtx taken_label = operands[3];
/* Operands were possibly modified, but msw_skip doesn't expect this.
Always use the original ones. */
if (msw_taken != LAST_AND_UNUSED_RTX_CODE)
{
operands[1] = op1h;
operands[2] = op2h;
if (reload_completed
&& ! arith_reg_or_0_operand (op2h, SImode)
&& (true_regnum (op1h) || (comparison != EQ && comparison != NE)))
{
emit_move_insn (scratch, operands[2]);
operands[2] = scratch;
}
}
operands[3] = skip_label = gen_label_rtx ();
expand_cbranchsi4 (operands, msw_skip, msw_skip_prob);
operands[3] = taken_label;
}
operands[1] = op1l;
operands[2] = op2l;
if (lsw_taken != LAST_AND_UNUSED_RTX_CODE)
{
if (reload_completed
&& ! arith_reg_or_0_operand (op2l, SImode)
&& (true_regnum (op1l) || (lsw_taken != EQ && lsw_taken != NE)))
{
emit_move_insn (scratch, operands[2]);
operands[2] = scratch;
}
expand_cbranchsi4 (operands, lsw_taken, lsw_taken_prob);
}
if (msw_skip != LAST_AND_UNUSED_RTX_CODE)
emit_label (skip_label);
return true;
}
/* Given an operand, return 1 if the evaluated operand plugged into an
if_then_else will result in a branch_true, 0 if branch_false, or
-1 if neither nor applies. The truth table goes like this:
op | cmpval | code | result
---------+--------+---------+--------------------
T (0) | 0 | EQ (1) | 0 = 0 ^ (0 == 1)
T (0) | 1 | EQ (1) | 1 = 0 ^ (1 == 1)
T (0) | 0 | NE (0) | 1 = 0 ^ (0 == 0)
T (0) | 1 | NE (0) | 0 = 0 ^ (1 == 0)
!T (1) | 0 | EQ (1) | 1 = 1 ^ (0 == 1)
!T (1) | 1 | EQ (1) | 0 = 1 ^ (1 == 1)
!T (1) | 0 | NE (0) | 0 = 1 ^ (0 == 0)
!T (1) | 1 | NE (0) | 1 = 1 ^ (1 == 0) */
int
sh_eval_treg_value (rtx op)
{
if (t_reg_operand (op, GET_MODE (op)))
return 1;
if (negt_reg_operand (op, GET_MODE (op)))
return 0;
rtx_code code = GET_CODE (op);
if ((code != EQ && code != NE) || !CONST_INT_P (XEXP (op, 1)))
return -1;
int cmpop = code == EQ ? 1 : 0;
int cmpval = INTVAL (XEXP (op, 1));
if (cmpval != 0 && cmpval != 1)
return -1;
int t;
if (t_reg_operand (XEXP (op, 0), GET_MODE (XEXP (op, 0))))
t = 0;
else if (negt_reg_operand (XEXP (op, 0), GET_MODE (XEXP (op, 0))))
t = 1;
else
return -1;
return t ^ (cmpval == cmpop);
}
/* Emit INSN, possibly in a PARALLEL with an USE/CLOBBER of FPSCR bits in case
of floating-point comparisons. */
static void
sh_emit_set_t_insn (rtx insn, machine_mode mode)
{
if (TARGET_FPU_ANY && GET_MODE_CLASS (mode) == MODE_FLOAT
&& GET_CODE (insn) != PARALLEL)
{
insn = gen_rtx_PARALLEL (VOIDmode,
gen_rtvec (3, insn,
gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (SImode, FPSCR_STAT_REG)),
gen_rtx_USE (VOIDmode, gen_rtx_REG (SImode, FPSCR_MODES_REG))));
}
emit_insn (insn);
}
/* Prepare the operands for an scc instruction; make sure that the
compare has been done and the result is in T_REG. */
void
sh_emit_scc_to_t (enum rtx_code code, rtx op0, rtx op1)
{
rtx t_reg = get_t_reg_rtx ();
enum rtx_code oldcode = code;
machine_mode mode;
/* First need a compare insn. */
switch (code)
{
case NE:
/* It isn't possible to handle this case. */
gcc_unreachable ();
case LT:
code = GT;
break;
case LE:
code = GE;
break;
case LTU:
code = GTU;
break;
case LEU:
code = GEU;
break;
default:
break;
}
if (code != oldcode)
std::swap (op0, op1);
mode = GET_MODE (op0);
if (mode == VOIDmode)
mode = GET_MODE (op1);
op0 = force_reg (mode, op0);
if ((code != EQ && code != NE
&& (op1 != const0_rtx
|| code == GTU || code == GEU || code == LTU || code == LEU))
|| (mode == DImode && op1 != const0_rtx)
|| (TARGET_SH2E && GET_MODE_CLASS (mode) == MODE_FLOAT))
op1 = force_reg (mode, op1);
sh_emit_set_t_insn (gen_rtx_SET (VOIDmode, t_reg,
gen_rtx_fmt_ee (code, SImode, op0, op1)),
mode);
}
rtx
sh_emit_cheap_store_flag (machine_mode mode, enum rtx_code code,
rtx op0, rtx op1)
{
rtx target = gen_reg_rtx (SImode);
rtx tmp;
gcc_assert (TARGET_SHMEDIA);
switch (code)
{
case EQ:
case GT:
case LT:
case UNORDERED:
case GTU:
case LTU:
tmp = gen_rtx_fmt_ee (code, SImode, op0, op1);
emit_insn (gen_cstore4_media (target, tmp, op0, op1));
code = NE;
break;
case NE:
case GE:
case LE:
case ORDERED:
case GEU:
case LEU:
tmp = gen_rtx_fmt_ee (reverse_condition (code), mode, op0, op1);
emit_insn (gen_cstore4_media (target, tmp, op0, op1));
code = EQ;
break;
case UNEQ:
case UNGE:
case UNGT:
case UNLE:
case UNLT:
case LTGT:
return NULL_RTX;
default:
gcc_unreachable ();
}
if (mode == DImode)
{
rtx t2 = gen_reg_rtx (DImode);
emit_insn (gen_extendsidi2 (t2, target));
target = t2;
}
return gen_rtx_fmt_ee (code, VOIDmode, target, const0_rtx);
}
/* Called from the md file, set up the operands of a compare instruction. */
void
sh_emit_compare_and_branch (rtx *operands, machine_mode mode)
{
enum rtx_code code = GET_CODE (operands[0]);
enum rtx_code branch_code;
rtx op0 = operands[1];
rtx op1 = operands[2];
rtx insn;
bool need_ccmpeq = false;
if (TARGET_SH2E && GET_MODE_CLASS (mode) == MODE_FLOAT)
{
op0 = force_reg (mode, op0);
op1 = force_reg (mode, op1);
}
else
{
if (code != EQ || mode == DImode)
{
/* Force args into regs, since we can't use constants here. */
op0 = force_reg (mode, op0);
if (op1 != const0_rtx || code == GTU || code == GEU)
op1 = force_reg (mode, op1);
}
}
if (GET_MODE_CLASS (mode) == MODE_FLOAT)
{
if (code == LT
|| (code == LE && TARGET_IEEE && TARGET_SH2E)
|| (code == GE && !(TARGET_IEEE && TARGET_SH2E)))
{
std::swap (op0, op1);
code = swap_condition (code);
}
/* GE becomes fcmp/gt+fcmp/eq, for SH2E and TARGET_IEEE only. */
if (code == GE)
{
gcc_assert (TARGET_IEEE && TARGET_SH2E);
need_ccmpeq = true;
code = GT;
}
/* Now we can have EQ, NE, GT, LE. NE and LE are then transformed
to EQ/GT respectively. */
gcc_assert (code == EQ || code == GT || code == NE || code == LE);
}
switch (code)
{
case EQ:
case GT:
case GE:
case GTU:
case GEU:
branch_code = code;
break;
case NE:
case LT:
case LE:
case LTU:
case LEU:
branch_code = reverse_condition (code);
break;
default:
gcc_unreachable ();
}
insn = gen_rtx_SET (VOIDmode,
get_t_reg_rtx (),
gen_rtx_fmt_ee (branch_code, SImode, op0, op1));
sh_emit_set_t_insn (insn, mode);
if (need_ccmpeq)
sh_emit_set_t_insn (gen_ieee_ccmpeqsf_t (op0, op1), mode);
if (branch_code == code)
emit_jump_insn (gen_branch_true (operands[3]));
else
emit_jump_insn (gen_branch_false (operands[3]));
}
void
sh_emit_compare_and_set (rtx *operands, machine_mode mode)
{
enum rtx_code code = GET_CODE (operands[1]);
rtx op0 = operands[2];
rtx op1 = operands[3];
rtx_code_label *lab = NULL;
bool invert = false;
op0 = force_reg (mode, op0);
if ((code != EQ && code != NE
&& (op1 != const0_rtx
|| code == GTU || code == GEU || code == LTU || code == LEU))
|| (mode == DImode && op1 != const0_rtx)
|| (TARGET_SH2E && GET_MODE_CLASS (mode) == MODE_FLOAT))
op1 = force_reg (mode, op1);
if (GET_MODE_CLASS (mode) == MODE_FLOAT)
{
if (code == LT || code == LE)
{
std::swap (op0, op1);
code = swap_condition (code);
}
if (code == GE)
{
if (TARGET_IEEE)
{
lab = gen_label_rtx ();
sh_emit_scc_to_t (EQ, op0, op1);
emit_jump_insn (gen_branch_true (lab));
code = GT;
}
else
{
code = LT;
invert = true;
}
}
}
if (code == NE)
{
code = EQ;
invert = true;
}
sh_emit_scc_to_t (code, op0, op1);
if (lab)
emit_label (lab);
if (invert)
emit_insn (gen_movnegt (operands[0], get_t_reg_rtx ()));
else
emit_move_insn (operands[0], get_t_reg_rtx ());
}
/* Functions to output assembly code. */
/* Return a sequence of instructions to perform DI or DF move.
Since the SH cannot move a DI or DF in one instruction, we have
to take care when we see overlapping source and dest registers. */
const char *
output_movedouble (rtx insn ATTRIBUTE_UNUSED, rtx operands[],
machine_mode mode)
{
rtx dst = operands[0];
rtx src = operands[1];
if (MEM_P (dst)
&& GET_CODE (XEXP (dst, 0)) == PRE_DEC)
return "mov.l %T1,%0" "\n"
" mov.l %1,%0";
if (register_operand (dst, mode)
&& register_operand (src, mode))
{
if (REGNO (src) == MACH_REG)
return "sts mach,%S0" "\n"
" sts macl,%R0";
/* When mov.d r1,r2 do r2->r3 then r1->r2;
when mov.d r1,r0 do r1->r0 then r2->r1. */
if (REGNO (src) + 1 == REGNO (dst))
return "mov %T1,%T0" "\n"
" mov %1,%0";
else
return "mov %1,%0" "\n"
" mov %T1,%T0";
}
else if (CONST_INT_P (src))
{
if (INTVAL (src) < 0)
output_asm_insn ("mov #-1,%S0", operands);
else
output_asm_insn ("mov #0,%S0", operands);
return "mov %1,%R0";
}
else if (MEM_P (src))
{
int ptrreg = -1;
int dreg = REGNO (dst);
rtx inside = XEXP (src, 0);
switch (GET_CODE (inside))
{
case REG:
ptrreg = REGNO (inside);
break;
case SUBREG:
ptrreg = subreg_regno (inside);
break;
case PLUS:
ptrreg = REGNO (XEXP (inside, 0));
/* ??? A r0+REG address shouldn't be possible here, because it isn't
an offsettable address. Unfortunately, offsettable addresses use
QImode to check the offset, and a QImode offsettable address
requires r0 for the other operand, which is not currently
supported, so we can't use the 'o' constraint.
Thus we must check for and handle r0+REG addresses here.
We punt for now, since this is likely very rare. */
gcc_assert (!REG_P (XEXP (inside, 1)));
break;
case LABEL_REF:
return "mov.l %1,%0" "\n"
" mov.l %1+4,%T0";
case POST_INC:
return "mov.l %1,%0" "\n"
" mov.l %1,%T0";
default:
gcc_unreachable ();
}
/* Work out the safe way to copy. Copy into the second half first. */
if (dreg == ptrreg)
return "mov.l %T1,%T0" "\n"
" mov.l %1,%0";
}
return "mov.l %1,%0" "\n"
" mov.l %T1,%T0";
}
/* Print an instruction which would have gone into a delay slot after
another instruction, but couldn't because the other instruction expanded
into a sequence where putting the slot insn at the end wouldn't work. */
static void
print_slot (rtx_sequence *seq)
{
final_scan_insn (seq->insn (1), asm_out_file, optimize, 1, NULL);
seq->insn (1)->set_deleted ();
}
const char *
output_far_jump (rtx_insn *insn, rtx op)
{
struct { rtx lab, reg, op; } this_jmp;
rtx_code_label *braf_base_lab = NULL;
const char *jump;
int far;
int offset = branch_dest (insn) - INSN_ADDRESSES (INSN_UID (insn));
rtx_insn *prev;
this_jmp.lab = gen_label_rtx ();
if (TARGET_SH2
&& offset >= -32764
&& offset - get_attr_length (insn) <= 32766)
{
far = 0;
jump = "mov.w %O0,%1" "\n"
" braf %1";
}
else
{
far = 1;
if (flag_pic)
{
if (TARGET_SH2)
jump = "mov.l %O0,%1" "\n"
" braf %1";
else
jump = "mov.l r0,@-r15" "\n"
" mova %O0,r0" "\n"
" mov.l @r0,%1" "\n"
" add r0,%1" "\n"
" mov.l @r15+,r0" "\n"
" jmp @%1";
}
else
jump = "mov.l %O0,%1" "\n"
" jmp @%1";
}
/* If we have a scratch register available, use it. */
if (NONJUMP_INSN_P ((prev = prev_nonnote_insn (insn)))
&& INSN_CODE (prev) == CODE_FOR_indirect_jump_scratch)
{
this_jmp.reg = SET_DEST (XVECEXP (PATTERN (prev), 0, 0));
if (REGNO (this_jmp.reg) == R0_REG && flag_pic && ! TARGET_SH2)
jump = "mov.l r1,@-r15" "\n"
" mova %O0,r0" "\n"
" mov.l @r0,r1" "\n"
" add r1,r0" "\n"
" mov.l @r15+,r1" "\n"
" jmp @%1";
output_asm_insn (jump, &this_jmp.lab);
if (dbr_sequence_length ())
print_slot (final_sequence);
else
output_asm_insn ("nop", 0);
}
else
{
/* Output the delay slot insn first if any. */
if (dbr_sequence_length ())
print_slot (final_sequence);
this_jmp.reg = gen_rtx_REG (SImode, 13);
/* We must keep the stack aligned to 8-byte boundaries on SH5.
Fortunately, MACL is fixed and call-clobbered, and we never
need its value across jumps, so save r13 in it instead of in
the stack. */
if (TARGET_SH5)
output_asm_insn ("lds r13,macl", 0);
else
output_asm_insn ("mov.l r13,@-r15", 0);
output_asm_insn (jump, &this_jmp.lab);
if (TARGET_SH5)
output_asm_insn ("sts macl,r13", 0);
else
output_asm_insn ("mov.l @r15+,r13", 0);
}
if (far && flag_pic && TARGET_SH2)
{
braf_base_lab = gen_label_rtx ();
(*targetm.asm_out.internal_label) (asm_out_file, "L",
CODE_LABEL_NUMBER (braf_base_lab));
}
if (far)
output_asm_insn (".align 2", 0);
(*targetm.asm_out.internal_label) (asm_out_file, "L", CODE_LABEL_NUMBER (this_jmp.lab));
this_jmp.op = op;
if (far && flag_pic)
{
if (TARGET_SH2)
this_jmp.lab = braf_base_lab;
output_asm_insn (".long %O2-%O0", &this_jmp.lab);
}
else
output_asm_insn (far ? ".long %O2" : ".word %O2-%O0", &this_jmp.lab);
return "";
}
/* Local label counter, used for constants in the pool and inside
pattern branches. */
static int lf = 100;
/* Output code for ordinary branches. */
const char *
output_branch (int logic, rtx_insn *insn, rtx *operands)
{
switch (get_attr_length (insn))
{
case 6:
/* This can happen if filling the delay slot has caused a forward
branch to exceed its range (we could reverse it, but only
when we know we won't overextend other branches; this should
best be handled by relaxation).
It can also happen when other condbranches hoist delay slot insn
from their destination, thus leading to code size increase.
But the branch will still be in the range -4092..+4098 bytes. */
if (! TARGET_RELAX)
{
int label = lf++;
/* The call to print_slot will clobber the operands. */
rtx op0 = operands[0];
/* If the instruction in the delay slot is annulled (true), then
there is no delay slot where we can put it now. The only safe
place for it is after the label. final will do that by default. */
if (final_sequence
&& ! INSN_ANNULLED_BRANCH_P (final_sequence->insn (0))
&& get_attr_length (final_sequence->insn (1)))
{
asm_fprintf (asm_out_file, "\tb%s%ss\t%LLF%d\n", logic ? "f" : "t",
ASSEMBLER_DIALECT ? "/" : ".", label);
print_slot (final_sequence);
}
else
asm_fprintf (asm_out_file, "\tb%s\t%LLF%d\n", logic ? "f" : "t", label);
output_asm_insn ("bra\t%l0", &op0);
fprintf (asm_out_file, "\tnop\n");
(*targetm.asm_out.internal_label) (asm_out_file, "LF", label);
return "";
}
/* When relaxing, handle this like a short branch. The linker
will fix it up if it still doesn't fit after relaxation. */
case 2:
return logic ? "bt%.\t%l0" : "bf%.\t%l0";
/* These are for SH2e, in which we have to account for the
extra nop because of the hardware bug in annulled branches. */
case 8:
if (! TARGET_RELAX)
{
int label = lf++;
gcc_assert (!final_sequence
|| !(INSN_ANNULLED_BRANCH_P
(XVECEXP (final_sequence, 0, 0))));
asm_fprintf (asm_out_file, "b%s%ss\t%LLF%d\n",
logic ? "f" : "t",
ASSEMBLER_DIALECT ? "/" : ".", label);
fprintf (asm_out_file, "\tnop\n");
output_asm_insn ("bra\t%l0", operands);
fprintf (asm_out_file, "\tnop\n");
(*targetm.asm_out.internal_label) (asm_out_file, "LF", label);
return "";
}
/* When relaxing, fall through. */
case 4:
{
char buffer[10];
sprintf (buffer, "b%s%ss\t%%l0",
logic ? "t" : "f",
ASSEMBLER_DIALECT ? "/" : ".");
output_asm_insn (buffer, &operands[0]);
return "nop";
}
default:
/* There should be no longer branches now - that would
indicate that something has destroyed the branches set
up in machine_dependent_reorg. */
gcc_unreachable ();
}
}
/* Output a code sequence for INSN using TEMPL with OPERANDS; but before,
fill in operands 9 as a label to the successor insn.
We try to use jump threading where possible.
IF CODE matches the comparison in the IF_THEN_ELSE of a following jump,
we assume the jump is taken. I.e. EQ means follow jmp and bf, NE means
follow jmp and bt, if the address is in range. */
const char *
output_branchy_insn (enum rtx_code code, const char *templ,
rtx_insn *insn, rtx *operands)
{
rtx_insn *next_insn = NEXT_INSN (insn);
if (next_insn && JUMP_P (next_insn) && condjump_p (next_insn))
{
rtx src = SET_SRC (PATTERN (next_insn));
if (GET_CODE (src) == IF_THEN_ELSE && GET_CODE (XEXP (src, 0)) != code)
{
/* Following branch not taken */
rtx_code_label *lab = gen_label_rtx ();
emit_label_after (lab, next_insn);
INSN_ADDRESSES_NEW (lab,
INSN_ADDRESSES (INSN_UID (next_insn))
+ get_attr_length (next_insn));
operands[9] = lab;
return templ;
}
else
{
int offset = (branch_dest (next_insn)
- INSN_ADDRESSES (INSN_UID (next_insn)) + 4);
if (offset >= -252 && offset <= 258)
{
if (GET_CODE (src) == IF_THEN_ELSE)
/* branch_true */
src = XEXP (src, 1);
operands[9] = src;
return templ;
}
}
}
rtx_code_label *lab = gen_label_rtx ();
emit_label_after (lab, insn);
INSN_ADDRESSES_NEW (lab,
INSN_ADDRESSES (INSN_UID (insn))
+ get_attr_length (insn));
operands[9] = lab;
return templ;
}
const char *
output_ieee_ccmpeq (rtx_insn *insn, rtx *operands)
{
return output_branchy_insn (NE, "bt %l9" "\n"
" fcmp/eq %1,%0",
insn, operands);
}
/* Output the start of the assembler file. */
static void
sh_file_start (void)
{
default_file_start ();
if (TARGET_ELF)
/* We need to show the text section with the proper
attributes as in TEXT_SECTION_ASM_OP, before dwarf2out
emits it without attributes in TEXT_SECTION_ASM_OP, else GAS
will complain. We can teach GAS specifically about the
default attributes for our choice of text section, but
then we would have to change GAS again if/when we change
the text section name. */
fprintf (asm_out_file, "%s\n", TEXT_SECTION_ASM_OP);
else
/* Switch to the data section so that the coffsem symbol
isn't in the text section. */
switch_to_section (data_section);
if (TARGET_LITTLE_ENDIAN)
fputs ("\t.little\n", asm_out_file);
if (!TARGET_ELF)
{
if (TARGET_SHCOMPACT)
fputs ("\t.mode\tSHcompact\n", asm_out_file);
else if (TARGET_SHMEDIA)
fprintf (asm_out_file, "\t.mode\tSHmedia\n\t.abi\t%i\n",
TARGET_SHMEDIA64 ? 64 : 32);
}
}
/* Check if PAT includes UNSPEC_CALLER unspec pattern. */
static bool
unspec_caller_rtx_p (rtx pat)
{
rtx base, offset;
int i;
split_const (pat, &base, &offset);