Copyright (C) 1993, 1994, 1995, 1997, 1997, 1998, 1999, 2000, 2001, 2002,
2003, 2004 Free Software Foundation, Inc.
Contributed by Steve Chamberlain (sac@cygnus.com).
- Improved by Jim Wilson (wilson@cygnus.com).
+ Improved by Jim Wilson (wilson@cygnus.com).
This file is part of GCC.
static void sh_output_function_epilogue (FILE *, HOST_WIDE_INT);
static void sh_insert_attributes (tree, tree *);
static int sh_adjust_cost (rtx, rtx, rtx, int);
-static int sh_use_dfa_interface (void);
static int sh_issue_rate (void);
static int sh_dfa_new_cycle (FILE *, int, rtx, int, int, int *sort_p);
static short find_set_regmode_weight (rtx, enum machine_mode);
static int sh_reorder2 (FILE *, int, rtx *, int *, int);
static void sh_md_init (FILE *, int, int);
static int sh_variable_issue (FILE *, int, rtx, int);
-
+
static bool sh_function_ok_for_sibcall (tree, tree);
static bool sh_cannot_modify_jumps_p (void);
static bool sh_pretend_outgoing_varargs_named (CUMULATIVE_ARGS *);
static tree sh_build_builtin_va_list (void);
static tree sh_gimplify_va_arg_expr (tree, tree, tree *, tree *);
+static bool sh_pass_by_reference (CUMULATIVE_ARGS *, enum machine_mode,
+ tree, bool);
\f
/* Initialize the GCC target structure. */
#undef TARGET_SCHED_ADJUST_COST
#define TARGET_SCHED_ADJUST_COST sh_adjust_cost
-#undef TARGET_SCHED_USE_DFA_PIPELINE_INTERFACE
-#define TARGET_SCHED_USE_DFA_PIPELINE_INTERFACE \
- sh_use_dfa_interface
#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
+ 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
+ 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
upto a max of 8 cycles so that such insns may move from Q -> R.
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.
+ find_insn_reg_weights does.
TARGET_SCHED_FINISH_GLOBAL: Corresponding cleanup hook.
TARGET_SCHED_DFA_NEW_CYCLE: Skip cycles if high register pressure is
#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_BUILD_BUILTIN_VA_LIST
#define TARGET_BUILD_BUILTIN_VA_LIST sh_build_builtin_va_list
#undef TARGET_GIMPLIFY_VA_ARG_EXPR
#define TARGET_GIMPLIFY_VA_ARG_EXPR sh_gimplify_va_arg_expr
+#undef TARGET_VECTOR_MODE_SUPPORTED_P
+#define TARGET_VECTOR_MODE_SUPPORTED_P sh_vector_mode_supported_p
+
#undef TARGET_PCH_VALID_P
#define TARGET_PCH_VALID_P sh_pch_valid_p
'T' print the next word of a dp value - same as 'R' in big endian mode.
'M' print an `x' if `m' will print `base,index'.
'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' prints the lowest 16 bits of CONST_INT, as an unsigned value.
'o' output an operator. */
}
break;
+ case 'd':
+ if (GET_CODE (x) != REG || GET_MODE (x) != V2SFmode)
+ abort ();
+
+ fprintf ((stream), "d%s", reg_names[REGNO (x)] + 1);
+ break;
+
case 'N':
if (x == CONST0_RTX (GET_MODE (x)))
{
case MEM:
output_address (XEXP (x, 0));
break;
-
+
case CONST:
if (TARGET_SHMEDIA
&& GET_CODE (XEXP (x, 0)) == SIGN_EXTEND
int constp = (GET_CODE (operands[2]) == CONST_INT);
int bytes = (constp ? INTVAL (operands[2]) : 0);
+ if (! constp)
+ return 0;
+
+ /* If we could use mov.l to move words and dest is word-aligned, we
+ can use movua.l for loads and still generate a relatively short
+ and efficient sequence. */
+ if (TARGET_SH4A_ARCH && align < 4
+ && MEM_ALIGN (operands[0]) >= 32
+ && can_move_by_pieces (bytes, 32))
+ {
+ rtx dest = copy_rtx (operands[0]);
+ rtx src = copy_rtx (operands[1]);
+ /* We could use different pseudos for each copied word, but
+ since movua can only load into r0, it's kind of
+ pointless. */
+ rtx temp = gen_reg_rtx (SImode);
+ rtx src_addr = copy_addr_to_reg (XEXP (src, 0));
+ int copied = 0;
+
+ while (copied + 4 <= bytes)
+ {
+ rtx to = adjust_address (dest, SImode, copied);
+ rtx from = adjust_automodify_address (src, SImode, src_addr, copied);
+
+ emit_insn (gen_movua (temp, from));
+ emit_move_insn (src_addr, plus_constant (src_addr, 4));
+ emit_move_insn (to, temp);
+ copied += 4;
+ }
+
+ if (copied < bytes)
+ move_by_pieces (adjust_address (dest, BLKmode, copied),
+ adjust_automodify_address (src, BLKmode,
+ src_addr, copied),
+ bytes - copied, align, 0);
+
+ return 1;
+ }
+
/* If it isn't a constant number of bytes, or if it doesn't have 4 byte
alignment, or if it isn't a multiple of 4 bytes, then fail. */
- if (! constp || align < 4 || (bytes % 4 != 0))
+ if (align < 4 || (bytes % 4 != 0))
return 0;
if (TARGET_HARD_SH4)
{
rtx tga_op1, tga_ret, tmp, tmp2;
-
+
switch (tls_kind)
{
case TLS_MODEL_GLOBAL_DYNAMIC:
emit_insn (gen_load_gbr (tmp2));
tmp = gen_reg_rtx (Pmode);
emit_insn (gen_symTPOFF2reg (tmp, op1));
- RTX_UNCHANGING_P (tmp) = 1;
if (register_operand (op0, Pmode))
op1 = op0;
|| (TARGET_SH2E && GET_MODE_CLASS (mode) == MODE_FLOAT))
sh_compare_op1 = force_reg (mode, sh_compare_op1);
- if (TARGET_SH4 && GET_MODE_CLASS (mode) == MODE_FLOAT)
+ if ((TARGET_SH4 || TARGET_SH2A) && GET_MODE_CLASS (mode) == MODE_FLOAT)
(mode == SFmode ? emit_sf_insn : emit_df_insn)
(gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2,
gen_rtx_SET (VOIDmode, t_reg,
gen_rtx_REG (SImode, T_REG),
gen_rtx_fmt_ee (code, SImode,
sh_compare_op0, sh_compare_op1));
- if (TARGET_SH4 && GET_MODE_CLASS (mode) == MODE_FLOAT)
+ if ((TARGET_SH4 || TARGET_SH2A) && GET_MODE_CLASS (mode) == MODE_FLOAT)
{
insn = gen_rtx_PARALLEL (VOIDmode,
gen_rtvec (2, insn,
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 (XVECEXP (final_sequence, 0, 0)))
{
}
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
case 4:
{
char buffer[10];
-
+
sprintf (buffer, "b%s%ss\t%%l0",
logic ? "t" : "f",
ASSEMBLER_DIALECT ? "/" : ".");
fputs ("\t.section .directive, \"SM\", @progbits, 1\n", asm_out_file);
fputs ("\t.asciz \"#<SYMEDIT>#\\n\"\n", asm_out_file);
#endif
-
+
if (TARGET_ELF)
/* We need to show the text section with the proper
attributes as in TEXT_SECTION_ASM_OP, before dwarf2out
/* Likewise, but for shift amounts < 16, up to three highmost bits
might be clobbered. This is typically used when combined with some
kind of sign or zero extension. */
-
+
static const char ext_shift_insns[] =
{ 0,1,1,2,2,3,2,2,1,2,2,3,3,3,2,2,1,2,2,3,3,4,3,3,2,3,3,4,4,4,3,3};
/* Output RTL to split a constant shift into its component SH constant
shift instructions. */
-
+
void
gen_shifty_op (int code, rtx *operands)
{
/* Truncate the shift count in case it is out of bounds. */
value = value & 0x1f;
-
+
if (value == 31)
{
if (code == LSHIFTRT)
for (i = 0; i < max; i++)
gen_ashift (code, shift_amounts[value][i], operands[0]);
}
-
+
/* Same as above, but optimized for values where the topmost bits don't
matter. */
return len + shift_insns[INTVAL (XEXP (op, 1))];
}
-/* Generating rtl? */
-extern int rtx_equal_function_value_matters;
-
/* Generate rtl for instructions for which shl_and_kind advised a particular
method of generating them, i.e. returned zero. */
case 2:
/* Don't expand fine-grained when combining, because that will
make the pattern fail. */
- if (rtx_equal_function_value_matters
+ if (currently_expanding_to_rtl
|| reload_in_progress || reload_completed)
{
rtx operands[3];
-
+
/* Cases 3 and 4 should be handled by this split
only while combining */
if (kind > 2)
/* Don't expand fine-grained when combining, because that will
make the pattern fail. */
- if (! rtx_equal_function_value_matters
+ if (! currently_expanding_to_rtl
&& ! reload_in_progress && ! reload_completed)
{
emit_insn (gen_shl_sext_ext (dest, source, left_rtx, size_rtx));
case 5:
{
int i = 16 - size;
- if (! rtx_equal_function_value_matters
+ if (! currently_expanding_to_rtl
&& ! reload_in_progress && ! reload_completed)
emit_insn (gen_shl_sext_ext (dest, source, left_rtx, size_rtx));
else
case 7:
/* Don't expand fine-grained when combining, because that will
make the pattern fail. */
- if (! rtx_equal_function_value_matters
+ if (! currently_expanding_to_rtl
&& ! reload_in_progress && ! reload_completed)
{
emit_insn (gen_shl_sext_ext (dest, source, left_rtx, size_rtx));
gen_rtx_UNSPEC (GET_MODE (sym),
gen_rtvec (1, sym),
UNSPEC_DATALABEL));
-
+
if (GET_CODE (sym) != SYMBOL_REF)
abort ();
pool_size = 0;
}
-
+
for (i = 0; i < pool_size; i++)
{
pool_node *p = &pool_vector[i];
== SCRATCH))
&& GET_CODE (SET_DEST (pat)) == REG
&& FP_REGISTER_P (REGNO (SET_DEST (pat))))
+ && ! (TARGET_SH2A
+ && GET_MODE (SET_DEST (pat)) == SImode
+ && GET_CODE (SET_SRC (pat)) == CONST_INT
+ && CONST_OK_FOR_I20 (INTVAL (SET_SRC (pat))))
&& (GET_CODE (SET_SRC (pat)) != CONST_INT
|| ! CONST_OK_FOR_I08 (INTVAL (SET_SRC (pat)))))
return 1;
case SUBREG:
{
rtx y = SUBREG_REG (x);
-
+
if (GET_CODE (y) != REG)
break;
if (REGNO (y) < 16)
pass 1. Pass 2 if a definite blocking insn is needed.
-1 is used internally to avoid deep recursion.
If a blocking instruction is made or recognized, return it. */
-
+
static rtx
gen_block_redirect (rtx jump, int addr, int need_block)
{
it would cause trouble if an interrupt occurred. */
unsigned try = 0x7fff, used;
int jump_left = flag_expensive_optimizations + 1;
-
+
/* It is likely that the most recent eligible instruction is wanted for
the delay slot. Therefore, find out which registers it uses, and
try to avoid using them. */
-
+
for (scan = jump; (scan = PREV_INSN (scan)); )
{
enum rtx_code code;
threading with a jump beyond the delay slot insn.
Don't check if we are called recursively; the jump has been or will be
checked in a different invocation then. */
-
+
else if (optimize && need_block >= 0)
{
rtx next = next_active_insn (next_active_insn (dest));
Hence, after delay slot scheduling, we'll have to expect
NOTE_INSN_BLOCK_END notes between the indirect_jump_scratch and
the jump. */
-
+
INSN_LOCATOR (insn) = INSN_LOCATOR (jump);
INSN_CODE (insn) = CODE_FOR_indirect_jump_scratch;
return insn;
if (GET_CODE (x) == LABEL_REF && XEXP (x, 0) == vec_lab)
break;
}
+ /* FIXME: This is a bug in the optimizer, but it seems harmless
+ to just avoid panicing. */
+ if (!prev)
+ continue;
/* Emit the reference label of the braf where it belongs, right after
the casesi_jump_2 (i.e. braf). */
{
rtx next = next_real_insn (barrier_or_label), pat, prev;
int slot, credit, jump_to_next = 0;
-
+
if (! next)
return 0;
an alignment, against that of fetching unneeded insn in front of the
branch target when there is no alignment. */
- /* There are two delay_slot cases to consider. One is the simple case
- where the preceding branch is to the insn beyond the barrier (simple
- delay slot filling), and the other is where the preceding branch has
- a delay slot that is a duplicate of the insn after the barrier
- (fill_eager_delay_slots) and the branch is to the insn after the insn
+ /* There are two delay_slot cases to consider. One is the simple case
+ where the preceding branch is to the insn beyond the barrier (simple
+ delay slot filling), and the other is where the preceding branch has
+ a delay slot that is a duplicate of the insn after the barrier
+ (fill_eager_delay_slots) and the branch is to the insn after the insn
after the barrier. */
/* PREV is presumed to be the JUMP_INSN for the barrier under
if (GET_CODE (PATTERN (prev)) == SEQUENCE)
{
prev = XVECEXP (PATTERN (prev), 0, 1);
- if (INSN_UID (prev) == INSN_UID (next))
+ if (INSN_UID (prev) == INSN_UID (next))
{
/* Delay slot was filled with insn at jump target. */
jump_to_next = 1;
/* There is no upper bound on redundant instructions
that might have been skipped, but we must not put an
alignment where none had been before. */
- || (x = (NEXT_INSN (NEXT_INSN (PREV_INSN (prev)))),
- (INSN_P (x)
+ || (x = (NEXT_INSN (NEXT_INSN (PREV_INSN (prev)))),
+ (INSN_P (x)
&& (INSN_CODE (x) == CODE_FOR_block_branch_redirect
|| INSN_CODE (x) == CODE_FOR_indirect_jump_scratch
|| INSN_CODE (x) == CODE_FOR_stuff_delay_slot))))
}
}
}
-
+
return align_jumps_log;
}
/* Remove the clobber of r0. */
*clobberp = gen_rtx_CLOBBER (GET_MODE (clobber),
gen_rtx_SCRATCH (Pmode));
- RTX_UNCHANGING_P (newsrc) = 1;
}
/* This is a mova needing a label. Create it. */
else if (GET_CODE (src) == UNSPEC
else
{
lab = add_constant (src, mode, 0);
- newsrc = gen_rtx_MEM (mode,
- gen_rtx_LABEL_REF (VOIDmode, lab));
- RTX_UNCHANGING_P (newsrc) = 1;
+ newsrc = gen_rtx_LABEL_REF (VOIDmode, lab);
+ newsrc = gen_const_mem (mode, newsrc);
}
*patp = gen_rtx_SET (VOIDmode, dst, newsrc);
INSN_CODE (scan) = -1;
if (type == TYPE_CBRANCH)
{
rtx next, beyond;
-
+
if (get_attr_length (insn) > 4)
{
rtx src = SET_SRC (PATTERN (insn));
rtx label = 0;
int dest_uid = get_dest_uid (olabel, max_uid);
struct far_branch *bp = uid_branch[dest_uid];
-
+
/* redirect_jump needs a valid JUMP_LABEL, and it might delete
the label if the LABEL_NUSES count drops to zero. There is
always a jump_optimize pass that sets these values, but it
beyond
= next_active_insn (XEXP (XEXP (SET_SRC (PATTERN (insn)), 1),
0));
-
+
if (beyond
&& (GET_CODE (beyond) == JUMP_INSN
|| ((beyond = next_active_insn (beyond))
gen_block_redirect (beyond,
INSN_ADDRESSES (INSN_UID (beyond)), 1);
}
-
+
next = next_active_insn (insn);
if ((GET_CODE (next) == JUMP_INSN
- || GET_CODE (next = next_active_insn (next)) == JUMP_INSN)
+ || ((next = next_active_insn (next))
+ && GET_CODE (next) == JUMP_INSN))
&& GET_CODE (PATTERN (next)) == SET
&& recog_memoized (next) == CODE_FOR_jump_compact
&& ((INSN_ADDRESSES
{
HOST_WIDE_INT align = STACK_BOUNDARY / BITS_PER_UNIT;
+/* This test is bogus, as output_stack_adjust is used to re-align the
+ stack. */
+#if 0
if (size % align)
abort ();
+#endif
if (CONST_OK_FOR_ADD (size))
emit_fn (GEN_ADD3 (reg, reg, GEN_INT (size)));
to handle this case, so just abort when we see it. */
if (epilogue_p < 0
|| current_function_interrupt
- || ! call_used_regs[temp] || fixed_regs[temp])
+ || ! call_really_used_regs[temp] || fixed_regs[temp])
temp = -1;
if (temp < 0 && ! current_function_interrupt
&& (TARGET_SHMEDIA || epilogue_p >= 0))
x = gen_push_fpul ();
else if (rn == FPSCR_REG)
x = gen_push_fpscr ();
- else if (TARGET_SH4 && TARGET_FMOVD && ! TARGET_FPU_SINGLE
+ else if ((TARGET_SH4 || TARGET_SH2A_DOUBLE) && TARGET_FMOVD && ! TARGET_FPU_SINGLE
&& FP_OR_XD_REGISTER_P (rn))
{
if (FP_REGISTER_P (rn) && (rn - FIRST_FP_REG) & 1)
x = gen_pop_fpul ();
else if (rn == FPSCR_REG)
x = gen_pop_fpscr ();
- else if (TARGET_SH4 && TARGET_FMOVD && ! TARGET_FPU_SINGLE
+ else if ((TARGET_SH4 || TARGET_SH2A_DOUBLE) && TARGET_FMOVD && ! TARGET_FPU_SINGLE
&& FP_OR_XD_REGISTER_P (rn))
{
if (FP_REGISTER_P (rn) && (rn - FIRST_FP_REG) & 1)
x = gen_pop_e (gen_rtx_REG (SFmode, rn));
else
x = gen_pop (gen_rtx_REG (SImode, rn));
-
+
x = emit_insn (x);
REG_NOTES (x)
= gen_rtx_EXPR_LIST (REG_INC,
int interrupt_handler = sh_cfun_interrupt_handler_p ();
for (reg = LAST_TARGET_REG; reg >= FIRST_TARGET_REG; reg--)
- if ((! call_used_regs[reg] || interrupt_handler)
+ if ((! call_really_used_regs[reg] || interrupt_handler)
&& ! TEST_HARD_REG_BIT (*live_regs_mask, reg))
/* Leave space to save this target register on the stack,
in case target register allocation wants to use it. */
stack_space += GET_MODE_SIZE (REGISTER_NATURAL_MODE (reg));
return stack_space;
}
-
+
/* Decide whether we should reserve space for callee-save target registers,
in case target register allocation wants to use them. REGS_SAVED is
the space, in bytes, that is already required for register saves.
interrupt_handler = sh_cfun_interrupt_handler_p ();
CLEAR_HARD_REG_SET (*live_regs_mask);
- if (TARGET_SH4 && TARGET_FMOVD && interrupt_handler
+ if ((TARGET_SH4 || TARGET_SH2A_DOUBLE) && TARGET_FMOVD && interrupt_handler
&& regs_ever_live[FPSCR_REG])
target_flags &= ~FPU_SINGLE_BIT;
/* If we can save a lot of saves by switching to double mode, do that. */
- else if (TARGET_SH4 && TARGET_FMOVD && TARGET_FPU_SINGLE)
+ else if ((TARGET_SH4 || TARGET_SH2A_DOUBLE) && TARGET_FMOVD && TARGET_FPU_SINGLE)
for (count = 0, reg = FIRST_FP_REG; reg <= LAST_FP_REG; reg += 2)
if (regs_ever_live[reg] && regs_ever_live[reg+1]
- && (! call_used_regs[reg] || (interrupt_handler && ! pragma_trapa))
+ && (! call_really_used_regs[reg]
+ || (interrupt_handler && ! pragma_trapa))
&& ++count > 2)
{
target_flags &= ~FPU_SINGLE_BIT;
: (interrupt_handler && ! pragma_trapa)
? (/* Need to save all the regs ever live. */
(regs_ever_live[reg]
- || (call_used_regs[reg]
- && (! fixed_regs[reg] || reg == MACH_REG || reg == MACL_REG)
+ || (call_really_used_regs[reg]
+ && (! fixed_regs[reg] || reg == MACH_REG || reg == MACL_REG
+ || reg == PIC_OFFSET_TABLE_REGNUM)
&& has_call)
|| (has_call && REGISTER_NATURAL_MODE (reg) == SImode
&& (GENERAL_REGISTER_P (reg) || TARGET_REGISTER_P (reg))))
&& flag_pic
&& current_function_args_info.call_cookie
&& reg == (int) PIC_OFFSET_TABLE_REGNUM)
- || (regs_ever_live[reg] && ! call_used_regs[reg])
+ || (regs_ever_live[reg] && ! call_really_used_regs[reg])
|| (current_function_calls_eh_return
&& (reg == (int) EH_RETURN_DATA_REGNO (0)
|| reg == (int) EH_RETURN_DATA_REGNO (1)
SET_HARD_REG_BIT (*live_regs_mask, reg);
count += GET_MODE_SIZE (REGISTER_NATURAL_MODE (reg));
- if ((TARGET_SH4 || TARGET_SH5) && TARGET_FMOVD
+ if ((TARGET_SH4 || TARGET_SH2A_DOUBLE || TARGET_SH5) && TARGET_FMOVD
&& GET_MODE_CLASS (REGISTER_NATURAL_MODE (reg)) == MODE_FLOAT)
{
if (FP_REGISTER_P (reg))
&& TARGET_SAVE_ALL_TARGET_REGS
&& shmedia_space_reserved_for_target_registers)
for (reg = LAST_TARGET_REG; reg >= FIRST_TARGET_REG; reg--)
- if ((! call_used_regs[reg] || interrupt_handler)
+ if ((! call_really_used_regs[reg] || interrupt_handler)
&& ! TEST_HARD_REG_BIT (*live_regs_mask, reg))
{
SET_HARD_REG_BIT (*live_regs_mask, reg);
tr0_used = flag_pic && regs_ever_live[PIC_OFFSET_TABLE_REGNUM];
for (regno = FIRST_TARGET_REG + tr0_used; regno <= LAST_TARGET_REG; regno++)
- if (call_used_regs[regno] && ! regs_ever_live[regno])
+ if (call_really_used_regs[regno] && ! regs_ever_live[regno])
return regno;
return -1;
use reverse order. Returns the last entry written to (not counting
the delimiter). OFFSET_BASE is a number to be added to all offset
entries. */
-
+
static save_entry *
sh5_schedule_saves (HARD_REG_SET *live_regs_mask, save_schedule *schedule,
int offset_base)
if (! current_function_interrupt)
for (i = FIRST_GENERAL_REG; tmpx < MAX_TEMPS && i <= LAST_GENERAL_REG; i++)
- if (call_used_regs[i] && ! fixed_regs[i] && i != PR_MEDIA_REG
+ if (call_really_used_regs[i] && ! fixed_regs[i] && i != PR_MEDIA_REG
&& ! FUNCTION_ARG_REGNO_P (i)
&& i != FIRST_RET_REG
&& ! (cfun->static_chain_decl != NULL && i == STATIC_CHAIN_REGNUM)
int rn = NPARM_REGS(SImode) + FIRST_PARM_REG - i - 1;
rtx insn;
- if (i >= (NPARM_REGS(SImode)
+ if (i >= (NPARM_REGS(SImode)
- current_function_args_info.arg_count[(int) SH_ARG_INT]
))
break;
that already happens to be at the function start into the prologue. */
if (target_flags != save_flags && ! current_function_interrupt)
emit_insn (gen_toggle_sz ());
-
+
if (TARGET_SH5)
{
int offset_base, offset;
save_entry *entry;
int *tmp_pnt;
- if (call_used_regs[R0_REG] && ! fixed_regs[R0_REG]
+ if (call_really_used_regs[R0_REG] && ! fixed_regs[R0_REG]
&& ! current_function_interrupt)
r0 = gen_rtx_REG (Pmode, R0_REG);
GEN_INT (offset - offset_in_r0)));
offset_in_r0 += offset - offset_in_r0;
}
-
+
if (pre_dec != NULL_RTX)
{
if (! sp_in_r0)
insn = emit_move_insn (mem_rtx, reg_rtx);
RTX_FRAME_RELATED_P (insn) = 1;
- if (TARGET_SHCOMPACT && (offset_in_r0 != -1))
+ if (TARGET_SHCOMPACT && (offset_in_r0 != -1))
{
rtx reg_rtx = gen_rtx_REG (mode, reg);
rtx set, note_rtx;
post_inc = NULL_RTX;
break;
-
+
post_inc_ok:
mem_rtx = NULL_RTX;
}
while (0);
-
+
if (mem_rtx != NULL_RTX)
goto addr_ok;
GEN_INT (offset - offset_in_r0)));
offset_in_r0 += offset - offset_in_r0;
}
-
+
if (post_inc != NULL_RTX)
{
if (! sp_in_r0)
(Pmode, r0, stack_pointer_rtx));
sp_in_r0 = 1;
}
-
+
mem_rtx = post_inc;
offset_in_r0 += GET_MODE_SIZE (mode);
for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
{
int j = (FIRST_PSEUDO_REGISTER - 1) - i;
-
+
if (j == FPSCR_REG && current_function_interrupt && TARGET_FMOVD
&& hard_regs_intersect_p (&live_regs_mask,
®_class_contents[DF_REGS]))
int offset;
save_schedule schedule;
save_entry *entry;
-
+
entry = sh5_schedule_saves (&live_regs_mask, &schedule, 0);
offset = entry[1].offset;
for (; entry->mode != VOIDmode; entry--)
if (TARGET_SHCOMPACT)
return const0_rtx;
}
-
+
if (! TARGET_SH2E && ! TARGET_SH4 && ! TARGET_SH5)
{
error ("__builtin_saveregs not supported by this subtarget");
emit_move_insn (fpregs, XEXP (regbuf, 0));
emit_insn (gen_addsi3 (fpregs, fpregs,
GEN_INT (n_floatregs * UNITS_PER_WORD)));
- if (TARGET_SH4)
+ if (TARGET_SH4 || TARGET_SH2A_DOUBLE)
{
rtx mem;
for (regno = NPARM_REGS (DFmode) - 2; regno >= first_floatreg; regno -= 2)
GEN_INT (-2 * UNITS_PER_WORD)));
mem = gen_rtx_MEM (DFmode, fpregs);
set_mem_alias_set (mem, alias_set);
- emit_move_insn (mem,
+ emit_move_insn (mem,
gen_rtx_REG (DFmode, BASE_ARG_REG (DFmode) + regno));
}
regno = first_floatreg;
else
nfp = 0;
u = fold (build (PLUS_EXPR, ptr_type_node, u,
- build_int_2 (UNITS_PER_WORD * nfp, 0)));
+ build_int_cst (NULL_TREE, UNITS_PER_WORD * nfp)));
t = build (MODIFY_EXPR, ptr_type_node, next_fp_limit, u);
TREE_SIDE_EFFECTS (t) = 1;
expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL);
else
nint = 0;
u = fold (build (PLUS_EXPR, ptr_type_node, u,
- build_int_2 (UNITS_PER_WORD * nint, 0)));
+ build_int_cst (NULL_TREE, UNITS_PER_WORD * nint)));
t = build (MODIFY_EXPR, ptr_type_node, next_o_limit, u);
TREE_SIDE_EFFECTS (t) = 1;
expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL);
HOST_WIDE_INT size, rsize;
tree tmp, pptr_type_node;
tree addr, lab_over, result = NULL;
- int pass_by_ref = targetm.calls.must_pass_in_stack (TYPE_MODE (type), type);
+ int pass_by_ref = pass_by_reference (NULL, TYPE_MODE (type), type, false);
if (pass_by_ref)
type = build_pointer_type (type);
return ! sh_attr_renesas_p (type);
}
+/* Whether an argument must be passed by reference. On SHcompact, we
+ pretend arguments wider than 32-bits that would have been passed in
+ registers are passed by reference, so that an SHmedia trampoline
+ loads them into the full 64-bits registers. */
+
+static int
+shcompact_byref (CUMULATIVE_ARGS *cum, enum machine_mode mode,
+ tree type, bool named)
+{
+ unsigned HOST_WIDE_INT size;
+
+ if (type)
+ size = int_size_in_bytes (type);
+ else
+ size = GET_MODE_SIZE (mode);
+
+ if (cum->arg_count[SH_ARG_INT] < NPARM_REGS (SImode)
+ && (!named
+ || GET_SH_ARG_CLASS (mode) == SH_ARG_INT
+ || (GET_SH_ARG_CLASS (mode) == SH_ARG_FLOAT
+ && cum->arg_count[SH_ARG_FLOAT] >= NPARM_REGS (SFmode)))
+ && size > 4
+ && !SHCOMPACT_FORCE_ON_STACK (mode, type)
+ && !SH5_WOULD_BE_PARTIAL_NREGS (*cum, mode, type, named))
+ return size;
+ else
+ return 0;
+}
+
+static bool
+sh_pass_by_reference (CUMULATIVE_ARGS *cum, enum machine_mode mode,
+ tree type, bool named)
+{
+ if (targetm.calls.must_pass_in_stack (mode, type))
+ return true;
+
+ if (TARGET_SHCOMPACT)
+ {
+ cum->byref = shcompact_byref (cum, mode, type, named);
+ return cum->byref != 0;
+ }
+
+ return false;
+}
+
/* Define where to put the arguments to a function.
Value is zero to push the argument on the stack,
or a hard register in which to store the argument.
return gen_rtx_REG (mode, regno);
}
-
+
if (TARGET_SH5)
{
if (mode == VOIDmode && TARGET_SHCOMPACT)
return 0;
}
-
+
/* Update the data in CUM to advance over an argument
of mode MODE and data type TYPE.
(TYPE is null for libcalls where that information may not be
}
}
- if (! (TARGET_SH4 || ca->renesas_abi)
+ if (! ((TARGET_SH4 || TARGET_SH2A) || ca->renesas_abi)
|| PASS_IN_REG_P (*ca, mode, type))
(ca->arg_count[(int) GET_SH_ARG_CLASS (mode)]
= (ROUND_REG (*ca, mode)
int pr_reg = TARGET_SHMEDIA ? PR_MEDIA_REG : PR_REG;
save_schedule schedule;
save_entry *entry;
-
+
n += total_auto_space;
/* If it wasn't saved, there's not much we can do. */
/* Symbian support adds three new attributes:
dllexport - for exporting a function/variable that will live in a dll
dllimport - for importing a function/variable from a dll
-
+
Microsoft allows multiple declspecs in one __declspec, separating
them with spaces. We do NOT support this. Instead, use __declspec
multiple times. */
int flag_mask
= (SH1_BIT | SH2_BIT | SH3_BIT | SH_E_BIT | HARD_SH4_BIT | FPU_SINGLE_BIT
| SH4_BIT | HITACHI_BIT | LITTLE_ENDIAN_BIT);
-
+
/* -fpic and -fpie also usually make a PCH invalid. */
if (data[0] != flag_pic)
return _("created and used with different settings of -fpic");
}
data += sizeof (target_flags);
len -= sizeof (target_flags);
-
+
/* Check string options. */
#ifdef TARGET_OPTIONS
for (i = 0; i < ARRAY_SIZE (target_options); i++)
#endif
return NULL;
-
+
make_message:
{
char *r;
int
fpscr_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
{
- return (GET_CODE (op) == REG && REGNO (op) == FPSCR_REG
+ return (GET_CODE (op) == REG
+ && (REGNO (op) == FPSCR_REG
+ || (REGNO (op) >= FIRST_PSEUDO_REGISTER
+ && !(reload_in_progress || reload_completed)))
&& GET_MODE (op) == PSImode);
}
}
\f
/* Return the destination address of a branch. */
-
+
static int
branch_dest (rtx branch)
{
if (set == 0 && reg_overlap_mentioned_p (reg, PATTERN (insn)))
return 0;
- if (code == CALL_INSN && call_used_regs[REGNO (reg)])
+ if (code == CALL_INSN && call_really_used_regs[REGNO (reg)])
return 1;
}
return 1;
remove assignments that are dead due to a following assignment in the
same basic block. */
-static void
+static void
mark_use (rtx x, rtx *reg_set_block)
{
enum rtx_code code;
abort ();
}
-/* This function will set the fpscr from memory.
+/* This function will set the fpscr from memory.
MODE is the mode we are setting it to. */
void
fpscr_set_from_mem (int mode, HARD_REG_SET regs_live)
return 2;
/* sh-dsp parallel processing insn take four bytes instead of two. */
-
+
if (GET_CODE (insn) == INSN)
{
int sum = 0;
/* Return true if it's possible to redirect BRANCH1 to the destination
of an unconditional jump BRANCH2. We only want to do this if the
resulting branch will have a short displacement. */
-int
+int
sh_can_redirect_branch (rtx branch1, rtx branch2)
{
if (flag_expensive_optimizations && simplejump_p (branch2))
rtx dest = XEXP (SET_SRC (single_set (branch2)), 0);
rtx insn;
int distance;
-
- for (distance = 0, insn = NEXT_INSN (branch1);
- insn && distance < 256;
+
+ for (distance = 0, insn = NEXT_INSN (branch1);
+ insn && distance < 256;
insn = PREV_INSN (insn))
{
- if (insn == dest)
+ if (insn == dest)
return 1;
else
distance += get_attr_length (insn);
}
- for (distance = 0, insn = NEXT_INSN (branch1);
- insn && distance < 256;
+ for (distance = 0, insn = NEXT_INSN (branch1);
+ insn && distance < 256;
insn = NEXT_INSN (insn))
{
- if (insn == dest)
+ if (insn == dest)
return 1;
else
distance += get_attr_length (insn);
return REG_N_SETS (TARGET_SHMEDIA ? PR_MEDIA_REG : PR_REG);
}
-/* This Function returns nonzero if the DFA based scheduler interface
- is to be used. At present this is only supported properly for the SH4.
- For the SH1 the current DFA model is just the converted form of the old
- pipeline model description. */
-static int
-sh_use_dfa_interface (void)
-{
- if (TARGET_SH1)
- return 1;
- else
- return 0;
-}
-
/* This function returns "2" to indicate dual issue for the SH4
processor. To be used by the DFA pipeline description. */
static int
if (SCHED_GROUP_P (tmp) != SCHED_GROUP_P (tmp2))
return SCHED_GROUP_P (tmp2) ? 1 : -1;
- /* If insns are equally good, sort by INSN_LUID (original insn order), This
+ /* If insns are equally good, sort by INSN_LUID (original insn order), This
minimizes instruction movement, thus minimizing sched's effect on
register pressure. */
return INSN_LUID (tmp) - INSN_LUID (tmp2);
}
/* Skip cycles if the current register pressure is high. */
-static int
+static int
sh_reorder2 (FILE *dump ATTRIBUTE_UNUSED,
int sched_verbose ATTRIBUTE_UNUSED,
rtx *ready ATTRIBUTE_UNUSED,
if (reload_completed)
return cached_can_issue_more;
- if (high_pressure(SFmode) || high_pressure (SImode))
+ if (high_pressure(SFmode) || high_pressure (SImode))
skip_cycles = 1;
return cached_can_issue_more;
/* Skip cycles without sorting the ready queue. This will move insn from
Q->R. If this is the last cycle we are skipping; allow sorting of ready
- queue by sh_reorder. */
+ queue by sh_reorder. */
-/* Generally, skipping these many cycles are sufficient for all insns to move
- from Q -> R. */
-#define MAX_SKIPS 8
+/* Generally, skipping these many cycles are sufficient for all insns to move
+ from Q -> R. */
+#define MAX_SKIPS 8
static int
sh_dfa_new_cycle (FILE *sched_dump ATTRIBUTE_UNUSED,
if (reload_completed)
return 0;
- if (skip_cycles)
+ if (skip_cycles)
{
if ((clock_var - last_clock_var) < MAX_SKIPS)
{
return (TARGET_SH5 || TARGET_HITACHI || sh_attr_renesas_p (record_type));
}
\f
-/*
+/*
On the SH1..SH4, the trampoline looks like
2 0002 D202 mov.l l2,r2
1 0000 D301 mov.l l1,r3
emit_move_insn (gen_rtx_MEM (Pmode,
plus_constant (tramp,
fixed_len
- + GET_MODE_SIZE (Pmode))),
+ + GET_MODE_SIZE (Pmode))),
cxt);
emit_insn (gen_ic_invalidate_line (tramp));
return;
static bool
sh_function_ok_for_sibcall (tree decl, tree exp ATTRIBUTE_UNUSED)
{
- return (decl
+ return (decl
&& (! TARGET_SHCOMPACT
|| current_function_args_info.stack_regs == 0)
&& ! sh_cfun_interrupt_handler_p ());
if (signature < SH_BLTIN_NUM_SHARED_SIGNATURES)
shared[signature] = type;
}
- builtin_function (d->name, type, d - bdesc, BUILT_IN_MD,
- NULL, NULL_TREE);
+ lang_hooks.builtin_function (d->name, type, d - bdesc, BUILT_IN_MD,
+ NULL, NULL_TREE);
}
}
+/* Implements target hook vector_mode_supported_p. */
+bool
+sh_vector_mode_supported_p (enum machine_mode mode)
+{
+ if (TARGET_FPU_ANY
+ && ((mode == V2SFmode)
+ || (mode == V4SFmode)
+ || (mode == V16SFmode)))
+ return true;
+
+ else if (TARGET_SHMEDIA
+ && ((mode == V8QImode)
+ || (mode == V2HImode)
+ || (mode == V4HImode)
+ || (mode == V2SImode)))
+ return true;
+
+ return false;
+}
+
static void
sh_init_builtins (void)
{
sh_cannot_change_mode_class (enum machine_mode from, enum machine_mode to,
enum reg_class class)
{
+ /* We want to enable the use of SUBREGs as a means to
+ VEC_SELECT a single element of a vector. */
+ if (to == SFmode && VECTOR_MODE_P (from) && GET_MODE_INNER (from) == SFmode)
+ return (reg_classes_intersect_p (GENERAL_REGS, class));
+
if (GET_MODE_SIZE (from) != GET_MODE_SIZE (to))
{
if (TARGET_LITTLE_ENDIAN)
if ((dstclass == FPUL_REGS
&& (srcclass == PR_REGS || srcclass == MAC_REGS || srcclass == T_REGS))
- || (srcclass == FPUL_REGS
+ || (srcclass == FPUL_REGS
&& (dstclass == PR_REGS || dstclass == MAC_REGS)))
return 7;
/* Find the "this" pointer. We have such a wide range of ABIs for the
SH that it's best to do this completely machine independently.
- "this" is passed as first argument, unless a structure return pointer
+ "this" is passed as first argument, unless a structure return pointer
comes first, in which case "this" comes second. */
INIT_CUMULATIVE_ARGS (cum, funtype, NULL_RTX, 0, 1);
#ifndef PCC_STATIC_STRUCT_RETURN
structure_value_byref = 1;
#endif /* not PCC_STATIC_STRUCT_RETURN */
if (structure_value_byref && sh_struct_value_rtx (function, 0) == 0)
- {
+ {
tree ptype = build_pointer_type (TREE_TYPE (funtype));
FUNCTION_ARG_ADVANCE (cum, Pmode, ptype, 1);
if (optimize > 0 && flag_schedule_insns_after_reload)
{
+ if (! basic_block_info)
+ init_flow ();
+ rtl_register_cfg_hooks ();
find_basic_blocks (insns, max_reg_num (), dump_file);
life_analysis (dump_file, PROP_FINAL);
break;
if (! INSN_P (insn))
continue;
-
+
if (GET_CODE (PATTERN (insn)) == SEQUENCE)
insn = XVECEXP (PATTERN (insn), 0, 0);
if (GET_CODE (PATTERN (insn)) != PARALLEL
abort ();
}
+/* Returns 1 if OP is a MEM that can be source of a simple move operation. */
+
+int
+unaligned_load_operand (rtx op, enum machine_mode mode)
+{
+ rtx inside;
+
+ if (GET_CODE (op) != MEM || GET_MODE (op) != mode)
+ return 0;
+
+ inside = XEXP (op, 0);
+
+ if (GET_CODE (inside) == POST_INC)
+ inside = XEXP (inside, 0);
+
+ if (GET_CODE (inside) == REG)
+ return 1;
+
+ return 0;
+}
+
+/* This function returns a constant rtx that represents pi / 2**15 in
+ SFmode. it's used to scale SFmode angles, in radians, to a
+ fixed-point signed 16.16-bit fraction of a full circle, i.e., 2*pi
+ maps to 0x10000). */
+
+static GTY(()) rtx sh_fsca_sf2int_rtx;
+
+rtx
+sh_fsca_sf2int (void)
+{
+ if (! sh_fsca_sf2int_rtx)
+ {
+ REAL_VALUE_TYPE rv;
+
+ real_from_string (&rv, "10430.378350470453");
+ sh_fsca_sf2int_rtx = const_double_from_real_value (rv, SFmode);
+ }
+
+ return sh_fsca_sf2int_rtx;
+}
+
+/* This function returns a constant rtx that represents pi / 2**15 in
+ DFmode. it's used to scale DFmode angles, in radians, to a
+ fixed-point signed 16.16-bit fraction of a full circle, i.e., 2*pi
+ maps to 0x10000). */
+
+static GTY(()) rtx sh_fsca_df2int_rtx;
+
+rtx
+sh_fsca_df2int (void)
+{
+ if (! sh_fsca_df2int_rtx)
+ {
+ REAL_VALUE_TYPE rv;
+
+ real_from_string (&rv, "10430.378350470453");
+ sh_fsca_df2int_rtx = const_double_from_real_value (rv, DFmode);
+ }
+
+ return sh_fsca_df2int_rtx;
+}
+
+/* This function returns a constant rtx that represents 2**15 / pi in
+ SFmode. it's used to scale a fixed-point signed 16.16-bit fraction
+ of a full circle back to a SFmode value, i.e., 0x10000 maps to
+ 2*pi). */
+
+static GTY(()) rtx sh_fsca_int2sf_rtx;
+
+rtx
+sh_fsca_int2sf (void)
+{
+ if (! sh_fsca_int2sf_rtx)
+ {
+ REAL_VALUE_TYPE rv;
+
+ real_from_string (&rv, "9.587379924285257e-5");
+ sh_fsca_int2sf_rtx = const_double_from_real_value (rv, SFmode);
+ }
+
+ return sh_fsca_int2sf_rtx;
+}
#include "gt-sh.h"
OSDN Git Service
