/* Output routines for GCC for Renesas / SuperH SH.
Copyright (C) 1993, 1994, 1995, 1996, 1997, 1998, 1999, 2000, 2001, 2002,
- 2003, 2004, 2005 Free Software Foundation, Inc.
+ 2003, 2004, 2005, 2006, 2007 Free Software Foundation, Inc.
Contributed by Steve Chamberlain (sac@cygnus.com).
Improved by Jim Wilson (wilson@cygnus.com).
#include "real.h"
#include "langhooks.h"
#include "basic-block.h"
+#include "df.h"
#include "cfglayout.h"
#include "intl.h"
#include "sched-int.h"
#include "ggc.h"
#include "tree-gimple.h"
#include "cfgloop.h"
+#include "alloc-pool.h"
+#include "tm-constrs.h"
int code_for_indirect_jump_scratch = CODE_FOR_indirect_jump_scratch;
/* Set to 1 by expand_prologue() when the function is an interrupt handler. */
int current_function_interrupt;
-/* ??? The pragma interrupt support will not work for SH3. */
-/* This is set by #pragma interrupt and #pragma trapa, and causes gcc to
- output code for the next function appropriate for an interrupt handler. */
-int pragma_interrupt;
-
-/* This is set by the trap_exit attribute for functions. It specifies
- a trap number to be used in a trapa instruction at function exit
- (instead of an rte instruction). */
-int trap_exit;
-
-/* This is used by the sp_switch attribute for functions. It specifies
- a variable holding the address of the stack the interrupt function
- should switch to/from at entry/exit. */
-rtx sp_switch;
-
-/* This is set by #pragma trapa, and is similar to the above, except that
- the compiler doesn't emit code to preserve all registers. */
-static int pragma_trapa;
-
-/* This is set by #pragma nosave_low_regs. This is useful on the SH3,
- which has a separate set of low regs for User and Supervisor modes.
- This should only be used for the lowest level of interrupts. Higher levels
- of interrupts must save the registers in case they themselves are
- interrupted. */
-int pragma_nosave_low_regs;
-
-/* This is used for communication between TARGET_SETUP_INCOMING_VARARGS and
- sh_expand_prologue. */
-int current_function_anonymous_args;
+tree sh_deferred_function_attributes;
+tree *sh_deferred_function_attributes_tail = &sh_deferred_function_attributes;
/* Global variables for machine-dependent things. */
/* 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;
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, GENERAL_REGS,
};
char sh_register_names[FIRST_PSEUDO_REGISTER] \
[MAX_ADDITIONAL_REGISTER_NAME_LENGTH + 1]
= SH_ADDITIONAL_REGISTER_NAMES_INITIALIZER;
-/* Provide reg_class from a letter such as appears in the machine
- description. *: target independently reserved letter.
- reg_class_from_letter['e' - 'a'] is set to NO_REGS for TARGET_FMOVD. */
-
-enum reg_class reg_class_from_letter[] =
-{
- /* a */ ALL_REGS, /* b */ TARGET_REGS, /* c */ FPSCR_REGS, /* d */ DF_REGS,
- /* e */ FP_REGS, /* f */ FP_REGS, /* g **/ NO_REGS, /* h */ NO_REGS,
- /* i **/ NO_REGS, /* j */ NO_REGS, /* k */ SIBCALL_REGS, /* l */ PR_REGS,
- /* m **/ NO_REGS, /* n **/ NO_REGS, /* o **/ NO_REGS, /* p **/ NO_REGS,
- /* q */ NO_REGS, /* r **/ NO_REGS, /* s **/ NO_REGS, /* t */ T_REGS,
- /* u */ NO_REGS, /* v */ NO_REGS, /* w */ FP0_REGS, /* x */ MAC_REGS,
- /* y */ FPUL_REGS, /* z */ R0_REGS
-};
-
int assembler_dialect;
static bool shmedia_space_reserved_for_target_registers;
static void pop (int);
static void push_regs (HARD_REG_SET *, int);
static int calc_live_regs (HARD_REG_SET *);
-static void mark_use (rtx, rtx *);
static HOST_WIDE_INT rounded_frame_size (int);
static rtx mark_constant_pool_use (rtx);
const struct attribute_spec sh_attribute_table[];
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 short find_insn_regmode_weight (rtx, enum machine_mode);
-static void find_regmode_weight (int, enum machine_mode);
+static void find_regmode_weight (basic_block, enum 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 bool sh_cannot_copy_insn_p (rtx);
static bool sh_rtx_costs (rtx, int, int, int *);
static int sh_address_cost (rtx);
-#ifdef TARGET_ADJUST_UNROLL_MAX
-static int sh_adjust_unroll_max (struct loop *, int, int, int, int);
-#endif
static int sh_pr_n_sets (void);
static rtx sh_allocate_initial_value (rtx);
static int shmedia_target_regs_stack_space (HARD_REG_SET *);
static int sh_arg_partial_bytes (CUMULATIVE_ARGS *, enum machine_mode,
tree, bool);
static int sh_dwarf_calling_convention (tree);
-static int hard_regs_intersect_p (HARD_REG_SET *, HARD_REG_SET *);
\f
/* Initialize the GCC target structure. */
#endif /* SYMBIAN */
-#ifdef TARGET_ADJUST_UNROLL_MAX
-#undef TARGET_ADJUST_UNROLL_MAX
-#define TARGET_ADJUST_UNROLL_MAX sh_adjust_unroll_max
-#endif
+#undef TARGET_SECONDARY_RELOAD
+#define TARGET_SECONDARY_RELOAD sh_secondary_reload
struct gcc_target targetm = TARGET_INITIALIZER;
\f
return true;
case OPT_m4:
+ case OPT_m4_100:
+ case OPT_m4_200:
+ case OPT_m4_300:
target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4;
return true;
case OPT_m4_nofpu:
+ case OPT_m4_100_nofpu:
+ case OPT_m4_200_nofpu:
+ case OPT_m4_300_nofpu:
+ case OPT_m4_340:
+ case OPT_m4_400:
+ case OPT_m4_500:
target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4_NOFPU;
return true;
case OPT_m4_single:
+ case OPT_m4_100_single:
+ case OPT_m4_200_single:
+ case OPT_m4_300_single:
target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4_SINGLE;
return true;
case OPT_m4_single_only:
+ case OPT_m4_100_single_only:
+ case OPT_m4_200_single_only:
+ case OPT_m4_300_single_only:
target_flags = (target_flags & ~MASK_ARCH) | SELECT_SH4_SINGLE_ONLY;
return true;
'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' print an `x' if `m' will print `base,index'.
+ '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.
switch (code)
{
+ tree trapa_attr;
+
case '.':
if (final_sequence
&& ! INSN_ANNULLED_BRANCH_P (XVECEXP (final_sequence, 0, 0))
fprintf (stream, "%s", LOCAL_LABEL_PREFIX);
break;
case '@':
- if (trap_exit)
- fprintf (stream, "trapa #%d", trap_exit);
+ 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 ())
fprintf (stream, "rte");
else
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':
- fputs (reg_names[REGNO (x) + LSW], (stream));
+ if (REG_P (x) || GET_CODE (x) == SUBREG)
+ {
+ regno = true_regnum (x);
+ regno += FP_REGISTER_P (regno) ? 1 : LSW;
+ fputs (reg_names[regno], (stream));
+ }
+ else if (MEM_P (x))
+ {
+ x = adjust_address (x, SImode, 4 * LSW);
+ 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 * LSW);
+ if (sub)
+ print_operand (stream, sub, 0);
+ else
+ output_operand_lossage ("invalid operand to %%R");
+ }
break;
case 'S':
- fputs (reg_names[REGNO (x) + MSW], (stream));
+ if (REG_P (x) || GET_CODE (x) == SUBREG)
+ {
+ regno = true_regnum (x);
+ regno += FP_REGISTER_P (regno) ? 0 : MSW;
+ fputs (reg_names[regno], (stream));
+ }
+ else if (MEM_P (x))
+ {
+ x = adjust_address (x, SImode, 4 * MSW);
+ 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 * MSW);
+ if (sub)
+ print_operand (stream, sub, 0);
+ else
+ output_operand_lossage ("invalid operand to %%S");
+ }
break;
case 'T':
/* Next word of a double. */
}
break;
case 'M':
- if (GET_CODE (x) == MEM
- && GET_CODE (XEXP (x, 0)) == PLUS
- && (GET_CODE (XEXP (XEXP (x, 0), 1)) == REG
- || GET_CODE (XEXP (XEXP (x, 0), 1)) == SUBREG))
- fputc ('x', stream);
+ if (TARGET_SHMEDIA)
+ {
+ if (GET_CODE (x) == MEM
+ && GET_CODE (XEXP (x, 0)) == PLUS
+ && (GET_CODE (XEXP (XEXP (x, 0), 1)) == REG
+ || GET_CODE (XEXP (XEXP (x, 0), 1)) == SUBREG))
+ fputc ('x', stream);
+ }
+ else
+ {
+ if (GET_CODE (x) == MEM)
+ {
+ 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':
case CONST:
if (TARGET_SHMEDIA
- && GET_CODE (XEXP (x, 0)) == SIGN_EXTEND
+ && (GET_CODE (XEXP (x, 0)) == SIGN_EXTEND
+ || GET_CODE (XEXP (x, 0)) == ZERO_EXTEND)
&& (GET_MODE (XEXP (x, 0)) == DImode
|| GET_MODE (XEXP (x, 0)) == SImode)
&& GET_CODE (XEXP (XEXP (x, 0), 0)) == TRUNCATE
&& GET_MODE (XEXP (XEXP (x, 0), 0)) == HImode)
{
rtx val = XEXP (XEXP (XEXP (x, 0), 0), 0);
+ rtx val2 = val;
+ bool nested_expr = false;
fputc ('(', stream);
if (GET_CODE (val) == ASHIFTRT)
{
fputc ('(', stream);
- if (GET_CODE (XEXP (val, 0)) == CONST)
- fputc ('(', stream);
- output_addr_const (stream, XEXP (val, 0));
- if (GET_CODE (XEXP (val, 0)) == CONST)
- fputc (')', stream);
+ val2 = XEXP (val, 0);
+ }
+ if (GET_CODE (val2) == CONST
+ || GET_RTX_CLASS (GET_CODE (val2)) != RTX_OBJ)
+ {
+ fputc ('(', stream);
+ nested_expr = true;
+ }
+ output_addr_const (stream, val2);
+ if (nested_expr)
+ fputc (')', stream);
+ if (GET_CODE (val) == ASHIFTRT)
+ {
fputs (" >> ", stream);
output_addr_const (stream, XEXP (val, 1));
fputc (')', stream);
}
- else
- {
- if (GET_CODE (val) == CONST)
- fputc ('(', stream);
- output_addr_const (stream, val);
- if (GET_CODE (val) == CONST)
- fputc (')', stream);
- }
fputs (" & 65535)", stream);
break;
}
while (copied + 4 <= bytes)
{
rtx to = adjust_address (dest, SImode, copied);
- rtx from = adjust_automodify_address (src, SImode, src_addr, copied);
+ rtx from = adjust_automodify_address (src, BLKmode,
+ src_addr, copied);
+ set_mem_size (from, GEN_INT (4));
emit_insn (gen_movua (temp, from));
emit_move_insn (src_addr, plus_constant (src_addr, 4));
emit_move_insn (to, temp);
/* It's ok. */;
else
{
- temp = no_new_pseudos ? operands[0] : gen_reg_rtx (Pmode);
+ temp = (!can_create_pseudo_p ()
+ ? operands[0]
+ : gen_reg_rtx (Pmode));
operands[1] = legitimize_pic_address (operands[1], mode, temp);
}
}
&& GET_CODE (XEXP (operands[1], 0)) == PLUS
&& SYMBOLIC_CONST_P (XEXP (XEXP (operands[1], 0), 0)))
{
- temp = no_new_pseudos ? operands[0] : gen_reg_rtx (Pmode);
+ 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),
- no_new_pseudos ? temp
- : gen_reg_rtx (Pmode),
+ (!can_create_pseudo_p ()
+ ? temp
+ : gen_reg_rtx (Pmode)),
0, OPTAB_LIB_WIDEN);
}
}
if (mode == Pmode || mode == ptr_mode)
{
- rtx op0, op1;
+ 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))
+ {
+ opc = XEXP (XEXP (op1, 0), 1);
+ op1 = XEXP (XEXP (op1, 0), 0);
+ }
+ else
+ opc = NULL_RTX;
+
if ((tls_kind = tls_symbolic_operand (op1, Pmode)))
{
rtx tga_op1, tga_ret, tmp, tmp2;
if (flag_schedule_insns)
emit_insn (gen_blockage ());
}
- tga_op1 = no_new_pseudos ? op0 : gen_reg_rtx (Pmode);
+ 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;
default:
gcc_unreachable ();
}
+ if (opc)
+ emit_insn (gen_addsi3 (op1, op1, force_reg (SImode, opc)));
operands[1] = op1;
}
}
return 0;
}
+enum rtx_code
+prepare_cbranch_operands (rtx *operands, enum machine_mode mode,
+ enum rtx_code comparison)
+{
+ rtx op1;
+ rtx scratch = NULL_RTX;
+
+ if (comparison == CODE_FOR_nothing)
+ comparison = GET_CODE (operands[0]);
+ else
+ scratch = operands[4];
+ if (GET_CODE (operands[1]) == CONST_INT
+ && GET_CODE (operands[2]) != CONST_INT)
+ {
+ rtx tmp = operands[1];
+
+ operands[1] = operands[2];
+ operands[2] = tmp;
+ comparison = swap_condition (comparison);
+ }
+ if (GET_CODE (operands[2]) == CONST_INT)
+ {
+ HOST_WIDE_INT val = INTVAL (operands[2]);
+ if ((val == -1 || val == -0x81)
+ && (comparison == GT || comparison == LE))
+ {
+ comparison = (comparison == GT) ? GE : LT;
+ operands[2] = gen_int_mode (val + 1, mode);
+ }
+ else if ((val == 1 || val == 0x80)
+ && (comparison == GE || comparison == LT))
+ {
+ comparison = (comparison == GE) ? GT : LE;
+ operands[2] = gen_int_mode (val - 1, mode);
+ }
+ else if (val == 1 && (comparison == GEU || comparison == LTU))
+ {
+ comparison = (comparison == GEU) ? NE : EQ;
+ operands[2] = CONST0_RTX (mode);
+ }
+ else if (val == 0x80 && (comparison == GEU || comparison == LTU))
+ {
+ comparison = (comparison == GEU) ? GTU : LEU;
+ operands[2] = gen_int_mode (val - 1, mode);
+ }
+ else if (val == 0 && (comparison == GTU || comparison == LEU))
+ comparison = (comparison == GTU) ? NE : EQ;
+ else if (mode == SImode
+ && ((val == 0x7fffffff
+ && (comparison == GTU || comparison == LEU))
+ || ((unsigned HOST_WIDE_INT) val
+ == (unsigned HOST_WIDE_INT) 0x7fffffff + 1
+ && (comparison == GEU || comparison == LTU))))
+ {
+ comparison = (comparison == GTU || comparison == GEU) ? LT : GE;
+ operands[2] = CONST0_RTX (mode);
+ }
+ }
+ 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])
+ && (GET_CODE (operands[2]) != CONST_INT
+ || (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;
+ rtx jump;
+
+ 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, gen_rtx_REG (SImode, T_REG),
+ gen_rtx_fmt_ee (comparison, SImode,
+ operands[1], operands[2])));
+ jump = emit_jump_insn (branch_expander (operands[3]));
+ if (probability >= 0)
+ REG_NOTES (jump)
+ = gen_rtx_EXPR_LIST (REG_BR_PROB, GEN_INT (probability),
+ REG_NOTES (jump));
+
+}
+
+/* ??? How should we distribute probabilities when more than one branch
+ is generated. So far we only have soem 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 skip_label = NULL_RTX;
+ 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 = CODE_FOR_nothing;
+ 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
+ {
+ gcc_assert (HOST_BITS_PER_WIDEST_INT >= 64);
+ lsw_taken_prob
+ = (prob
+ ? (REG_BR_PROB_BASE
+ - ((HOST_WIDEST_INT) REG_BR_PROB_BASE * rev_prob
+ / ((HOST_WIDEST_INT) 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 (GET_CODE (op2l) == CONST_INT && 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 (GET_CODE (op2l) == CONST_INT && 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
+ break;
+ msw_skip = swap_condition (msw_taken);
+ }
+ break;
+ default: return false;
+ }
+ num_branches = ((msw_taken != CODE_FOR_nothing)
+ + (msw_skip != CODE_FOR_nothing)
+ + (lsw_taken != CODE_FOR_nothing));
+ 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)
+ && (msw_taken != CODE_FOR_nothing || msw_skip != CODE_FOR_nothing))
+ {
+ emit_move_insn (scratch, operands[2]);
+ operands[2] = scratch;
+ }
+ if (msw_taken != CODE_FOR_nothing)
+ expand_cbranchsi4 (operands, msw_taken, msw_taken_prob);
+ if (msw_skip != CODE_FOR_nothing)
+ {
+ rtx taken_label = operands[3];
+
+ 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 != CODE_FOR_nothing)
+ {
+ if (reload_completed
+ && ! arith_reg_or_0_operand (op2l, SImode) && true_regnum (op1l))
+ operands[4] = scratch;
+ expand_cbranchsi4 (operands, lsw_taken, lsw_taken_prob);
+ }
+ if (msw_skip != CODE_FOR_nothing)
+ emit_label (skip_label);
+ return true;
+}
+
/* Prepare the operands for an scc instruction; make sure that the
compare has been done. */
rtx
}
}
+/* Output a code sequence for INSN using TEMPLATE 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 *template,
rtx insn, rtx *operands)
const char *
output_ieee_ccmpeq (rtx insn, rtx *operands)
{
- return output_branchy_insn (NE, "bt\t%l9\\;fcmp/eq\t%1,%0", insn, operands);
+ return output_branchy_insn (NE, "bt\t%l9\n\tfcmp/eq\t%1,%0",
+ insn, operands);
}
\f
/* Output the start of the assembler file. */
else
/* Switch to the data section so that the coffsem symbol
isn't in the text section. */
- data_section ();
+ switch_to_section (data_section);
if (TARGET_LITTLE_ENDIAN)
fputs ("\t.little\n", asm_out_file);
}
/* Indicate that INSN cannot be duplicated. This is true for insn
- that generates an unique label. */
+ that generates a unique label. */
static bool
sh_cannot_copy_insn_p (rtx insn)
return 2;
/* Everything else is invalid, because there is no pattern for it. */
- return 10000;
+ return MAX_COST;
}
/* If shift by a non constant, then this will be expensive. */
if (GET_CODE (XEXP (x, 1)) != CONST_INT)
if (TARGET_SHMEDIA)
{
- if ((GET_CODE (XEXP (x, 1)) == CONST_INT
- && CONST_OK_FOR_I16 (INTVAL (XEXP (x, 1))))
- || EXTRA_CONSTRAINT_C16 (XEXP (x, 1)))
+ if (satisfies_constraint_I10 (XEXP (x, 1))
+ || satisfies_constraint_J16 (XEXP (x, 1)))
return 1;
else
- return 2;
+ return 1 + rtx_cost (XEXP (x, 1), AND);
}
/* These constants are single cycle extu.[bw] instructions. */
else if (CONST_OK_FOR_I16 (INTVAL (x) >> 16))
*total = COSTS_N_INSNS ((outer_code != SET) + 1);
else if (CONST_OK_FOR_I16 ((INTVAL (x) >> 16) >> 16))
- *total = COSTS_N_INSNS (3);
+ *total = COSTS_N_INSNS ((outer_code != SET) + 2);
else
- *total = COSTS_N_INSNS (4);
+ *total = COSTS_N_INSNS ((outer_code != SET) + 3);
return true;
}
if (CONST_OK_FOR_I08 (INTVAL (x)))
else if ((outer_code == AND || outer_code == IOR || outer_code == XOR)
&& CONST_OK_FOR_K08 (INTVAL (x)))
*total = 1;
+ /* prepare_cmp_insn will force costly constants int registers before
+ the cbrach[sd]i4 patterns can see them, so preserve potentially
+ interesting ones not covered by I08 above. */
+ else if (outer_code == COMPARE
+ && ((unsigned HOST_WIDE_INT) INTVAL (x)
+ == (unsigned HOST_WIDE_INT) 0x7fffffff + 1
+ || INTVAL (x) == 0x7fffffff
+ || INTVAL (x) == 0x80 || INTVAL (x) == -0x81))
+ *total = 1;
else
*total = 8;
return true;
case CONST_DOUBLE:
if (TARGET_SHMEDIA)
*total = COSTS_N_INSNS (4);
+ /* prepare_cmp_insn will force costly constants int registers before
+ the cbrachdi4 pattern can see them, so preserve potentially
+ interesting ones. */
+ else if (outer_code == COMPARE && GET_MODE (x) == DImode)
+ *total = 1;
else
*total = 10;
return true;
}
\f
+static alloc_pool label_ref_list_pool;
+
+typedef struct label_ref_list_d
+{
+ rtx label;
+ struct label_ref_list_d *next;
+} *label_ref_list_t;
+
/* The SH cannot load a large constant into a register, constants have to
come from a pc relative load. The reference of a pc relative load
instruction must be less than 1k in front of the instruction. This
{
rtx value; /* Value in table. */
rtx label; /* Label of value. */
- rtx wend; /* End of window. */
+ label_ref_list_t wend; /* End of window. */
enum machine_mode mode; /* Mode of value. */
/* True if this constant is accessed as part of a post-increment
static rtx pool_window_label;
static int pool_window_last;
+static int max_labelno_before_reorg;
+
/* ??? If we need a constant in HImode which is the truncated value of a
constant we need in SImode, we could combine the two entries thus saving
two bytes. Is this common enough to be worth the effort of implementing
add_constant (rtx x, enum machine_mode mode, rtx last_value)
{
int i;
- rtx lab, new, ref, newref;
+ rtx lab, new;
+ label_ref_list_t ref, newref;
/* First see if we've already got it. */
for (i = 0; i < pool_size; i++)
}
if (lab && pool_window_label)
{
- newref = gen_rtx_LABEL_REF (VOIDmode, pool_window_label);
+ newref = (label_ref_list_t) pool_alloc (label_ref_list_pool);
+ newref->label = pool_window_label;
ref = pool_vector[pool_window_last].wend;
- LABEL_NEXTREF (newref) = ref;
+ newref->next = ref;
pool_vector[pool_window_last].wend = newref;
}
if (new)
lab = gen_label_rtx ();
pool_vector[pool_size].mode = mode;
pool_vector[pool_size].label = lab;
- pool_vector[pool_size].wend = NULL_RTX;
+ pool_vector[pool_size].wend = NULL;
pool_vector[pool_size].part_of_sequence_p = (lab == 0);
if (lab && pool_window_label)
{
- newref = gen_rtx_LABEL_REF (VOIDmode, pool_window_label);
+ newref = (label_ref_list_t) pool_alloc (label_ref_list_pool);
+ newref->label = pool_window_label;
ref = pool_vector[pool_window_last].wend;
- LABEL_NEXTREF (newref) = ref;
+ newref->next = ref;
pool_vector[pool_window_last].wend = newref;
}
if (lab)
rtx scan = barrier;
int i;
int need_align = 1;
- rtx lab, ref;
+ rtx lab;
+ label_ref_list_t ref;
int have_df = 0;
/* Do two passes, first time dump out the HI sized constants. */
scan = emit_label_after (lab, scan);
scan = emit_insn_after (gen_consttable_2 (p->value, const0_rtx),
scan);
- for (ref = p->wend; ref; ref = LABEL_NEXTREF (ref))
+ for (ref = p->wend; ref; ref = ref->next)
{
- lab = XEXP (ref, 0);
+ lab = ref->label;
scan = emit_insn_after (gen_consttable_window_end (lab), scan);
}
}
emit_label_before (lab, align_insn);
emit_insn_before (gen_consttable_4 (p->value, const0_rtx),
align_insn);
- for (ref = p->wend; ref; ref = LABEL_NEXTREF (ref))
+ for (ref = p->wend; ref; ref = ref->next)
{
- lab = XEXP (ref, 0);
+ lab = ref->label;
emit_insn_before (gen_consttable_window_end (lab),
align_insn);
}
if (p->mode != HImode)
{
- for (ref = p->wend; ref; ref = LABEL_NEXTREF (ref))
+ for (ref = p->wend; ref; ref = ref->next)
{
- lab = XEXP (ref, 0);
+ lab = ref->label;
scan = emit_insn_after (gen_consttable_window_end (lab),
scan);
}
if (p->mode != HImode)
{
- for (ref = p->wend; ref; ref = LABEL_NEXTREF (ref))
+ for (ref = p->wend; ref; ref = ref->next)
{
- lab = XEXP (ref, 0);
+ lab = ref->label;
scan = emit_insn_after (gen_consttable_window_end (lab), scan);
}
}
&& INTVAL (src) <= 32767);
}
+#define MOVA_LABELREF(mova) XVECEXP (SET_SRC (PATTERN (mova)), 0, 0)
+
/* Nonzero if the insn is a move instruction which needs to be fixed. */
/* ??? For a DImode/DFmode moves, we don't need to fix it if each half of the
if (GET_CODE (pat) == PARALLEL)
pat = XVECEXP (pat, 0, 0);
if (GET_CODE (pat) == SET
- /* We can load any 8 bit value if we don't care what the high
+ /* We can load any 8-bit value if we don't care what the high
order bits end up as. */
&& GET_MODE (SET_DEST (pat)) != QImode
&& (CONSTANT_P (SET_SRC (pat))
&& 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)))))
+ && satisfies_constraint_I20 (SET_SRC (pat)))
+ && ! satisfies_constraint_I08 (SET_SRC (pat)))
return 1;
}
&& GET_CODE (SET_SRC (PATTERN (insn))) == UNSPEC
&& XINT (SET_SRC (PATTERN (insn)), 1) == UNSPEC_MOVA
/* Don't match mova_const. */
- && GET_CODE (XVECEXP (SET_SRC (PATTERN (insn)), 0, 0)) == LABEL_REF);
+ && GET_CODE (MOVA_LABELREF (insn)) == LABEL_REF);
}
/* Fix up a mova from a switch that went out of range. */
static void
fixup_mova (rtx mova)
{
+ PUT_MODE (XEXP (MOVA_LABELREF (mova), 0), QImode);
if (! flag_pic)
{
- SET_SRC (PATTERN (mova)) = XVECEXP (SET_SRC (PATTERN (mova)), 0, 0);
+ SET_SRC (PATTERN (mova)) = MOVA_LABELREF (mova);
INSN_CODE (mova) = -1;
}
else
gcc_assert (worker
&& GET_CODE (worker) != CODE_LABEL
&& GET_CODE (worker) != JUMP_INSN);
- } while (recog_memoized (worker) != CODE_FOR_casesi_worker_1);
+ } while (GET_CODE (worker) == NOTE
+ || recog_memoized (worker) != CODE_FOR_casesi_worker_1);
wpat = PATTERN (worker);
wpat0 = XVECEXP (wpat, 0, 0);
wpat1 = XVECEXP (wpat, 0, 1);
}
}
+/* NEW_MOVA is a mova we've just encountered while scanning forward. Update
+ *num_mova, and check if the new mova is not nested within the first one.
+ return 0 if *first_mova was replaced, 1 if new_mova was replaced,
+ 2 if new_mova has been assigned to *first_mova, -1 otherwise.. */
+static int
+untangle_mova (int *num_mova, rtx *first_mova, rtx new_mova)
+{
+ int n_addr = 0; /* Initialization to shut up spurious warning. */
+ int f_target, n_target = 0; /* Likewise. */
+
+ if (optimize)
+ {
+ n_addr = INSN_ADDRESSES (INSN_UID (new_mova));
+ n_target = INSN_ADDRESSES (INSN_UID (XEXP (MOVA_LABELREF (new_mova), 0)));
+ if (n_addr > n_target || n_addr + 1022 < n_target)
+ {
+ /* Change the mova into a load.
+ broken_move will then return true for it. */
+ fixup_mova (new_mova);
+ return 1;
+ }
+ }
+ if (!(*num_mova)++)
+ {
+ *first_mova = new_mova;
+ return 2;
+ }
+ if (!optimize
+ || ((f_target
+ = INSN_ADDRESSES (INSN_UID (XEXP (MOVA_LABELREF (*first_mova), 0))))
+ >= n_target))
+ return -1;
+
+ (*num_mova)--;
+ if (f_target - INSN_ADDRESSES (INSN_UID (*first_mova))
+ > n_target - n_addr)
+ {
+ fixup_mova (*first_mova);
+ return 0;
+ }
+ else
+ {
+ fixup_mova (new_mova);
+ return 1;
+ }
+}
+
/* Find the last barrier from insn FROM which is close enough to hold the
constant pool. If we can't find one, then create one near the end of
the range. */
int inc = get_attr_length (from);
int new_align = 1;
- if (GET_CODE (from) == CODE_LABEL)
+ /* If this is a label that existed at the time of the compute_alignments
+ call, determine the alignment. N.B. When find_barrier recurses for
+ an out-of-reach mova, we might see labels at the start of previously
+ inserted constant tables. */
+ if (GET_CODE (from) == CODE_LABEL
+ && CODE_LABEL_NUMBER (from) <= max_labelno_before_reorg)
{
if (optimize)
new_align = 1 << label_to_alignment (from);
new_align = 1;
inc = 0;
}
+ /* In case we are scanning a constant table because of recursion, check
+ for explicit alignments. If the table is long, we might be forced
+ to emit the new table in front of it; the length of the alignment
+ might be the last straw. */
+ else if (GET_CODE (from) == INSN
+ && GET_CODE (PATTERN (from)) == UNSPEC_VOLATILE
+ && XINT (PATTERN (from), 1) == UNSPECV_ALIGN)
+ new_align = INTVAL (XVECEXP (PATTERN (from), 0, 0));
+ /* When we find the end of a constant table, paste the new constant
+ at the end. That is better than putting it in front because
+ this way, we don't need extra alignment for adding a 4-byte-aligned
+ mov(a) label to a 2/4 or 8/4 byte aligned table. */
+ else if (GET_CODE (from) == INSN
+ && GET_CODE (PATTERN (from)) == UNSPEC_VOLATILE
+ && XINT (PATTERN (from), 1) == UNSPECV_CONST_END)
+ return from;
if (GET_CODE (from) == BARRIER)
{
if (mova_p (from))
{
- if (! num_mova++)
+ switch (untangle_mova (&num_mova, &mova, from))
{
- leading_mova = 0;
- mova = from;
- barrier_before_mova = good_barrier ? good_barrier : found_barrier;
+ case 0: return find_barrier (0, 0, mova);
+ case 2:
+ {
+ leading_mova = 0;
+ barrier_before_mova
+ = good_barrier ? good_barrier : found_barrier;
+ }
+ default: break;
}
if (found_si > count_si)
count_si = found_si;
&& (GET_CODE (PATTERN (from)) == ADDR_VEC
|| GET_CODE (PATTERN (from)) == ADDR_DIFF_VEC))
{
- if (num_mova)
+ if ((num_mova > 1 && GET_MODE (prev_nonnote_insn (from)) == VOIDmode)
+ || (num_mova
+ && (prev_nonnote_insn (from)
+ == XEXP (MOVA_LABELREF (mova), 0))))
num_mova--;
if (barrier_align (next_real_insn (from)) == align_jumps_log)
{
rtx r0_inc_rtx = gen_rtx_POST_INC (Pmode, r0_rtx);
first = get_insns ();
+ max_labelno_before_reorg = max_label_num ();
/* We must split call insns before introducing `mova's. If we're
optimizing, they'll have already been split. Otherwise, make
if (GET_CODE (reg) != REG)
continue;
- /* This is a function call via REG. If the only uses of REG
- between the time that it is set and the time that it dies
- are in function calls, then we can associate all the
- function calls with the setting of REG. */
-
- for (link = LOG_LINKS (insn); link; link = XEXP (link, 1))
- {
- if (REG_NOTE_KIND (link) != 0)
- continue;
- set = single_set (XEXP (link, 0));
- if (set && rtx_equal_p (reg, SET_DEST (set)))
- {
- link = XEXP (link, 0);
- break;
- }
- }
-
- if (! link)
+ /* Try scanning backward to find where the register is set. */
+ link = NULL;
+ for (scan = PREV_INSN (insn);
+ scan && GET_CODE (scan) != CODE_LABEL;
+ scan = PREV_INSN (scan))
{
- /* ??? Sometimes global register allocation will have
- deleted the insn pointed to by LOG_LINKS. Try
- scanning backward to find where the register is set. */
- for (scan = PREV_INSN (insn);
- scan && GET_CODE (scan) != CODE_LABEL;
- scan = PREV_INSN (scan))
- {
- if (! INSN_P (scan))
- continue;
+ if (! INSN_P (scan))
+ continue;
- if (! reg_mentioned_p (reg, scan))
- continue;
+ if (! reg_mentioned_p (reg, scan))
+ continue;
- if (noncall_uses_reg (reg, scan, &set))
- break;
+ if (noncall_uses_reg (reg, scan, &set))
+ break;
- if (set)
- {
- link = scan;
- break;
- }
+ if (set)
+ {
+ link = scan;
+ break;
}
}
/* Don't try to trace forward past a CODE_LABEL if we haven't
seen INSN yet. Ordinarily, we will only find the setting insn
- in LOG_LINKS if it is in the same basic block. However,
+ if it is in the same basic block. However,
cross-jumping can insert code labels in between the load and
the call, and can result in situations where a single call
insn may have two targets depending on where we came from. */
later insn. */
/* ??? We shouldn't have to use FOUNDINSN here.
- However, the LOG_LINKS fields are apparently not
- entirely reliable around libcalls;
- newlib/libm/math/e_pow.c is a test case. Sometimes
- an insn will appear in LOG_LINKS even though it is
- not the most recent insn which sets the register. */
+ This dates back to when we used LOG_LINKS to find
+ the most recent insn which sets the register. */
if (foundinsn
&& (scanset
mdep_reorg_phase = SH_SHORTEN_BRANCHES0;
shorten_branches (first);
}
+
/* Scan the function looking for move instructions which have to be
changed to pc-relative loads and insert the literal tables. */
-
+ label_ref_list_pool = create_alloc_pool ("label references list",
+ sizeof (struct label_ref_list_d),
+ 30);
mdep_reorg_phase = SH_FIXUP_PCLOAD;
for (insn = first, num_mova = 0; insn; insn = NEXT_INSN (insn))
{
below the switch table. Check if that has happened.
We only have the addresses available when optimizing; but then,
this check shouldn't be needed when not optimizing. */
- rtx label_ref = XVECEXP (SET_SRC (PATTERN (insn)), 0, 0);
- if (optimize
- && (INSN_ADDRESSES (INSN_UID (insn))
- > INSN_ADDRESSES (INSN_UID (XEXP (label_ref, 0)))))
+ if (!untangle_mova (&num_mova, &mova, insn))
{
- /* Change the mova into a load.
- broken_move will then return true for it. */
- fixup_mova (insn);
+ insn = mova;
+ num_mova = 0;
}
- else if (! num_mova++)
- mova = insn;
}
else if (GET_CODE (insn) == JUMP_INSN
&& GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC
- && num_mova)
+ && num_mova
+ /* ??? loop invariant motion can also move a mova out of a
+ loop. Since loop does this code motion anyway, maybe we
+ should wrap UNSPEC_MOVA into a CONST, so that reload can
+ move it back. */
+ && ((num_mova > 1
+ && GET_MODE (prev_nonnote_insn (insn)) == VOIDmode)
+ || (prev_nonnote_insn (insn)
+ == XEXP (MOVA_LABELREF (mova), 0))))
{
rtx scan;
int total;
}
last_float_move = scan;
last_float = src;
- newsrc = gen_rtx_MEM (mode,
+ newsrc = gen_const_mem (mode,
(((TARGET_SH4 && ! TARGET_FMOVD)
|| REGNO (dst) == FPUL_REG)
? r0_inc_rtx
insn = barrier;
}
}
+ free_alloc_pool (label_ref_list_pool);
+ for (insn = first; insn; insn = NEXT_INSN (insn))
+ PUT_MODE (insn, VOIDmode);
mdep_reorg_phase = SH_SHORTEN_BRANCHES1;
INSN_ADDRESSES_FREE ();
{
/* Shorten_branches would split this instruction again,
so transform it into a note. */
- PUT_CODE (insn, NOTE);
- NOTE_LINE_NUMBER (insn) = NOTE_INSN_DELETED;
- NOTE_SOURCE_FILE (insn) = 0;
+ SET_INSN_DELETED (insn);
}
else if (GET_CODE (insn) == JUMP_INSN
/* Don't mess with ADDR_DIFF_VEC */
rtx pattern;
pattern = PATTERN (insn);
+ if (GET_CODE (pattern) == PARALLEL)
+ pattern = XVECEXP (pattern, 0, 0);
switch (GET_CODE (pattern))
{
- case PARALLEL:
- pattern = XVECEXP (pattern, 0, 0);
- break;
-
case SET:
if (GET_CODE (SET_SRC (pattern)) != CALL
&& get_attr_type (insn) != TYPE_SFUNC)
temp = scavenge_reg (&temps);
}
if (temp < 0 && live_regs_mask)
- temp = scavenge_reg (live_regs_mask);
+ {
+ HARD_REG_SET temps;
+
+ COPY_HARD_REG_SET (temps, *live_regs_mask);
+ CLEAR_HARD_REG_BIT (temps, REGNO (reg));
+ temp = scavenge_reg (&temps);
+ }
if (temp < 0)
{
rtx adj_reg, tmp_reg, mem;
adj_reg = gen_rtx_REG (GET_MODE (reg), 4);
tmp_reg = gen_rtx_REG (GET_MODE (reg), 5);
- emit_move_insn (gen_rtx_MEM (Pmode, reg), adj_reg);
+ emit_move_insn (gen_tmp_stack_mem (Pmode, reg), adj_reg);
emit_insn (GEN_MOV (adj_reg, GEN_INT (size)));
emit_insn (GEN_ADD3 (adj_reg, adj_reg, reg));
- mem = gen_rtx_MEM (Pmode, gen_rtx_PRE_DEC (Pmode, adj_reg));
+ mem = gen_tmp_stack_mem (Pmode, gen_rtx_PRE_DEC (Pmode, adj_reg));
emit_move_insn (mem, tmp_reg);
- emit_move_insn (tmp_reg, gen_rtx_MEM (Pmode, reg));
- mem = gen_rtx_MEM (Pmode, gen_rtx_PRE_DEC (Pmode, adj_reg));
- emit_move_insn (mem, tmp_reg);
- emit_move_insn (reg, adj_reg);
- mem = gen_rtx_MEM (Pmode, gen_rtx_POST_INC (Pmode, reg));
- emit_move_insn (adj_reg, mem);
- mem = gen_rtx_MEM (Pmode, gen_rtx_POST_INC (Pmode, reg));
- emit_move_insn (tmp_reg, mem);
- return;
+ emit_move_insn (tmp_reg, gen_tmp_stack_mem (Pmode, reg));
+ mem = gen_tmp_stack_mem (Pmode, gen_rtx_PRE_DEC (Pmode, adj_reg));
+ emit_move_insn (mem, tmp_reg);
+ emit_move_insn (reg, adj_reg);
+ mem = gen_tmp_stack_mem (Pmode, gen_rtx_POST_INC (Pmode, reg));
+ emit_move_insn (adj_reg, mem);
+ mem = gen_tmp_stack_mem (Pmode, gen_rtx_POST_INC (Pmode, reg));
+ emit_move_insn (tmp_reg, mem);
+ /* Tell flow the insns that pop r4/r5 aren't dead. */
+ emit_insn (gen_rtx_USE (VOIDmode, tmp_reg));
+ emit_insn (gen_rtx_USE (VOIDmode, adj_reg));
+ return;
}
const_reg = gen_rtx_REG (GET_MODE (reg), temp);
static void
push_regs (HARD_REG_SET *mask, int interrupt_handler)
{
- int i;
+ int i = interrupt_handler ? LAST_BANKED_REG + 1 : 0;
int skip_fpscr = 0;
/* Push PR last; this gives better latencies after the prologue, and
candidates for the return delay slot when there are no general
registers pushed. */
- for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
+ for (; i < FIRST_PSEUDO_REGISTER; i++)
{
/* If this is an interrupt handler, and the SZ bit varies,
and we have to push any floating point register, we need
to switch to the correct precision first. */
if (i == FIRST_FP_REG && interrupt_handler && TARGET_FMOVD
- && hard_regs_intersect_p (mask, ®_class_contents[DF_REGS]))
+ && hard_reg_set_intersect_p (*mask, reg_class_contents[DF_REGS]))
{
HARD_REG_SET unsaved;
&& TEST_HARD_REG_BIT (*mask, i))
push (i);
}
+
+ /* Push banked registers last to improve delay slot opportunities. */
+ if (interrupt_handler)
+ for (i = FIRST_BANKED_REG; i <= LAST_BANKED_REG; i++)
+ if (TEST_HARD_REG_BIT (*mask, i))
+ push (i);
+
if (TEST_HARD_REG_BIT (*mask, PR_REG))
push (PR_REG);
}
{
unsigned int reg;
int count;
- int interrupt_handler;
+ tree attrs;
+ bool interrupt_or_trapa_handler, trapa_handler, interrupt_handler;
+ bool nosave_low_regs;
int pr_live, has_call;
- interrupt_handler = sh_cfun_interrupt_handler_p ();
+ attrs = DECL_ATTRIBUTES (current_function_decl);
+ interrupt_or_trapa_handler = sh_cfun_interrupt_handler_p ();
+ trapa_handler = lookup_attribute ("trapa_handler", attrs) != NULL_TREE;
+ interrupt_handler = interrupt_or_trapa_handler && ! trapa_handler;
+ nosave_low_regs = lookup_attribute ("nosave_low_regs", attrs) != NULL_TREE;
CLEAR_HARD_REG_SET (*live_regs_mask);
if ((TARGET_SH4 || TARGET_SH2A_DOUBLE) && TARGET_FMOVD && interrupt_handler
- && regs_ever_live[FPSCR_REG])
+ && df_regs_ever_live_p (FPSCR_REG))
target_flags &= ~MASK_FPU_SINGLE;
/* If we can save a lot of saves by switching to double mode, do that. */
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]
+ if (df_regs_ever_live_p (reg) && df_regs_ever_live_p (reg+1)
&& (! call_really_used_regs[reg]
- || (interrupt_handler && ! pragma_trapa))
+ || interrupt_handler)
&& ++count > 2)
{
target_flags &= ~MASK_FPU_SINGLE;
pr_live = (pr_initial
? (GET_CODE (pr_initial) != REG
|| REGNO (pr_initial) != (PR_REG))
- : regs_ever_live[PR_REG]);
+ : df_regs_ever_live_p (PR_REG));
/* For Shcompact, if not optimizing, we end up with a memory reference
using the return address pointer for __builtin_return_address even
though there is no actual need to put the PR register on the stack. */
- pr_live |= regs_ever_live[RETURN_ADDRESS_POINTER_REGNUM];
+ pr_live |= df_regs_ever_live_p (RETURN_ADDRESS_POINTER_REGNUM);
}
/* Force PR to be live if the prologue has to call the SHmedia
argument decoder or register saver. */
{
if (reg == (TARGET_SHMEDIA ? PR_MEDIA_REG : PR_REG)
? pr_live
- : (interrupt_handler && ! pragma_trapa)
+ : interrupt_handler
? (/* Need to save all the regs ever live. */
- (regs_ever_live[reg]
+ (df_regs_ever_live_p (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
+ || (TARGET_SHMEDIA && has_call
+ && REGISTER_NATURAL_MODE (reg) == SImode
&& (GENERAL_REGISTER_P (reg) || TARGET_REGISTER_P (reg))))
&& reg != STACK_POINTER_REGNUM && reg != ARG_POINTER_REGNUM
&& reg != RETURN_ADDRESS_POINTER_REGNUM
&& flag_pic
&& current_function_args_info.call_cookie
&& reg == PIC_OFFSET_TABLE_REGNUM)
- || (regs_ever_live[reg] && ! call_really_used_regs[reg])
+ || (df_regs_ever_live_p (reg)
+ && (!call_really_used_regs[reg]
+ || (trapa_handler && reg == FPSCR_REG && TARGET_FPU_ANY)))
|| (current_function_calls_eh_return
&& (reg == EH_RETURN_DATA_REGNO (0)
|| reg == EH_RETURN_DATA_REGNO (1)
|| reg == EH_RETURN_DATA_REGNO (2)
|| reg == EH_RETURN_DATA_REGNO (3)))
|| ((reg == MACL_REG || reg == MACH_REG)
- && regs_ever_live[reg]
+ && df_regs_ever_live_p (reg)
&& sh_cfun_attr_renesas_p ())
))
{
{
if (FP_REGISTER_P (reg))
{
- if (! TARGET_FPU_SINGLE && ! regs_ever_live[reg ^ 1])
+ if (! TARGET_FPU_SINGLE && ! df_regs_ever_live_p (reg ^ 1))
{
SET_HARD_REG_BIT (*live_regs_mask, (reg ^ 1));
count += GET_MODE_SIZE (REGISTER_NATURAL_MODE (reg ^ 1));
}
}
}
+ if (nosave_low_regs && reg == R8_REG)
+ break;
}
/* If we have a target register optimization pass after prologue / epilogue
threading, we need to assume all target registers will be live even if
Make sure we save at least one general purpose register when we need
to save target registers. */
if (interrupt_handler
- && hard_regs_intersect_p (live_regs_mask,
- ®_class_contents[TARGET_REGS])
- && ! hard_regs_intersect_p (live_regs_mask,
- ®_class_contents[GENERAL_REGS]))
+ && hard_reg_set_intersect_p (*live_regs_mask,
+ reg_class_contents[TARGET_REGS])
+ && ! hard_reg_set_intersect_p (*live_regs_mask,
+ reg_class_contents[GENERAL_REGS]))
{
SET_HARD_REG_BIT (*live_regs_mask, R0_REG);
count += GET_MODE_SIZE (REGISTER_NATURAL_MODE (R0_REG));
if (sh_cfun_interrupt_handler_p ())
return -1;
- tr0_used = flag_pic && regs_ever_live[PIC_OFFSET_TABLE_REGNUM];
+ tr0_used = flag_pic && df_regs_ever_live_p (PIC_OFFSET_TABLE_REGNUM);
for (regno = FIRST_TARGET_REG + tr0_used; regno <= LAST_TARGET_REG; regno++)
- if (call_really_used_regs[regno] && ! regs_ever_live[regno])
+ if (call_really_used_regs[regno] && ! df_regs_ever_live_p (regno))
return regno;
return -1;
int d_rounding = 0;
int save_flags = target_flags;
int pretend_args;
+ tree sp_switch_attr
+ = lookup_attribute ("sp_switch", DECL_ATTRIBUTES (current_function_decl));
current_function_interrupt = sh_cfun_interrupt_handler_p ();
incoming-argument decoder and/or of the return trampoline from
the GOT, so make sure the PIC register is preserved and
initialized. */
- regs_ever_live[PIC_OFFSET_TABLE_REGNUM] = 1;
+ df_set_regs_ever_live (PIC_OFFSET_TABLE_REGNUM, true);
if (TARGET_SHCOMPACT
&& (current_function_args_info.call_cookie & ~ CALL_COOKIE_RET_TRAMP(1)))
int tr = sh_media_register_for_return ();
if (tr >= 0)
- {
- rtx insn = emit_move_insn (gen_rtx_REG (DImode, tr),
- gen_rtx_REG (DImode, PR_MEDIA_REG));
-
- /* ??? We should suppress saving pr when we don't need it, but this
- is tricky because of builtin_return_address. */
-
- /* If this function only exits with sibcalls, this copy
- will be flagged as dead. */
- REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_MAYBE_DEAD,
- const0_rtx,
- REG_NOTES (insn));
- }
+ emit_move_insn (gen_rtx_REG (DImode, tr),
+ gen_rtx_REG (DImode, PR_MEDIA_REG));
}
/* Emit the code for SETUP_VARARGS. */
}
/* If we're supposed to switch stacks at function entry, do so now. */
- if (sp_switch)
- emit_insn (gen_sp_switch_1 ());
+ if (sp_switch_attr)
+ {
+ /* The argument specifies a variable holding the address of the
+ stack the interrupt function should switch to/from at entry/exit. */
+ const char *s
+ = ggc_strdup (TREE_STRING_POINTER (TREE_VALUE (sp_switch_attr)));
+ rtx sp_switch = gen_rtx_SYMBOL_REF (Pmode, s);
+
+ emit_insn (gen_sp_switch_1 (sp_switch));
+ }
d = calc_live_regs (&live_regs_mask);
/* ??? Maybe we could save some switching if we can move a mode switch
reg_rtx = gen_rtx_REG (mode, reg);
- mem_rtx = gen_rtx_MEM (mode,
- gen_rtx_PLUS (Pmode,
- stack_pointer_rtx,
- GEN_INT (offset)));
+ mem_rtx = gen_frame_mem (mode,
+ gen_rtx_PLUS (Pmode,
+ stack_pointer_rtx,
+ GEN_INT (offset)));
GO_IF_LEGITIMATE_ADDRESS (mode, XEXP (mem_rtx, 0), try_pre_dec);
|| mem_rtx == NULL_RTX
|| reg == PR_REG || SPECIAL_REGISTER_P (reg)))
{
- pre_dec = gen_rtx_MEM (mode,
- gen_rtx_PRE_DEC (Pmode, r0));
+ pre_dec = gen_frame_mem (mode, gen_rtx_PRE_DEC (Pmode, r0));
GO_IF_LEGITIMATE_ADDRESS (mode, XEXP (pre_dec, 0),
pre_dec_ok);
mem_rtx = pre_dec;
}
else if (sp_in_r0)
- mem_rtx = gen_rtx_MEM (mode, r0);
+ mem_rtx = gen_frame_mem (mode, r0);
else
- mem_rtx = gen_rtx_MEM (mode,
- gen_rtx_PLUS (Pmode,
- stack_pointer_rtx,
- r0));
+ mem_rtx = gen_frame_mem (mode,
+ gen_rtx_PLUS (Pmode,
+ stack_pointer_rtx,
+ r0));
/* We must not use an r0-based address for target-branch
registers or for special registers without pre-dec
{
rtx reg_rtx = gen_rtx_REG (mode, reg);
rtx set, note_rtx;
- rtx mem_rtx = gen_rtx_MEM (mode,
- gen_rtx_PLUS (Pmode,
- stack_pointer_rtx,
- GEN_INT (offset)));
+ rtx mem_rtx = gen_frame_mem (mode,
+ gen_rtx_PLUS (Pmode,
+ stack_pointer_rtx,
+ GEN_INT (offset)));
set = gen_rtx_SET (VOIDmode, mem_rtx, reg_rtx);
note_rtx = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR, set,
else
push_regs (&live_regs_mask, current_function_interrupt);
- if (flag_pic && regs_ever_live[PIC_OFFSET_TABLE_REGNUM])
- {
- rtx insn = get_last_insn ();
- rtx last = emit_insn (gen_GOTaddr2picreg ());
-
- /* Mark these insns as possibly dead. Sometimes, flow2 may
- delete all uses of the PIC register. In this case, let it
- delete the initialization too. */
- do
- {
- insn = NEXT_INSN (insn);
-
- REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_MAYBE_DEAD,
- const0_rtx,
- REG_NOTES (insn));
- }
- while (insn != last);
- }
+ if (flag_pic && df_regs_ever_live_p (PIC_OFFSET_TABLE_REGNUM))
+ emit_insn (gen_GOTaddr2picreg ());
if (SHMEDIA_REGS_STACK_ADJUST ())
{
}
if (target_flags != save_flags && ! current_function_interrupt)
- {
- rtx insn = emit_insn (gen_toggle_sz ());
-
- /* If we're lucky, a mode switch in the function body will
- overwrite fpscr, turning this insn dead. Tell flow this
- insn is ok to delete. */
- REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_MAYBE_DEAD,
- const0_rtx,
- REG_NOTES (insn));
- }
+ emit_insn (gen_toggle_sz ());
target_flags = save_flags;
stack_pointer_rtx, 0, NULL);
if (frame_pointer_needed)
- frame_insn (GEN_MOV (frame_pointer_rtx, stack_pointer_rtx));
+ frame_insn (GEN_MOV (hard_frame_pointer_rtx, stack_pointer_rtx));
if (TARGET_SHCOMPACT
&& (current_function_args_info.call_cookie & ~ CALL_COOKIE_RET_TRAMP(1)))
when exception handling is enabled. See PR/18032. */
if (flag_exceptions)
emit_insn (gen_blockage ());
- output_stack_adjust (frame_size, frame_pointer_rtx, e, &live_regs_mask);
+ output_stack_adjust (frame_size, hard_frame_pointer_rtx, e,
+ &live_regs_mask);
/* We must avoid moving the stack pointer adjustment past code
which reads from the local frame, else an interrupt could
occur after the SP adjustment and clobber data in the local
frame. */
emit_insn (gen_blockage ());
- emit_insn (GEN_MOV (stack_pointer_rtx, frame_pointer_rtx));
+ emit_insn (GEN_MOV (stack_pointer_rtx, hard_frame_pointer_rtx));
}
else if (frame_size)
{
offset = offset_base + entry->offset;
reg_rtx = gen_rtx_REG (mode, reg);
- mem_rtx = gen_rtx_MEM (mode,
- gen_rtx_PLUS (Pmode,
- stack_pointer_rtx,
- GEN_INT (offset)));
+ mem_rtx = gen_frame_mem (mode,
+ gen_rtx_PLUS (Pmode,
+ stack_pointer_rtx,
+ GEN_INT (offset)));
GO_IF_LEGITIMATE_ADDRESS (mode, XEXP (mem_rtx, 0), try_post_inc);
&& mem_rtx == NULL_RTX)
|| reg == PR_REG || SPECIAL_REGISTER_P (reg)))
{
- post_inc = gen_rtx_MEM (mode,
- gen_rtx_POST_INC (Pmode, r0));
+ post_inc = gen_frame_mem (mode, gen_rtx_POST_INC (Pmode, r0));
GO_IF_LEGITIMATE_ADDRESS (mode, XEXP (post_inc, 0),
post_inc_ok);
offset_in_r0 += GET_MODE_SIZE (mode);
}
else if (sp_in_r0)
- mem_rtx = gen_rtx_MEM (mode, r0);
+ mem_rtx = gen_frame_mem (mode, r0);
else
- mem_rtx = gen_rtx_MEM (mode,
- gen_rtx_PLUS (Pmode,
- stack_pointer_rtx,
- r0));
+ mem_rtx = gen_frame_mem (mode,
+ gen_rtx_PLUS (Pmode,
+ stack_pointer_rtx,
+ r0));
gcc_assert ((reg != PR_REG && !SPECIAL_REGISTER_P (reg))
|| mem_rtx == post_inc);
}
insn = emit_move_insn (reg_rtx, mem_rtx);
- if (reg == PR_MEDIA_REG && sh_media_register_for_return () >= 0)
- /* This is dead, unless we return with a sibcall. */
- REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_MAYBE_DEAD,
- const0_rtx,
- REG_NOTES (insn));
}
gcc_assert (entry->offset + offset_base == d + d_rounding);
}
else /* ! TARGET_SH5 */
{
+ int last_reg;
+
save_size = 0;
if (TEST_HARD_REG_BIT (live_regs_mask, PR_REG))
- pop (PR_REG);
- for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
+ {
+ if (!frame_pointer_needed)
+ emit_insn (gen_blockage ());
+ pop (PR_REG);
+ }
+
+ /* Banked registers are poped first to avoid being scheduled in the
+ delay slot. RTE switches banks before the ds instruction. */
+ if (current_function_interrupt)
+ {
+ for (i = FIRST_BANKED_REG; i <= LAST_BANKED_REG; i++)
+ if (TEST_HARD_REG_BIT (live_regs_mask, i))
+ pop (LAST_BANKED_REG - i);
+
+ last_reg = FIRST_PSEUDO_REGISTER - LAST_BANKED_REG - 1;
+ }
+ else
+ last_reg = FIRST_PSEUDO_REGISTER;
+
+ for (i = 0; i < last_reg; 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]))
+ && hard_reg_set_intersect_p (live_regs_mask,
+ reg_class_contents[DF_REGS]))
fpscr_deferred = 1;
else if (j != PR_REG && TEST_HARD_REG_BIT (live_regs_mask, j))
pop (j);
+
if (j == FIRST_FP_REG && fpscr_deferred)
pop (FPSCR_REG);
-
}
}
if (target_flags != save_flags && ! current_function_interrupt)
EH_RETURN_STACKADJ_RTX));
/* Switch back to the normal stack if necessary. */
- if (sp_switch)
+ if (lookup_attribute ("sp_switch", DECL_ATTRIBUTES (current_function_decl)))
emit_insn (gen_sp_switch_2 ());
/* Tell flow the insn that pops PR isn't dead. */
pr_offset = rounded_frame_size (d);
emit_insn (GEN_MOV (tmp, GEN_INT (pr_offset)));
- emit_insn (GEN_ADD3 (tmp, tmp, frame_pointer_rtx));
+ emit_insn (GEN_ADD3 (tmp, tmp, hard_frame_pointer_rtx));
- tmp = gen_rtx_MEM (Pmode, tmp);
+ tmp = gen_frame_mem (Pmode, tmp);
emit_insn (GEN_MOV (tmp, ra));
}
sh_output_function_epilogue (FILE *file ATTRIBUTE_UNUSED,
HOST_WIDE_INT size ATTRIBUTE_UNUSED)
{
- trap_exit = pragma_interrupt = pragma_trapa = pragma_nosave_low_regs = 0;
sh_need_epilogue_known = 0;
- sp_switch = NULL_RTX;
}
static rtx
bufsize = (n_intregs * UNITS_PER_WORD) + (n_floatregs * UNITS_PER_WORD);
if (TARGET_SHMEDIA)
- regbuf = gen_rtx_MEM (BLKmode,
- gen_rtx_REG (Pmode, ARG_POINTER_REGNUM));
+ regbuf = gen_frame_mem (BLKmode, gen_rtx_REG (Pmode, ARG_POINTER_REGNUM));
else if (n_floatregs & 1)
{
rtx addr;
emit_insn (gen_iorsi3 (addr, addr, GEN_INT (UNITS_PER_WORD)));
regbuf = change_address (regbuf, BLKmode, addr);
}
+ else if (STACK_BOUNDARY < 64 && TARGET_FPU_DOUBLE && n_floatregs)
+ {
+ rtx addr, mask;
+
+ regbuf = assign_stack_local (BLKmode, bufsize + UNITS_PER_WORD, 0);
+ addr = copy_to_mode_reg (Pmode, plus_constant (XEXP (regbuf, 0), 4));
+ mask = copy_to_mode_reg (Pmode, GEN_INT (-8));
+ emit_insn (gen_andsi3 (addr, addr, mask));
+ regbuf = change_address (regbuf, BLKmode, addr);
+ }
else
- regbuf = assign_stack_local (BLKmode, bufsize, 0);
+ regbuf = assign_stack_local (BLKmode, bufsize, TARGET_FPU_DOUBLE ? 64 : 0);
alias_set = get_varargs_alias_set ();
set_mem_alias_set (regbuf, alias_set);
{
emit_insn (gen_addsi3 (fpregs, fpregs,
GEN_INT (-2 * UNITS_PER_WORD)));
- mem = gen_rtx_MEM (DFmode, fpregs);
- set_mem_alias_set (mem, alias_set);
+ mem = change_address (regbuf, DFmode, fpregs);
emit_move_insn (mem,
gen_rtx_REG (DFmode, BASE_ARG_REG (DFmode) + regno));
}
if (regno & 1)
{
emit_insn (gen_addsi3 (fpregs, fpregs, GEN_INT (-UNITS_PER_WORD)));
- mem = gen_rtx_MEM (SFmode, fpregs);
- set_mem_alias_set (mem, alias_set);
+ mem = change_address (regbuf, SFmode, fpregs);
emit_move_insn (mem,
gen_rtx_REG (SFmode, BASE_ARG_REG (SFmode) + regno
- (TARGET_LITTLE_ENDIAN != 0)));
rtx mem;
emit_insn (gen_addsi3 (fpregs, fpregs, GEN_INT (-UNITS_PER_WORD)));
- mem = gen_rtx_MEM (SFmode, fpregs);
- set_mem_alias_set (mem, alias_set);
+ mem = change_address (regbuf, SFmode, fpregs);
emit_move_insn (mem,
gen_rtx_REG (SFmode, BASE_ARG_REG (SFmode) + regno));
}
f_next_fp_limit = TREE_CHAIN (f_next_fp);
f_next_stack = TREE_CHAIN (f_next_fp_limit);
- next_o = build (COMPONENT_REF, TREE_TYPE (f_next_o), valist, f_next_o,
- NULL_TREE);
- next_o_limit = build (COMPONENT_REF, TREE_TYPE (f_next_o_limit),
- valist, f_next_o_limit, NULL_TREE);
- next_fp = build (COMPONENT_REF, TREE_TYPE (f_next_fp), valist, f_next_fp,
+ next_o = build3 (COMPONENT_REF, TREE_TYPE (f_next_o), valist, f_next_o,
NULL_TREE);
- next_fp_limit = build (COMPONENT_REF, TREE_TYPE (f_next_fp_limit),
- valist, f_next_fp_limit, NULL_TREE);
- next_stack = build (COMPONENT_REF, TREE_TYPE (f_next_stack),
- valist, f_next_stack, NULL_TREE);
+ next_o_limit = build3 (COMPONENT_REF, TREE_TYPE (f_next_o_limit),
+ valist, f_next_o_limit, NULL_TREE);
+ next_fp = build3 (COMPONENT_REF, TREE_TYPE (f_next_fp), valist, f_next_fp,
+ NULL_TREE);
+ next_fp_limit = build3 (COMPONENT_REF, TREE_TYPE (f_next_fp_limit),
+ valist, f_next_fp_limit, NULL_TREE);
+ next_stack = build3 (COMPONENT_REF, TREE_TYPE (f_next_stack),
+ valist, f_next_stack, NULL_TREE);
/* Call __builtin_saveregs. */
- u = make_tree (ptr_type_node, expand_builtin_saveregs ());
- t = build (MODIFY_EXPR, ptr_type_node, next_fp, u);
+ u = make_tree (sizetype, expand_builtin_saveregs ());
+ u = fold_convert (ptr_type_node, u);
+ t = build2 (GIMPLE_MODIFY_STMT, ptr_type_node, next_fp, u);
TREE_SIDE_EFFECTS (t) = 1;
expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL);
nfp = 8 - nfp;
else
nfp = 0;
- u = fold (build (PLUS_EXPR, ptr_type_node, u,
- build_int_cst (NULL_TREE, UNITS_PER_WORD * nfp)));
- t = build (MODIFY_EXPR, ptr_type_node, next_fp_limit, u);
+ u = fold_build2 (POINTER_PLUS_EXPR, ptr_type_node, u,
+ size_int (UNITS_PER_WORD * nfp));
+ t = build2 (GIMPLE_MODIFY_STMT, ptr_type_node, next_fp_limit, u);
TREE_SIDE_EFFECTS (t) = 1;
expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL);
- t = build (MODIFY_EXPR, ptr_type_node, next_o, u);
+ t = build2 (GIMPLE_MODIFY_STMT, ptr_type_node, next_o, u);
TREE_SIDE_EFFECTS (t) = 1;
expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL);
nint = 4 - nint;
else
nint = 0;
- u = fold (build (PLUS_EXPR, ptr_type_node, u,
- build_int_cst (NULL_TREE, UNITS_PER_WORD * nint)));
- t = build (MODIFY_EXPR, ptr_type_node, next_o_limit, u);
+ u = fold_build2 (POINTER_PLUS_EXPR, ptr_type_node, u,
+ size_int (UNITS_PER_WORD * nint));
+ t = build2 (GIMPLE_MODIFY_STMT, ptr_type_node, next_o_limit, u);
TREE_SIDE_EFFECTS (t) = 1;
expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL);
u = make_tree (ptr_type_node, nextarg);
- t = build (MODIFY_EXPR, ptr_type_node, next_stack, u);
+ t = build2 (GIMPLE_MODIFY_STMT, ptr_type_node, next_stack, u);
TREE_SIDE_EFFECTS (t) = 1;
expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL);
}
+/* TYPE is a RECORD_TYPE. If there is only a single nonzero-sized
+ member, return it. */
+static tree
+find_sole_member (tree type)
+{
+ tree field, member = NULL_TREE;
+
+ for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
+ {
+ if (TREE_CODE (field) != FIELD_DECL)
+ continue;
+ if (!DECL_SIZE (field))
+ return NULL_TREE;
+ if (integer_zerop (DECL_SIZE (field)))
+ continue;
+ if (member)
+ return NULL_TREE;
+ member = field;
+ }
+ return member;
+}
/* Implement `va_arg'. */
static tree
tree tmp, pptr_type_node;
tree addr, lab_over = NULL, result = NULL;
int pass_by_ref = targetm.calls.must_pass_in_stack (TYPE_MODE (type), type);
+ tree eff_type;
if (pass_by_ref)
type = build_pointer_type (type);
tree next_o, next_o_limit, next_fp, next_fp_limit, next_stack;
int pass_as_float;
tree lab_false;
+ tree member;
f_next_o = TYPE_FIELDS (va_list_type_node);
f_next_o_limit = TREE_CHAIN (f_next_o);
f_next_fp_limit = TREE_CHAIN (f_next_fp);
f_next_stack = TREE_CHAIN (f_next_fp_limit);
- next_o = build (COMPONENT_REF, TREE_TYPE (f_next_o), valist, f_next_o,
- NULL_TREE);
- next_o_limit = build (COMPONENT_REF, TREE_TYPE (f_next_o_limit),
- valist, f_next_o_limit, NULL_TREE);
- next_fp = build (COMPONENT_REF, TREE_TYPE (f_next_fp),
- valist, f_next_fp, NULL_TREE);
- next_fp_limit = build (COMPONENT_REF, TREE_TYPE (f_next_fp_limit),
- valist, f_next_fp_limit, NULL_TREE);
- next_stack = build (COMPONENT_REF, TREE_TYPE (f_next_stack),
- valist, f_next_stack, NULL_TREE);
+ next_o = build3 (COMPONENT_REF, TREE_TYPE (f_next_o), valist, f_next_o,
+ NULL_TREE);
+ next_o_limit = build3 (COMPONENT_REF, TREE_TYPE (f_next_o_limit),
+ valist, f_next_o_limit, NULL_TREE);
+ next_fp = build3 (COMPONENT_REF, TREE_TYPE (f_next_fp),
+ valist, f_next_fp, NULL_TREE);
+ next_fp_limit = build3 (COMPONENT_REF, TREE_TYPE (f_next_fp_limit),
+ valist, f_next_fp_limit, NULL_TREE);
+ next_stack = build3 (COMPONENT_REF, TREE_TYPE (f_next_stack),
+ valist, f_next_stack, NULL_TREE);
/* Structures with a single member with a distinct mode are passed
like their member. This is relevant if the latter has a REAL_TYPE
or COMPLEX_TYPE type. */
- if (TREE_CODE (type) == RECORD_TYPE
- && TYPE_FIELDS (type)
- && TREE_CODE (TYPE_FIELDS (type)) == FIELD_DECL
- && (TREE_CODE (TREE_TYPE (TYPE_FIELDS (type))) == REAL_TYPE
- || TREE_CODE (TREE_TYPE (TYPE_FIELDS (type))) == COMPLEX_TYPE)
- && TREE_CHAIN (TYPE_FIELDS (type)) == NULL_TREE)
- type = TREE_TYPE (TYPE_FIELDS (type));
+ eff_type = type;
+ while (TREE_CODE (eff_type) == RECORD_TYPE
+ && (member = find_sole_member (eff_type))
+ && (TREE_CODE (TREE_TYPE (member)) == REAL_TYPE
+ || TREE_CODE (TREE_TYPE (member)) == COMPLEX_TYPE
+ || TREE_CODE (TREE_TYPE (member)) == RECORD_TYPE))
+ {
+ tree field_type = TREE_TYPE (member);
+
+ if (TYPE_MODE (eff_type) == TYPE_MODE (field_type))
+ eff_type = field_type;
+ else
+ {
+ gcc_assert ((TYPE_ALIGN (eff_type)
+ < GET_MODE_ALIGNMENT (TYPE_MODE (field_type)))
+ || (TYPE_ALIGN (eff_type)
+ > GET_MODE_BITSIZE (TYPE_MODE (field_type))));
+ break;
+ }
+ }
if (TARGET_SH4)
{
- pass_as_float = ((TREE_CODE (type) == REAL_TYPE && size <= 8)
- || (TREE_CODE (type) == COMPLEX_TYPE
- && TREE_CODE (TREE_TYPE (type)) == REAL_TYPE
+ pass_as_float = ((TREE_CODE (eff_type) == REAL_TYPE && size <= 8)
+ || (TREE_CODE (eff_type) == COMPLEX_TYPE
+ && TREE_CODE (TREE_TYPE (eff_type)) == REAL_TYPE
&& size <= 16));
}
else
{
- pass_as_float = (TREE_CODE (type) == REAL_TYPE && size == 4);
+ pass_as_float = (TREE_CODE (eff_type) == REAL_TYPE && size == 4);
}
addr = create_tmp_var (pptr_type_node, NULL);
if (pass_as_float)
{
- int first_floatreg
- = current_function_args_info.arg_count[(int) SH_ARG_FLOAT];
- int n_floatregs = MAX (0, NPARM_REGS (SFmode) - first_floatreg);
+ tree next_fp_tmp = create_tmp_var (TREE_TYPE (f_next_fp), NULL);
+ tree cmp;
+ bool is_double = size == 8 && TREE_CODE (eff_type) == REAL_TYPE;
- tmp = build (GE_EXPR, boolean_type_node, next_fp, next_fp_limit);
- tmp = build (COND_EXPR, void_type_node, tmp,
- build (GOTO_EXPR, void_type_node, lab_false),
- NULL);
+ tmp = build1 (ADDR_EXPR, pptr_type_node, next_fp);
+ tmp = build2 (GIMPLE_MODIFY_STMT, void_type_node, addr, tmp);
gimplify_and_add (tmp, pre_p);
- if (TYPE_ALIGN (type) > BITS_PER_WORD
- || (((TREE_CODE (type) == REAL_TYPE && size == 8) || size == 16)
- && (n_floatregs & 1)))
+ tmp = build2 (GIMPLE_MODIFY_STMT, ptr_type_node, next_fp_tmp, valist);
+ gimplify_and_add (tmp, pre_p);
+ tmp = next_fp_limit;
+ if (size > 4 && !is_double)
+ tmp = build2 (POINTER_PLUS_EXPR, TREE_TYPE (tmp), tmp,
+ size_int (4 - size));
+ tmp = build2 (GE_EXPR, boolean_type_node, next_fp_tmp, tmp);
+ cmp = build3 (COND_EXPR, void_type_node, tmp,
+ build1 (GOTO_EXPR, void_type_node, lab_false),
+ NULL_TREE);
+ if (!is_double)
+ gimplify_and_add (cmp, pre_p);
+
+ if (TYPE_ALIGN (eff_type) > BITS_PER_WORD
+ || (is_double || size == 16))
{
- tmp = fold_convert (ptr_type_node, size_int (UNITS_PER_WORD));
- tmp = build (BIT_AND_EXPR, ptr_type_node, next_fp, tmp);
- tmp = build (PLUS_EXPR, ptr_type_node, next_fp, tmp);
- tmp = build (MODIFY_EXPR, ptr_type_node, next_fp, tmp);
+ tmp = fold_convert (sizetype, next_fp_tmp);
+ tmp = build2 (BIT_AND_EXPR, sizetype, tmp,
+ size_int (UNITS_PER_WORD));
+ tmp = build2 (POINTER_PLUS_EXPR, ptr_type_node,
+ next_fp_tmp, tmp);
+ tmp = build2 (GIMPLE_MODIFY_STMT, ptr_type_node,
+ next_fp_tmp, tmp);
gimplify_and_add (tmp, pre_p);
}
-
- tmp = build1 (ADDR_EXPR, pptr_type_node, next_fp);
- tmp = build (MODIFY_EXPR, void_type_node, addr, tmp);
- gimplify_and_add (tmp, pre_p);
+ if (is_double)
+ gimplify_and_add (cmp, pre_p);
#ifdef FUNCTION_ARG_SCmode_WART
- if (TYPE_MODE (type) == SCmode && TARGET_SH4 && TARGET_LITTLE_ENDIAN)
+ if (TYPE_MODE (eff_type) == SCmode
+ && TARGET_SH4 && TARGET_LITTLE_ENDIAN)
{
- tree subtype = TREE_TYPE (type);
+ tree subtype = TREE_TYPE (eff_type);
tree real, imag;
- imag = std_gimplify_va_arg_expr (valist, subtype, pre_p, NULL);
+ imag
+ = std_gimplify_va_arg_expr (next_fp_tmp, subtype, pre_p, NULL);
imag = get_initialized_tmp_var (imag, pre_p, NULL);
- real = std_gimplify_va_arg_expr (valist, subtype, pre_p, NULL);
+ real
+ = std_gimplify_va_arg_expr (next_fp_tmp, subtype, pre_p, NULL);
real = get_initialized_tmp_var (real, pre_p, NULL);
- result = build (COMPLEX_EXPR, type, real, imag);
+ result = build2 (COMPLEX_EXPR, type, real, imag);
result = get_initialized_tmp_var (result, pre_p, NULL);
}
#endif /* FUNCTION_ARG_SCmode_WART */
- tmp = build (GOTO_EXPR, void_type_node, lab_over);
+ tmp = build1 (GOTO_EXPR, void_type_node, lab_over);
gimplify_and_add (tmp, pre_p);
- tmp = build (LABEL_EXPR, void_type_node, lab_false);
+ tmp = build1 (LABEL_EXPR, void_type_node, lab_false);
gimplify_and_add (tmp, pre_p);
tmp = build1 (ADDR_EXPR, pptr_type_node, next_stack);
- tmp = build (MODIFY_EXPR, void_type_node, addr, tmp);
+ tmp = build2 (GIMPLE_MODIFY_STMT, void_type_node, addr, tmp);
+ gimplify_and_add (tmp, pre_p);
+ tmp = build2 (GIMPLE_MODIFY_STMT, ptr_type_node, next_fp_tmp, valist);
gimplify_and_add (tmp, pre_p);
+
+ tmp = build2 (GIMPLE_MODIFY_STMT, ptr_type_node, valist, next_fp_tmp);
+ gimplify_and_add (tmp, post_p);
+ valist = next_fp_tmp;
}
else
{
- tmp = fold_convert (ptr_type_node, size_int (rsize));
- tmp = build (PLUS_EXPR, ptr_type_node, next_o, tmp);
- tmp = build (GT_EXPR, boolean_type_node, tmp, next_o_limit);
- tmp = build (COND_EXPR, void_type_node, tmp,
- build (GOTO_EXPR, void_type_node, lab_false),
- NULL);
+ tmp = build2 (POINTER_PLUS_EXPR, ptr_type_node, next_o,
+ size_int (rsize));
+ tmp = build2 (GT_EXPR, boolean_type_node, tmp, next_o_limit);
+ tmp = build3 (COND_EXPR, void_type_node, tmp,
+ build1 (GOTO_EXPR, void_type_node, lab_false),
+ NULL_TREE);
gimplify_and_add (tmp, pre_p);
tmp = build1 (ADDR_EXPR, pptr_type_node, next_o);
- tmp = build (MODIFY_EXPR, void_type_node, addr, tmp);
+ tmp = build2 (GIMPLE_MODIFY_STMT, void_type_node, addr, tmp);
gimplify_and_add (tmp, pre_p);
- tmp = build (GOTO_EXPR, void_type_node, lab_over);
+ tmp = build1 (GOTO_EXPR, void_type_node, lab_over);
gimplify_and_add (tmp, pre_p);
- tmp = build (LABEL_EXPR, void_type_node, lab_false);
+ tmp = build1 (LABEL_EXPR, void_type_node, lab_false);
gimplify_and_add (tmp, pre_p);
if (size > 4 && ! TARGET_SH4)
{
- tmp = build (MODIFY_EXPR, ptr_type_node, next_o, next_o_limit);
+ tmp = build2 (GIMPLE_MODIFY_STMT, ptr_type_node,
+ next_o, next_o_limit);
gimplify_and_add (tmp, pre_p);
}
tmp = build1 (ADDR_EXPR, pptr_type_node, next_stack);
- tmp = build (MODIFY_EXPR, void_type_node, addr, tmp);
+ tmp = build2 (GIMPLE_MODIFY_STMT, void_type_node, addr, tmp);
gimplify_and_add (tmp, pre_p);
}
if (!result)
{
- tmp = build (LABEL_EXPR, void_type_node, lab_over);
+ tmp = build1 (LABEL_EXPR, void_type_node, lab_over);
gimplify_and_add (tmp, pre_p);
}
}
tmp = std_gimplify_va_arg_expr (valist, type, pre_p, NULL);
if (result)
{
- tmp = build (MODIFY_EXPR, void_type_node, result, tmp);
+ tmp = build2 (GIMPLE_MODIFY_STMT, void_type_node, result, tmp);
gimplify_and_add (tmp, pre_p);
- tmp = build (LABEL_EXPR, void_type_node, lab_over);
+ tmp = build1 (LABEL_EXPR, void_type_node, lab_over);
gimplify_and_add (tmp, pre_p);
}
else
result = tmp;
if (pass_by_ref)
- result = build_fold_indirect_ref (result);
+ result = build_va_arg_indirect_ref (result);
return result;
}
total_saved_regs_space = regs_saved + regs_saved_rounding;
- if (from == ARG_POINTER_REGNUM && to == FRAME_POINTER_REGNUM)
+ if (from == ARG_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM)
return total_saved_regs_space + total_auto_space
+ current_function_args_info.byref_regs * 8;
+ current_function_args_info.byref_regs * 8;
/* Initial gap between fp and sp is 0. */
- if (from == FRAME_POINTER_REGNUM && to == STACK_POINTER_REGNUM)
+ if (from == HARD_FRAME_POINTER_REGNUM && to == STACK_POINTER_REGNUM)
return 0;
+ if (from == FRAME_POINTER_REGNUM && to == STACK_POINTER_REGNUM)
+ return rounded_frame_size (0);
+
+ if (from == FRAME_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM)
+ return rounded_frame_size (0);
+
gcc_assert (from == RETURN_ADDRESS_POINTER_REGNUM
- && (to == FRAME_POINTER_REGNUM || to == STACK_POINTER_REGNUM));
+ && (to == HARD_FRAME_POINTER_REGNUM
+ || to == STACK_POINTER_REGNUM));
if (TARGET_SH5)
{
int n = total_saved_regs_space;
return total_auto_space;
}
\f
-/* Handle machine specific pragmas to be semi-compatible with Renesas
- compiler. */
-
-void
-sh_pr_interrupt (struct cpp_reader *pfile ATTRIBUTE_UNUSED)
-{
- pragma_interrupt = 1;
-}
-
-void
-sh_pr_trapa (struct cpp_reader *pfile ATTRIBUTE_UNUSED)
-{
- pragma_interrupt = pragma_trapa = 1;
-}
-
-void
-sh_pr_nosave_low_regs (struct cpp_reader *pfile ATTRIBUTE_UNUSED)
-{
- pragma_nosave_low_regs = 1;
-}
-
-/* Generate 'handle_interrupt' attribute for decls */
-
+/* Insert any deferred function attributes from earlier pragmas. */
static void
sh_insert_attributes (tree node, tree *attributes)
{
- if (! pragma_interrupt
- || TREE_CODE (node) != FUNCTION_DECL)
+ tree attrs;
+
+ if (TREE_CODE (node) != FUNCTION_DECL)
return;
/* We are only interested in fields. */
if (!DECL_P (node))
return;
- /* Add a 'handle_interrupt' attribute. */
- * attributes = tree_cons (get_identifier ("interrupt_handler"), NULL, * attributes);
+ /* Append the attributes to the deferred attributes. */
+ *sh_deferred_function_attributes_tail = *attributes;
+ attrs = sh_deferred_function_attributes;
+ if (!attrs)
+ return;
+
+ /* Some attributes imply or require the interrupt attribute. */
+ if (!lookup_attribute ("interrupt_handler", attrs)
+ && !lookup_attribute ("interrupt_handler", DECL_ATTRIBUTES (node)))
+ {
+ /* If we have a trapa_handler, but no interrupt_handler attribute,
+ insert an interrupt_handler attribute. */
+ if (lookup_attribute ("trapa_handler", attrs) != NULL_TREE)
+ /* We can't use sh_pr_interrupt here because that's not in the
+ java frontend. */
+ attrs
+ = tree_cons (get_identifier("interrupt_handler"), NULL_TREE, attrs);
+ /* However, for sp_switch, trap_exit and nosave_low_regs, if the
+ interrupt attribute is missing, we ignore the attribute and warn. */
+ else if (lookup_attribute ("sp_switch", attrs)
+ || lookup_attribute ("trap_exit", attrs)
+ || lookup_attribute ("nosave_low_regs", attrs))
+ {
+ tree *tail;
+
+ for (tail = attributes; attrs; attrs = TREE_CHAIN (attrs))
+ {
+ if (is_attribute_p ("sp_switch", TREE_PURPOSE (attrs))
+ || is_attribute_p ("trap_exit", TREE_PURPOSE (attrs))
+ || is_attribute_p ("nosave_low_regs", TREE_PURPOSE (attrs)))
+ warning (OPT_Wattributes,
+ "%qs attribute only applies to interrupt functions",
+ IDENTIFIER_POINTER (TREE_PURPOSE (attrs)));
+ else
+ {
+ *tail = tree_cons (TREE_PURPOSE (attrs), NULL_TREE,
+ NULL_TREE);
+ tail = &TREE_CHAIN (*tail);
+ }
+ }
+ attrs = *attributes;
+ }
+ }
+
+ /* Install the processed list. */
+ *attributes = attrs;
+
+ /* Clear deferred attributes. */
+ sh_deferred_function_attributes = NULL_TREE;
+ sh_deferred_function_attributes_tail = &sh_deferred_function_attributes;
return;
}
interrupt_handler -- specifies this function is an interrupt handler.
+ trapa_handler - like above, but don't save all registers.
+
sp_switch -- specifies an alternate stack for an interrupt handler
to run on.
trap_exit -- use a trapa to exit an interrupt function instead of
an rte instruction.
+ nosave_low_regs - don't save r0..r7 in an interrupt handler.
+ This is useful on the SH3 and upwards,
+ which has a separate set of low regs for User and Supervisor modes.
+ This should only be used for the lowest level of interrupts. Higher levels
+ of interrupts must save the registers in case they themselves are
+ interrupted.
+
renesas -- use Renesas calling/layout conventions (functions and
structures).
{ "sp_switch", 1, 1, true, false, false, sh_handle_sp_switch_attribute },
{ "trap_exit", 1, 1, true, false, false, sh_handle_trap_exit_attribute },
{ "renesas", 0, 0, false, true, false, sh_handle_renesas_attribute },
+ { "trapa_handler", 0, 0, true, false, false, sh_handle_interrupt_handler_attribute },
+ { "nosave_low_regs", 0, 0, true, false, false, sh_handle_interrupt_handler_attribute },
#ifdef SYMBIAN
/* Symbian support adds three new attributes:
dllexport - for exporting a function/variable that will live in a dll
IDENTIFIER_POINTER (name));
*no_add_attrs = true;
}
- else if (!pragma_interrupt)
- {
- /* The sp_switch attribute only has meaning for interrupt functions. */
- warning (OPT_Wattributes, "%qs attribute only applies to "
- "interrupt functions", IDENTIFIER_POINTER (name));
- *no_add_attrs = true;
- }
else if (TREE_CODE (TREE_VALUE (args)) != STRING_CST)
{
/* The argument must be a constant string. */
IDENTIFIER_POINTER (name));
*no_add_attrs = true;
}
- else
- {
- const char *s = ggc_strdup (TREE_STRING_POINTER (TREE_VALUE (args)));
- sp_switch = gen_rtx_SYMBOL_REF (VOIDmode, s);
- }
return NULL_TREE;
}
IDENTIFIER_POINTER (name));
*no_add_attrs = true;
}
- else if (!pragma_interrupt)
- {
- /* The trap_exit attribute only has meaning for interrupt functions. */
- warning (OPT_Wattributes, "%qs attribute only applies to "
- "interrupt functions", IDENTIFIER_POINTER (name));
- *no_add_attrs = true;
- }
+ /* The argument specifies a trap number to be used in a trapa instruction
+ at function exit (instead of an rte instruction). */
else if (TREE_CODE (TREE_VALUE (args)) != INTEGER_CST)
{
/* The argument must be a constant integer. */
"integer constant", IDENTIFIER_POINTER (name));
*no_add_attrs = true;
}
- else
- {
- trap_exit = TREE_INT_CST_LOW (TREE_VALUE (args));
- }
return NULL_TREE;
}
return 0;
if (DECL_P (td))
td = TREE_TYPE (td);
+ if (td == error_mark_node)
+ return 0;
return (lookup_attribute ("renesas", TYPE_ATTRIBUTES (td))
!= NULL_TREE);
}
return fpscr_rtx;
}
+static GTY(()) tree fpscr_values;
+
+static void
+emit_fpu_switch (rtx scratch, int index)
+{
+ rtx dst, src;
+
+ if (fpscr_values == NULL)
+ {
+ tree t;
+
+ t = build_index_type (integer_one_node);
+ t = build_array_type (integer_type_node, t);
+ t = build_decl (VAR_DECL, get_identifier ("__fpscr_values"), t);
+ DECL_ARTIFICIAL (t) = 1;
+ DECL_IGNORED_P (t) = 1;
+ DECL_EXTERNAL (t) = 1;
+ TREE_STATIC (t) = 1;
+ TREE_PUBLIC (t) = 1;
+ TREE_USED (t) = 1;
+
+ fpscr_values = t;
+ }
+
+ src = DECL_RTL (fpscr_values);
+ if (!can_create_pseudo_p ())
+ {
+ emit_move_insn (scratch, XEXP (src, 0));
+ if (index != 0)
+ emit_insn (gen_addsi3 (scratch, scratch, GEN_INT (index * 4)));
+ src = adjust_automodify_address (src, PSImode, scratch, index * 4);
+ }
+ else
+ src = adjust_address (src, PSImode, index * 4);
+
+ dst = get_fpscr_rtx ();
+ emit_move_insn (dst, src);
+}
+
void
emit_sf_insn (rtx pat)
{
get_fpscr_rtx ()));
}
\f
-/* ??? gcc does flow analysis strictly after common subexpression
- elimination. As a result, common subexpression elimination fails
- when there are some intervening statements setting the same register.
- If we did nothing about this, this would hurt the precision switching
- for SH4 badly. There is some cse after reload, but it is unable to
- undo the extra register pressure from the unused instructions, and
- it cannot remove auto-increment loads.
-
- A C code example that shows this flow/cse weakness for (at least) SH
- and sparc (as of gcc ss-970706) is this:
-
-double
-f(double a)
-{
- double d;
- d = 0.1;
- a += d;
- d = 1.1;
- d = 0.1;
- a *= d;
- return a;
-}
-
- So we add another pass before common subexpression elimination, to
- remove assignments that are dead due to a following assignment in the
- same basic block. */
-
-static void
-mark_use (rtx x, rtx *reg_set_block)
-{
- enum rtx_code code;
-
- if (! x)
- return;
- code = GET_CODE (x);
- switch (code)
- {
- case REG:
- {
- int regno = REGNO (x);
- int nregs = (regno < FIRST_PSEUDO_REGISTER
- ? HARD_REGNO_NREGS (regno, GET_MODE (x))
- : 1);
- do
- {
- reg_set_block[regno + nregs - 1] = 0;
- }
- while (--nregs);
- break;
- }
- case SET:
- {
- rtx dest = SET_DEST (x);
-
- if (GET_CODE (dest) == SUBREG)
- dest = SUBREG_REG (dest);
- if (GET_CODE (dest) != REG)
- mark_use (dest, reg_set_block);
- mark_use (SET_SRC (x), reg_set_block);
- break;
- }
- case CLOBBER:
- break;
- default:
- {
- const char *fmt = GET_RTX_FORMAT (code);
- int i, j;
- for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
- {
- if (fmt[i] == 'e')
- mark_use (XEXP (x, i), reg_set_block);
- else if (fmt[i] == 'E')
- for (j = XVECLEN (x, i) - 1; j >= 0; j--)
- mark_use (XVECEXP (x, i, j), reg_set_block);
- }
- break;
- }
- }
-}
-\f
static rtx get_free_reg (HARD_REG_SET);
/* This function returns a register to use to load the address to load
fpscr_set_from_mem (int mode, HARD_REG_SET regs_live)
{
enum attr_fp_mode fp_mode = mode;
- rtx addr_reg = get_free_reg (regs_live);
+ enum attr_fp_mode norm_mode = ACTUAL_NORMAL_MODE (FP_MODE);
+ rtx addr_reg;
- if (fp_mode == (enum attr_fp_mode) ACTUAL_NORMAL_MODE (FP_MODE))
- emit_insn (gen_fpu_switch1 (addr_reg));
- else
- emit_insn (gen_fpu_switch0 (addr_reg));
+ addr_reg = !can_create_pseudo_p () ? get_free_reg (regs_live) : NULL_RTX;
+ emit_fpu_switch (addr_reg, fp_mode == norm_mode);
}
/* Is the given character a logical line separator for the assembler? */
template = XSTR (body, 0);
else if (asm_noperands (body) >= 0)
template
- = decode_asm_operands (body, NULL, NULL, NULL, NULL);
+ = decode_asm_operands (body, NULL, NULL, NULL, NULL, NULL);
else
return 0;
do
saved by the prologue, even if they would normally be
call-clobbered. */
- if (sh_cfun_interrupt_handler_p () && !regs_ever_live[new_reg])
+ if (sh_cfun_interrupt_handler_p () && !df_regs_ever_live_p (new_reg))
return 0;
return 1;
}
else if (REG_NOTE_KIND (link) == 0)
{
- enum attr_type dep_type, type;
+ enum attr_type type;
+ rtx dep_set;
if (recog_memoized (insn) < 0
|| recog_memoized (dep_insn) < 0)
return cost;
- dep_type = get_attr_type (dep_insn);
- if (dep_type == TYPE_FLOAD || dep_type == TYPE_PCFLOAD)
- cost--;
- if ((dep_type == TYPE_LOAD_SI || dep_type == TYPE_PCLOAD_SI)
- && (type = get_attr_type (insn)) != TYPE_CALL
- && type != TYPE_SFUNC)
- cost--;
+ dep_set = single_set (dep_insn);
+
+ /* The latency that we specify in the scheduling description refers
+ to the actual output, not to an auto-increment register; for that,
+ the latency is one. */
+ if (dep_set && MEM_P (SET_SRC (dep_set)) && cost > 1)
+ {
+ rtx set = single_set (insn);
+ if (set
+ && !reg_mentioned_p (SET_DEST (dep_set), SET_SRC (set))
+ && (!MEM_P (SET_DEST (set))
+ || !reg_mentioned_p (SET_DEST (dep_set),
+ XEXP (SET_DEST (set), 0))))
+ cost = 1;
+ }
/* The only input for a call that is timing-critical is the
function's address. */
- if (GET_CODE(insn) == CALL_INSN)
+ if (GET_CODE (insn) == CALL_INSN)
{
rtx call = PATTERN (insn);
/* sibcalli_thunk uses a symbol_ref in an unspec. */
&& (GET_CODE (XEXP (XEXP (call, 0), 0)) == UNSPEC
|| ! reg_set_p (XEXP (XEXP (call, 0), 0), dep_insn)))
- cost = 0;
+ cost -= TARGET_SH4_300 ? 3 : 6;
}
/* Likewise, the most timing critical input for an sfuncs call
is the function address. However, sfuncs typically start
using their arguments pretty quickly.
- Assume a four cycle delay before they are needed. */
+ Assume a four cycle delay for SH4 before they are needed.
+ Cached ST40-300 calls are quicker, so assume only a one
+ cycle delay there.
+ ??? Maybe we should encode the delays till input registers
+ are needed by sfuncs into the sfunc call insn. */
/* All sfunc calls are parallels with at least four components.
Exploit this to avoid unnecessary calls to sfunc_uses_reg. */
else if (GET_CODE (PATTERN (insn)) == PARALLEL
&& (reg = sfunc_uses_reg (insn)))
{
if (! reg_set_p (reg, dep_insn))
- cost -= 4;
- }
- /* When the preceding instruction loads the shift amount of
- the following SHAD/SHLD, the latency of the load is increased
- by 1 cycle. */
- else if (TARGET_SH4
- && get_attr_type (insn) == TYPE_DYN_SHIFT
- && get_attr_any_int_load (dep_insn) == ANY_INT_LOAD_YES
- && reg_overlap_mentioned_p (SET_DEST (PATTERN (dep_insn)),
- XEXP (SET_SRC (single_set (insn)),
- 1)))
- cost++;
- /* When an LS group instruction with a latency of less than
- 3 cycles is followed by a double-precision floating-point
- instruction, FIPR, or FTRV, the latency of the first
- instruction is increased to 3 cycles. */
- else if (cost < 3
- && get_attr_insn_class (dep_insn) == INSN_CLASS_LS_GROUP
- && get_attr_dfp_comp (insn) == DFP_COMP_YES)
- cost = 3;
- /* The lsw register of a double-precision computation is ready one
- cycle earlier. */
- else if (reload_completed
- && get_attr_dfp_comp (dep_insn) == DFP_COMP_YES
- && (use_pat = single_set (insn))
- && ! regno_use_in (REGNO (SET_DEST (single_set (dep_insn))),
- SET_SRC (use_pat)))
- cost -= 1;
-
- if (get_attr_any_fp_comp (dep_insn) == ANY_FP_COMP_YES
- && get_attr_late_fp_use (insn) == LATE_FP_USE_YES)
- cost -= 1;
+ cost -= TARGET_SH4_300 ? 1 : 4;
+ }
+ if (TARGET_HARD_SH4 && !TARGET_SH4_300)
+ {
+ enum attr_type dep_type = get_attr_type (dep_insn);
+
+ if (dep_type == TYPE_FLOAD || dep_type == TYPE_PCFLOAD)
+ cost--;
+ else if ((dep_type == TYPE_LOAD_SI || dep_type == TYPE_PCLOAD_SI)
+ && (type = get_attr_type (insn)) != TYPE_CALL
+ && type != TYPE_SFUNC)
+ cost--;
+ /* When the preceding instruction loads the shift amount of
+ the following SHAD/SHLD, the latency of the load is increased
+ by 1 cycle. */
+ if (get_attr_type (insn) == TYPE_DYN_SHIFT
+ && get_attr_any_int_load (dep_insn) == ANY_INT_LOAD_YES
+ && reg_overlap_mentioned_p (SET_DEST (dep_set),
+ XEXP (SET_SRC (single_set (insn)),
+ 1)))
+ cost++;
+ /* When an LS group instruction with a latency of less than
+ 3 cycles is followed by a double-precision floating-point
+ instruction, FIPR, or FTRV, the latency of the first
+ instruction is increased to 3 cycles. */
+ else if (cost < 3
+ && get_attr_insn_class (dep_insn) == INSN_CLASS_LS_GROUP
+ && get_attr_dfp_comp (insn) == DFP_COMP_YES)
+ cost = 3;
+ /* The lsw register of a double-precision computation is ready one
+ cycle earlier. */
+ else if (reload_completed
+ && get_attr_dfp_comp (dep_insn) == DFP_COMP_YES
+ && (use_pat = single_set (insn))
+ && ! regno_use_in (REGNO (SET_DEST (single_set (dep_insn))),
+ SET_SRC (use_pat)))
+ cost -= 1;
+
+ if (get_attr_any_fp_comp (dep_insn) == ANY_FP_COMP_YES
+ && get_attr_late_fp_use (insn) == LATE_FP_USE_YES)
+ cost -= 1;
+ }
+ else if (TARGET_SH4_300)
+ {
+ /* Stores need their input register two cycles later. */
+ if (dep_set && cost >= 1
+ && ((type = get_attr_type (insn)) == TYPE_STORE
+ || type == TYPE_PSTORE
+ || type == TYPE_FSTORE || type == TYPE_MAC_MEM))
+ {
+ rtx set = single_set (insn);
+
+ if (!reg_mentioned_p (SET_SRC (set), XEXP (SET_DEST (set), 0))
+ && rtx_equal_p (SET_SRC (set), SET_DEST (dep_set)))
+ {
+ cost -= 2;
+ /* But don't reduce the cost below 1 if the address depends
+ on a side effect of dep_insn. */
+ if (cost < 1
+ && modified_in_p (XEXP (SET_DEST (set), 0), dep_insn))
+ cost = 1;
+ }
+ }
+ }
}
/* An anti-dependence penalty of two applies if the first insn is a double
precision fadd / fsub / fmul. */
- else if (REG_NOTE_KIND (link) == REG_DEP_ANTI
+ else if (!TARGET_SH4_300
+ && REG_NOTE_KIND (link) == REG_DEP_ANTI
&& recog_memoized (dep_insn) >= 0
- && get_attr_type (dep_insn) == TYPE_DFP_ARITH
+ && (get_attr_type (dep_insn) == TYPE_DFP_ARITH
+ || get_attr_type (dep_insn) == TYPE_DFP_MUL)
/* A lot of alleged anti-flow dependences are fake,
so check this one is real. */
&& flow_dependent_p (dep_insn, insn))
cost = 2;
-
return cost;
}
static int
sh_pr_n_sets (void)
{
- return REG_N_SETS (TARGET_SHMEDIA ? PR_MEDIA_REG : PR_REG);
+ return DF_REG_DEF_COUNT (TARGET_SHMEDIA ? PR_MEDIA_REG : PR_REG);
}
/* Return where to allocate pseudo for a given hard register initial
|| current_function_has_nonlocal_label)))
x = hard_reg;
else
- x = gen_rtx_MEM (Pmode, return_address_pointer_rtx);
+ x = gen_frame_mem (Pmode, return_address_pointer_rtx);
}
else
x = NULL_RTX;
/* Calculate regmode weights for all insns of a basic block. */
static void
-find_regmode_weight (int b, enum machine_mode mode)
+find_regmode_weight (basic_block b, enum machine_mode mode)
{
rtx insn, next_tail, head, tail;
- get_block_head_tail (b, &head, &tail);
+ get_ebb_head_tail (b, b, &head, &tail);
next_tail = NEXT_INSN (tail);
for (insn = head; insn != next_tail; insn = NEXT_INSN (insn))
SCHED_REORDER (ready, nready);
}
+/* Count life regions of r0 for a block. */
+static int
+find_r0_life_regions (basic_block b)
+{
+ rtx end, insn;
+ rtx pset;
+ rtx r0_reg;
+ int live;
+ int set;
+ int death = 0;
+
+ if (REGNO_REG_SET_P (df_get_live_in (b), R0_REG))
+ {
+ set = 1;
+ live = 1;
+ }
+ else
+ {
+ set = 0;
+ live = 0;
+ }
+
+ insn = BB_HEAD (b);
+ end = BB_END (b);
+ r0_reg = gen_rtx_REG (SImode, R0_REG);
+ while (1)
+ {
+ if (INSN_P (insn))
+ {
+ if (find_regno_note (insn, REG_DEAD, R0_REG))
+ {
+ death++;
+ live = 0;
+ }
+ if (!live
+ && (pset = single_set (insn))
+ && reg_overlap_mentioned_p (r0_reg, SET_DEST (pset))
+ && !find_regno_note (insn, REG_UNUSED, R0_REG))
+ {
+ set++;
+ live = 1;
+ }
+ }
+ if (insn == end)
+ break;
+ insn = NEXT_INSN (insn);
+ }
+ return set - death;
+}
+
/* Calculate regmode weights for all insns of all basic block. */
static void
sh_md_init_global (FILE *dump ATTRIBUTE_UNUSED,
regmode_weight[0] = (short *) xcalloc (old_max_uid, sizeof (short));
regmode_weight[1] = (short *) xcalloc (old_max_uid, sizeof (short));
+ r0_life_regions = 0;
FOR_EACH_BB_REVERSE (b)
{
- find_regmode_weight (b->index, SImode);
- find_regmode_weight (b->index, SFmode);
+ find_regmode_weight (b, SImode);
+ find_regmode_weight (b, SFmode);
+ if (!reload_completed)
+ r0_life_regions += find_r0_life_regions (b);
}
CURR_REGMODE_PRESSURE (SImode) = 0;
/* Pressure on register r0 can lead to spill failures. so avoid sched1 for
functions that already have high pressure on r0. */
#define R0_MAX_LIFE_REGIONS 2
-#define R0_MAX_LIVE_LENGTH 12
/* Register Pressure thresholds for SImode and SFmode registers. */
#define SIMODE_MAX_WEIGHT 5
#define SFMODE_MAX_WEIGHT 10
{
/* Pressure on register r0 can lead to spill failures. so avoid sched1 for
functions that already have high pressure on r0. */
- if ((REG_N_SETS (0) - REG_N_DEATHS (0)) >= R0_MAX_LIFE_REGIONS
- && REG_LIVE_LENGTH (0) >= R0_MAX_LIVE_LENGTH)
- return 1;
+ if (r0_life_regions >= R0_MAX_LIFE_REGIONS)
+ return 1;
if (mode == SFmode)
return (CURR_REGMODE_PRESSURE (SFmode) > SFMODE_MAX_WEIGHT);
sh_optimize_target_register_callee_saved (bool after_prologue_epilogue_gen)
{
HARD_REG_SET dummy;
+#if 0
rtx insn;
+#endif
if (! shmedia_space_reserved_for_target_registers)
return 0;
return 0;
if (calc_live_regs (&dummy) >= 6 * 8)
return 1;
- /* This is a borderline case. See if we got a nested loop, or a loop
- with a call, or with more than 4 labels inside. */
- for (insn = get_insns(); insn; insn = NEXT_INSN (insn))
- {
- if (GET_CODE (insn) == NOTE
- && NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG)
- {
- int labels = 0;
-
- do
- {
- insn = NEXT_INSN (insn);
- if ((GET_CODE (insn) == NOTE
- && NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG)
- || GET_CODE (insn) == CALL_INSN
- || (GET_CODE (insn) == CODE_LABEL && ++labels > 4))
- return 1;
- }
- while (GET_CODE (insn) != NOTE
- || NOTE_LINE_NUMBER (insn) != NOTE_INSN_LOOP_END);
- }
- }
return 0;
}
void
sh_initialize_trampoline (rtx tramp, rtx fnaddr, rtx cxt)
{
+ rtx tramp_mem = gen_frame_mem (BLKmode, tramp);
+
if (TARGET_SHMEDIA64)
{
rtx tramp_templ;
insn = gen_rtx_IOR (DImode, insn, gen_int_mode (0xec000010, SImode));
insn = force_operand (insn, NULL_RTX);
insn = gen_lowpart (SImode, insn);
- emit_move_insn (gen_rtx_MEM (SImode, tramp), insn);
+ emit_move_insn (change_address (tramp_mem, SImode, NULL_RTX), insn);
insn = gen_rtx_LSHIFTRT (DImode, fnaddr, GEN_INT (38));
insn = gen_rtx_AND (DImode, insn, mask);
insn = force_operand (gen_rtx_IOR (DImode, movi1, insn), NULL_RTX);
insn = gen_lowpart (SImode, insn);
- emit_move_insn (gen_rtx_MEM (SImode, plus_constant (tramp, 4)), insn);
+ emit_move_insn (adjust_address (tramp_mem, SImode, 4), insn);
insn = gen_rtx_LSHIFTRT (DImode, fnaddr, GEN_INT (22));
insn = gen_rtx_AND (DImode, insn, mask);
insn = force_operand (gen_rtx_IOR (DImode, shori1, insn), NULL_RTX);
insn = gen_lowpart (SImode, insn);
- emit_move_insn (gen_rtx_MEM (SImode, plus_constant (tramp, 8)), insn);
+ emit_move_insn (adjust_address (tramp_mem, SImode, 8), insn);
insn = gen_rtx_LSHIFTRT (DImode, fnaddr, GEN_INT (6));
insn = gen_rtx_AND (DImode, insn, mask);
insn = force_operand (gen_rtx_IOR (DImode, shori1, insn), NULL_RTX);
insn = gen_lowpart (SImode, insn);
- emit_move_insn (gen_rtx_MEM (SImode, plus_constant (tramp, 12)),
- insn);
+ emit_move_insn (adjust_address (tramp_mem, SImode, 12), insn);
insn = gen_rtx_ASHIFT (DImode, fnaddr, GEN_INT (10));
insn = gen_rtx_AND (DImode, insn, mask);
insn = force_operand (gen_rtx_IOR (DImode, shori1, insn), NULL_RTX);
insn = gen_lowpart (SImode, insn);
- emit_move_insn (gen_rtx_MEM (SImode, plus_constant (tramp, 16)),
- insn);
- emit_move_insn (gen_rtx_MEM (SImode, plus_constant (tramp, 20)),
+ emit_move_insn (adjust_address (tramp_mem, SImode, 16), insn);
+ emit_move_insn (adjust_address (tramp_mem, SImode, 20),
GEN_INT (0x6bf10600));
- emit_move_insn (gen_rtx_MEM (SImode, plus_constant (tramp, 24)),
+ emit_move_insn (adjust_address (tramp_mem, SImode, 24),
GEN_INT (0x4415fc10));
- emit_move_insn (gen_rtx_MEM (SImode, plus_constant (tramp, 28)),
+ emit_move_insn (adjust_address (tramp_mem, SImode, 28),
GEN_INT (0x4401fff0));
emit_insn (gen_ic_invalidate_line (tramp));
return;
fixed_len = TRAMPOLINE_SIZE - 2 * GET_MODE_SIZE (Pmode);
tramp_templ = gen_datalabel_ref (tramp_templ);
- dst = gen_rtx_MEM (BLKmode, tramp);
- src = gen_rtx_MEM (BLKmode, tramp_templ);
+ dst = tramp_mem;
+ src = gen_const_mem (BLKmode, tramp_templ);
set_mem_align (dst, 256);
set_mem_align (src, 64);
emit_block_move (dst, src, GEN_INT (fixed_len), BLOCK_OP_NORMAL);
- emit_move_insn (gen_rtx_MEM (Pmode, plus_constant (tramp, fixed_len)),
- fnaddr);
- emit_move_insn (gen_rtx_MEM (Pmode,
- plus_constant (tramp,
- fixed_len
- + GET_MODE_SIZE (Pmode))),
+ emit_move_insn (adjust_address (tramp_mem, Pmode, fixed_len), fnaddr);
+ emit_move_insn (adjust_address (tramp_mem, Pmode,
+ fixed_len + GET_MODE_SIZE (Pmode)),
cxt);
emit_insn (gen_ic_invalidate_line (tramp));
return;
emit_insn (gen_rotrdi3_mextr (quad0, quad0,
GEN_INT (TARGET_LITTLE_ENDIAN ? 24 : 56)));
emit_insn (gen_ashldi3_media (quad0, quad0, const2_rtx));
- emit_move_insn (gen_rtx_MEM (DImode, tramp), quad0);
+ emit_move_insn (change_address (tramp_mem, DImode, NULL_RTX), quad0);
emit_insn (gen_mshflo_w_x (gen_rtx_SUBREG (V4HImode, cxtload, 0),
gen_rtx_SUBREG (V2HImode, cxt, 0),
movishori));
emit_insn (gen_mextr4 (quad1, cxtload, ptabs));
emit_insn (gen_mshflo_l_di (quad2, blink, cxtload));
}
- emit_move_insn (gen_rtx_MEM (DImode, plus_constant (tramp, 8)), quad1);
- emit_move_insn (gen_rtx_MEM (DImode, plus_constant (tramp, 16)), quad2);
+ emit_move_insn (adjust_address (tramp_mem, DImode, 8), quad1);
+ emit_move_insn (adjust_address (tramp_mem, DImode, 16), quad2);
emit_insn (gen_ic_invalidate_line (tramp));
return;
}
emit_insn (gen_initialize_trampoline (tramp, cxt, fnaddr));
return;
}
- emit_move_insn (gen_rtx_MEM (SImode, tramp),
+ emit_move_insn (change_address (tramp_mem, SImode, NULL_RTX),
gen_int_mode (TARGET_LITTLE_ENDIAN ? 0xd301d202 : 0xd202d301,
SImode));
- emit_move_insn (gen_rtx_MEM (SImode, plus_constant (tramp, 4)),
+ emit_move_insn (adjust_address (tramp_mem, SImode, 4),
gen_int_mode (TARGET_LITTLE_ENDIAN ? 0x0009422b : 0x422b0009,
SImode));
- emit_move_insn (gen_rtx_MEM (SImode, plus_constant (tramp, 8)),
- cxt);
- emit_move_insn (gen_rtx_MEM (SImode, plus_constant (tramp, 12)),
- fnaddr);
+ emit_move_insn (adjust_address (tramp_mem, SImode, 8), cxt);
+ emit_move_insn (adjust_address (tramp_mem, SImode, 12), fnaddr);
if (TARGET_HARVARD)
{
- if (TARGET_USERMODE)
+ if (!TARGET_INLINE_IC_INVALIDATE
+ || (!(TARGET_SH4A_ARCH || TARGET_SH4_300) && TARGET_USERMODE))
emit_library_call (function_symbol (NULL, "__ic_invalidate",
FUNCTION_ORDINARY),
0, VOIDmode, 1, tramp, SImode);
/* describe number and signedness of arguments; arg[0] == result
(1: unsigned, 2: signed, 4: don't care, 8: pointer 0: no argument */
-/* 9: 64 bit pointer, 10: 32 bit pointer */
+/* 9: 64-bit pointer, 10: 32-bit pointer */
static const char signature_args[][4] =
{
#define SH_BLTIN_V2SI2 0
if (signature < SH_BLTIN_NUM_SHARED_SIGNATURES)
shared[signature] = type;
}
- lang_hooks.builtin_function (d->name, type, d - bdesc, BUILT_IN_MD,
- NULL, NULL_TREE);
+ add_builtin_function (d->name, type, d - bdesc, BUILT_IN_MD,
+ NULL, NULL_TREE);
}
}
sh_expand_builtin (tree exp, rtx target, rtx subtarget ATTRIBUTE_UNUSED,
enum machine_mode mode ATTRIBUTE_UNUSED, int ignore)
{
- tree fndecl = TREE_OPERAND (TREE_OPERAND (exp, 0), 0);
- tree arglist = TREE_OPERAND (exp, 1);
+ tree fndecl = TREE_OPERAND (CALL_EXPR_FN (exp), 0);
unsigned int fcode = DECL_FUNCTION_CODE (fndecl);
const struct builtin_description *d = &bdesc[fcode];
enum insn_code icode = d->icode;
if (! signature_args[signature][i])
break;
- arg = TREE_VALUE (arglist);
+ arg = CALL_EXPR_ARG (exp, i - 1);
if (arg == error_mark_node)
return const0_rtx;
- arglist = TREE_CHAIN (arglist);
if (signature_args[signature][i] & 8)
{
opmode = ptr_mode;
&& REGCLASS_HAS_FP_REG (dstclass))
return 4;
+ if (REGCLASS_HAS_FP_REG (dstclass) && srcclass == T_REGS)
+ return ((TARGET_HARD_SH4 && !optimize_size) ? 10 : 7);
+
if ((REGCLASS_HAS_FP_REG (dstclass) && srcclass == MAC_REGS)
|| (dstclass == MAC_REGS && REGCLASS_HAS_FP_REG (srcclass)))
return 9;
static rtx
emit_load_ptr (rtx reg, rtx addr)
{
- rtx mem = gen_rtx_MEM (ptr_mode, addr);
+ rtx mem = gen_const_mem (ptr_mode, addr);
if (Pmode != ptr_mode)
mem = gen_rtx_SIGN_EXTEND (Pmode, mem);
reload_completed = 1;
epilogue_completed = 1;
- no_new_pseudos = 1;
current_function_uses_only_leaf_regs = 1;
- reset_block_changes ();
emit_note (NOTE_INSN_PROLOGUE_END);
the insns emitted. Note that use_thunk calls
assemble_start_function and assemble_end_function. */
- insn_locators_initialize ();
+ insn_locators_alloc ();
insns = get_insns ();
- if (optimize > 0 && flag_schedule_insns_after_reload)
+#if 0
+ if (optimize > 0)
{
/* Initialize the bitmap obstacks. */
bitmap_obstack_initialize (NULL);
if (! cfun->cfg)
init_flow ();
rtl_register_cfg_hooks ();
+ init_rtl_bb_info (ENTRY_BLOCK_PTR);
+ init_rtl_bb_info (EXIT_BLOCK_PTR);
+ ENTRY_BLOCK_PTR->flags |= BB_RTL;
+ EXIT_BLOCK_PTR->flags |= BB_RTL;
find_basic_blocks (insns);
- life_analysis (dump_file, PROP_FINAL);
- split_all_insns (1);
+ if (flag_schedule_insns_after_reload)
+ {
+ life_analysis (PROP_FINAL);
+
+ split_all_insns (1);
- schedule_insns (dump_file);
+ schedule_insns ();
+ }
+ /* We must split jmp insn in PIC case. */
+ else if (flag_pic)
+ split_all_insns_noflow ();
+ }
+#else
+ if (optimize > 0)
+ {
+ if (! cfun->cfg)
+ init_flow ();
+ split_all_insns_noflow ();
}
+#endif
sh_reorg ();
if (optimize > 0 && flag_delayed_branch)
- {
- if (! cfun->cfg)
- {
- init_flow ();
- find_basic_blocks (insns);
- }
- dbr_schedule (insns, dump_file);
- }
+ dbr_schedule (insns);
+
shorten_branches (insns);
final_start_function (insns, file, 1);
final (insns, file, 1);
final_end_function ();
- if (optimize > 0 && flag_schedule_insns_after_reload)
+#if 0
+ if (optimize > 0)
{
- /* Release all memory allocated by flow. */
- free_basic_block_vars ();
+ /* Release all memory allocated by df. */
+ if (rtl_df)
+ {
+ df_finish (rtl_df);
+ rtl_df = NULL;
+ }
/* Release the bitmap obstacks. */
bitmap_obstack_release (®_obstack);
bitmap_obstack_release (NULL);
}
+#endif
reload_completed = 0;
epilogue_completed = 0;
- no_new_pseudos = 0;
}
rtx
&& ((current_function_args_info.call_cookie
& ~ CALL_COOKIE_RET_TRAMP (1))
|| current_function_has_nonlocal_label))
- return gen_rtx_MEM (SImode, return_address_pointer_rtx);
+ return gen_frame_mem (SImode, return_address_pointer_rtx);
/* If we haven't finished rtl generation, there might be a nonlocal label
that we haven't seen yet.
- ??? get_hard_reg_initial_val fails if it is called while no_new_pseudos
- is set, unless it has been called before for the same register. And even
- then, we end in trouble if we didn't use the register in the same
- basic block before. So call get_hard_reg_initial_val now and wrap it
- in an unspec if we might need to replace it. */
+ ??? get_hard_reg_initial_val fails if it is called after register
+ allocation has started, unless it has been called before for the
+ same register. And even then, we end in trouble if we didn't use
+ the register in the same basic block before. So call
+ get_hard_reg_initial_val now and wrap it in an unspec if we might
+ need to replace it. */
/* ??? We also must do this for TARGET_SH1 in general, because otherwise
combine can put the pseudo returned by get_hard_reg_initial_val into
instructions that need a general purpose registers, which will fail to
}
}
-/* Determine if two hard register sets intersect.
- Return 1 if they do. */
-
-static int
-hard_regs_intersect_p (HARD_REG_SET *a, HARD_REG_SET *b)
-{
- HARD_REG_SET c;
- COPY_HARD_REG_SET (c, *a);
- AND_HARD_REG_SET (c, *b);
- GO_IF_HARD_REG_SUBSET (c, reg_class_contents[(int) NO_REGS], lose);
- return 1;
-lose:
- return 0;
-}
-
-#ifdef TARGET_ADJUST_UNROLL_MAX
-static int
-sh_adjust_unroll_max (struct loop * loop, int insn_count,
- int max_unrolled_insns, int strength_reduce_p,
- int unroll_type)
-{
-/* This doesn't work in 4.0 because the old unroller & loop.h is gone. */
- if (TARGET_ADJUST_UNROLL && TARGET_SHMEDIA)
- {
- /* Throttle back loop unrolling so that the costs of using more
- targets than the eight target register we have don't outweigh
- the benefits of unrolling. */
- rtx insn;
- int n_labels = 0, n_calls = 0, n_exit_dest = 0, n_inner_loops = -1;
- int n_barriers = 0;
- rtx dest;
- int i;
- rtx exit_dest[8];
- int threshold;
- int unroll_benefit = 0, mem_latency = 0;
- int base_cost, best_cost, cost;
- int factor, best_factor;
- int n_dest;
- unsigned max_iterations = 32767;
- int n_iterations;
- int need_precond = 0, precond = 0;
- basic_block * bbs = get_loop_body (loop);
- struct niter_desc *desc;
-
- /* Assume that all labels inside the loop are used from inside the
- loop. If the loop has multiple entry points, it is unlikely to
- be unrolled anyways.
- Also assume that all calls are to different functions. That is
- somewhat pessimistic, but if you have lots of calls, unrolling the
- loop is not likely to gain you much in the first place. */
- i = loop->num_nodes - 1;
- for (insn = BB_HEAD (bbs[i]); ; )
- {
- if (GET_CODE (insn) == CODE_LABEL)
- n_labels++;
- else if (GET_CODE (insn) == CALL_INSN)
- n_calls++;
- else if (GET_CODE (insn) == NOTE
- && NOTE_LINE_NUMBER (insn) == NOTE_INSN_LOOP_BEG)
- n_inner_loops++;
- else if (GET_CODE (insn) == BARRIER)
- n_barriers++;
- if (insn != BB_END (bbs[i]))
- insn = NEXT_INSN (insn);
- else if (--i >= 0)
- insn = BB_HEAD (bbs[i]);
- else
- break;
- }
- free (bbs);
- /* One label for the loop top is normal, and it won't be duplicated by
- unrolling. */
- if (n_labels <= 1)
- return max_unrolled_insns;
- if (n_inner_loops > 0)
- return 0;
- for (dest = loop->exit_labels; dest && n_exit_dest < 8;
- dest = LABEL_NEXTREF (dest))
- {
- for (i = n_exit_dest - 1;
- i >= 0 && XEXP (dest, 0) != XEXP (exit_dest[i], 0); i--);
- if (i < 0)
- exit_dest[n_exit_dest++] = dest;
- }
- /* If the loop top and call and exit destinations are enough to fill up
- the target registers, we're unlikely to do any more damage by
- unrolling. */
- if (n_calls + n_exit_dest >= 7)
- return max_unrolled_insns;
-
- /* ??? In the new loop unroller, there is no longer any strength
- reduction information available. Thus, when it comes to unrolling,
- we know the cost of everything, but we know the value of nothing. */
-#if 0
- if (strength_reduce_p
- && (unroll_type == LPT_UNROLL_RUNTIME
- || unroll_type == LPT_UNROLL_CONSTANT
- || unroll_type == LPT_PEEL_COMPLETELY))
- {
- struct loop_ivs *ivs = LOOP_IVS (loop);
- struct iv_class *bl;
-
- /* We'll save one compare-and-branch in each loop body copy
- but the last one. */
- unroll_benefit = 1;
- /* Assess the benefit of removing biv & giv updates. */
- for (bl = ivs->list; bl; bl = bl->next)
- {
- rtx increment = biv_total_increment (bl);
- struct induction *v;
-
- if (increment && GET_CODE (increment) == CONST_INT)
- {
- unroll_benefit++;
- for (v = bl->giv; v; v = v->next_iv)
- {
- if (! v->ignore && v->same == 0
- && GET_CODE (v->mult_val) == CONST_INT)
- unroll_benefit++;
- /* If this giv uses an array, try to determine
- a maximum iteration count from the size of the
- array. This need not be correct all the time,
- but should not be too far off the mark too often. */
- while (v->giv_type == DEST_ADDR)
- {
- rtx mem = PATTERN (v->insn);
- tree mem_expr, type, size_tree;
-
- if (GET_CODE (SET_SRC (mem)) == MEM)
- mem = SET_SRC (mem);
- else if (GET_CODE (SET_DEST (mem)) == MEM)
- mem = SET_DEST (mem);
- else
- break;
- mem_expr = MEM_EXPR (mem);
- if (! mem_expr)
- break;
- type = TREE_TYPE (mem_expr);
- if (TREE_CODE (type) != ARRAY_TYPE
- || ! TYPE_SIZE (type) || ! TYPE_SIZE_UNIT (type))
- break;
- size_tree = fold (build (TRUNC_DIV_EXPR,
- bitsizetype,
- TYPE_SIZE (type),
- TYPE_SIZE_UNIT (type)));
- if (TREE_CODE (size_tree) == INTEGER_CST
- && ! TREE_INT_CST_HIGH (size_tree)
- && TREE_INT_CST_LOW (size_tree) < max_iterations)
- max_iterations = TREE_INT_CST_LOW (size_tree);
- break;
- }
- }
- }
- }
- }
-#else /* 0 */
- /* Assume there is at least some benefit. */
- unroll_benefit = 1;
-#endif /* 0 */
-
- desc = get_simple_loop_desc (loop);
- n_iterations = desc->const_iter ? desc->niter : 0;
- max_iterations
- = max_iterations < desc->niter_max ? max_iterations : desc->niter_max;
-
- if (! strength_reduce_p || ! n_iterations)
- need_precond = 1;
- if (! n_iterations)
- {
- n_iterations
- = max_iterations < 3 ? max_iterations : max_iterations * 3 / 4;
- if (! n_iterations)
- return 0;
- }
-#if 0 /* ??? See above - missing induction variable information. */
- while (unroll_benefit > 1) /* no loop */
- {
- /* We include the benefit of biv/ giv updates. Check if some or
- all of these updates are likely to fit into a scheduling
- bubble of a load.
- We check for the following case:
- - All the insns leading to the first JUMP_INSN are in a strict
- dependency chain.
- - there is at least one memory reference in them.
-
- When we find such a pattern, we assume that we can hide as many
- updates as the total of the load latency is, if we have an
- unroll factor of at least two. We might or might not also do
- this without unrolling, so rather than considering this as an
- extra unroll benefit, discount it in the unroll benefits of unroll
- factors higher than two. */
-
- rtx set, last_set;
-
- insn = next_active_insn (loop->start);
- last_set = single_set (insn);
- if (! last_set)
- break;
- if (GET_CODE (SET_SRC (last_set)) == MEM)
- mem_latency += 2;
- for (insn = NEXT_INSN (insn); insn != end; insn = NEXT_INSN (insn))
- {
- if (! INSN_P (insn))
- continue;
- if (GET_CODE (insn) == JUMP_INSN)
- break;
- if (! reg_referenced_p (SET_DEST (last_set), PATTERN (insn)))
- {
- /* Check if this is a to-be-reduced giv insn. */
- struct loop_ivs *ivs = LOOP_IVS (loop);
- struct iv_class *bl;
- struct induction *v;
- for (bl = ivs->list; bl; bl = bl->next)
- {
- if (bl->biv->insn == insn)
- goto is_biv;
- for (v = bl->giv; v; v = v->next_iv)
- if (v->insn == insn)
- goto is_giv;
- }
- mem_latency--;
- is_biv:
- is_giv:
- continue;
- }
- set = single_set (insn);
- if (! set)
- continue;
- if (GET_CODE (SET_SRC (set)) == MEM)
- mem_latency += 2;
- last_set = set;
- }
- if (mem_latency < 0)
- mem_latency = 0;
- else if (mem_latency > unroll_benefit - 1)
- mem_latency = unroll_benefit - 1;
- break;
- }
-#endif /* 0 */
- if (n_labels + (unroll_benefit + n_labels * 8) / n_iterations
- <= unroll_benefit)
- return max_unrolled_insns;
-
- n_dest = n_labels + n_calls + n_exit_dest;
- base_cost = n_dest <= 8 ? 0 : n_dest - 7;
- best_cost = 0;
- best_factor = 1;
- if (n_barriers * 2 > n_labels - 1)
- n_barriers = (n_labels - 1) / 2;
- for (factor = 2; factor <= 8; factor++)
- {
- /* Bump up preconditioning cost for each power of two. */
- if (! (factor & (factor-1)))
- precond += 4;
- /* When preconditioning, only powers of two will be considered. */
- else if (need_precond)
- continue;
- n_dest = ((unroll_type != LPT_PEEL_COMPLETELY)
- + (n_labels - 1) * factor + n_calls + n_exit_dest
- - (n_barriers * factor >> 1)
- + need_precond);
- cost
- = ((n_dest <= 8 ? 0 : n_dest - 7)
- - base_cost * factor
- - ((factor > 2 ? unroll_benefit - mem_latency : unroll_benefit)
- * (factor - (unroll_type != LPT_PEEL_COMPLETELY)))
- + ((unroll_benefit + 1 + (n_labels - 1) * factor)
- / n_iterations));
- if (need_precond)
- cost += (precond + unroll_benefit * factor / 2) / n_iterations;
- if (cost < best_cost)
- {
- best_cost = cost;
- best_factor = factor;
- }
- }
- threshold = best_factor * insn_count;
- if (max_unrolled_insns > threshold)
- max_unrolled_insns = threshold;
- }
- return max_unrolled_insns;
-}
-#endif /* TARGET_ADJUST_UNROLL_MAX */
-
/* Replace any occurrence of FROM(n) in X with TO(n). The function does
not enter into CONST_DOUBLE for the replace.
return (GET_CODE (*loc) == MEM);
}
-/* Return non-zero iff INSN contains a MEM. */
+/* Return nonzero iff INSN contains a MEM. */
int
sh_contains_memref_p (rtx insn)
{
return for_each_rtx (&PATTERN (insn), &sh_contains_memref_p_1, NULL);
}
+/* Return nonzero iff INSN loads a banked register. */
+int
+sh_loads_bankedreg_p (rtx insn)
+{
+ if (GET_CODE (PATTERN (insn)) == SET)
+ {
+ rtx op = SET_DEST (PATTERN(insn));
+ if (REG_P (op) && BANKED_REGISTER_P (REGNO (op)))
+ return 1;
+ }
+
+ return 0;
+}
+
/* FNADDR is the MEM expression from a call expander. Return an address
to use in an SHmedia insn pattern. */
rtx
return fnaddr;
}
-enum sh_divide_strategy_e sh_div_strategy = SH_DIV_STRATEGY_DEFAULT;
+enum reg_class
+sh_secondary_reload (bool in_p, rtx x, enum reg_class class,
+ enum machine_mode mode, secondary_reload_info *sri)
+{
+ if (in_p)
+ {
+ if (REGCLASS_HAS_FP_REG (class)
+ && ! TARGET_SHMEDIA
+ && immediate_operand ((x), mode)
+ && ! ((fp_zero_operand (x) || fp_one_operand (x))
+ && mode == SFmode && fldi_ok ()))
+ switch (mode)
+ {
+ case SFmode:
+ sri->icode = CODE_FOR_reload_insf__frn;
+ return NO_REGS;
+ case DFmode:
+ sri->icode = CODE_FOR_reload_indf__frn;
+ return NO_REGS;
+ case SImode:
+ /* ??? If we knew that we are in the appropriate mode -
+ single precision - we could use a reload pattern directly. */
+ return FPUL_REGS;
+ default:
+ abort ();
+ }
+ if (class == FPUL_REGS
+ && ((GET_CODE (x) == REG
+ && (REGNO (x) == MACL_REG || REGNO (x) == MACH_REG
+ || REGNO (x) == T_REG))
+ || GET_CODE (x) == PLUS))
+ return GENERAL_REGS;
+ if (class == FPUL_REGS && immediate_operand (x, mode))
+ {
+ if (satisfies_constraint_I08 (x))
+ return GENERAL_REGS;
+ sri->icode = CODE_FOR_reload_insi__i_fpul;
+ return NO_REGS;
+ }
+ if (class == FPSCR_REGS
+ && ((GET_CODE (x) == REG && REGNO (x) >= FIRST_PSEUDO_REGISTER)
+ || (GET_CODE (x) == MEM && GET_CODE (XEXP (x, 0)) == PLUS)))
+ return GENERAL_REGS;
+ if (REGCLASS_HAS_FP_REG (class)
+ && TARGET_SHMEDIA
+ && immediate_operand (x, mode)
+ && x != CONST0_RTX (GET_MODE (x))
+ && GET_MODE (x) != V4SFmode)
+ return GENERAL_REGS;
+ if ((mode == QImode || mode == HImode)
+ && TARGET_SHMEDIA && inqhi_operand (x, mode))
+ {
+ sri->icode = ((mode == QImode)
+ ? CODE_FOR_reload_inqi : CODE_FOR_reload_inhi);
+ return NO_REGS;
+ }
+ if (TARGET_SHMEDIA && class == GENERAL_REGS
+ && (GET_CODE (x) == LABEL_REF || PIC_DIRECT_ADDR_P (x)))
+ return TARGET_REGS;
+ } /* end of input-only processing. */
+
+ if (((REGCLASS_HAS_FP_REG (class)
+ && (GET_CODE (x) == REG
+ && (GENERAL_OR_AP_REGISTER_P (REGNO (x))
+ || (FP_REGISTER_P (REGNO (x)) && mode == SImode
+ && TARGET_FMOVD))))
+ || (REGCLASS_HAS_GENERAL_REG (class)
+ && GET_CODE (x) == REG
+ && FP_REGISTER_P (REGNO (x))))
+ && ! TARGET_SHMEDIA
+ && (mode == SFmode || mode == SImode))
+ return FPUL_REGS;
+ if ((class == FPUL_REGS
+ || (REGCLASS_HAS_FP_REG (class)
+ && ! TARGET_SHMEDIA && mode == SImode))
+ && (GET_CODE (x) == MEM
+ || (GET_CODE (x) == REG
+ && (REGNO (x) >= FIRST_PSEUDO_REGISTER
+ || REGNO (x) == T_REG
+ || system_reg_operand (x, VOIDmode)))))
+ {
+ if (class == FPUL_REGS)
+ return GENERAL_REGS;
+ return FPUL_REGS;
+ }
+ if ((class == TARGET_REGS
+ || (TARGET_SHMEDIA && class == SIBCALL_REGS))
+ && !satisfies_constraint_Csy (x)
+ && (GET_CODE (x) != REG || ! GENERAL_REGISTER_P (REGNO (x))))
+ return GENERAL_REGS;
+ if ((class == MAC_REGS || class == PR_REGS)
+ && GET_CODE (x) == REG && ! GENERAL_REGISTER_P (REGNO (x))
+ && class != REGNO_REG_CLASS (REGNO (x)))
+ return GENERAL_REGS;
+ if (class != GENERAL_REGS && GET_CODE (x) == REG
+ && TARGET_REGISTER_P (REGNO (x)))
+ return GENERAL_REGS;
+ return NO_REGS;
+}
-/* This defines the storage for the variable part of a -mboard= option.
- It is only required when using the sh-superh-elf target */
-#ifdef _SUPERH_H
-const char * boardtype = "7750p2";
-const char * osruntime = "bare";
-#endif
+enum sh_divide_strategy_e sh_div_strategy = SH_DIV_STRATEGY_DEFAULT;
#include "gt-sh.h"
OSDN Git Service
