/* Subroutines for insn-output.c for Motorola 68000 family.
Copyright (C) 1987, 1993, 1994, 1995, 1996, 1997, 1998, 1999, 2000,
- 2001, 2003, 2004, 2005, 2006
+ 2001, 2003, 2004, 2005, 2006, 2007, 2008, 2009
Free Software Foundation, Inc.
This file is part of GCC.
GCC is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
-the Free Software Foundation; either version 2, or (at your option)
+the Free Software Foundation; either version 3, or (at your option)
any later version.
GCC is distributed in the hope that it will be useful,
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
-along with GCC; see the file COPYING. If not, write to
-the Free Software Foundation, 51 Franklin Street, Fifth Floor,
-Boston, MA 02110-1301, USA. */
+along with GCC; see the file COPYING3. If not see
+<http://www.gnu.org/licenses/>. */
#include "config.h"
#include "system.h"
#include "target-def.h"
#include "debug.h"
#include "flags.h"
+#include "df.h"
+/* ??? Need to add a dependency between m68k.o and sched-int.h. */
+#include "sched-int.h"
+#include "insn-codes.h"
+#include "ggc.h"
enum reg_class regno_reg_class[] =
{
};
-/* The ASM_DOT macro allows easy string pasting to handle the differences
- between MOTOROLA and MIT syntaxes in asm_fprintf(), which doesn't
- support the %. option. */
-#if MOTOROLA
-# define ASM_DOT "."
-# define ASM_DOTW ".w"
-# define ASM_DOTL ".l"
-#else
-# define ASM_DOT ""
-# define ASM_DOTW ""
-# define ASM_DOTL ""
-#endif
+/* The minimum number of integer registers that we want to save with the
+ movem instruction. Using two movel instructions instead of a single
+ moveml is about 15% faster for the 68020 and 68030 at no expense in
+ code size. */
+#define MIN_MOVEM_REGS 3
+/* The minimum number of floating point registers that we want to save
+ with the fmovem instruction. */
+#define MIN_FMOVEM_REGS 1
/* Structure describing stack frame layout. */
struct m68k_frame
/* Data and address register. */
int reg_no;
unsigned int reg_mask;
- unsigned int reg_rev_mask;
/* FPU registers. */
int fpu_no;
unsigned int fpu_mask;
- unsigned int fpu_rev_mask;
/* Offsets relative to ARG_POINTER. */
HOST_WIDE_INT frame_pointer_offset;
int scale;
};
+static int m68k_sched_adjust_cost (rtx, rtx, rtx, int);
+static int m68k_sched_issue_rate (void);
+static int m68k_sched_variable_issue (FILE *, int, rtx, int);
+static void m68k_sched_md_init_global (FILE *, int, int);
+static void m68k_sched_md_finish_global (FILE *, int);
+static void m68k_sched_md_init (FILE *, int, int);
+static void m68k_sched_dfa_pre_advance_cycle (void);
+static void m68k_sched_dfa_post_advance_cycle (void);
+static int m68k_sched_first_cycle_multipass_dfa_lookahead (void);
+
+static bool m68k_legitimate_address_p (enum machine_mode, rtx, bool);
static bool m68k_handle_option (size_t, const char *, int);
static rtx find_addr_reg (rtx);
static const char *singlemove_string (rtx *);
-static void m68k_output_function_prologue (FILE *, HOST_WIDE_INT);
-static void m68k_output_function_epilogue (FILE *, HOST_WIDE_INT);
-#ifdef M68K_TARGET_COFF
-static void m68k_coff_asm_named_section (const char *, unsigned int, tree);
-#endif /* M68K_TARGET_COFF */
static void m68k_output_mi_thunk (FILE *, tree, HOST_WIDE_INT,
HOST_WIDE_INT, tree);
static rtx m68k_struct_value_rtx (tree, int);
-static bool m68k_interrupt_function_p (tree func);
static tree m68k_handle_fndecl_attribute (tree *node, tree name,
tree args, int flags,
bool *no_add_attrs);
static void m68k_compute_frame_layout (void);
static bool m68k_save_reg (unsigned int regno, bool interrupt_handler);
-static bool m68k_rtx_costs (rtx, int, int, int *);
+static bool m68k_ok_for_sibcall_p (tree, tree);
+static bool m68k_tls_symbol_p (rtx);
+static rtx m68k_legitimize_address (rtx, rtx, enum machine_mode);
+static bool m68k_rtx_costs (rtx, int, int, int *, bool);
+#if M68K_HONOR_TARGET_STRICT_ALIGNMENT
+static bool m68k_return_in_memory (const_tree, const_tree);
+#endif
+static void m68k_output_dwarf_dtprel (FILE *, int, rtx) ATTRIBUTE_UNUSED;
\f
/* Specify the identification number of the library being built */
const char *m68k_library_id_string = "_current_shared_library_a5_offset_";
-
-/* Nonzero if the last compare/test insn had FP operands. The
- sCC expanders peek at this to determine what to do for the
- 68060, which has no fsCC instructions. */
-int m68k_last_compare_had_fp_operands;
\f
/* Initialize the GCC target structure. */
#undef TARGET_ASM_UNALIGNED_SI_OP
#define TARGET_ASM_UNALIGNED_SI_OP TARGET_ASM_ALIGNED_SI_OP
-#undef TARGET_ASM_FUNCTION_PROLOGUE
-#define TARGET_ASM_FUNCTION_PROLOGUE m68k_output_function_prologue
-#undef TARGET_ASM_FUNCTION_EPILOGUE
-#define TARGET_ASM_FUNCTION_EPILOGUE m68k_output_function_epilogue
-
#undef TARGET_ASM_OUTPUT_MI_THUNK
#define TARGET_ASM_OUTPUT_MI_THUNK m68k_output_mi_thunk
#undef TARGET_ASM_CAN_OUTPUT_MI_THUNK
-#define TARGET_ASM_CAN_OUTPUT_MI_THUNK default_can_output_mi_thunk_no_vcall
+#define TARGET_ASM_CAN_OUTPUT_MI_THUNK hook_bool_const_tree_hwi_hwi_const_tree_true
#undef TARGET_ASM_FILE_START_APP_OFF
#define TARGET_ASM_FILE_START_APP_OFF true
-#undef TARGET_DEFAULT_TARGET_FLAGS
-#define TARGET_DEFAULT_TARGET_FLAGS MASK_STRICT_ALIGNMENT
+#undef TARGET_LEGITIMIZE_ADDRESS
+#define TARGET_LEGITIMIZE_ADDRESS m68k_legitimize_address
+
+#undef TARGET_SCHED_ADJUST_COST
+#define TARGET_SCHED_ADJUST_COST m68k_sched_adjust_cost
+
+#undef TARGET_SCHED_ISSUE_RATE
+#define TARGET_SCHED_ISSUE_RATE m68k_sched_issue_rate
+
+#undef TARGET_SCHED_VARIABLE_ISSUE
+#define TARGET_SCHED_VARIABLE_ISSUE m68k_sched_variable_issue
+
+#undef TARGET_SCHED_INIT_GLOBAL
+#define TARGET_SCHED_INIT_GLOBAL m68k_sched_md_init_global
+
+#undef TARGET_SCHED_FINISH_GLOBAL
+#define TARGET_SCHED_FINISH_GLOBAL m68k_sched_md_finish_global
+
+#undef TARGET_SCHED_INIT
+#define TARGET_SCHED_INIT m68k_sched_md_init
+
+#undef TARGET_SCHED_DFA_PRE_ADVANCE_CYCLE
+#define TARGET_SCHED_DFA_PRE_ADVANCE_CYCLE m68k_sched_dfa_pre_advance_cycle
+
+#undef TARGET_SCHED_DFA_POST_ADVANCE_CYCLE
+#define TARGET_SCHED_DFA_POST_ADVANCE_CYCLE m68k_sched_dfa_post_advance_cycle
+
+#undef TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD
+#define TARGET_SCHED_FIRST_CYCLE_MULTIPASS_DFA_LOOKAHEAD \
+ m68k_sched_first_cycle_multipass_dfa_lookahead
+
#undef TARGET_HANDLE_OPTION
#define TARGET_HANDLE_OPTION m68k_handle_option
#define TARGET_ATTRIBUTE_TABLE m68k_attribute_table
#undef TARGET_PROMOTE_PROTOTYPES
-#define TARGET_PROMOTE_PROTOTYPES hook_bool_tree_true
+#define TARGET_PROMOTE_PROTOTYPES hook_bool_const_tree_true
#undef TARGET_STRUCT_VALUE_RTX
#define TARGET_STRUCT_VALUE_RTX m68k_struct_value_rtx
#undef TARGET_CANNOT_FORCE_CONST_MEM
#define TARGET_CANNOT_FORCE_CONST_MEM m68k_illegitimate_symbolic_constant_p
+#undef TARGET_FUNCTION_OK_FOR_SIBCALL
+#define TARGET_FUNCTION_OK_FOR_SIBCALL m68k_ok_for_sibcall_p
+
+#if M68K_HONOR_TARGET_STRICT_ALIGNMENT
+#undef TARGET_RETURN_IN_MEMORY
+#define TARGET_RETURN_IN_MEMORY m68k_return_in_memory
+#endif
+
+#ifdef HAVE_AS_TLS
+#undef TARGET_HAVE_TLS
+#define TARGET_HAVE_TLS (true)
+
+#undef TARGET_ASM_OUTPUT_DWARF_DTPREL
+#define TARGET_ASM_OUTPUT_DWARF_DTPREL m68k_output_dwarf_dtprel
+#endif
+
+#undef TARGET_LEGITIMATE_ADDRESS_P
+#define TARGET_LEGITIMATE_ADDRESS_P m68k_legitimate_address_p
+
static const struct attribute_spec m68k_attribute_table[] =
{
/* { name, min_len, max_len, decl_req, type_req, fn_type_req, handler } */
+ { "interrupt", 0, 0, true, false, false, m68k_handle_fndecl_attribute },
{ "interrupt_handler", 0, 0, true, false, false, m68k_handle_fndecl_attribute },
+ { "interrupt_thread", 0, 0, true, false, false, m68k_handle_fndecl_attribute },
{ NULL, 0, 0, false, false, false, NULL }
};
#define FL_FOR_isa_aplus (FL_FOR_isa_a | FL_ISA_APLUS | FL_CF_USP)
/* Note ISA_B doesn't necessarily include USP (user stack pointer) support. */
#define FL_FOR_isa_b (FL_FOR_isa_a | FL_ISA_B | FL_CF_HWDIV)
-#define FL_FOR_isa_c (FL_FOR_isa_b | FL_ISA_C | FL_CF_USP)
+/* ISA_C is not upwardly compatible with ISA_B. */
+#define FL_FOR_isa_c (FL_FOR_isa_a | FL_ISA_C | FL_CF_USP)
enum m68k_isa
{
| FL_CF_HWDIV) },
{ "isab", mcf5407, NULL, ucfv4, isa_b, FL_FOR_isa_b },
{ "isac", unk_device, NULL, ucfv4, isa_c, (FL_FOR_isa_c
- | FL_CF_FPU
- | FL_CF_EMAC) },
+ | FL_CF_HWDIV) },
{ NULL, unk_device, NULL, unk_arch, isa_max, 0 }
};
{ "68040", m68040, NULL, u68040, isa_40, FL_FOR_isa_40 },
{ "68060", m68060, NULL, u68060, isa_40, FL_FOR_isa_40 },
{ "cpu32", cpu32, NULL, ucpu32, isa_20, FL_FOR_isa_cpu32 },
+ { "cfv1", mcf51qe, NULL, ucfv1, isa_c, FL_FOR_isa_c },
{ "cfv2", mcf5206, NULL, ucfv2, isa_a, FL_FOR_isa_a },
{ "cfv3", mcf5307, NULL, ucfv3, isa_a, (FL_FOR_isa_a
| FL_CF_HWDIV) },
/* The set of FL_* flags that apply to the target processor. */
unsigned int m68k_cpu_flags;
+/* The set of FL_* flags that apply to the processor to be tuned for. */
+unsigned int m68k_tune_flags;
+
/* Asm templates for calling or jumping to an arbitrary symbolic address,
or NULL if such calls or jumps are not supported. The address is held
in operand 0. */
const char *m68k_symbolic_call;
const char *m68k_symbolic_jump;
+
+/* Enum variable that corresponds to m68k_symbolic_call values. */
+enum M68K_SYMBOLIC_CALL m68k_symbolic_call_var;
+
\f
/* See whether TABLE has an entry with name NAME. Return true and
store the entry in *ENTRY if so, otherwise return false and
error ("-mshared-library-id=%s is not between 0 and %d",
arg, MAX_LIBRARY_ID);
else
- asprintf ((char **) &m68k_library_id_string, "%d", (value * -4) - 4);
+ {
+ char *tmp;
+ asprintf (&tmp, "%d", (value * -4) - 4);
+ m68k_library_id_string = tmp;
+ }
return true;
default:
/* Use the architecture setting to derive default values for
certain flags. */
target_mask = 0;
+
+ /* ColdFire is lenient about alignment. */
+ if (!TARGET_COLDFIRE)
+ target_mask |= MASK_STRICT_ALIGNMENT;
+
if ((m68k_cpu_flags & FL_BITFIELD) != 0)
target_mask |= MASK_BITFIELD;
if ((m68k_cpu_flags & FL_CF_HWDIV) != 0)
/* Set the directly-usable versions of the -mcpu and -mtune settings. */
m68k_cpu = entry->device;
if (m68k_tune_entry)
- m68k_tune = m68k_tune_entry->microarch;
+ {
+ m68k_tune = m68k_tune_entry->microarch;
+ m68k_tune_flags = m68k_tune_entry->flags;
+ }
#ifdef M68K_DEFAULT_TUNE
else if (!m68k_cpu_entry && !m68k_arch_entry)
- m68k_tune = M68K_DEFAULT_TUNE;
+ {
+ enum target_device dev;
+ dev = all_microarchs[M68K_DEFAULT_TUNE].device;
+ m68k_tune_flags = all_devices[dev]->flags;
+ }
#endif
else
- m68k_tune = entry->microarch;
+ {
+ m68k_tune = entry->microarch;
+ m68k_tune_flags = entry->flags;
+ }
/* Set the type of FPU. */
m68k_fpu = (!TARGET_HARD_FLOAT ? FPUTYPE_NONE
: (m68k_cpu_flags & FL_COLDFIRE) != 0 ? FPUTYPE_COLDFIRE
: FPUTYPE_68881);
- if (TARGET_COLDFIRE_FPU)
- {
- REAL_MODE_FORMAT (SFmode) = &coldfire_single_format;
- REAL_MODE_FORMAT (DFmode) = &coldfire_double_format;
- }
-
/* Sanity check to ensure that msep-data and mid-sahred-library are not
* both specified together. Doing so simply doesn't make sense.
*/
if (!flag_pic)
{
-#if MOTOROLA && !defined (USE_GAS)
- m68k_symbolic_call = "jsr %a0";
- m68k_symbolic_jump = "jmp %a0";
-#else
- m68k_symbolic_call = "jbsr %a0";
+ m68k_symbolic_call_var = M68K_SYMBOLIC_CALL_JSR;
+
m68k_symbolic_jump = "jra %a0";
-#endif
}
else if (TARGET_ID_SHARED_LIBRARY)
/* All addresses must be loaded from the GOT. */
;
- else if (TARGET_68020 || TARGET_ISAB)
+ else if (TARGET_68020 || TARGET_ISAB || TARGET_ISAC)
{
if (TARGET_PCREL)
- {
- m68k_symbolic_call = "bsr.l %c0";
- m68k_symbolic_jump = "bra.l %c0";
- }
+ m68k_symbolic_call_var = M68K_SYMBOLIC_CALL_BSR_C;
else
- {
-#if defined(USE_GAS)
- m68k_symbolic_call = "bsr.l %p0";
- m68k_symbolic_jump = "bra.l %p0";
-#else
- m68k_symbolic_call = "bsr %p0";
- m68k_symbolic_jump = "bra %p0";
-#endif
- }
+ m68k_symbolic_call_var = M68K_SYMBOLIC_CALL_BSR_P;
+
+ if (TARGET_ISAC)
+ /* No unconditional long branch */;
+ else if (TARGET_PCREL)
+ m68k_symbolic_jump = "bra%.l %c0";
+ else
+ m68k_symbolic_jump = "bra%.l %p0";
/* Turn off function cse if we are doing PIC. We always want
function call to be done as `bsr foo@PLTPC'. */
/* ??? It's traditional to do this for -mpcrel too, but it isn't
flag_no_function_cse = 1;
}
+ switch (m68k_symbolic_call_var)
+ {
+ case M68K_SYMBOLIC_CALL_JSR:
+ m68k_symbolic_call = "jsr %a0";
+ break;
+
+ case M68K_SYMBOLIC_CALL_BSR_C:
+ m68k_symbolic_call = "bsr%.l %c0";
+ break;
+
+ case M68K_SYMBOLIC_CALL_BSR_P:
+ m68k_symbolic_call = "bsr%.l %p0";
+ break;
+
+ case M68K_SYMBOLIC_CALL_NONE:
+ gcc_assert (m68k_symbolic_call == NULL);
+ break;
+
+ default:
+ gcc_unreachable ();
+ }
+
+#ifndef ASM_OUTPUT_ALIGN_WITH_NOP
+ if (align_labels > 2)
+ {
+ warning (0, "-falign-labels=%d is not supported", align_labels);
+ align_labels = 0;
+ }
+ if (align_loops > 2)
+ {
+ warning (0, "-falign-loops=%d is not supported", align_loops);
+ align_loops = 0;
+ }
+#endif
+
SUBTARGET_OVERRIDE_OPTIONS;
+
+ /* Setup scheduling options. */
+ if (TUNE_CFV1)
+ m68k_sched_cpu = CPU_CFV1;
+ else if (TUNE_CFV2)
+ m68k_sched_cpu = CPU_CFV2;
+ else if (TUNE_CFV3)
+ m68k_sched_cpu = CPU_CFV3;
+ else if (TUNE_CFV4)
+ m68k_sched_cpu = CPU_CFV4;
+ else
+ {
+ m68k_sched_cpu = CPU_UNKNOWN;
+ flag_schedule_insns = 0;
+ flag_schedule_insns_after_reload = 0;
+ flag_modulo_sched = 0;
+ }
+
+ if (m68k_sched_cpu != CPU_UNKNOWN)
+ {
+ if ((m68k_cpu_flags & (FL_CF_EMAC | FL_CF_EMAC_B)) != 0)
+ m68k_sched_mac = MAC_CF_EMAC;
+ else if ((m68k_cpu_flags & FL_CF_MAC) != 0)
+ m68k_sched_mac = MAC_CF_MAC;
+ else
+ m68k_sched_mac = MAC_NO;
+ }
}
/* Generate a macro of the form __mPREFIX_cpu_NAME, where PREFIX is the
return concat ("__m", prefix, "_family_", m68k_cpu_entry->family, NULL);
}
\f
-/* Return nonzero if FUNC is an interrupt function as specified by the
- "interrupt_handler" attribute. */
-static bool
-m68k_interrupt_function_p(tree func)
+/* Return m68k_fk_interrupt_handler if FUNC has an "interrupt" or
+ "interrupt_handler" attribute and interrupt_thread if FUNC has an
+ "interrupt_thread" attribute. Otherwise, return
+ m68k_fk_normal_function. */
+
+enum m68k_function_kind
+m68k_get_function_kind (tree func)
{
tree a;
- if (TREE_CODE (func) != FUNCTION_DECL)
- return false;
+ gcc_assert (TREE_CODE (func) == FUNCTION_DECL);
+
+ a = lookup_attribute ("interrupt", DECL_ATTRIBUTES (func));
+ if (a != NULL_TREE)
+ return m68k_fk_interrupt_handler;
a = lookup_attribute ("interrupt_handler", DECL_ATTRIBUTES (func));
- return (a != NULL_TREE);
+ if (a != NULL_TREE)
+ return m68k_fk_interrupt_handler;
+
+ a = lookup_attribute ("interrupt_thread", DECL_ATTRIBUTES (func));
+ if (a != NULL_TREE)
+ return m68k_fk_interrupt_thread;
+
+ return m68k_fk_normal_function;
}
/* Handle an attribute requiring a FUNCTION_DECL; arguments as in
{
if (TREE_CODE (*node) != FUNCTION_DECL)
{
- warning (OPT_Wattributes, "%qs attribute only applies to functions",
- IDENTIFIER_POINTER (name));
+ warning (OPT_Wattributes, "%qE attribute only applies to functions",
+ name);
+ *no_add_attrs = true;
+ }
+
+ if (m68k_get_function_kind (*node) != m68k_fk_normal_function)
+ {
+ error ("multiple interrupt attributes not allowed");
+ *no_add_attrs = true;
+ }
+
+ if (!TARGET_FIDOA
+ && !strcmp (IDENTIFIER_POINTER (name), "interrupt_thread"))
+ {
+ error ("interrupt_thread is available only on fido");
*no_add_attrs = true;
}
m68k_compute_frame_layout (void)
{
int regno, saved;
- unsigned int mask, rmask;
- bool interrupt_handler = m68k_interrupt_function_p (current_function_decl);
+ unsigned int mask;
+ enum m68k_function_kind func_kind =
+ m68k_get_function_kind (current_function_decl);
+ bool interrupt_handler = func_kind == m68k_fk_interrupt_handler;
+ bool interrupt_thread = func_kind == m68k_fk_interrupt_thread;
/* Only compute the frame once per function.
Don't cache information until reload has been completed. */
current_frame.size = (get_frame_size () + 3) & -4;
- mask = rmask = saved = 0;
- for (regno = 0; regno < 16; regno++)
- if (m68k_save_reg (regno, interrupt_handler))
- {
- mask |= 1 << regno;
- rmask |= 1 << (15 - regno);
- saved++;
- }
+ mask = saved = 0;
+
+ /* Interrupt thread does not need to save any register. */
+ if (!interrupt_thread)
+ for (regno = 0; regno < 16; regno++)
+ if (m68k_save_reg (regno, interrupt_handler))
+ {
+ mask |= 1 << (regno - D0_REG);
+ saved++;
+ }
current_frame.offset = saved * 4;
current_frame.reg_no = saved;
current_frame.reg_mask = mask;
- current_frame.reg_rev_mask = rmask;
current_frame.foffset = 0;
- mask = rmask = saved = 0;
+ mask = saved = 0;
if (TARGET_HARD_FLOAT)
{
- for (regno = 16; regno < 24; regno++)
- if (m68k_save_reg (regno, interrupt_handler))
- {
- mask |= 1 << (regno - 16);
- rmask |= 1 << (23 - regno);
- saved++;
- }
+ /* Interrupt thread does not need to save any register. */
+ if (!interrupt_thread)
+ for (regno = 16; regno < 24; regno++)
+ if (m68k_save_reg (regno, interrupt_handler))
+ {
+ mask |= 1 << (regno - FP0_REG);
+ saved++;
+ }
current_frame.foffset = saved * TARGET_FP_REG_SIZE;
current_frame.offset += current_frame.foffset;
}
current_frame.fpu_no = saved;
current_frame.fpu_mask = mask;
- current_frame.fpu_rev_mask = rmask;
/* Remember what function this frame refers to. */
current_frame.funcdef_no = current_function_funcdef_no;
static bool
m68k_save_reg (unsigned int regno, bool interrupt_handler)
{
- if (flag_pic && regno == PIC_OFFSET_TABLE_REGNUM)
+ if (flag_pic && regno == PIC_REG)
{
- if (current_function_uses_pic_offset_table)
+ if (crtl->saves_all_registers)
+ return true;
+ if (crtl->uses_pic_offset_table)
return true;
- if (!current_function_is_leaf && TARGET_ID_SHARED_LIBRARY)
+ /* Reload may introduce constant pool references into a function
+ that thitherto didn't need a PIC register. Note that the test
+ above will not catch that case because we will only set
+ crtl->uses_pic_offset_table when emitting
+ the address reloads. */
+ if (crtl->uses_const_pool)
return true;
}
- if (current_function_calls_eh_return)
+ if (crtl->calls_eh_return)
{
unsigned int i;
for (i = 0; ; i++)
if they are live or when calling nested functions. */
if (interrupt_handler)
{
- if (regs_ever_live[regno])
+ if (df_regs_ever_live_p (regno))
return true;
if (!current_function_is_leaf && call_used_regs[regno])
}
/* Never need to save registers that aren't touched. */
- if (!regs_ever_live[regno])
+ if (!df_regs_ever_live_p (regno))
return false;
/* Otherwise save everything that isn't call-clobbered. */
return !call_used_regs[regno];
}
-/* This function generates the assembly code for function entry.
- STREAM is a stdio stream to output the code to.
- SIZE is an int: how many units of temporary storage to allocate. */
+/* Emit RTL for a MOVEM or FMOVEM instruction. BASE + OFFSET represents
+ the lowest memory address. COUNT is the number of registers to be
+ moved, with register REGNO + I being moved if bit I of MASK is set.
+ STORE_P specifies the direction of the move and ADJUST_STACK_P says
+ whether or not this is pre-decrement (if STORE_P) or post-increment
+ (if !STORE_P) operation. */
+
+static rtx
+m68k_emit_movem (rtx base, HOST_WIDE_INT offset,
+ unsigned int count, unsigned int regno,
+ unsigned int mask, bool store_p, bool adjust_stack_p)
+{
+ int i;
+ rtx body, addr, src, operands[2];
+ enum machine_mode mode;
+
+ body = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (adjust_stack_p + count));
+ mode = reg_raw_mode[regno];
+ i = 0;
+
+ if (adjust_stack_p)
+ {
+ src = plus_constant (base, (count
+ * GET_MODE_SIZE (mode)
+ * (HOST_WIDE_INT) (store_p ? -1 : 1)));
+ XVECEXP (body, 0, i++) = gen_rtx_SET (VOIDmode, base, src);
+ }
+
+ for (; mask != 0; mask >>= 1, regno++)
+ if (mask & 1)
+ {
+ addr = plus_constant (base, offset);
+ operands[!store_p] = gen_frame_mem (mode, addr);
+ operands[store_p] = gen_rtx_REG (mode, regno);
+ XVECEXP (body, 0, i++)
+ = gen_rtx_SET (VOIDmode, operands[0], operands[1]);
+ offset += GET_MODE_SIZE (mode);
+ }
+ gcc_assert (i == XVECLEN (body, 0));
+
+ return emit_insn (body);
+}
+
+/* Make INSN a frame-related instruction. */
static void
-m68k_output_function_prologue (FILE *stream,
- HOST_WIDE_INT size ATTRIBUTE_UNUSED)
+m68k_set_frame_related (rtx insn)
+{
+ rtx body;
+ int i;
+
+ RTX_FRAME_RELATED_P (insn) = 1;
+ body = PATTERN (insn);
+ if (GET_CODE (body) == PARALLEL)
+ for (i = 0; i < XVECLEN (body, 0); i++)
+ RTX_FRAME_RELATED_P (XVECEXP (body, 0, i)) = 1;
+}
+
+/* Emit RTL for the "prologue" define_expand. */
+
+void
+m68k_expand_prologue (void)
{
HOST_WIDE_INT fsize_with_regs;
- HOST_WIDE_INT cfa_offset = INCOMING_FRAME_SP_OFFSET;
+ rtx limit, src, dest, insn;
- m68k_compute_frame_layout();
+ m68k_compute_frame_layout ();
/* If the stack limit is a symbol, we can check it here,
before actually allocating the space. */
- if (current_function_limit_stack
+ if (crtl->limit_stack
&& GET_CODE (stack_limit_rtx) == SYMBOL_REF)
- asm_fprintf (stream, "\tcmp" ASM_DOT "l %I%s+%wd,%Rsp\n\ttrapcs\n",
- XSTR (stack_limit_rtx, 0), current_frame.size + 4);
+ {
+ limit = plus_constant (stack_limit_rtx, current_frame.size + 4);
+ if (!LEGITIMATE_CONSTANT_P (limit))
+ {
+ emit_move_insn (gen_rtx_REG (Pmode, D0_REG), limit);
+ limit = gen_rtx_REG (Pmode, D0_REG);
+ }
+ emit_insn (gen_ctrapsi4 (gen_rtx_LTU (VOIDmode,
+ stack_pointer_rtx, limit),
+ stack_pointer_rtx, limit,
+ const1_rtx));
+ }
- /* On ColdFire add register save into initial stack frame setup, if possible. */
fsize_with_regs = current_frame.size;
if (TARGET_COLDFIRE)
{
- if (current_frame.reg_no > 2)
- fsize_with_regs += current_frame.reg_no * 4;
- if (current_frame.fpu_no)
- fsize_with_regs += current_frame.fpu_no * 8;
+ /* ColdFire's move multiple instructions do not allow pre-decrement
+ addressing. Add the size of movem saves to the initial stack
+ allocation instead. */
+ if (current_frame.reg_no >= MIN_MOVEM_REGS)
+ fsize_with_regs += current_frame.reg_no * GET_MODE_SIZE (SImode);
+ if (current_frame.fpu_no >= MIN_FMOVEM_REGS)
+ fsize_with_regs += current_frame.fpu_no * GET_MODE_SIZE (DFmode);
}
if (frame_pointer_needed)
{
- if (current_frame.size == 0 && TUNE_68040)
- /* on the 68040, pea + move is faster than link.w 0 */
- fprintf (stream, (MOTOROLA
- ? "\tpea (%s)\n\tmove.l %s,%s\n"
- : "\tpea %s@\n\tmovel %s,%s\n"),
- M68K_REGNAME (FRAME_POINTER_REGNUM),
- M68K_REGNAME (STACK_POINTER_REGNUM),
- M68K_REGNAME (FRAME_POINTER_REGNUM));
- else if (fsize_with_regs < 0x8000)
- asm_fprintf (stream, "\tlink" ASM_DOTW " %s,%I%wd\n",
- M68K_REGNAME (FRAME_POINTER_REGNUM), -fsize_with_regs);
- else if (TARGET_68020)
- asm_fprintf (stream, "\tlink" ASM_DOTL " %s,%I%wd\n",
- M68K_REGNAME (FRAME_POINTER_REGNUM), -fsize_with_regs);
- else
- /* Adding negative number is faster on the 68040. */
- asm_fprintf (stream,
- "\tlink" ASM_DOTW " %s,%I0\n"
- "\tadd" ASM_DOT "l %I%wd,%Rsp\n",
- M68K_REGNAME (FRAME_POINTER_REGNUM), -fsize_with_regs);
- }
- else if (fsize_with_regs) /* !frame_pointer_needed */
- {
- if (fsize_with_regs < 0x8000)
+ if (fsize_with_regs == 0 && TUNE_68040)
{
- if (fsize_with_regs <= 8)
- {
- if (!TARGET_COLDFIRE)
- asm_fprintf (stream, "\tsubq" ASM_DOT "w %I%wd,%Rsp\n",
- fsize_with_regs);
- else
- asm_fprintf (stream, "\tsubq" ASM_DOT "l %I%wd,%Rsp\n",
- fsize_with_regs);
- }
- else if (fsize_with_regs <= 16 && TUNE_CPU32)
- /* On the CPU32 it is faster to use two subqw instructions to
- subtract a small integer (8 < N <= 16) to a register. */
- asm_fprintf (stream,
- "\tsubq" ASM_DOT "w %I8,%Rsp\n"
- "\tsubq" ASM_DOT "w %I%wd,%Rsp\n",
- fsize_with_regs - 8);
- else if (TUNE_68040)
- /* Adding negative number is faster on the 68040. */
- asm_fprintf (stream, "\tadd" ASM_DOT "w %I%wd,%Rsp\n",
- -fsize_with_regs);
- else
- asm_fprintf (stream, (MOTOROLA
- ? "\tlea (%wd,%Rsp),%Rsp\n"
- : "\tlea %Rsp@(%wd),%Rsp\n"),
- -fsize_with_regs);
+ /* On the 68040, two separate moves are faster than link.w 0. */
+ dest = gen_frame_mem (Pmode,
+ gen_rtx_PRE_DEC (Pmode, stack_pointer_rtx));
+ m68k_set_frame_related (emit_move_insn (dest, frame_pointer_rtx));
+ m68k_set_frame_related (emit_move_insn (frame_pointer_rtx,
+ stack_pointer_rtx));
}
- else /* fsize_with_regs >= 0x8000 */
- asm_fprintf (stream, "\tadd" ASM_DOT "l %I%wd,%Rsp\n",
- -fsize_with_regs);
- } /* !frame_pointer_needed */
-
- if (dwarf2out_do_frame ())
- {
- if (frame_pointer_needed)
- {
- char *l;
- l = (char *) dwarf2out_cfi_label ();
- cfa_offset += 4;
- dwarf2out_reg_save (l, FRAME_POINTER_REGNUM, -cfa_offset);
- dwarf2out_def_cfa (l, FRAME_POINTER_REGNUM, cfa_offset);
- cfa_offset += current_frame.size;
- }
+ else if (fsize_with_regs < 0x8000 || TARGET_68020)
+ m68k_set_frame_related
+ (emit_insn (gen_link (frame_pointer_rtx,
+ GEN_INT (-4 - fsize_with_regs))));
else
- {
- cfa_offset += current_frame.size;
- dwarf2out_def_cfa ("", STACK_POINTER_REGNUM, cfa_offset);
- }
+ {
+ m68k_set_frame_related
+ (emit_insn (gen_link (frame_pointer_rtx, GEN_INT (-4))));
+ m68k_set_frame_related
+ (emit_insn (gen_addsi3 (stack_pointer_rtx,
+ stack_pointer_rtx,
+ GEN_INT (-fsize_with_regs))));
+ }
+
+ /* If the frame pointer is needed, emit a special barrier that
+ will prevent the scheduler from moving stores to the frame
+ before the stack adjustment. */
+ emit_insn (gen_stack_tie (stack_pointer_rtx, frame_pointer_rtx));
}
+ else if (fsize_with_regs != 0)
+ m68k_set_frame_related
+ (emit_insn (gen_addsi3 (stack_pointer_rtx,
+ stack_pointer_rtx,
+ GEN_INT (-fsize_with_regs))));
if (current_frame.fpu_mask)
{
+ gcc_assert (current_frame.fpu_no >= MIN_FMOVEM_REGS);
if (TARGET_68881)
- {
- asm_fprintf (stream, (MOTOROLA
- ? "\tfmovm %I0x%x,-(%Rsp)\n"
- : "\tfmovem %I0x%x,%Rsp@-\n"),
- current_frame.fpu_mask);
- }
+ m68k_set_frame_related
+ (m68k_emit_movem (stack_pointer_rtx,
+ current_frame.fpu_no * -GET_MODE_SIZE (XFmode),
+ current_frame.fpu_no, FP0_REG,
+ current_frame.fpu_mask, true, true));
else
{
int offset;
- /* stack already has registers in it. Find the offset from
- the bottom of stack to where the FP registers go */
- if (current_frame.reg_no <= 2)
+ /* If we're using moveml to save the integer registers,
+ the stack pointer will point to the bottom of the moveml
+ save area. Find the stack offset of the first FP register. */
+ if (current_frame.reg_no < MIN_MOVEM_REGS)
offset = 0;
else
- offset = current_frame.reg_no * 4;
- if (offset)
- asm_fprintf (stream,
- "\tfmovem %I0x%x,%d(%Rsp)\n",
- current_frame.fpu_rev_mask,
- offset);
- else
- asm_fprintf (stream,
- "\tfmovem %I0x%x,(%Rsp)\n",
- current_frame.fpu_rev_mask);
- }
-
- if (dwarf2out_do_frame ())
- {
- char *l = (char *) dwarf2out_cfi_label ();
- int n_regs, regno;
-
- cfa_offset += current_frame.fpu_no * TARGET_FP_REG_SIZE;
- if (! frame_pointer_needed)
- dwarf2out_def_cfa (l, STACK_POINTER_REGNUM, cfa_offset);
- for (regno = 16, n_regs = 0; regno < 24; regno++)
- if (current_frame.fpu_mask & (1 << (regno - 16)))
- dwarf2out_reg_save (l, regno, -cfa_offset
- + n_regs++ * TARGET_FP_REG_SIZE);
+ offset = current_frame.reg_no * GET_MODE_SIZE (SImode);
+ m68k_set_frame_related
+ (m68k_emit_movem (stack_pointer_rtx, offset,
+ current_frame.fpu_no, FP0_REG,
+ current_frame.fpu_mask, true, false));
}
}
/* If the stack limit is not a symbol, check it here.
This has the disadvantage that it may be too late... */
- if (current_function_limit_stack)
+ if (crtl->limit_stack)
{
if (REG_P (stack_limit_rtx))
- asm_fprintf (stream, "\tcmp" ASM_DOT "l %s,%Rsp\n\ttrapcs\n",
- M68K_REGNAME (REGNO (stack_limit_rtx)));
+ emit_insn (gen_ctrapsi4 (gen_rtx_LTU (VOIDmode, stack_pointer_rtx,
+ stack_limit_rtx),
+ stack_pointer_rtx, stack_limit_rtx,
+ const1_rtx));
+
else if (GET_CODE (stack_limit_rtx) != SYMBOL_REF)
warning (0, "stack limit expression is not supported");
}
- if (current_frame.reg_no <= 2)
+ if (current_frame.reg_no < MIN_MOVEM_REGS)
{
- /* Store each separately in the same order moveml uses.
- Using two movel instructions instead of a single moveml
- is about 15% faster for the 68020 and 68030 at no expense
- in code size. */
-
+ /* Store each register separately in the same order moveml does. */
int i;
- for (i = 0; i < 16; i++)
- if (current_frame.reg_rev_mask & (1 << i))
+ for (i = 16; i-- > 0; )
+ if (current_frame.reg_mask & (1 << i))
{
- asm_fprintf (stream, (MOTOROLA
- ? "\t%Omove.l %s,-(%Rsp)\n"
- : "\tmovel %s,%Rsp@-\n"),
- M68K_REGNAME (15 - i));
- if (dwarf2out_do_frame ())
- {
- char *l = (char *) dwarf2out_cfi_label ();
-
- cfa_offset += 4;
- if (! frame_pointer_needed)
- dwarf2out_def_cfa (l, STACK_POINTER_REGNUM, cfa_offset);
- dwarf2out_reg_save (l, 15 - i, -cfa_offset);
- }
+ src = gen_rtx_REG (SImode, D0_REG + i);
+ dest = gen_frame_mem (SImode,
+ gen_rtx_PRE_DEC (Pmode, stack_pointer_rtx));
+ m68k_set_frame_related (emit_insn (gen_movsi (dest, src)));
}
}
- else if (current_frame.reg_rev_mask)
+ else
{
if (TARGET_COLDFIRE)
- /* The ColdFire does not support the predecrement form of the
- MOVEM instruction, so we must adjust the stack pointer and
- then use the plain address register indirect mode.
- The required register save space was combined earlier with
- the fsize_with_regs amount. */
-
- asm_fprintf (stream, (MOTOROLA
- ? "\tmovm.l %I0x%x,(%Rsp)\n"
- : "\tmoveml %I0x%x,%Rsp@\n"),
- current_frame.reg_mask);
- else
- asm_fprintf (stream, (MOTOROLA
- ? "\tmovm.l %I0x%x,-(%Rsp)\n"
- : "\tmoveml %I0x%x,%Rsp@-\n"),
- current_frame.reg_rev_mask);
- if (dwarf2out_do_frame ())
- {
- char *l = (char *) dwarf2out_cfi_label ();
- int n_regs, regno;
-
- cfa_offset += current_frame.reg_no * 4;
- if (! frame_pointer_needed)
- dwarf2out_def_cfa (l, STACK_POINTER_REGNUM, cfa_offset);
- for (regno = 0, n_regs = 0; regno < 16; regno++)
- if (current_frame.reg_mask & (1 << regno))
- dwarf2out_reg_save (l, regno, -cfa_offset + n_regs++ * 4);
- }
- }
- if (!TARGET_SEP_DATA && flag_pic
- && (current_function_uses_pic_offset_table
- || (!current_function_is_leaf && TARGET_ID_SHARED_LIBRARY)))
- {
- if (TARGET_ID_SHARED_LIBRARY)
- {
- asm_fprintf (stream, "\tmovel %s@(%s), %s\n",
- M68K_REGNAME (PIC_OFFSET_TABLE_REGNUM),
- m68k_library_id_string,
- M68K_REGNAME (PIC_OFFSET_TABLE_REGNUM));
- }
+ /* The required register save space has already been allocated.
+ The first register should be stored at (%sp). */
+ m68k_set_frame_related
+ (m68k_emit_movem (stack_pointer_rtx, 0,
+ current_frame.reg_no, D0_REG,
+ current_frame.reg_mask, true, false));
else
- {
- if (MOTOROLA)
- asm_fprintf (stream,
- "\t%Olea (%Rpc, %U_GLOBAL_OFFSET_TABLE_@GOTPC), %s\n",
- M68K_REGNAME (PIC_OFFSET_TABLE_REGNUM));
- else
- {
- asm_fprintf (stream, "\tmovel %I%U_GLOBAL_OFFSET_TABLE_, %s\n",
- M68K_REGNAME (PIC_OFFSET_TABLE_REGNUM));
- asm_fprintf (stream, "\tlea %Rpc@(0,%s:l),%s\n",
- M68K_REGNAME (PIC_OFFSET_TABLE_REGNUM),
- M68K_REGNAME (PIC_OFFSET_TABLE_REGNUM));
- }
- }
+ m68k_set_frame_related
+ (m68k_emit_movem (stack_pointer_rtx,
+ current_frame.reg_no * -GET_MODE_SIZE (SImode),
+ current_frame.reg_no, D0_REG,
+ current_frame.reg_mask, true, true));
}
+
+ if (!TARGET_SEP_DATA
+ && crtl->uses_pic_offset_table)
+ insn = emit_insn (gen_load_got (pic_offset_table_rtx));
}
\f
/* Return true if a simple (return) instruction is sufficient for this
return current_frame.offset == 0;
}
-/* This function generates the assembly code for function exit,
- on machines that need it.
+/* Emit RTL for the "epilogue" or "sibcall_epilogue" define_expand;
+ SIBCALL_P says which.
The function epilogue should not depend on the current stack pointer!
It should use the frame pointer only, if there is a frame pointer.
This is mandatory because of alloca; we also take advantage of it to
omit stack adjustments before returning. */
-static void
-m68k_output_function_epilogue (FILE *stream,
- HOST_WIDE_INT size ATTRIBUTE_UNUSED)
+void
+m68k_expand_epilogue (bool sibcall_p)
{
HOST_WIDE_INT fsize, fsize_with_regs;
- bool big = false;
- bool restore_from_sp = false;
- rtx insn = get_last_insn ();
+ bool big, restore_from_sp;
m68k_compute_frame_layout ();
- /* If the last insn was a BARRIER, we don't have to write any code. */
- if (GET_CODE (insn) == NOTE)
- insn = prev_nonnote_insn (insn);
- if (insn && GET_CODE (insn) == BARRIER)
- return;
-
fsize = current_frame.size;
+ big = false;
+ restore_from_sp = false;
- /* FIXME: leaf_function_p below is too strong.
+ /* FIXME : current_function_is_leaf below is too strong.
What we really need to know there is if there could be pending
stack adjustment needed at that point. */
- restore_from_sp
- = (! frame_pointer_needed
- || (! current_function_calls_alloca && leaf_function_p ()));
+ restore_from_sp = (!frame_pointer_needed
+ || (!cfun->calls_alloca
+ && current_function_is_leaf));
/* fsize_with_regs is the size we need to adjust the sp when
popping the frame. */
fsize_with_regs = fsize;
-
- /* Because the ColdFire doesn't support moveml with
- complex address modes, we must adjust the stack manually
- after restoring registers. When the frame pointer isn't used,
- we can merge movem adjustment into frame unlinking
- made immediately after it. */
if (TARGET_COLDFIRE && restore_from_sp)
{
- if (current_frame.reg_no > 2)
- fsize_with_regs += current_frame.reg_no * 4;
- if (current_frame.fpu_no)
- fsize_with_regs += current_frame.fpu_no * 8;
+ /* ColdFire's move multiple instructions do not allow post-increment
+ addressing. Add the size of movem loads to the final deallocation
+ instead. */
+ if (current_frame.reg_no >= MIN_MOVEM_REGS)
+ fsize_with_regs += current_frame.reg_no * GET_MODE_SIZE (SImode);
+ if (current_frame.fpu_no >= MIN_FMOVEM_REGS)
+ fsize_with_regs += current_frame.fpu_no * GET_MODE_SIZE (DFmode);
}
if (current_frame.offset + fsize >= 0x8000
- && ! restore_from_sp
+ && !restore_from_sp
&& (current_frame.reg_mask || current_frame.fpu_mask))
{
- /* Because the ColdFire doesn't support moveml with
- complex address modes we make an extra correction here. */
- if (TARGET_COLDFIRE)
- fsize += current_frame.offset;
-
- asm_fprintf (stream, "\t%Omove" ASM_DOT "l %I%wd,%Ra1\n", -fsize);
- fsize = 0, big = true;
+ if (TARGET_COLDFIRE
+ && (current_frame.reg_no >= MIN_MOVEM_REGS
+ || current_frame.fpu_no >= MIN_FMOVEM_REGS))
+ {
+ /* ColdFire's move multiple instructions do not support the
+ (d8,Ax,Xi) addressing mode, so we're as well using a normal
+ stack-based restore. */
+ emit_move_insn (gen_rtx_REG (Pmode, A1_REG),
+ GEN_INT (-(current_frame.offset + fsize)));
+ emit_insn (gen_addsi3 (stack_pointer_rtx,
+ gen_rtx_REG (Pmode, A1_REG),
+ frame_pointer_rtx));
+ restore_from_sp = true;
+ }
+ else
+ {
+ emit_move_insn (gen_rtx_REG (Pmode, A1_REG), GEN_INT (-fsize));
+ fsize = 0;
+ big = true;
+ }
}
- if (current_frame.reg_no <= 2)
- {
- /* Restore each separately in the same order moveml does.
- Using two movel instructions instead of a single moveml
- is about 15% faster for the 68020 and 68030 at no expense
- in code size. */
+ if (current_frame.reg_no < MIN_MOVEM_REGS)
+ {
+ /* Restore each register separately in the same order moveml does. */
int i;
- HOST_WIDE_INT offset = current_frame.offset + fsize;
+ HOST_WIDE_INT offset;
+ offset = current_frame.offset + fsize;
for (i = 0; i < 16; i++)
if (current_frame.reg_mask & (1 << i))
{
- if (big)
- {
- if (MOTOROLA)
- asm_fprintf (stream, "\t%Omove.l -%wd(%s,%Ra1.l),%s\n",
- offset,
- M68K_REGNAME (FRAME_POINTER_REGNUM),
- M68K_REGNAME (i));
- else
- asm_fprintf (stream, "\tmovel %s@(-%wd,%Ra1:l),%s\n",
- M68K_REGNAME (FRAME_POINTER_REGNUM),
- offset,
- M68K_REGNAME (i));
- }
- else if (restore_from_sp)
- asm_fprintf (stream, (MOTOROLA
- ? "\t%Omove.l (%Rsp)+,%s\n"
- : "\tmovel %Rsp@+,%s\n"),
- M68K_REGNAME (i));
- else
+ rtx addr;
+
+ if (big)
{
- if (MOTOROLA)
- asm_fprintf (stream, "\t%Omove.l -%wd(%s),%s\n",
- offset,
- M68K_REGNAME (FRAME_POINTER_REGNUM),
- M68K_REGNAME (i));
- else
- asm_fprintf (stream, "\tmovel %s@(-%wd),%s\n",
- M68K_REGNAME (FRAME_POINTER_REGNUM),
- offset,
- M68K_REGNAME (i));
+ /* Generate the address -OFFSET(%fp,%a1.l). */
+ addr = gen_rtx_REG (Pmode, A1_REG);
+ addr = gen_rtx_PLUS (Pmode, addr, frame_pointer_rtx);
+ addr = plus_constant (addr, -offset);
}
- offset -= 4;
- }
+ else if (restore_from_sp)
+ addr = gen_rtx_POST_INC (Pmode, stack_pointer_rtx);
+ else
+ addr = plus_constant (frame_pointer_rtx, -offset);
+ emit_move_insn (gen_rtx_REG (SImode, D0_REG + i),
+ gen_frame_mem (SImode, addr));
+ offset -= GET_MODE_SIZE (SImode);
+ }
}
else if (current_frame.reg_mask)
{
- /* The ColdFire requires special handling due to its limited moveml
- insn. */
- if (TARGET_COLDFIRE)
- {
- if (big)
- {
- asm_fprintf (stream, "\tadd" ASM_DOT "l %s,%Ra1\n",
- M68K_REGNAME (FRAME_POINTER_REGNUM));
- asm_fprintf (stream, (MOTOROLA
- ? "\tmovm.l (%Ra1),%I0x%x\n"
- : "\tmoveml %Ra1@,%I0x%x\n"),
- current_frame.reg_mask);
- }
- else if (restore_from_sp)
- asm_fprintf (stream, (MOTOROLA
- ? "\tmovm.l (%Rsp),%I0x%x\n"
- : "\tmoveml %Rsp@,%I0x%x\n"),
- current_frame.reg_mask);
- else
- {
- if (MOTOROLA)
- asm_fprintf (stream, "\tmovm.l -%wd(%s),%I0x%x\n",
- current_frame.offset + fsize,
- M68K_REGNAME (FRAME_POINTER_REGNUM),
- current_frame.reg_mask);
- else
- asm_fprintf (stream, "\tmoveml %s@(-%wd),%I0x%x\n",
- M68K_REGNAME (FRAME_POINTER_REGNUM),
- current_frame.offset + fsize,
- current_frame.reg_mask);
- }
- }
- else /* !TARGET_COLDFIRE */
- {
- if (big)
- {
- if (MOTOROLA)
- asm_fprintf (stream, "\tmovm.l -%wd(%s,%Ra1.l),%I0x%x\n",
- current_frame.offset + fsize,
- M68K_REGNAME (FRAME_POINTER_REGNUM),
- current_frame.reg_mask);
- else
- asm_fprintf (stream, "\tmoveml %s@(-%wd,%Ra1:l),%I0x%x\n",
- M68K_REGNAME (FRAME_POINTER_REGNUM),
- current_frame.offset + fsize,
- current_frame.reg_mask);
- }
- else if (restore_from_sp)
- {
- asm_fprintf (stream, (MOTOROLA
- ? "\tmovm.l (%Rsp)+,%I0x%x\n"
- : "\tmoveml %Rsp@+,%I0x%x\n"),
- current_frame.reg_mask);
- }
- else
- {
- if (MOTOROLA)
- asm_fprintf (stream, "\tmovm.l -%wd(%s),%I0x%x\n",
- current_frame.offset + fsize,
- M68K_REGNAME (FRAME_POINTER_REGNUM),
- current_frame.reg_mask);
- else
- asm_fprintf (stream, "\tmoveml %s@(-%wd),%I0x%x\n",
- M68K_REGNAME (FRAME_POINTER_REGNUM),
- current_frame.offset + fsize,
- current_frame.reg_mask);
- }
- }
+ if (big)
+ m68k_emit_movem (gen_rtx_PLUS (Pmode,
+ gen_rtx_REG (Pmode, A1_REG),
+ frame_pointer_rtx),
+ -(current_frame.offset + fsize),
+ current_frame.reg_no, D0_REG,
+ current_frame.reg_mask, false, false);
+ else if (restore_from_sp)
+ m68k_emit_movem (stack_pointer_rtx, 0,
+ current_frame.reg_no, D0_REG,
+ current_frame.reg_mask, false,
+ !TARGET_COLDFIRE);
+ else
+ m68k_emit_movem (frame_pointer_rtx,
+ -(current_frame.offset + fsize),
+ current_frame.reg_no, D0_REG,
+ current_frame.reg_mask, false, false);
}
- if (current_frame.fpu_rev_mask)
+
+ if (current_frame.fpu_no > 0)
{
if (big)
- {
- if (TARGET_COLDFIRE)
- {
- if (current_frame.reg_no)
- asm_fprintf (stream, MOTOROLA ?
- "\tfmovem.d %d(%Ra1),%I0x%x\n" :
- "\tfmovmd (%d,%Ra1),%I0x%x\n",
- current_frame.reg_no * 4,
- current_frame.fpu_rev_mask);
- else
- asm_fprintf (stream, MOTOROLA ?
- "\tfmovem.d (%Ra1),%I0x%x\n" :
- "\tfmovmd (%Ra1),%I0x%x\n",
- current_frame.fpu_rev_mask);
- }
- else if (MOTOROLA)
- asm_fprintf (stream, "\tfmovm -%wd(%s,%Ra1.l),%I0x%x\n",
- current_frame.foffset + fsize,
- M68K_REGNAME (FRAME_POINTER_REGNUM),
- current_frame.fpu_rev_mask);
- else
- asm_fprintf (stream, "\tfmovem %s@(-%wd,%Ra1:l),%I0x%x\n",
- M68K_REGNAME (FRAME_POINTER_REGNUM),
- current_frame.foffset + fsize,
- current_frame.fpu_rev_mask);
- }
+ m68k_emit_movem (gen_rtx_PLUS (Pmode,
+ gen_rtx_REG (Pmode, A1_REG),
+ frame_pointer_rtx),
+ -(current_frame.foffset + fsize),
+ current_frame.fpu_no, FP0_REG,
+ current_frame.fpu_mask, false, false);
else if (restore_from_sp)
{
if (TARGET_COLDFIRE)
{
int offset;
- /* Stack already has registers in it. Find the offset from
- the bottom of stack to where the FP registers go. */
- if (current_frame.reg_no <= 2)
+ /* If we used moveml to restore the integer registers, the
+ stack pointer will still point to the bottom of the moveml
+ save area. Find the stack offset of the first FP
+ register. */
+ if (current_frame.reg_no < MIN_MOVEM_REGS)
offset = 0;
else
- offset = current_frame.reg_no * 4;
- if (offset)
- asm_fprintf (stream,
- "\tfmovem %Rsp@(%d), %I0x%x\n",
- offset, current_frame.fpu_rev_mask);
- else
- asm_fprintf (stream,
- "\tfmovem %Rsp@, %I0x%x\n",
- current_frame.fpu_rev_mask);
+ offset = current_frame.reg_no * GET_MODE_SIZE (SImode);
+ m68k_emit_movem (stack_pointer_rtx, offset,
+ current_frame.fpu_no, FP0_REG,
+ current_frame.fpu_mask, false, false);
}
else
- asm_fprintf (stream, MOTOROLA ?
- "\tfmovm (%Rsp)+,%I0x%x\n" :
- "\tfmovem %Rsp@+,%I0x%x\n",
- current_frame.fpu_rev_mask);
+ m68k_emit_movem (stack_pointer_rtx, 0,
+ current_frame.fpu_no, FP0_REG,
+ current_frame.fpu_mask, false, true);
}
else
- {
- if (MOTOROLA && !TARGET_COLDFIRE)
- asm_fprintf (stream, "\tfmovm -%wd(%s),%I0x%x\n",
- current_frame.foffset + fsize,
- M68K_REGNAME (FRAME_POINTER_REGNUM),
- current_frame.fpu_rev_mask);
- else
- asm_fprintf (stream, "\tfmovem %s@(-%wd),%I0x%x\n",
- M68K_REGNAME (FRAME_POINTER_REGNUM),
- current_frame.foffset + fsize,
- current_frame.fpu_rev_mask);
- }
+ m68k_emit_movem (frame_pointer_rtx,
+ -(current_frame.foffset + fsize),
+ current_frame.fpu_no, FP0_REG,
+ current_frame.fpu_mask, false, false);
}
+
if (frame_pointer_needed)
- fprintf (stream, "\tunlk %s\n", M68K_REGNAME (FRAME_POINTER_REGNUM));
+ emit_insn (gen_unlink (frame_pointer_rtx));
else if (fsize_with_regs)
- {
- if (fsize_with_regs <= 8)
- {
- if (!TARGET_COLDFIRE)
- asm_fprintf (stream, "\taddq" ASM_DOT "w %I%wd,%Rsp\n",
- fsize_with_regs);
- else
- asm_fprintf (stream, "\taddq" ASM_DOT "l %I%wd,%Rsp\n",
- fsize_with_regs);
- }
- else if (fsize_with_regs <= 16 && TUNE_CPU32)
- {
- /* On the CPU32 it is faster to use two addqw instructions to
- add a small integer (8 < N <= 16) to a register. */
- asm_fprintf (stream,
- "\taddq" ASM_DOT "w %I8,%Rsp\n"
- "\taddq" ASM_DOT "w %I%wd,%Rsp\n",
- fsize_with_regs - 8);
- }
- else if (fsize_with_regs < 0x8000)
- {
- if (TUNE_68040)
- asm_fprintf (stream, "\tadd" ASM_DOT "w %I%wd,%Rsp\n",
- fsize_with_regs);
- else
- asm_fprintf (stream, (MOTOROLA
- ? "\tlea (%wd,%Rsp),%Rsp\n"
- : "\tlea %Rsp@(%wd),%Rsp\n"),
- fsize_with_regs);
- }
- else
- asm_fprintf (stream, "\tadd" ASM_DOT "l %I%wd,%Rsp\n", fsize_with_regs);
- }
- if (current_function_calls_eh_return)
- asm_fprintf (stream, "\tadd" ASM_DOT "l %Ra0,%Rsp\n");
- if (m68k_interrupt_function_p (current_function_decl))
- fprintf (stream, "\trte\n");
- else if (current_function_pops_args)
- asm_fprintf (stream, "\trtd %I%d\n", current_function_pops_args);
- else
- fprintf (stream, "\trts\n");
+ emit_insn (gen_addsi3 (stack_pointer_rtx,
+ stack_pointer_rtx,
+ GEN_INT (fsize_with_regs)));
+
+ if (crtl->calls_eh_return)
+ emit_insn (gen_addsi3 (stack_pointer_rtx,
+ stack_pointer_rtx,
+ EH_RETURN_STACKADJ_RTX));
+
+ if (!sibcall_p)
+ emit_jump_insn (gen_rtx_RETURN (VOIDmode));
}
\f
/* Return true if X is a valid comparison operator for the dbcc
return cc_status.flags & CC_IN_68881;
}
+/* Implement TARGET_FUNCTION_OK_FOR_SIBCALL_P. */
+
+static bool
+m68k_ok_for_sibcall_p (tree decl, tree exp)
+{
+ enum m68k_function_kind kind;
+
+ /* We cannot use sibcalls for nested functions because we use the
+ static chain register for indirect calls. */
+ if (CALL_EXPR_STATIC_CHAIN (exp))
+ return false;
+
+ kind = m68k_get_function_kind (current_function_decl);
+ if (kind == m68k_fk_normal_function)
+ /* We can always sibcall from a normal function, because it's
+ undefined if it is calling an interrupt function. */
+ return true;
+
+ /* Otherwise we can only sibcall if the function kind is known to be
+ the same. */
+ if (decl && m68k_get_function_kind (decl) == kind)
+ return true;
+
+ return false;
+}
+
/* Convert X to a legitimate function call memory reference and return the
result. */
return replace_equiv_address (x, force_reg (Pmode, XEXP (x, 0)));
}
-/* Output a dbCC; jCC sequence. Note we do not handle the
- floating point version of this sequence (Fdbcc). We also
- do not handle alternative conditions when CC_NO_OVERFLOW is
- set. It is assumed that valid_dbcc_comparison_p and flags_in_68881 will
- kick those out before we get here. */
+/* Likewise for sibling calls. */
-void
-output_dbcc_and_branch (rtx *operands)
+rtx
+m68k_legitimize_sibcall_address (rtx x)
{
- switch (GET_CODE (operands[3]))
- {
- case EQ:
- output_asm_insn (MOTOROLA
- ? "dbeq %0,%l1\n\tjbeq %l2"
- : "dbeq %0,%l1\n\tjeq %l2",
- operands);
- break;
-
+ gcc_assert (MEM_P (x));
+ if (sibcall_operand (XEXP (x, 0), VOIDmode))
+ return x;
+
+ emit_move_insn (gen_rtx_REG (Pmode, STATIC_CHAIN_REGNUM), XEXP (x, 0));
+ return replace_equiv_address (x, gen_rtx_REG (Pmode, STATIC_CHAIN_REGNUM));
+}
+
+/* Convert X to a legitimate address and return it if successful. Otherwise
+ return X.
+
+ For the 68000, we handle X+REG by loading X into a register R and
+ using R+REG. R will go in an address reg and indexing will be used.
+ However, if REG is a broken-out memory address or multiplication,
+ nothing needs to be done because REG can certainly go in an address reg. */
+
+rtx
+m68k_legitimize_address (rtx x, rtx oldx, enum machine_mode mode)
+{
+ if (m68k_tls_symbol_p (x))
+ return m68k_legitimize_tls_address (x);
+
+ if (GET_CODE (x) == PLUS)
+ {
+ int ch = (x) != (oldx);
+ int copied = 0;
+
+#define COPY_ONCE(Y) if (!copied) { Y = copy_rtx (Y); copied = ch = 1; }
+
+ if (GET_CODE (XEXP (x, 0)) == MULT)
+ {
+ COPY_ONCE (x);
+ XEXP (x, 0) = force_operand (XEXP (x, 0), 0);
+ }
+ if (GET_CODE (XEXP (x, 1)) == MULT)
+ {
+ COPY_ONCE (x);
+ XEXP (x, 1) = force_operand (XEXP (x, 1), 0);
+ }
+ if (ch)
+ {
+ if (GET_CODE (XEXP (x, 1)) == REG
+ && GET_CODE (XEXP (x, 0)) == REG)
+ {
+ if (TARGET_COLDFIRE_FPU && GET_MODE_CLASS (mode) == MODE_FLOAT)
+ {
+ COPY_ONCE (x);
+ x = force_operand (x, 0);
+ }
+ return x;
+ }
+ if (memory_address_p (mode, x))
+ return x;
+ }
+ if (GET_CODE (XEXP (x, 0)) == REG
+ || (GET_CODE (XEXP (x, 0)) == SIGN_EXTEND
+ && GET_CODE (XEXP (XEXP (x, 0), 0)) == REG
+ && GET_MODE (XEXP (XEXP (x, 0), 0)) == HImode))
+ {
+ rtx temp = gen_reg_rtx (Pmode);
+ rtx val = force_operand (XEXP (x, 1), 0);
+ emit_move_insn (temp, val);
+ COPY_ONCE (x);
+ XEXP (x, 1) = temp;
+ if (TARGET_COLDFIRE_FPU && GET_MODE_CLASS (mode) == MODE_FLOAT
+ && GET_CODE (XEXP (x, 0)) == REG)
+ x = force_operand (x, 0);
+ }
+ else if (GET_CODE (XEXP (x, 1)) == REG
+ || (GET_CODE (XEXP (x, 1)) == SIGN_EXTEND
+ && GET_CODE (XEXP (XEXP (x, 1), 0)) == REG
+ && GET_MODE (XEXP (XEXP (x, 1), 0)) == HImode))
+ {
+ rtx temp = gen_reg_rtx (Pmode);
+ rtx val = force_operand (XEXP (x, 0), 0);
+ emit_move_insn (temp, val);
+ COPY_ONCE (x);
+ XEXP (x, 0) = temp;
+ if (TARGET_COLDFIRE_FPU && GET_MODE_CLASS (mode) == MODE_FLOAT
+ && GET_CODE (XEXP (x, 1)) == REG)
+ x = force_operand (x, 0);
+ }
+ }
+
+ return x;
+}
+
+
+/* Output a dbCC; jCC sequence. Note we do not handle the
+ floating point version of this sequence (Fdbcc). We also
+ do not handle alternative conditions when CC_NO_OVERFLOW is
+ set. It is assumed that valid_dbcc_comparison_p and flags_in_68881 will
+ kick those out before we get here. */
+
+void
+output_dbcc_and_branch (rtx *operands)
+{
+ switch (GET_CODE (operands[3]))
+ {
+ case EQ:
+ output_asm_insn ("dbeq %0,%l1\n\tjeq %l2", operands);
+ break;
+
case NE:
- output_asm_insn (MOTOROLA
- ? "dbne %0,%l1\n\tjbne %l2"
- : "dbne %0,%l1\n\tjne %l2",
- operands);
+ output_asm_insn ("dbne %0,%l1\n\tjne %l2", operands);
break;
case GT:
- output_asm_insn (MOTOROLA
- ? "dbgt %0,%l1\n\tjbgt %l2"
- : "dbgt %0,%l1\n\tjgt %l2",
- operands);
+ output_asm_insn ("dbgt %0,%l1\n\tjgt %l2", operands);
break;
case GTU:
- output_asm_insn (MOTOROLA
- ? "dbhi %0,%l1\n\tjbhi %l2"
- : "dbhi %0,%l1\n\tjhi %l2",
- operands);
+ output_asm_insn ("dbhi %0,%l1\n\tjhi %l2", operands);
break;
case LT:
- output_asm_insn (MOTOROLA
- ? "dblt %0,%l1\n\tjblt %l2"
- : "dblt %0,%l1\n\tjlt %l2",
- operands);
+ output_asm_insn ("dblt %0,%l1\n\tjlt %l2", operands);
break;
case LTU:
- output_asm_insn (MOTOROLA
- ? "dbcs %0,%l1\n\tjbcs %l2"
- : "dbcs %0,%l1\n\tjcs %l2",
- operands);
+ output_asm_insn ("dbcs %0,%l1\n\tjcs %l2", operands);
break;
case GE:
- output_asm_insn (MOTOROLA
- ? "dbge %0,%l1\n\tjbge %l2"
- : "dbge %0,%l1\n\tjge %l2",
- operands);
+ output_asm_insn ("dbge %0,%l1\n\tjge %l2", operands);
break;
case GEU:
- output_asm_insn (MOTOROLA
- ? "dbcc %0,%l1\n\tjbcc %l2"
- : "dbcc %0,%l1\n\tjcc %l2",
- operands);
+ output_asm_insn ("dbcc %0,%l1\n\tjcc %l2", operands);
break;
case LE:
- output_asm_insn (MOTOROLA
- ? "dble %0,%l1\n\tjble %l2"
- : "dble %0,%l1\n\tjle %l2",
- operands);
+ output_asm_insn ("dble %0,%l1\n\tjle %l2", operands);
break;
case LEU:
- output_asm_insn (MOTOROLA
- ? "dbls %0,%l1\n\tjbls %l2"
- : "dbls %0,%l1\n\tjls %l2",
- operands);
+ output_asm_insn ("dbls %0,%l1\n\tjls %l2", operands);
break;
default:
switch (GET_MODE (operands[0]))
{
case SImode:
- output_asm_insn (MOTOROLA
- ? "clr%.w %0\n\tsubq%.l #1,%0\n\tjbpl %l1"
- : "clr%.w %0\n\tsubq%.l #1,%0\n\tjpl %l1",
- operands);
+ output_asm_insn ("clr%.w %0\n\tsubq%.l #1,%0\n\tjpl %l1", operands);
break;
case HImode:
}
loperands[4] = gen_label_rtx ();
if (operand2 != const0_rtx)
- {
- output_asm_insn (MOTOROLA
- ? "cmp%.l %2,%0\n\tjbne %l4\n\tcmp%.l %3,%1"
- : "cmp%.l %2,%0\n\tjne %l4\n\tcmp%.l %3,%1",
- loperands);
- }
+ output_asm_insn ("cmp%.l %2,%0\n\tjne %l4\n\tcmp%.l %3,%1", loperands);
else
{
if (TARGET_68020 || TARGET_COLDFIRE || ! ADDRESS_REG_P (loperands[0]))
else
output_asm_insn ("cmp%.w #0,%0", loperands);
- output_asm_insn (MOTOROLA ? "jbne %l4" : "jne %l4", loperands);
+ output_asm_insn ("jne %l4", loperands);
if (TARGET_68020 || TARGET_COLDFIRE || ! ADDRESS_REG_P (loperands[1]))
output_asm_insn ("tst%.l %1", loperands);
case GT:
loperands[6] = gen_label_rtx ();
- output_asm_insn (MOTOROLA ? "shi %5\n\tjbra %l6" : "shi %5\n\tjra %l6",
- loperands);
+ output_asm_insn ("shi %5\n\tjra %l6", loperands);
(*targetm.asm_out.internal_label) (asm_out_file, "L",
CODE_LABEL_NUMBER (loperands[4]));
output_asm_insn ("sgt %5", loperands);
case LT:
loperands[6] = gen_label_rtx ();
- output_asm_insn (MOTOROLA ? "scs %5\n\tjbra %l6" : "scs %5\n\tjra %l6",
- loperands);
+ output_asm_insn ("scs %5\n\tjra %l6", loperands);
(*targetm.asm_out.internal_label) (asm_out_file, "L",
CODE_LABEL_NUMBER (loperands[4]));
output_asm_insn ("slt %5", loperands);
case GE:
loperands[6] = gen_label_rtx ();
- output_asm_insn (MOTOROLA ? "scc %5\n\tjbra %l6" : "scc %5\n\tjra %l6",
- loperands);
+ output_asm_insn ("scc %5\n\tjra %l6", loperands);
(*targetm.asm_out.internal_label) (asm_out_file, "L",
CODE_LABEL_NUMBER (loperands[4]));
output_asm_insn ("sge %5", loperands);
case LE:
loperands[6] = gen_label_rtx ();
- output_asm_insn (MOTOROLA ? "sls %5\n\tjbra %l6" : "sls %5\n\tjra %l6",
- loperands);
+ output_asm_insn ("sls %5\n\tjra %l6", loperands);
(*targetm.asm_out.internal_label) (asm_out_file, "L",
CODE_LABEL_NUMBER (loperands[4]));
output_asm_insn ("sle %5", loperands);
if (count == 7
&& next_insn_tests_no_inequality (insn))
return "tst%.b %1";
+ /* Try to use `movew to ccr' followed by the appropriate branch insn.
+ On some m68k variants unfortunately that's slower than btst.
+ On 68000 and higher, that should also work for all HImode operands. */
+ if (TUNE_CPU32 || TARGET_COLDFIRE || optimize_size)
+ {
+ if (count == 3 && DATA_REG_P (operands[1])
+ && next_insn_tests_no_inequality (insn))
+ {
+ cc_status.flags = CC_NOT_NEGATIVE | CC_Z_IN_NOT_N | CC_NO_OVERFLOW;
+ return "move%.w %1,%%ccr";
+ }
+ if (count == 2 && DATA_REG_P (operands[1])
+ && next_insn_tests_no_inequality (insn))
+ {
+ cc_status.flags = CC_NOT_NEGATIVE | CC_INVERTED | CC_NO_OVERFLOW;
+ return "move%.w %1,%%ccr";
+ }
+ /* count == 1 followed by bvc/bvs and
+ count == 0 followed by bcc/bcs are also possible, but need
+ m68k-specific CC_Z_IN_NOT_V and CC_Z_IN_NOT_C flags. */
+ }
cc_status.flags = CC_NOT_NEGATIVE;
}
return (REG_P (x)
&& (strict_p
? REGNO_OK_FOR_BASE_P (REGNO (x))
- : !DATA_REGNO_P (REGNO (x)) && !FP_REGNO_P (REGNO (x))));
+ : REGNO_OK_FOR_BASE_NONSTRICT_P (REGNO (x))));
}
/* Return true if X is a legitimate index register. STRICT_P says
return (REG_P (x)
&& (strict_p
? REGNO_OK_FOR_INDEX_P (REGNO (x))
- : !FP_REGNO_P (REGNO (x))));
+ : REGNO_OK_FOR_INDEX_NONSTRICT_P (REGNO (x))));
}
/* Return true if X is a legitimate index expression for a (d8,An,Xn) or
&& !offset_within_block_p (base, INTVAL (offset)))
return true;
}
- return false;
+ return m68k_tls_reference_p (x, false);
}
/* Return true if X is a legitimate constant address that can reach
return false;
}
- return true;
+ return !m68k_tls_reference_p (x, false);
}
/* Return true if X is a LABEL_REF for a jump table. Assume that unplaced
/* Check for GOT loads. These are (bd,An,Xn) addresses if
TARGET_68020 && flag_pic == 2, otherwise they are (d16,An)
addresses. */
- if (flag_pic
- && GET_CODE (x) == PLUS
- && XEXP (x, 0) == pic_offset_table_rtx
- && (GET_CODE (XEXP (x, 1)) == SYMBOL_REF
- || GET_CODE (XEXP (x, 1)) == LABEL_REF))
+ if (GET_CODE (x) == PLUS
+ && XEXP (x, 0) == pic_offset_table_rtx)
{
- address->base = XEXP (x, 0);
- address->offset = XEXP (x, 1);
- return true;
+ /* As we are processing a PLUS, do not unwrap RELOC32 symbols --
+ they are invalid in this context. */
+ if (m68k_unwrap_symbol (XEXP (x, 1), false) != XEXP (x, 1))
+ {
+ address->base = XEXP (x, 0);
+ address->offset = XEXP (x, 1);
+ return true;
+ }
}
/* The ColdFire FPU only accepts addressing modes 2-5. */
&& !address.index);
}
+/* Return GOT pointer. */
+
+static rtx
+m68k_get_gp (void)
+{
+ if (pic_offset_table_rtx == NULL_RTX)
+ pic_offset_table_rtx = gen_rtx_REG (Pmode, PIC_REG);
+
+ crtl->uses_pic_offset_table = 1;
+
+ return pic_offset_table_rtx;
+}
+
+/* M68K relocations, used to distinguish GOT and TLS relocations in UNSPEC
+ wrappers. */
+enum m68k_reloc { RELOC_GOT, RELOC_TLSGD, RELOC_TLSLDM, RELOC_TLSLDO,
+ RELOC_TLSIE, RELOC_TLSLE };
+
+#define TLS_RELOC_P(RELOC) ((RELOC) != RELOC_GOT)
+
+/* Wrap symbol X into unspec representing relocation RELOC.
+ BASE_REG - register that should be added to the result.
+ TEMP_REG - if non-null, temporary register. */
+
+static rtx
+m68k_wrap_symbol (rtx x, enum m68k_reloc reloc, rtx base_reg, rtx temp_reg)
+{
+ bool use_x_p;
+
+ use_x_p = (base_reg == pic_offset_table_rtx) ? TARGET_XGOT : TARGET_XTLS;
+
+ if (TARGET_COLDFIRE && use_x_p)
+ /* When compiling with -mx{got, tls} switch the code will look like this:
+
+ move.l <X>@<RELOC>,<TEMP_REG>
+ add.l <BASE_REG>,<TEMP_REG> */
+ {
+ /* Wrap X in UNSPEC_??? to tip m68k_output_addr_const_extra
+ to put @RELOC after reference. */
+ x = gen_rtx_UNSPEC (Pmode, gen_rtvec (2, x, GEN_INT (reloc)),
+ UNSPEC_RELOC32);
+ x = gen_rtx_CONST (Pmode, x);
+
+ if (temp_reg == NULL)
+ {
+ gcc_assert (can_create_pseudo_p ());
+ temp_reg = gen_reg_rtx (Pmode);
+ }
+
+ emit_move_insn (temp_reg, x);
+ emit_insn (gen_addsi3 (temp_reg, temp_reg, base_reg));
+ x = temp_reg;
+ }
+ else
+ {
+ x = gen_rtx_UNSPEC (Pmode, gen_rtvec (2, x, GEN_INT (reloc)),
+ UNSPEC_RELOC16);
+ x = gen_rtx_CONST (Pmode, x);
+
+ x = gen_rtx_PLUS (Pmode, base_reg, x);
+ }
+
+ return x;
+}
+
+/* Helper for m68k_unwrap_symbol.
+ Also, if unwrapping was successful (that is if (ORIG != <return value>)),
+ sets *RELOC_PTR to relocation type for the symbol. */
+
+static rtx
+m68k_unwrap_symbol_1 (rtx orig, bool unwrap_reloc32_p,
+ enum m68k_reloc *reloc_ptr)
+{
+ if (GET_CODE (orig) == CONST)
+ {
+ rtx x;
+ enum m68k_reloc dummy;
+
+ x = XEXP (orig, 0);
+
+ if (reloc_ptr == NULL)
+ reloc_ptr = &dummy;
+
+ /* Handle an addend. */
+ if ((GET_CODE (x) == PLUS || GET_CODE (x) == MINUS)
+ && CONST_INT_P (XEXP (x, 1)))
+ x = XEXP (x, 0);
+
+ if (GET_CODE (x) == UNSPEC)
+ {
+ switch (XINT (x, 1))
+ {
+ case UNSPEC_RELOC16:
+ orig = XVECEXP (x, 0, 0);
+ *reloc_ptr = (enum m68k_reloc) INTVAL (XVECEXP (x, 0, 1));
+ break;
+
+ case UNSPEC_RELOC32:
+ if (unwrap_reloc32_p)
+ {
+ orig = XVECEXP (x, 0, 0);
+ *reloc_ptr = (enum m68k_reloc) INTVAL (XVECEXP (x, 0, 1));
+ }
+ break;
+
+ default:
+ break;
+ }
+ }
+ }
+
+ return orig;
+}
+
+/* Unwrap symbol from UNSPEC_RELOC16 and, if unwrap_reloc32_p,
+ UNSPEC_RELOC32 wrappers. */
+
+rtx
+m68k_unwrap_symbol (rtx orig, bool unwrap_reloc32_p)
+{
+ return m68k_unwrap_symbol_1 (orig, unwrap_reloc32_p, NULL);
+}
+
+/* Helper for m68k_final_prescan_insn. */
+
+static int
+m68k_final_prescan_insn_1 (rtx *x_ptr, void *data ATTRIBUTE_UNUSED)
+{
+ rtx x = *x_ptr;
+
+ if (m68k_unwrap_symbol (x, true) != x)
+ /* For rationale of the below, see comment in m68k_final_prescan_insn. */
+ {
+ rtx plus;
+
+ gcc_assert (GET_CODE (x) == CONST);
+ plus = XEXP (x, 0);
+
+ if (GET_CODE (plus) == PLUS || GET_CODE (plus) == MINUS)
+ {
+ rtx unspec;
+ rtx addend;
+
+ unspec = XEXP (plus, 0);
+ gcc_assert (GET_CODE (unspec) == UNSPEC);
+ addend = XEXP (plus, 1);
+ gcc_assert (CONST_INT_P (addend));
+
+ /* We now have all the pieces, rearrange them. */
+
+ /* Move symbol to plus. */
+ XEXP (plus, 0) = XVECEXP (unspec, 0, 0);
+
+ /* Move plus inside unspec. */
+ XVECEXP (unspec, 0, 0) = plus;
+
+ /* Move unspec to top level of const. */
+ XEXP (x, 0) = unspec;
+ }
+
+ return -1;
+ }
+
+ return 0;
+}
+
+/* Prescan insn before outputing assembler for it. */
+
+void
+m68k_final_prescan_insn (rtx insn ATTRIBUTE_UNUSED,
+ rtx *operands, int n_operands)
+{
+ int i;
+
+ /* Combine and, possibly, other optimizations may do good job
+ converting
+ (const (unspec [(symbol)]))
+ into
+ (const (plus (unspec [(symbol)])
+ (const_int N))).
+ The problem with this is emitting @TLS or @GOT decorations.
+ The decoration is emitted when processing (unspec), so the
+ result would be "#symbol@TLSLE+N" instead of "#symbol+N@TLSLE".
+
+ It seems that the easiest solution to this is to convert such
+ operands to
+ (const (unspec [(plus (symbol)
+ (const_int N))])).
+ Note, that the top level of operand remains intact, so we don't have
+ to patch up anything outside of the operand. */
+
+ for (i = 0; i < n_operands; ++i)
+ {
+ rtx op;
+
+ op = operands[i];
+
+ for_each_rtx (&op, m68k_final_prescan_insn_1, NULL);
+ }
+}
+
+/* Move X to a register and add REG_EQUAL note pointing to ORIG.
+ If REG is non-null, use it; generate new pseudo otherwise. */
+
+static rtx
+m68k_move_to_reg (rtx x, rtx orig, rtx reg)
+{
+ rtx insn;
+
+ if (reg == NULL_RTX)
+ {
+ gcc_assert (can_create_pseudo_p ());
+ reg = gen_reg_rtx (Pmode);
+ }
+
+ insn = emit_move_insn (reg, x);
+ /* Put a REG_EQUAL note on this insn, so that it can be optimized
+ by loop. */
+ set_unique_reg_note (insn, REG_EQUAL, orig);
+
+ return reg;
+}
+
+/* Does the same as m68k_wrap_symbol, but returns a memory reference to
+ GOT slot. */
+
+static rtx
+m68k_wrap_symbol_into_got_ref (rtx x, enum m68k_reloc reloc, rtx temp_reg)
+{
+ x = m68k_wrap_symbol (x, reloc, m68k_get_gp (), temp_reg);
+
+ x = gen_rtx_MEM (Pmode, x);
+ MEM_READONLY_P (x) = 1;
+
+ return x;
+}
+
/* Legitimize PIC addresses. If the address is already
position-independent, we return ORIG. Newly generated
position-independent addresses go to REG. If we need more
{
gcc_assert (reg);
- pic_ref = gen_rtx_MEM (Pmode,
- gen_rtx_PLUS (Pmode,
- pic_offset_table_rtx, orig));
- current_function_uses_pic_offset_table = 1;
- MEM_READONLY_P (pic_ref) = 1;
- emit_move_insn (reg, pic_ref);
- return reg;
+ pic_ref = m68k_wrap_symbol_into_got_ref (orig, RELOC_GOT, reg);
+ pic_ref = m68k_move_to_reg (pic_ref, orig, reg);
}
else if (GET_CODE (orig) == CONST)
{
rtx base;
/* Make sure this has not already been legitimized. */
- if (GET_CODE (XEXP (orig, 0)) == PLUS
- && XEXP (XEXP (orig, 0), 0) == pic_offset_table_rtx)
+ if (m68k_unwrap_symbol (orig, true) != orig)
return orig;
gcc_assert (reg);
base == reg ? 0 : reg);
if (GET_CODE (orig) == CONST_INT)
- return plus_constant (base, INTVAL (orig));
- pic_ref = gen_rtx_PLUS (Pmode, base, orig);
- /* Likewise, should we set special REG_NOTEs here? */
+ pic_ref = plus_constant (base, INTVAL (orig));
+ else
+ pic_ref = gen_rtx_PLUS (Pmode, base, orig);
}
+
return pic_ref;
}
-\f
-typedef enum { MOVL, SWAP, NEGW, NOTW, NOTB, MOVQ, MVS, MVZ } CONST_METHOD;
+/* The __tls_get_addr symbol. */
+static GTY(()) rtx m68k_tls_get_addr;
-#define USE_MOVQ(i) ((unsigned) ((i) + 128) <= 255)
+/* Return SYMBOL_REF for __tls_get_addr. */
-/* Return the type of move that should be used for integer I. */
+static rtx
+m68k_get_tls_get_addr (void)
+{
+ if (m68k_tls_get_addr == NULL_RTX)
+ m68k_tls_get_addr = init_one_libfunc ("__tls_get_addr");
+
+ return m68k_tls_get_addr;
+}
+
+/* Return libcall result in A0 instead of usual D0. */
+static bool m68k_libcall_value_in_a0_p = false;
+
+/* Emit instruction sequence that calls __tls_get_addr. X is
+ the TLS symbol we are referencing and RELOC is the symbol type to use
+ (either TLSGD or TLSLDM). EQV is the REG_EQUAL note for the sequence
+ emitted. A pseudo register with result of __tls_get_addr call is
+ returned. */
-static CONST_METHOD
-const_method (HOST_WIDE_INT i)
+static rtx
+m68k_call_tls_get_addr (rtx x, rtx eqv, enum m68k_reloc reloc)
{
- unsigned u;
+ rtx a0;
+ rtx insns;
+ rtx dest;
- if (USE_MOVQ (i))
- return MOVQ;
+ /* Emit the call sequence. */
+ start_sequence ();
+
+ /* FIXME: Unfortunately, emit_library_call_value does not
+ consider (plus (%a5) (const (unspec))) to be a good enough
+ operand for push, so it forces it into a register. The bad
+ thing about this is that combiner, due to copy propagation and other
+ optimizations, sometimes can not later fix this. As a consequence,
+ additional register may be allocated resulting in a spill.
+ For reference, see args processing loops in
+ calls.c:emit_library_call_value_1.
+ For testcase, see gcc.target/m68k/tls-{gd, ld}.c */
+ x = m68k_wrap_symbol (x, reloc, m68k_get_gp (), NULL_RTX);
+
+ /* __tls_get_addr() is not a libcall, but emitting a libcall_value
+ is the simpliest way of generating a call. The difference between
+ __tls_get_addr() and libcall is that the result is returned in D0
+ instead of A0. To workaround this, we use m68k_libcall_value_in_a0_p
+ which temporarily switches returning the result to A0. */
+
+ m68k_libcall_value_in_a0_p = true;
+ a0 = emit_library_call_value (m68k_get_tls_get_addr (), NULL_RTX, LCT_PURE,
+ Pmode, 1, x, Pmode);
+ m68k_libcall_value_in_a0_p = false;
+
+ insns = get_insns ();
+ end_sequence ();
- /* The ColdFire doesn't have byte or word operations. */
- /* FIXME: This may not be useful for the m68060 either. */
- if (!TARGET_COLDFIRE)
- {
- /* if -256 < N < 256 but N is not in range for a moveq
- N^ff will be, so use moveq #N^ff, dreg; not.b dreg. */
- if (USE_MOVQ (i ^ 0xff))
- return NOTB;
- /* Likewise, try with not.w */
- if (USE_MOVQ (i ^ 0xffff))
- return NOTW;
- /* This is the only value where neg.w is useful */
- if (i == -65408)
- return NEGW;
- }
+ gcc_assert (can_create_pseudo_p ());
+ dest = gen_reg_rtx (Pmode);
+ emit_libcall_block (insns, dest, a0, eqv);
- /* Try also with swap. */
- u = i;
- if (USE_MOVQ ((u >> 16) | (u << 16)))
- return SWAP;
+ return dest;
+}
- if (TARGET_ISAB)
- {
- /* Try using MVZ/MVS with an immediate value to load constants. */
- if (i >= 0 && i <= 65535)
- return MVZ;
- if (i >= -32768 && i <= 32767)
- return MVS;
- }
+/* The __tls_get_addr symbol. */
+static GTY(()) rtx m68k_read_tp;
- /* Otherwise, use move.l */
- return MOVL;
+/* Return SYMBOL_REF for __m68k_read_tp. */
+
+static rtx
+m68k_get_m68k_read_tp (void)
+{
+ if (m68k_read_tp == NULL_RTX)
+ m68k_read_tp = init_one_libfunc ("__m68k_read_tp");
+
+ return m68k_read_tp;
}
-/* Return the cost of moving constant I into a data register. */
+/* Emit instruction sequence that calls __m68k_read_tp.
+ A pseudo register with result of __m68k_read_tp call is returned. */
-static int
-const_int_cost (HOST_WIDE_INT i)
+static rtx
+m68k_call_m68k_read_tp (void)
{
- switch (const_method (i))
- {
- case MOVQ:
- /* Constants between -128 and 127 are cheap due to moveq. */
- return 0;
- case MVZ:
- case MVS:
- case NOTB:
- case NOTW:
- case NEGW:
- case SWAP:
- /* Constants easily generated by moveq + not.b/not.w/neg.w/swap. */
- return 1;
- case MOVL:
- return 2;
- default:
- gcc_unreachable ();
- }
+ rtx a0;
+ rtx eqv;
+ rtx insns;
+ rtx dest;
+
+ start_sequence ();
+
+ /* __m68k_read_tp() is not a libcall, but emitting a libcall_value
+ is the simpliest way of generating a call. The difference between
+ __m68k_read_tp() and libcall is that the result is returned in D0
+ instead of A0. To workaround this, we use m68k_libcall_value_in_a0_p
+ which temporarily switches returning the result to A0. */
+
+ /* Emit the call sequence. */
+ m68k_libcall_value_in_a0_p = true;
+ a0 = emit_library_call_value (m68k_get_m68k_read_tp (), NULL_RTX, LCT_PURE,
+ Pmode, 0);
+ m68k_libcall_value_in_a0_p = false;
+ insns = get_insns ();
+ end_sequence ();
+
+ /* Attach a unique REG_EQUIV, to allow the RTL optimizers to
+ share the m68k_read_tp result with other IE/LE model accesses. */
+ eqv = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, const1_rtx), UNSPEC_RELOC32);
+
+ gcc_assert (can_create_pseudo_p ());
+ dest = gen_reg_rtx (Pmode);
+ emit_libcall_block (insns, dest, a0, eqv);
+
+ return dest;
}
-static bool
-m68k_rtx_costs (rtx x, int code, int outer_code, int *total)
+/* Return a legitimized address for accessing TLS SYMBOL_REF X.
+ For explanations on instructions sequences see TLS/NPTL ABI for m68k and
+ ColdFire. */
+
+rtx
+m68k_legitimize_tls_address (rtx orig)
{
- switch (code)
+ switch (SYMBOL_REF_TLS_MODEL (orig))
{
- case CONST_INT:
- /* Constant zero is super cheap due to clr instruction. */
- if (x == const0_rtx)
- *total = 0;
- else
- *total = const_int_cost (INTVAL (x));
- return true;
+ case TLS_MODEL_GLOBAL_DYNAMIC:
+ orig = m68k_call_tls_get_addr (orig, orig, RELOC_TLSGD);
+ break;
- case CONST:
- case LABEL_REF:
+ case TLS_MODEL_LOCAL_DYNAMIC:
+ {
+ rtx eqv;
+ rtx a0;
+ rtx x;
+
+ /* Attach a unique REG_EQUIV, to allow the RTL optimizers to
+ share the LDM result with other LD model accesses. */
+ eqv = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, const0_rtx),
+ UNSPEC_RELOC32);
+
+ a0 = m68k_call_tls_get_addr (orig, eqv, RELOC_TLSLDM);
+
+ x = m68k_wrap_symbol (orig, RELOC_TLSLDO, a0, NULL_RTX);
+
+ if (can_create_pseudo_p ())
+ x = m68k_move_to_reg (x, orig, NULL_RTX);
+
+ orig = x;
+ break;
+ }
+
+ case TLS_MODEL_INITIAL_EXEC:
+ {
+ rtx a0;
+ rtx x;
+
+ a0 = m68k_call_m68k_read_tp ();
+
+ x = m68k_wrap_symbol_into_got_ref (orig, RELOC_TLSIE, NULL_RTX);
+ x = gen_rtx_PLUS (Pmode, x, a0);
+
+ if (can_create_pseudo_p ())
+ x = m68k_move_to_reg (x, orig, NULL_RTX);
+
+ orig = x;
+ break;
+ }
+
+ case TLS_MODEL_LOCAL_EXEC:
+ {
+ rtx a0;
+ rtx x;
+
+ a0 = m68k_call_m68k_read_tp ();
+
+ x = m68k_wrap_symbol (orig, RELOC_TLSLE, a0, NULL_RTX);
+
+ if (can_create_pseudo_p ())
+ x = m68k_move_to_reg (x, orig, NULL_RTX);
+
+ orig = x;
+ break;
+ }
+
+ default:
+ gcc_unreachable ();
+ }
+
+ return orig;
+}
+
+/* Return true if X is a TLS symbol. */
+
+static bool
+m68k_tls_symbol_p (rtx x)
+{
+ if (!TARGET_HAVE_TLS)
+ return false;
+
+ if (GET_CODE (x) != SYMBOL_REF)
+ return false;
+
+ return SYMBOL_REF_TLS_MODEL (x) != 0;
+}
+
+/* Helper for m68k_tls_referenced_p. */
+
+static int
+m68k_tls_reference_p_1 (rtx *x_ptr, void *data ATTRIBUTE_UNUSED)
+{
+ /* Note: this is not the same as m68k_tls_symbol_p. */
+ if (GET_CODE (*x_ptr) == SYMBOL_REF)
+ return SYMBOL_REF_TLS_MODEL (*x_ptr) != 0 ? 1 : 0;
+
+ /* Don't recurse into legitimate TLS references. */
+ if (m68k_tls_reference_p (*x_ptr, true))
+ return -1;
+
+ return 0;
+}
+
+/* If !LEGITIMATE_P, return true if X is a TLS symbol reference,
+ though illegitimate one.
+ If LEGITIMATE_P, return true if X is a legitimate TLS symbol reference. */
+
+bool
+m68k_tls_reference_p (rtx x, bool legitimate_p)
+{
+ if (!TARGET_HAVE_TLS)
+ return false;
+
+ if (!legitimate_p)
+ return for_each_rtx (&x, m68k_tls_reference_p_1, NULL) == 1 ? true : false;
+ else
+ {
+ enum m68k_reloc reloc = RELOC_GOT;
+
+ return (m68k_unwrap_symbol_1 (x, true, &reloc) != x
+ && TLS_RELOC_P (reloc));
+ }
+}
+
+\f
+
+#define USE_MOVQ(i) ((unsigned) ((i) + 128) <= 255)
+
+/* Return the type of move that should be used for integer I. */
+
+M68K_CONST_METHOD
+m68k_const_method (HOST_WIDE_INT i)
+{
+ unsigned u;
+
+ if (USE_MOVQ (i))
+ return MOVQ;
+
+ /* The ColdFire doesn't have byte or word operations. */
+ /* FIXME: This may not be useful for the m68060 either. */
+ if (!TARGET_COLDFIRE)
+ {
+ /* if -256 < N < 256 but N is not in range for a moveq
+ N^ff will be, so use moveq #N^ff, dreg; not.b dreg. */
+ if (USE_MOVQ (i ^ 0xff))
+ return NOTB;
+ /* Likewise, try with not.w */
+ if (USE_MOVQ (i ^ 0xffff))
+ return NOTW;
+ /* This is the only value where neg.w is useful */
+ if (i == -65408)
+ return NEGW;
+ }
+
+ /* Try also with swap. */
+ u = i;
+ if (USE_MOVQ ((u >> 16) | (u << 16)))
+ return SWAP;
+
+ if (TARGET_ISAB)
+ {
+ /* Try using MVZ/MVS with an immediate value to load constants. */
+ if (i >= 0 && i <= 65535)
+ return MVZ;
+ if (i >= -32768 && i <= 32767)
+ return MVS;
+ }
+
+ /* Otherwise, use move.l */
+ return MOVL;
+}
+
+/* Return the cost of moving constant I into a data register. */
+
+static int
+const_int_cost (HOST_WIDE_INT i)
+{
+ switch (m68k_const_method (i))
+ {
+ case MOVQ:
+ /* Constants between -128 and 127 are cheap due to moveq. */
+ return 0;
+ case MVZ:
+ case MVS:
+ case NOTB:
+ case NOTW:
+ case NEGW:
+ case SWAP:
+ /* Constants easily generated by moveq + not.b/not.w/neg.w/swap. */
+ return 1;
+ case MOVL:
+ return 2;
+ default:
+ gcc_unreachable ();
+ }
+}
+
+static bool
+m68k_rtx_costs (rtx x, int code, int outer_code, int *total,
+ bool speed ATTRIBUTE_UNUSED)
+{
+ switch (code)
+ {
+ case CONST_INT:
+ /* Constant zero is super cheap due to clr instruction. */
+ if (x == const0_rtx)
+ *total = 0;
+ else
+ *total = const_int_cost (INTVAL (x));
+ return true;
+
+ case CONST:
+ case LABEL_REF:
case SYMBOL_REF:
*total = 3;
return true;
#define MULL_COST \
(TUNE_68060 ? 2 \
: TUNE_68040 ? 5 \
- : TUNE_CFV2 ? 10 \
+ : (TUNE_CFV2 && TUNE_EMAC) ? 3 \
+ : (TUNE_CFV2 && TUNE_MAC) ? 4 \
+ : TUNE_CFV2 ? 8 \
: TARGET_COLDFIRE ? 3 : 13)
#define MULW_COST \
(TUNE_68060 ? 2 \
: TUNE_68040 ? 3 \
- : TUNE_68000_10 || TUNE_CFV2 ? 5 \
+ : TUNE_68000_10 ? 5 \
+ : (TUNE_CFV2 && TUNE_EMAC) ? 3 \
+ : (TUNE_CFV2 && TUNE_MAC) ? 2 \
+ : TUNE_CFV2 ? 8 \
: TARGET_COLDFIRE ? 2 : 8)
#define DIVW_COST \
*total = COSTS_N_INSNS (43); /* div.l */
return true;
+ case ZERO_EXTRACT:
+ if (outer_code == COMPARE)
+ *total = 0;
+ return false;
+
default:
return false;
}
HOST_WIDE_INT i;
i = INTVAL (operands[1]);
- switch (const_method (i))
+ switch (m68k_const_method (i))
{
case MVZ:
return "mvzw %1,%0";
}
else if (CONSTANT_P (operands[1]))
return "move%.l %1,%0";
- /* Recognize the insn before a tablejump, one that refers
- to a table of offsets. Such an insn will need to refer
- to a label on the insn. So output one. Use the label-number
- of the table of offsets to generate this label. This code,
- and similar code below, assumes that there will be at most one
- reference to each table. */
- if (GET_CODE (operands[1]) == MEM
- && GET_CODE (XEXP (operands[1], 0)) == PLUS
- && GET_CODE (XEXP (XEXP (operands[1], 0), 1)) == LABEL_REF
- && GET_CODE (XEXP (XEXP (operands[1], 0), 0)) != PLUS)
- {
- rtx labelref = XEXP (XEXP (operands[1], 0), 1);
- if (MOTOROLA)
- asm_fprintf (asm_out_file, "\t.set %LLI%d,.+2\n",
- CODE_LABEL_NUMBER (XEXP (labelref, 0)));
- else
- (*targetm.asm_out.internal_label) (asm_out_file, "LI",
- CODE_LABEL_NUMBER (XEXP (labelref, 0)));
- }
return "move%.w %1,%0";
}
}
-/* Output assembler code to perform a doubleword move insn
- with operands OPERANDS. */
+/* Output assembler or rtl code to perform a doubleword move insn
+ with operands OPERANDS.
+ Pointers to 3 helper functions should be specified:
+ HANDLE_REG_ADJUST to adjust a register by a small value,
+ HANDLE_COMPADR to compute an address and
+ HANDLE_MOVSI to move 4 bytes. */
-const char *
-output_move_double (rtx *operands)
+static void
+handle_move_double (rtx operands[2],
+ void (*handle_reg_adjust) (rtx, int),
+ void (*handle_compadr) (rtx [2]),
+ void (*handle_movsi) (rtx [2]))
{
enum
{
if (optype0 == PUSHOP && optype1 == POPOP)
{
operands[0] = XEXP (XEXP (operands[0], 0), 0);
- if (size == 12)
- output_asm_insn ("sub%.l #12,%0", operands);
- else
- output_asm_insn ("subq%.l #8,%0", operands);
+
+ handle_reg_adjust (operands[0], -size);
+
if (GET_MODE (operands[1]) == XFmode)
operands[0] = gen_rtx_MEM (XFmode, operands[0]);
else if (GET_MODE (operands[0]) == DFmode)
if (optype0 == POPOP && optype1 == PUSHOP)
{
operands[1] = XEXP (XEXP (operands[1], 0), 0);
- if (size == 12)
- output_asm_insn ("sub%.l #12,%1", operands);
- else
- output_asm_insn ("subq%.l #8,%1", operands);
+
+ handle_reg_adjust (operands[1], -size);
+
if (GET_MODE (operands[1]) == XFmode)
operands[1] = gen_rtx_MEM (XFmode, operands[1]);
else if (GET_MODE (operands[1]) == DFmode)
}
else
{
- middlehalf[0] = operands[0];
- latehalf[0] = operands[0];
+ middlehalf[0] = adjust_address (operands[0], SImode, 0);
+ latehalf[0] = adjust_address (operands[0], SImode, 0);
}
if (optype1 == REGOP)
}
else
{
- middlehalf[1] = operands[1];
- latehalf[1] = operands[1];
+ middlehalf[1] = adjust_address (operands[1], SImode, 0);
+ latehalf[1] = adjust_address (operands[1], SImode, 0);
}
}
else
else if (optype0 == OFFSOP)
latehalf[0] = adjust_address (operands[0], SImode, size - 4);
else
- latehalf[0] = operands[0];
+ latehalf[0] = adjust_address (operands[0], SImode, 0);
if (optype1 == REGOP)
latehalf[1] = gen_rtx_REG (SImode, REGNO (operands[1]) + 1);
else if (optype1 == CNSTOP)
split_double (operands[1], &operands[1], &latehalf[1]);
else
- latehalf[1] = operands[1];
+ latehalf[1] = adjust_address (operands[1], SImode, 0);
}
/* If insn is effectively movd N(sp),-(sp) then we will do the
compadr:
xops[0] = latehalf[0];
xops[1] = XEXP (operands[1], 0);
- output_asm_insn ("lea %a1,%0", xops);
- if (GET_MODE (operands[1]) == XFmode )
+
+ handle_compadr (xops);
+ if (GET_MODE (operands[1]) == XFmode)
{
operands[1] = gen_rtx_MEM (XFmode, latehalf[0]);
middlehalf[1] = adjust_address (operands[1], DImode, size - 8);
gcc_assert (!addreg0 && !addreg1);
/* Only the middle reg conflicts; simply put it last. */
- output_asm_insn (singlemove_string (operands), operands);
- output_asm_insn (singlemove_string (latehalf), latehalf);
- output_asm_insn (singlemove_string (middlehalf), middlehalf);
- return "";
+ handle_movsi (operands);
+ handle_movsi (latehalf);
+ handle_movsi (middlehalf);
+
+ return;
}
else if (reg_overlap_mentioned_p (testlow, XEXP (operands[1], 0)))
/* If the low half of dest is mentioned in the source memory
{
/* Make any unoffsettable addresses point at high-numbered word. */
if (addreg0)
- {
- if (size == 12)
- output_asm_insn ("addq%.l #8,%0", &addreg0);
- else
- output_asm_insn ("addq%.l #4,%0", &addreg0);
- }
+ handle_reg_adjust (addreg0, size - 4);
if (addreg1)
- {
- if (size == 12)
- output_asm_insn ("addq%.l #8,%0", &addreg1);
- else
- output_asm_insn ("addq%.l #4,%0", &addreg1);
- }
+ handle_reg_adjust (addreg1, size - 4);
/* Do that word. */
- output_asm_insn (singlemove_string (latehalf), latehalf);
+ handle_movsi (latehalf);
/* Undo the adds we just did. */
if (addreg0)
- output_asm_insn ("subq%.l #4,%0", &addreg0);
+ handle_reg_adjust (addreg0, -4);
if (addreg1)
- output_asm_insn ("subq%.l #4,%0", &addreg1);
+ handle_reg_adjust (addreg1, -4);
if (size == 12)
{
- output_asm_insn (singlemove_string (middlehalf), middlehalf);
+ handle_movsi (middlehalf);
+
if (addreg0)
- output_asm_insn ("subq%.l #4,%0", &addreg0);
+ handle_reg_adjust (addreg0, -4);
if (addreg1)
- output_asm_insn ("subq%.l #4,%0", &addreg1);
+ handle_reg_adjust (addreg1, -4);
}
/* Do low-numbered word. */
- return singlemove_string (operands);
+
+ handle_movsi (operands);
+ return;
}
/* Normal case: do the two words, low-numbered first. */
- output_asm_insn (singlemove_string (operands), operands);
+ handle_movsi (operands);
/* Do the middle one of the three words for long double */
if (size == 12)
{
if (addreg0)
- output_asm_insn ("addq%.l #4,%0", &addreg0);
+ handle_reg_adjust (addreg0, 4);
if (addreg1)
- output_asm_insn ("addq%.l #4,%0", &addreg1);
+ handle_reg_adjust (addreg1, 4);
- output_asm_insn (singlemove_string (middlehalf), middlehalf);
+ handle_movsi (middlehalf);
}
/* Make any unoffsettable addresses point at high-numbered word. */
if (addreg0)
- output_asm_insn ("addq%.l #4,%0", &addreg0);
+ handle_reg_adjust (addreg0, 4);
if (addreg1)
- output_asm_insn ("addq%.l #4,%0", &addreg1);
+ handle_reg_adjust (addreg1, 4);
/* Do that word. */
- output_asm_insn (singlemove_string (latehalf), latehalf);
+ handle_movsi (latehalf);
/* Undo the adds we just did. */
if (addreg0)
+ handle_reg_adjust (addreg0, -(size - 4));
+ if (addreg1)
+ handle_reg_adjust (addreg1, -(size - 4));
+
+ return;
+}
+
+/* Output assembler code to adjust REG by N. */
+static void
+output_reg_adjust (rtx reg, int n)
+{
+ const char *s;
+
+ gcc_assert (GET_MODE (reg) == SImode
+ && -12 <= n && n != 0 && n <= 12);
+
+ switch (n)
{
- if (size == 12)
- output_asm_insn ("subq%.l #8,%0", &addreg0);
- else
- output_asm_insn ("subq%.l #4,%0", &addreg0);
+ case 12:
+ s = "add%.l #12,%0";
+ break;
+
+ case 8:
+ s = "addq%.l #8,%0";
+ break;
+
+ case 4:
+ s = "addq%.l #4,%0";
+ break;
+
+ case -12:
+ s = "sub%.l #12,%0";
+ break;
+
+ case -8:
+ s = "subq%.l #8,%0";
+ break;
+
+ case -4:
+ s = "subq%.l #4,%0";
+ break;
+
+ default:
+ gcc_unreachable ();
+ s = NULL;
}
- if (addreg1)
+
+ output_asm_insn (s, ®);
+}
+
+/* Emit rtl code to adjust REG by N. */
+static void
+emit_reg_adjust (rtx reg1, int n)
+{
+ rtx reg2;
+
+ gcc_assert (GET_MODE (reg1) == SImode
+ && -12 <= n && n != 0 && n <= 12);
+
+ reg1 = copy_rtx (reg1);
+ reg2 = copy_rtx (reg1);
+
+ if (n < 0)
+ emit_insn (gen_subsi3 (reg1, reg2, GEN_INT (-n)));
+ else if (n > 0)
+ emit_insn (gen_addsi3 (reg1, reg2, GEN_INT (n)));
+ else
+ gcc_unreachable ();
+}
+
+/* Output assembler to load address OPERANDS[0] to register OPERANDS[1]. */
+static void
+output_compadr (rtx operands[2])
+{
+ output_asm_insn ("lea %a1,%0", operands);
+}
+
+/* Output the best assembler insn for moving operands[1] into operands[0]
+ as a fullword. */
+static void
+output_movsi (rtx operands[2])
+{
+ output_asm_insn (singlemove_string (operands), operands);
+}
+
+/* Copy OP and change its mode to MODE. */
+static rtx
+copy_operand (rtx op, enum machine_mode mode)
+{
+ /* ??? This looks really ugly. There must be a better way
+ to change a mode on the operand. */
+ if (GET_MODE (op) != VOIDmode)
{
- if (size == 12)
- output_asm_insn ("subq%.l #8,%0", &addreg1);
+ if (REG_P (op))
+ op = gen_rtx_REG (mode, REGNO (op));
else
- output_asm_insn ("subq%.l #4,%0", &addreg1);
+ {
+ op = copy_rtx (op);
+ PUT_MODE (op, mode);
+ }
}
+ return op;
+}
+
+/* Emit rtl code for moving operands[1] into operands[0] as a fullword. */
+static void
+emit_movsi (rtx operands[2])
+{
+ operands[0] = copy_operand (operands[0], SImode);
+ operands[1] = copy_operand (operands[1], SImode);
+
+ emit_insn (gen_movsi (operands[0], operands[1]));
+}
+
+/* Output assembler code to perform a doubleword move insn
+ with operands OPERANDS. */
+const char *
+output_move_double (rtx *operands)
+{
+ handle_move_double (operands,
+ output_reg_adjust, output_compadr, output_movsi);
+
return "";
}
+/* Output rtl code to perform a doubleword move insn
+ with operands OPERANDS. */
+void
+m68k_emit_move_double (rtx operands[2])
+{
+ handle_move_double (operands, emit_reg_adjust, emit_movsi, emit_movsi);
+}
/* Ensure mode of ORIG, a REG rtx, is MODE. Returns either ORIG or a
new rtx with the correct mode. */
}
}
-/* Return a REG that occurs in ADDR with coefficient 1.
- ADDR can be effectively incremented by incrementing REG. */
+/* Split X into a base and a constant offset, storing them in *BASE
+ and *OFFSET respectively. */
-static rtx
-find_addr_reg (rtx addr)
+static void
+m68k_split_offset (rtx x, rtx *base, HOST_WIDE_INT *offset)
{
- while (GET_CODE (addr) == PLUS)
+ *offset = 0;
+ if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 1)) == CONST_INT)
{
- if (GET_CODE (XEXP (addr, 0)) == REG)
- addr = XEXP (addr, 0);
- else if (GET_CODE (XEXP (addr, 1)) == REG)
- addr = XEXP (addr, 1);
- else if (CONSTANT_P (XEXP (addr, 0)))
- addr = XEXP (addr, 1);
- else if (CONSTANT_P (XEXP (addr, 1)))
- addr = XEXP (addr, 0);
- else
- gcc_unreachable ();
+ *offset += INTVAL (XEXP (x, 1));
+ x = XEXP (x, 0);
}
- gcc_assert (GET_CODE (addr) == REG);
- return addr;
+ *base = x;
}
-/* Output assembler code to perform a 32-bit 3-operand add. */
+/* Return true if PATTERN is a PARALLEL suitable for a movem or fmovem
+ instruction. STORE_P says whether the move is a load or store.
-const char *
-output_addsi3 (rtx *operands)
-{
+ If the instruction uses post-increment or pre-decrement addressing,
+ AUTOMOD_BASE is the base register and AUTOMOD_OFFSET is the total
+ adjustment. This adjustment will be made by the first element of
+ PARALLEL, with the loads or stores starting at element 1. If the
+ instruction does not use post-increment or pre-decrement addressing,
+ AUTOMOD_BASE is null, AUTOMOD_OFFSET is 0, and the loads or stores
+ start at element 0. */
+
+bool
+m68k_movem_pattern_p (rtx pattern, rtx automod_base,
+ HOST_WIDE_INT automod_offset, bool store_p)
+{
+ rtx base, mem_base, set, mem, reg, last_reg;
+ HOST_WIDE_INT offset, mem_offset;
+ int i, first, len;
+ enum reg_class rclass;
+
+ len = XVECLEN (pattern, 0);
+ first = (automod_base != NULL);
+
+ if (automod_base)
+ {
+ /* Stores must be pre-decrement and loads must be post-increment. */
+ if (store_p != (automod_offset < 0))
+ return false;
+
+ /* Work out the base and offset for lowest memory location. */
+ base = automod_base;
+ offset = (automod_offset < 0 ? automod_offset : 0);
+ }
+ else
+ {
+ /* Allow any valid base and offset in the first access. */
+ base = NULL;
+ offset = 0;
+ }
+
+ last_reg = NULL;
+ rclass = NO_REGS;
+ for (i = first; i < len; i++)
+ {
+ /* We need a plain SET. */
+ set = XVECEXP (pattern, 0, i);
+ if (GET_CODE (set) != SET)
+ return false;
+
+ /* Check that we have a memory location... */
+ mem = XEXP (set, !store_p);
+ if (!MEM_P (mem) || !memory_operand (mem, VOIDmode))
+ return false;
+
+ /* ...with the right address. */
+ if (base == NULL)
+ {
+ m68k_split_offset (XEXP (mem, 0), &base, &offset);
+ /* The ColdFire instruction only allows (An) and (d16,An) modes.
+ There are no mode restrictions for 680x0 besides the
+ automodification rules enforced above. */
+ if (TARGET_COLDFIRE
+ && !m68k_legitimate_base_reg_p (base, reload_completed))
+ return false;
+ }
+ else
+ {
+ m68k_split_offset (XEXP (mem, 0), &mem_base, &mem_offset);
+ if (!rtx_equal_p (base, mem_base) || offset != mem_offset)
+ return false;
+ }
+
+ /* Check that we have a register of the required mode and class. */
+ reg = XEXP (set, store_p);
+ if (!REG_P (reg)
+ || !HARD_REGISTER_P (reg)
+ || GET_MODE (reg) != reg_raw_mode[REGNO (reg)])
+ return false;
+
+ if (last_reg)
+ {
+ /* The register must belong to RCLASS and have a higher number
+ than the register in the previous SET. */
+ if (!TEST_HARD_REG_BIT (reg_class_contents[rclass], REGNO (reg))
+ || REGNO (last_reg) >= REGNO (reg))
+ return false;
+ }
+ else
+ {
+ /* Work out which register class we need. */
+ if (INT_REGNO_P (REGNO (reg)))
+ rclass = GENERAL_REGS;
+ else if (FP_REGNO_P (REGNO (reg)))
+ rclass = FP_REGS;
+ else
+ return false;
+ }
+
+ last_reg = reg;
+ offset += GET_MODE_SIZE (GET_MODE (reg));
+ }
+
+ /* If we have an automodification, check whether the final offset is OK. */
+ if (automod_base && offset != (automod_offset < 0 ? 0 : automod_offset))
+ return false;
+
+ /* Reject unprofitable cases. */
+ if (len < first + (rclass == FP_REGS ? MIN_FMOVEM_REGS : MIN_MOVEM_REGS))
+ return false;
+
+ return true;
+}
+
+/* Return the assembly code template for a movem or fmovem instruction
+ whose pattern is given by PATTERN. Store the template's operands
+ in OPERANDS.
+
+ If the instruction uses post-increment or pre-decrement addressing,
+ AUTOMOD_OFFSET is the total adjustment, otherwise it is 0. STORE_P
+ is true if this is a store instruction. */
+
+const char *
+m68k_output_movem (rtx *operands, rtx pattern,
+ HOST_WIDE_INT automod_offset, bool store_p)
+{
+ unsigned int mask;
+ int i, first;
+
+ gcc_assert (GET_CODE (pattern) == PARALLEL);
+ mask = 0;
+ first = (automod_offset != 0);
+ for (i = first; i < XVECLEN (pattern, 0); i++)
+ {
+ /* When using movem with pre-decrement addressing, register X + D0_REG
+ is controlled by bit 15 - X. For all other addressing modes,
+ register X + D0_REG is controlled by bit X. Confusingly, the
+ register mask for fmovem is in the opposite order to that for
+ movem. */
+ unsigned int regno;
+
+ gcc_assert (MEM_P (XEXP (XVECEXP (pattern, 0, i), !store_p)));
+ gcc_assert (REG_P (XEXP (XVECEXP (pattern, 0, i), store_p)));
+ regno = REGNO (XEXP (XVECEXP (pattern, 0, i), store_p));
+ if (automod_offset < 0)
+ {
+ if (FP_REGNO_P (regno))
+ mask |= 1 << (regno - FP0_REG);
+ else
+ mask |= 1 << (15 - (regno - D0_REG));
+ }
+ else
+ {
+ if (FP_REGNO_P (regno))
+ mask |= 1 << (7 - (regno - FP0_REG));
+ else
+ mask |= 1 << (regno - D0_REG);
+ }
+ }
+ CC_STATUS_INIT;
+
+ if (automod_offset == 0)
+ operands[0] = XEXP (XEXP (XVECEXP (pattern, 0, first), !store_p), 0);
+ else if (automod_offset < 0)
+ operands[0] = gen_rtx_PRE_DEC (Pmode, SET_DEST (XVECEXP (pattern, 0, 0)));
+ else
+ operands[0] = gen_rtx_POST_INC (Pmode, SET_DEST (XVECEXP (pattern, 0, 0)));
+ operands[1] = GEN_INT (mask);
+ if (FP_REGNO_P (REGNO (XEXP (XVECEXP (pattern, 0, first), store_p))))
+ {
+ if (store_p)
+ return "fmovem %1,%a0";
+ else
+ return "fmovem %a0,%1";
+ }
+ else
+ {
+ if (store_p)
+ return "movem%.l %1,%a0";
+ else
+ return "movem%.l %a0,%1";
+ }
+}
+
+/* Return a REG that occurs in ADDR with coefficient 1.
+ ADDR can be effectively incremented by incrementing REG. */
+
+static rtx
+find_addr_reg (rtx addr)
+{
+ while (GET_CODE (addr) == PLUS)
+ {
+ if (GET_CODE (XEXP (addr, 0)) == REG)
+ addr = XEXP (addr, 0);
+ else if (GET_CODE (XEXP (addr, 1)) == REG)
+ addr = XEXP (addr, 1);
+ else if (CONSTANT_P (XEXP (addr, 0)))
+ addr = XEXP (addr, 1);
+ else if (CONSTANT_P (XEXP (addr, 1)))
+ addr = XEXP (addr, 0);
+ else
+ gcc_unreachable ();
+ }
+ gcc_assert (GET_CODE (addr) == REG);
+ return addr;
+}
+
+/* Output assembler code to perform a 32-bit 3-operand add. */
+
+const char *
+output_addsi3 (rtx *operands)
+{
if (! operands_match_p (operands[0], operands[1]))
{
if (!ADDRESS_REG_P (operands[1]))
case ROTATE: case ROTATERT:
/* These instructions always clear the overflow bit, and set
the carry to the bit shifted out. */
- /* ??? We don't currently have a way to signal carry not valid,
- nor do we check for it in the branch insns. */
- CC_STATUS_INIT;
+ cc_status.flags |= CC_OVERFLOW_UNUSABLE | CC_NO_CARRY;
break;
case PLUS: case MINUS: case MULT:
if (((cc_status.value1 && FP_REG_P (cc_status.value1))
|| (cc_status.value2 && FP_REG_P (cc_status.value2))))
cc_status.flags = CC_IN_68881;
+ if (cc_status.value2 && GET_CODE (cc_status.value2) == COMPARE
+ && GET_MODE_CLASS (GET_MODE (XEXP (cc_status.value2, 0))) == MODE_FLOAT)
+ {
+ cc_status.flags = CC_IN_68881;
+ if (!FP_REG_P (XEXP (cc_status.value2, 0)))
+ cc_status.flags |= CC_REVERSED;
+ }
}
\f
const char *
return 0;
exp = real_exponent (&r);
- real_2expN (&r1, exp);
+ real_2expN (&r1, exp, DFmode);
if (REAL_VALUES_EQUAL (r1, r))
return exp;
'$' for the letter `s' in an op code, but only on the 68040.
'&' for the letter `d' in an op code, but only on the 68040.
'/' for register prefix needed by longlong.h.
+ '?' for m68k_library_id_string
'b' for byte insn (no effect, on the Sun; this is for the ISI).
'd' to force memory addressing to be absolute, not relative.
}
else if (letter == '/')
asm_fprintf (file, "%R");
+ else if (letter == '?')
+ asm_fprintf (file, m68k_library_id_string);
else if (letter == 'p')
{
output_addr_const (file, op);
else if (GET_CODE (op) == CONST_DOUBLE && GET_MODE (op) == SFmode)
{
REAL_VALUE_TYPE r;
+ long l;
REAL_VALUE_FROM_CONST_DOUBLE (r, op);
- ASM_OUTPUT_FLOAT_OPERAND (letter, file, r);
+ REAL_VALUE_TO_TARGET_SINGLE (r, l);
+ asm_fprintf (file, "%I0x%lx", l & 0xFFFFFFFF);
}
else if (GET_CODE (op) == CONST_DOUBLE && GET_MODE (op) == XFmode)
{
REAL_VALUE_TYPE r;
+ long l[3];
REAL_VALUE_FROM_CONST_DOUBLE (r, op);
- ASM_OUTPUT_LONG_DOUBLE_OPERAND (file, r);
+ REAL_VALUE_TO_TARGET_LONG_DOUBLE (r, l);
+ asm_fprintf (file, "%I0x%lx%08lx%08lx", l[0] & 0xFFFFFFFF,
+ l[1] & 0xFFFFFFFF, l[2] & 0xFFFFFFFF);
}
else if (GET_CODE (op) == CONST_DOUBLE && GET_MODE (op) == DFmode)
{
REAL_VALUE_TYPE r;
+ long l[2];
REAL_VALUE_FROM_CONST_DOUBLE (r, op);
- ASM_OUTPUT_DOUBLE_OPERAND (file, r);
+ REAL_VALUE_TO_TARGET_DOUBLE (r, l);
+ asm_fprintf (file, "%I0x%lx%08lx", l[0] & 0xFFFFFFFF, l[1] & 0xFFFFFFFF);
}
else
{
}
}
+/* Return string for TLS relocation RELOC. */
+
+static const char *
+m68k_get_reloc_decoration (enum m68k_reloc reloc)
+{
+ /* To my knowledge, !MOTOROLA assemblers don't support TLS. */
+ gcc_assert (MOTOROLA || reloc == RELOC_GOT);
+
+ switch (reloc)
+ {
+ case RELOC_GOT:
+ if (MOTOROLA)
+ {
+ if (flag_pic == 1 && TARGET_68020)
+ return "@GOT.w";
+ else
+ return "@GOT";
+ }
+ else
+ {
+ if (TARGET_68020)
+ {
+ switch (flag_pic)
+ {
+ case 1:
+ return ":w";
+ case 2:
+ return ":l";
+ default:
+ return "";
+ }
+ }
+ }
+
+ case RELOC_TLSGD:
+ return "@TLSGD";
+
+ case RELOC_TLSLDM:
+ return "@TLSLDM";
+
+ case RELOC_TLSLDO:
+ return "@TLSLDO";
+
+ case RELOC_TLSIE:
+ return "@TLSIE";
+
+ case RELOC_TLSLE:
+ return "@TLSLE";
+
+ default:
+ gcc_unreachable ();
+ }
+}
+
+/* m68k implementation of OUTPUT_ADDR_CONST_EXTRA. */
+
+bool
+m68k_output_addr_const_extra (FILE *file, rtx x)
+{
+ if (GET_CODE (x) == UNSPEC)
+ {
+ switch (XINT (x, 1))
+ {
+ case UNSPEC_RELOC16:
+ case UNSPEC_RELOC32:
+ output_addr_const (file, XVECEXP (x, 0, 0));
+ fputs (m68k_get_reloc_decoration (INTVAL (XVECEXP (x, 0, 1))), file);
+ return true;
+
+ default:
+ break;
+ }
+ }
+
+ return false;
+}
+
+/* M68K implementation of TARGET_ASM_OUTPUT_DWARF_DTPREL. */
+
+static void
+m68k_output_dwarf_dtprel (FILE *file, int size, rtx x)
+{
+ gcc_assert (size == 4);
+ fputs ("\t.long\t", file);
+ output_addr_const (file, x);
+ fputs ("@TLSLDO+0x8000", file);
+}
+
\f
/* A C compound statement to output to stdio stream STREAM the
assembler syntax for an instruction operand that is a memory
It is possible for PIC to generate a (plus (label_ref...) (reg...))
and we handle that just like we would a (plus (symbol_ref...) (reg...)).
- Some SGS assemblers have a bug such that "Lnnn-LInnn-2.b(pc,d0.l*2)"
- fails to assemble. Luckily "Lnnn(pc,d0.l*2)" produces the results
- we want. This difference can be accommodated by using an assembler
- define such "LDnnn" to be either "Lnnn-LInnn-2.b", "Lnnn", or any other
- string, as necessary. This is accomplished via the ASM_OUTPUT_CASE_END
- macro. See m68k/sgs.h for an example; for versions without the bug.
- Some assemblers refuse all the above solutions. The workaround is to
- emit "K(pc,d0.l*2)" with K being a small constant known to give the
- right behavior.
-
- They also do not like things like "pea 1.w", so we simple leave off
- the .w on small constants.
-
This routine is responsible for distinguishing between -fpic and -fPIC
style relocations in an address. When generating -fpic code the
offset is output in word mode (e.g. movel a5@(_foo:w), a0). When generating
int labelno;
/* If ADDR is a (d8,pc,Xn) address, this is the number of the
- label being acceesed, otherwise it is -1. */
+ label being accessed, otherwise it is -1. */
labelno = (address.offset
&& !address.base
&& GET_CODE (address.offset) == LABEL_REF
{
/* Print the "offset(base" component. */
if (labelno >= 0)
- asm_fprintf (file, "%LL%d-%LLI%d.b(%Rpc,", labelno, labelno);
+ asm_fprintf (file, "%LL%d(%Rpc,", labelno);
else
{
if (address.offset)
- {
- output_addr_const (file, address.offset);
- if (flag_pic && address.base == pic_offset_table_rtx)
- {
- fprintf (file, "@GOT");
- if (flag_pic == 1 && TARGET_68020)
- fprintf (file, ".w");
- }
- }
+ output_addr_const (file, address.offset);
+
putc ('(', file);
if (address.base)
fputs (M68K_REGNAME (REGNO (address.base)), file);
{
/* Print the "base@(offset" component. */
if (labelno >= 0)
- asm_fprintf (file, "%Rpc@(%LL%d-%LLI%d-2:b", labelno, labelno);
+ asm_fprintf (file, "%Rpc@(%LL%d", labelno);
else
{
if (address.base)
fputs (M68K_REGNAME (REGNO (address.base)), file);
fprintf (file, "@(");
if (address.offset)
- {
- output_addr_const (file, address.offset);
- if (address.base == pic_offset_table_rtx && TARGET_68020)
- switch (flag_pic)
- {
- case 1:
- fprintf (file, ":w"); break;
- case 2:
- fprintf (file, ":l"); break;
- default:
- break;
- }
- }
+ output_addr_const (file, address.offset);
}
/* Print the ",index" component, if any. */
if (address.index)
strict_low_part_peephole_ok (enum machine_mode mode, rtx first_insn,
rtx target)
{
- rtx p;
-
- p = prev_nonnote_insn (first_insn);
+ rtx p = first_insn;
- while (p)
+ while ((p = PREV_INSN (p)))
{
+ if (NOTE_INSN_BASIC_BLOCK_P (p))
+ return false;
+
+ if (NOTE_P (p))
+ continue;
+
/* If it isn't an insn, then give up. */
- if (GET_CODE (p) != INSN)
+ if (!INSN_P (p))
return false;
if (reg_set_p (target, p))
else
return false;
}
-
- p = prev_nonnote_insn (p);
}
return false;
simple fact that the m68k does not allow a pc-relative addressing
mode as a destination. gcc does not distinguish between source and
destination addresses. Hence, if we claim that pc-relative address
- modes are valid, e.g. GO_IF_LEGITIMATE_ADDRESS accepts them, then we
+ modes are valid, e.g. TARGET_LEGITIMATE_ADDRESS_P accepts them, then we
end up with invalid code. To get around this problem, we left
pc-relative modes as invalid addresses, and then added special
predicates and constraints to accept them.
return "jsr %a0";
}
-#ifdef M68K_TARGET_COFF
+/* Likewise sibling calls. */
-/* Output assembly to switch to section NAME with attribute FLAGS. */
-
-static void
-m68k_coff_asm_named_section (const char *name, unsigned int flags,
- tree decl ATTRIBUTE_UNUSED)
+const char *
+output_sibcall (rtx x)
{
- char flagchar;
-
- if (flags & SECTION_WRITE)
- flagchar = 'd';
+ if (symbolic_operand (x, VOIDmode))
+ return m68k_symbolic_jump;
else
- flagchar = 'x';
-
- fprintf (asm_out_file, "\t.section\t%s,\"%c\"\n", name, flagchar);
+ return "jmp %a0";
}
-#endif /* M68K_TARGET_COFF */
-
static void
m68k_output_mi_thunk (FILE *file, tree thunk ATTRIBUTE_UNUSED,
- HOST_WIDE_INT delta,
- HOST_WIDE_INT vcall_offset ATTRIBUTE_UNUSED,
+ HOST_WIDE_INT delta, HOST_WIDE_INT vcall_offset,
tree function)
{
- rtx xops[1];
- const char *fmt;
-
- if (delta > 0 && delta <= 8)
- asm_fprintf (file, (MOTOROLA
- ? "\taddq.l %I%d,4(%Rsp)\n"
- : "\taddql %I%d,%Rsp@(4)\n"),
- (int) delta);
- else if (delta < 0 && delta >= -8)
- asm_fprintf (file, (MOTOROLA
- ? "\tsubq.l %I%d,4(%Rsp)\n"
- : "\tsubql %I%d,%Rsp@(4)\n"),
- (int) -delta);
- else if (TARGET_COLDFIRE)
- {
- /* ColdFire can't add/sub a constant to memory unless it is in
- the range of addq/subq. So load the value into %d0 and
- then add it to 4(%sp). */
- if (delta >= -128 && delta <= 127)
- asm_fprintf (file, (MOTOROLA
- ? "\tmoveq.l %I%wd,%Rd0\n"
- : "\tmoveql %I%wd,%Rd0\n"),
- delta);
- else
- asm_fprintf (file, (MOTOROLA
- ? "\tmove.l %I%wd,%Rd0\n"
- : "\tmovel %I%wd,%Rd0\n"),
- delta);
- asm_fprintf (file, (MOTOROLA
- ? "\tadd.l %Rd0,4(%Rsp)\n"
- : "\taddl %Rd0,%Rsp@(4)\n"));
+ rtx this_slot, offset, addr, mem, insn;
+
+ /* Pretend to be a post-reload pass while generating rtl. */
+ reload_completed = 1;
+
+ /* The "this" pointer is stored at 4(%sp). */
+ this_slot = gen_rtx_MEM (Pmode, plus_constant (stack_pointer_rtx, 4));
+
+ /* Add DELTA to THIS. */
+ if (delta != 0)
+ {
+ /* Make the offset a legitimate operand for memory addition. */
+ offset = GEN_INT (delta);
+ if ((delta < -8 || delta > 8)
+ && (TARGET_COLDFIRE || USE_MOVQ (delta)))
+ {
+ emit_move_insn (gen_rtx_REG (Pmode, D0_REG), offset);
+ offset = gen_rtx_REG (Pmode, D0_REG);
+ }
+ emit_insn (gen_add3_insn (copy_rtx (this_slot),
+ copy_rtx (this_slot), offset));
}
- else
- asm_fprintf (file, (MOTOROLA
- ? "\tadd.l %I%wd,4(%Rsp)\n"
- : "\taddl %I%wd,%Rsp@(4)\n"),
- delta);
- xops[0] = DECL_RTL (function);
+ /* If needed, add *(*THIS + VCALL_OFFSET) to THIS. */
+ if (vcall_offset != 0)
+ {
+ /* Set the static chain register to *THIS. */
+ emit_move_insn (static_chain_rtx, this_slot);
+ emit_move_insn (static_chain_rtx, gen_rtx_MEM (Pmode, static_chain_rtx));
+
+ /* Set ADDR to a legitimate address for *THIS + VCALL_OFFSET. */
+ addr = plus_constant (static_chain_rtx, vcall_offset);
+ if (!m68k_legitimate_address_p (Pmode, addr, true))
+ {
+ emit_insn (gen_rtx_SET (VOIDmode, static_chain_rtx, addr));
+ addr = static_chain_rtx;
+ }
- gcc_assert (MEM_P (xops[0])
- && symbolic_operand (XEXP (xops[0], 0), VOIDmode));
- xops[0] = XEXP (xops[0], 0);
+ /* Load the offset into %d0 and add it to THIS. */
+ emit_move_insn (gen_rtx_REG (Pmode, D0_REG),
+ gen_rtx_MEM (Pmode, addr));
+ emit_insn (gen_add3_insn (copy_rtx (this_slot),
+ copy_rtx (this_slot),
+ gen_rtx_REG (Pmode, D0_REG)));
+ }
- fmt = m68k_symbolic_jump;
- if (m68k_symbolic_jump == NULL)
- fmt = "move.l %%a1@GOT(%%a5), %%a1\n\tjmp (%%a1)";
+ /* Jump to the target function. Use a sibcall if direct jumps are
+ allowed, otherwise load the address into a register first. */
+ mem = DECL_RTL (function);
+ if (!sibcall_operand (XEXP (mem, 0), VOIDmode))
+ {
+ gcc_assert (flag_pic);
- output_asm_insn (fmt, xops);
+ if (!TARGET_SEP_DATA)
+ {
+ /* Use the static chain register as a temporary (call-clobbered)
+ GOT pointer for this function. We can use the static chain
+ register because it isn't live on entry to the thunk. */
+ SET_REGNO (pic_offset_table_rtx, STATIC_CHAIN_REGNUM);
+ emit_insn (gen_load_got (pic_offset_table_rtx));
+ }
+ legitimize_pic_address (XEXP (mem, 0), Pmode, static_chain_rtx);
+ mem = replace_equiv_address (mem, static_chain_rtx);
+ }
+ insn = emit_call_insn (gen_sibcall (mem, const0_rtx));
+ SIBLING_CALL_P (insn) = 1;
+
+ /* Run just enough of rest_of_compilation. */
+ insn = get_insns ();
+ split_all_insns_noflow ();
+ final_start_function (insn, file, 1);
+ final (insn, file, 1);
+ final_end_function ();
+
+ /* Clean up the vars set above. */
+ reload_completed = 0;
+
+ /* Restore the original PIC register. */
+ if (flag_pic)
+ SET_REGNO (pic_offset_table_rtx, PIC_REG);
+ free_after_compilation (cfun);
}
/* Worker function for TARGET_STRUCT_VALUE_RTX. */
saved by the prologue, even if they would normally be
call-clobbered. */
- if (m68k_interrupt_function_p (current_function_decl)
- && !regs_ever_live[new_reg])
+ if ((m68k_get_function_kind (current_function_decl)
+ == m68k_fk_interrupt_handler)
+ && !df_regs_ever_live_p (new_reg))
return 0;
return 1;
}
-/* Value is true if hard register REGNO can hold a value of machine-mode MODE.
- On the 68000, the cpu registers can hold any mode except bytes in address
- registers, but the 68881 registers can hold only SFmode or DFmode. */
+/* Value is true if hard register REGNO can hold a value of machine-mode
+ MODE. On the 68000, we let the cpu registers can hold any mode, but
+ restrict the 68881 registers to floating-point modes. */
+
bool
m68k_regno_mode_ok (int regno, enum machine_mode mode)
{
}
else if (ADDRESS_REGNO_P (regno))
{
- /* Address Registers, can't hold bytes, can hold aggregate if
- fits in. */
- if (GET_MODE_SIZE (mode) == 1)
- return false;
if (regno + GET_MODE_SIZE (mode) / 4 <= 16)
return true;
}
return false;
}
+/* Implement SECONDARY_RELOAD_CLASS. */
+
+enum reg_class
+m68k_secondary_reload_class (enum reg_class rclass,
+ enum machine_mode mode, rtx x)
+{
+ int regno;
+
+ regno = true_regnum (x);
+
+ /* If one operand of a movqi is an address register, the other
+ operand must be a general register or constant. Other types
+ of operand must be reloaded through a data register. */
+ if (GET_MODE_SIZE (mode) == 1
+ && reg_classes_intersect_p (rclass, ADDR_REGS)
+ && !(INT_REGNO_P (regno) || CONSTANT_P (x)))
+ return DATA_REGS;
+
+ /* PC-relative addresses must be loaded into an address register first. */
+ if (TARGET_PCREL
+ && !reg_class_subset_p (rclass, ADDR_REGS)
+ && symbolic_operand (x, VOIDmode))
+ return ADDR_REGS;
+
+ return NO_REGS;
+}
+
+/* Implement PREFERRED_RELOAD_CLASS. */
+
+enum reg_class
+m68k_preferred_reload_class (rtx x, enum reg_class rclass)
+{
+ enum reg_class secondary_class;
+
+ /* If RCLASS might need a secondary reload, try restricting it to
+ a class that doesn't. */
+ secondary_class = m68k_secondary_reload_class (rclass, GET_MODE (x), x);
+ if (secondary_class != NO_REGS
+ && reg_class_subset_p (secondary_class, rclass))
+ return secondary_class;
+
+ /* Prefer to use moveq for in-range constants. */
+ if (GET_CODE (x) == CONST_INT
+ && reg_class_subset_p (DATA_REGS, rclass)
+ && IN_RANGE (INTVAL (x), -0x80, 0x7f))
+ return DATA_REGS;
+
+ /* ??? Do we really need this now? */
+ if (GET_CODE (x) == CONST_DOUBLE
+ && GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT)
+ {
+ if (TARGET_HARD_FLOAT && reg_class_subset_p (FP_REGS, rclass))
+ return FP_REGS;
+
+ return NO_REGS;
+ }
+
+ return rclass;
+}
+
/* Return floating point values in a 68881 register. This makes 68881 code
a little bit faster. It also makes -msoft-float code incompatible with
hard-float code, so people have to be careful not to mix the two.
case DFmode:
case XFmode:
if (TARGET_68881)
- return gen_rtx_REG (mode, 16);
+ return gen_rtx_REG (mode, FP0_REG);
break;
default:
break;
}
- return gen_rtx_REG (mode, 0);
+
+ return gen_rtx_REG (mode, m68k_libcall_value_in_a0_p ? A0_REG : D0_REG);
}
rtx
-m68k_function_value (tree valtype, tree func ATTRIBUTE_UNUSED)
+m68k_function_value (const_tree valtype, const_tree func ATTRIBUTE_UNUSED)
{
enum machine_mode mode;
case DFmode:
case XFmode:
if (TARGET_68881)
- return gen_rtx_REG (mode, 16);
+ return gen_rtx_REG (mode, FP0_REG);
break;
default:
break;
else
return gen_rtx_REG (mode, D0_REG);
}
+
+/* Worker function for TARGET_RETURN_IN_MEMORY. */
+#if M68K_HONOR_TARGET_STRICT_ALIGNMENT
+static bool
+m68k_return_in_memory (const_tree type, const_tree fntype ATTRIBUTE_UNUSED)
+{
+ enum machine_mode mode = TYPE_MODE (type);
+
+ if (mode == BLKmode)
+ return true;
+
+ /* If TYPE's known alignment is less than the alignment of MODE that
+ would contain the structure, then return in memory. We need to
+ do so to maintain the compatibility between code compiled with
+ -mstrict-align and that compiled with -mno-strict-align. */
+ if (AGGREGATE_TYPE_P (type)
+ && TYPE_ALIGN (type) < GET_MODE_ALIGNMENT (mode))
+ return true;
+
+ return false;
+}
+#endif
+
+/* CPU to schedule the program for. */
+enum attr_cpu m68k_sched_cpu;
+
+/* MAC to schedule the program for. */
+enum attr_mac m68k_sched_mac;
+
+/* Operand type. */
+enum attr_op_type
+ {
+ /* No operand. */
+ OP_TYPE_NONE,
+
+ /* Integer register. */
+ OP_TYPE_RN,
+
+ /* FP register. */
+ OP_TYPE_FPN,
+
+ /* Implicit mem reference (e.g. stack). */
+ OP_TYPE_MEM1,
+
+ /* Memory without offset or indexing. EA modes 2, 3 and 4. */
+ OP_TYPE_MEM234,
+
+ /* Memory with offset but without indexing. EA mode 5. */
+ OP_TYPE_MEM5,
+
+ /* Memory with indexing. EA mode 6. */
+ OP_TYPE_MEM6,
+
+ /* Memory referenced by absolute address. EA mode 7. */
+ OP_TYPE_MEM7,
+
+ /* Immediate operand that doesn't require extension word. */
+ OP_TYPE_IMM_Q,
+
+ /* Immediate 16 bit operand. */
+ OP_TYPE_IMM_W,
+
+ /* Immediate 32 bit operand. */
+ OP_TYPE_IMM_L
+ };
+
+/* Return type of memory ADDR_RTX refers to. */
+static enum attr_op_type
+sched_address_type (enum machine_mode mode, rtx addr_rtx)
+{
+ struct m68k_address address;
+
+ if (symbolic_operand (addr_rtx, VOIDmode))
+ return OP_TYPE_MEM7;
+
+ if (!m68k_decompose_address (mode, addr_rtx,
+ reload_completed, &address))
+ {
+ gcc_assert (!reload_completed);
+ /* Reload will likely fix the address to be in the register. */
+ return OP_TYPE_MEM234;
+ }
+
+ if (address.scale != 0)
+ return OP_TYPE_MEM6;
+
+ if (address.base != NULL_RTX)
+ {
+ if (address.offset == NULL_RTX)
+ return OP_TYPE_MEM234;
+
+ return OP_TYPE_MEM5;
+ }
+
+ gcc_assert (address.offset != NULL_RTX);
+
+ return OP_TYPE_MEM7;
+}
+
+/* Return X or Y (depending on OPX_P) operand of INSN. */
+static rtx
+sched_get_operand (rtx insn, bool opx_p)
+{
+ int i;
+
+ if (recog_memoized (insn) < 0)
+ gcc_unreachable ();
+
+ extract_constrain_insn_cached (insn);
+
+ if (opx_p)
+ i = get_attr_opx (insn);
+ else
+ i = get_attr_opy (insn);
+
+ if (i >= recog_data.n_operands)
+ return NULL;
+
+ return recog_data.operand[i];
+}
+
+/* Return type of INSN's operand X (if OPX_P) or operand Y (if !OPX_P).
+ If ADDRESS_P is true, return type of memory location operand refers to. */
+static enum attr_op_type
+sched_attr_op_type (rtx insn, bool opx_p, bool address_p)
+{
+ rtx op;
+
+ op = sched_get_operand (insn, opx_p);
+
+ if (op == NULL)
+ {
+ gcc_assert (!reload_completed);
+ return OP_TYPE_RN;
+ }
+
+ if (address_p)
+ return sched_address_type (QImode, op);
+
+ if (memory_operand (op, VOIDmode))
+ return sched_address_type (GET_MODE (op), XEXP (op, 0));
+
+ if (register_operand (op, VOIDmode))
+ {
+ if ((!reload_completed && FLOAT_MODE_P (GET_MODE (op)))
+ || (reload_completed && FP_REG_P (op)))
+ return OP_TYPE_FPN;
+
+ return OP_TYPE_RN;
+ }
+
+ if (GET_CODE (op) == CONST_INT)
+ {
+ int ival;
+
+ ival = INTVAL (op);
+
+ /* Check for quick constants. */
+ switch (get_attr_type (insn))
+ {
+ case TYPE_ALUQ_L:
+ if (IN_RANGE (ival, 1, 8) || IN_RANGE (ival, -8, -1))
+ return OP_TYPE_IMM_Q;
+
+ gcc_assert (!reload_completed);
+ break;
+
+ case TYPE_MOVEQ_L:
+ if (USE_MOVQ (ival))
+ return OP_TYPE_IMM_Q;
+
+ gcc_assert (!reload_completed);
+ break;
+
+ case TYPE_MOV3Q_L:
+ if (valid_mov3q_const (ival))
+ return OP_TYPE_IMM_Q;
+
+ gcc_assert (!reload_completed);
+ break;
+
+ default:
+ break;
+ }
+
+ if (IN_RANGE (ival, -0x8000, 0x7fff))
+ return OP_TYPE_IMM_W;
+
+ return OP_TYPE_IMM_L;
+ }
+
+ if (GET_CODE (op) == CONST_DOUBLE)
+ {
+ switch (GET_MODE (op))
+ {
+ case SFmode:
+ return OP_TYPE_IMM_W;
+
+ case VOIDmode:
+ case DFmode:
+ return OP_TYPE_IMM_L;
+
+ default:
+ gcc_unreachable ();
+ }
+ }
+
+ if (GET_CODE (op) == CONST
+ || symbolic_operand (op, VOIDmode)
+ || LABEL_P (op))
+ {
+ switch (GET_MODE (op))
+ {
+ case QImode:
+ return OP_TYPE_IMM_Q;
+
+ case HImode:
+ return OP_TYPE_IMM_W;
+
+ case SImode:
+ return OP_TYPE_IMM_L;
+
+ default:
+ if (symbolic_operand (m68k_unwrap_symbol (op, false), VOIDmode))
+ /* Just a guess. */
+ return OP_TYPE_IMM_W;
+
+ return OP_TYPE_IMM_L;
+ }
+ }
+
+ gcc_assert (!reload_completed);
+
+ if (FLOAT_MODE_P (GET_MODE (op)))
+ return OP_TYPE_FPN;
+
+ return OP_TYPE_RN;
+}
+
+/* Implement opx_type attribute.
+ Return type of INSN's operand X.
+ If ADDRESS_P is true, return type of memory location operand refers to. */
+enum attr_opx_type
+m68k_sched_attr_opx_type (rtx insn, int address_p)
+{
+ switch (sched_attr_op_type (insn, true, address_p != 0))
+ {
+ case OP_TYPE_RN:
+ return OPX_TYPE_RN;
+
+ case OP_TYPE_FPN:
+ return OPX_TYPE_FPN;
+
+ case OP_TYPE_MEM1:
+ return OPX_TYPE_MEM1;
+
+ case OP_TYPE_MEM234:
+ return OPX_TYPE_MEM234;
+
+ case OP_TYPE_MEM5:
+ return OPX_TYPE_MEM5;
+
+ case OP_TYPE_MEM6:
+ return OPX_TYPE_MEM6;
+
+ case OP_TYPE_MEM7:
+ return OPX_TYPE_MEM7;
+
+ case OP_TYPE_IMM_Q:
+ return OPX_TYPE_IMM_Q;
+
+ case OP_TYPE_IMM_W:
+ return OPX_TYPE_IMM_W;
+
+ case OP_TYPE_IMM_L:
+ return OPX_TYPE_IMM_L;
+
+ default:
+ gcc_unreachable ();
+ return 0;
+ }
+}
+
+/* Implement opy_type attribute.
+ Return type of INSN's operand Y.
+ If ADDRESS_P is true, return type of memory location operand refers to. */
+enum attr_opy_type
+m68k_sched_attr_opy_type (rtx insn, int address_p)
+{
+ switch (sched_attr_op_type (insn, false, address_p != 0))
+ {
+ case OP_TYPE_RN:
+ return OPY_TYPE_RN;
+
+ case OP_TYPE_FPN:
+ return OPY_TYPE_FPN;
+
+ case OP_TYPE_MEM1:
+ return OPY_TYPE_MEM1;
+
+ case OP_TYPE_MEM234:
+ return OPY_TYPE_MEM234;
+
+ case OP_TYPE_MEM5:
+ return OPY_TYPE_MEM5;
+
+ case OP_TYPE_MEM6:
+ return OPY_TYPE_MEM6;
+
+ case OP_TYPE_MEM7:
+ return OPY_TYPE_MEM7;
+
+ case OP_TYPE_IMM_Q:
+ return OPY_TYPE_IMM_Q;
+
+ case OP_TYPE_IMM_W:
+ return OPY_TYPE_IMM_W;
+
+ case OP_TYPE_IMM_L:
+ return OPY_TYPE_IMM_L;
+
+ default:
+ gcc_unreachable ();
+ return 0;
+ }
+}
+
+/* Return size of INSN as int. */
+static int
+sched_get_attr_size_int (rtx insn)
+{
+ int size;
+
+ switch (get_attr_type (insn))
+ {
+ case TYPE_IGNORE:
+ /* There should be no references to m68k_sched_attr_size for 'ignore'
+ instructions. */
+ gcc_unreachable ();
+ return 0;
+
+ case TYPE_MUL_L:
+ size = 2;
+ break;
+
+ default:
+ size = 1;
+ break;
+ }
+
+ switch (get_attr_opx_type (insn))
+ {
+ case OPX_TYPE_NONE:
+ case OPX_TYPE_RN:
+ case OPX_TYPE_FPN:
+ case OPX_TYPE_MEM1:
+ case OPX_TYPE_MEM234:
+ case OPY_TYPE_IMM_Q:
+ break;
+
+ case OPX_TYPE_MEM5:
+ case OPX_TYPE_MEM6:
+ /* Here we assume that most absolute references are short. */
+ case OPX_TYPE_MEM7:
+ case OPY_TYPE_IMM_W:
+ ++size;
+ break;
+
+ case OPY_TYPE_IMM_L:
+ size += 2;
+ break;
+
+ default:
+ gcc_unreachable ();
+ }
+
+ switch (get_attr_opy_type (insn))
+ {
+ case OPY_TYPE_NONE:
+ case OPY_TYPE_RN:
+ case OPY_TYPE_FPN:
+ case OPY_TYPE_MEM1:
+ case OPY_TYPE_MEM234:
+ case OPY_TYPE_IMM_Q:
+ break;
+
+ case OPY_TYPE_MEM5:
+ case OPY_TYPE_MEM6:
+ /* Here we assume that most absolute references are short. */
+ case OPY_TYPE_MEM7:
+ case OPY_TYPE_IMM_W:
+ ++size;
+ break;
+
+ case OPY_TYPE_IMM_L:
+ size += 2;
+ break;
+
+ default:
+ gcc_unreachable ();
+ }
+
+ if (size > 3)
+ {
+ gcc_assert (!reload_completed);
+
+ size = 3;
+ }
+
+ return size;
+}
+
+/* Return size of INSN as attribute enum value. */
+enum attr_size
+m68k_sched_attr_size (rtx insn)
+{
+ switch (sched_get_attr_size_int (insn))
+ {
+ case 1:
+ return SIZE_1;
+
+ case 2:
+ return SIZE_2;
+
+ case 3:
+ return SIZE_3;
+
+ default:
+ gcc_unreachable ();
+ return 0;
+ }
+}
+
+/* Return operand X or Y (depending on OPX_P) of INSN,
+ if it is a MEM, or NULL overwise. */
+static enum attr_op_type
+sched_get_opxy_mem_type (rtx insn, bool opx_p)
+{
+ if (opx_p)
+ {
+ switch (get_attr_opx_type (insn))
+ {
+ case OPX_TYPE_NONE:
+ case OPX_TYPE_RN:
+ case OPX_TYPE_FPN:
+ case OPX_TYPE_IMM_Q:
+ case OPX_TYPE_IMM_W:
+ case OPX_TYPE_IMM_L:
+ return OP_TYPE_RN;
+
+ case OPX_TYPE_MEM1:
+ case OPX_TYPE_MEM234:
+ case OPX_TYPE_MEM5:
+ case OPX_TYPE_MEM7:
+ return OP_TYPE_MEM1;
+
+ case OPX_TYPE_MEM6:
+ return OP_TYPE_MEM6;
+
+ default:
+ gcc_unreachable ();
+ return 0;
+ }
+ }
+ else
+ {
+ switch (get_attr_opy_type (insn))
+ {
+ case OPY_TYPE_NONE:
+ case OPY_TYPE_RN:
+ case OPY_TYPE_FPN:
+ case OPY_TYPE_IMM_Q:
+ case OPY_TYPE_IMM_W:
+ case OPY_TYPE_IMM_L:
+ return OP_TYPE_RN;
+
+ case OPY_TYPE_MEM1:
+ case OPY_TYPE_MEM234:
+ case OPY_TYPE_MEM5:
+ case OPY_TYPE_MEM7:
+ return OP_TYPE_MEM1;
+
+ case OPY_TYPE_MEM6:
+ return OP_TYPE_MEM6;
+
+ default:
+ gcc_unreachable ();
+ return 0;
+ }
+ }
+}
+
+/* Implement op_mem attribute. */
+enum attr_op_mem
+m68k_sched_attr_op_mem (rtx insn)
+{
+ enum attr_op_type opx;
+ enum attr_op_type opy;
+
+ opx = sched_get_opxy_mem_type (insn, true);
+ opy = sched_get_opxy_mem_type (insn, false);
+
+ if (opy == OP_TYPE_RN && opx == OP_TYPE_RN)
+ return OP_MEM_00;
+
+ if (opy == OP_TYPE_RN && opx == OP_TYPE_MEM1)
+ {
+ switch (get_attr_opx_access (insn))
+ {
+ case OPX_ACCESS_R:
+ return OP_MEM_10;
+
+ case OPX_ACCESS_W:
+ return OP_MEM_01;
+
+ case OPX_ACCESS_RW:
+ return OP_MEM_11;
+
+ default:
+ gcc_unreachable ();
+ return 0;
+ }
+ }
+
+ if (opy == OP_TYPE_RN && opx == OP_TYPE_MEM6)
+ {
+ switch (get_attr_opx_access (insn))
+ {
+ case OPX_ACCESS_R:
+ return OP_MEM_I0;
+
+ case OPX_ACCESS_W:
+ return OP_MEM_0I;
+
+ case OPX_ACCESS_RW:
+ return OP_MEM_I1;
+
+ default:
+ gcc_unreachable ();
+ return 0;
+ }
+ }
+
+ if (opy == OP_TYPE_MEM1 && opx == OP_TYPE_RN)
+ return OP_MEM_10;
+
+ if (opy == OP_TYPE_MEM1 && opx == OP_TYPE_MEM1)
+ {
+ switch (get_attr_opx_access (insn))
+ {
+ case OPX_ACCESS_W:
+ return OP_MEM_11;
+
+ default:
+ gcc_assert (!reload_completed);
+ return OP_MEM_11;
+ }
+ }
+
+ if (opy == OP_TYPE_MEM1 && opx == OP_TYPE_MEM6)
+ {
+ switch (get_attr_opx_access (insn))
+ {
+ case OPX_ACCESS_W:
+ return OP_MEM_1I;
+
+ default:
+ gcc_assert (!reload_completed);
+ return OP_MEM_1I;
+ }
+ }
+
+ if (opy == OP_TYPE_MEM6 && opx == OP_TYPE_RN)
+ return OP_MEM_I0;
+
+ if (opy == OP_TYPE_MEM6 && opx == OP_TYPE_MEM1)
+ {
+ switch (get_attr_opx_access (insn))
+ {
+ case OPX_ACCESS_W:
+ return OP_MEM_I1;
+
+ default:
+ gcc_assert (!reload_completed);
+ return OP_MEM_I1;
+ }
+ }
+
+ gcc_assert (opy == OP_TYPE_MEM6 && opx == OP_TYPE_MEM6);
+ gcc_assert (!reload_completed);
+ return OP_MEM_I1;
+}
+
+/* Jump instructions types. Indexed by INSN_UID.
+ The same rtl insn can be expanded into different asm instructions
+ depending on the cc0_status. To properly determine type of jump
+ instructions we scan instruction stream and map jumps types to this
+ array. */
+static enum attr_type *sched_branch_type;
+
+/* Return the type of the jump insn. */
+enum attr_type
+m68k_sched_branch_type (rtx insn)
+{
+ enum attr_type type;
+
+ type = sched_branch_type[INSN_UID (insn)];
+
+ gcc_assert (type != 0);
+
+ return type;
+}
+
+/* Data for ColdFire V4 index bypass.
+ Producer modifies register that is used as index in consumer with
+ specified scale. */
+static struct
+{
+ /* Producer instruction. */
+ rtx pro;
+
+ /* Consumer instruction. */
+ rtx con;
+
+ /* Scale of indexed memory access within consumer.
+ Or zero if bypass should not be effective at the moment. */
+ int scale;
+} sched_cfv4_bypass_data;
+
+/* An empty state that is used in m68k_sched_adjust_cost. */
+static state_t sched_adjust_cost_state;
+
+/* Implement adjust_cost scheduler hook.
+ Return adjusted COST of dependency LINK between DEF_INSN and INSN. */
+static int
+m68k_sched_adjust_cost (rtx insn, rtx link ATTRIBUTE_UNUSED, rtx def_insn,
+ int cost)
+{
+ int delay;
+
+ if (recog_memoized (def_insn) < 0
+ || recog_memoized (insn) < 0)
+ return cost;
+
+ if (sched_cfv4_bypass_data.scale == 1)
+ /* Handle ColdFire V4 bypass for indexed address with 1x scale. */
+ {
+ /* haifa-sched.c: insn_cost () calls bypass_p () just before
+ targetm.sched.adjust_cost (). Hence, we can be relatively sure
+ that the data in sched_cfv4_bypass_data is up to date. */
+ gcc_assert (sched_cfv4_bypass_data.pro == def_insn
+ && sched_cfv4_bypass_data.con == insn);
+
+ if (cost < 3)
+ cost = 3;
+
+ sched_cfv4_bypass_data.pro = NULL;
+ sched_cfv4_bypass_data.con = NULL;
+ sched_cfv4_bypass_data.scale = 0;
+ }
+ else
+ gcc_assert (sched_cfv4_bypass_data.pro == NULL
+ && sched_cfv4_bypass_data.con == NULL
+ && sched_cfv4_bypass_data.scale == 0);
+
+ /* Don't try to issue INSN earlier than DFA permits.
+ This is especially useful for instructions that write to memory,
+ as their true dependence (default) latency is better to be set to 0
+ to workaround alias analysis limitations.
+ This is, in fact, a machine independent tweak, so, probably,
+ it should be moved to haifa-sched.c: insn_cost (). */
+ delay = min_insn_conflict_delay (sched_adjust_cost_state, def_insn, insn);
+ if (delay > cost)
+ cost = delay;
+
+ return cost;
+}
+
+/* Return maximal number of insns that can be scheduled on a single cycle. */
+static int
+m68k_sched_issue_rate (void)
+{
+ switch (m68k_sched_cpu)
+ {
+ case CPU_CFV1:
+ case CPU_CFV2:
+ case CPU_CFV3:
+ return 1;
+
+ case CPU_CFV4:
+ return 2;
+
+ default:
+ gcc_unreachable ();
+ return 0;
+ }
+}
+
+/* Maximal length of instruction for current CPU.
+ E.g. it is 3 for any ColdFire core. */
+static int max_insn_size;
+
+/* Data to model instruction buffer of CPU. */
+struct _sched_ib
+{
+ /* True if instruction buffer model is modeled for current CPU. */
+ bool enabled_p;
+
+ /* Size of the instruction buffer in words. */
+ int size;
+
+ /* Number of filled words in the instruction buffer. */
+ int filled;
+
+ /* Additional information about instruction buffer for CPUs that have
+ a buffer of instruction records, rather then a plain buffer
+ of instruction words. */
+ struct _sched_ib_records
+ {
+ /* Size of buffer in records. */
+ int n_insns;
+
+ /* Array to hold data on adjustements made to the size of the buffer. */
+ int *adjust;
+
+ /* Index of the above array. */
+ int adjust_index;
+ } records;
+
+ /* An insn that reserves (marks empty) one word in the instruction buffer. */
+ rtx insn;
+};
+
+static struct _sched_ib sched_ib;
+
+/* ID of memory unit. */
+static int sched_mem_unit_code;
+
+/* Implementation of the targetm.sched.variable_issue () hook.
+ It is called after INSN was issued. It returns the number of insns
+ that can possibly get scheduled on the current cycle.
+ It is used here to determine the effect of INSN on the instruction
+ buffer. */
+static int
+m68k_sched_variable_issue (FILE *sched_dump ATTRIBUTE_UNUSED,
+ int sched_verbose ATTRIBUTE_UNUSED,
+ rtx insn, int can_issue_more)
+{
+ int insn_size;
+
+ if (recog_memoized (insn) >= 0 && get_attr_type (insn) != TYPE_IGNORE)
+ {
+ switch (m68k_sched_cpu)
+ {
+ case CPU_CFV1:
+ case CPU_CFV2:
+ insn_size = sched_get_attr_size_int (insn);
+ break;
+
+ case CPU_CFV3:
+ insn_size = sched_get_attr_size_int (insn);
+
+ /* ColdFire V3 and V4 cores have instruction buffers that can
+ accumulate up to 8 instructions regardless of instructions'
+ sizes. So we should take care not to "prefetch" 24 one-word
+ or 12 two-words instructions.
+ To model this behavior we temporarily decrease size of the
+ buffer by (max_insn_size - insn_size) for next 7 instructions. */
+ {
+ int adjust;
+
+ adjust = max_insn_size - insn_size;
+ sched_ib.size -= adjust;
+
+ if (sched_ib.filled > sched_ib.size)
+ sched_ib.filled = sched_ib.size;
+
+ sched_ib.records.adjust[sched_ib.records.adjust_index] = adjust;
+ }
+
+ ++sched_ib.records.adjust_index;
+ if (sched_ib.records.adjust_index == sched_ib.records.n_insns)
+ sched_ib.records.adjust_index = 0;
+
+ /* Undo adjustement we did 7 instructions ago. */
+ sched_ib.size
+ += sched_ib.records.adjust[sched_ib.records.adjust_index];
+
+ break;
+
+ case CPU_CFV4:
+ gcc_assert (!sched_ib.enabled_p);
+ insn_size = 0;
+ break;
+
+ default:
+ gcc_unreachable ();
+ }
+
+ gcc_assert (insn_size <= sched_ib.filled);
+ --can_issue_more;
+ }
+ else if (GET_CODE (PATTERN (insn)) == ASM_INPUT
+ || asm_noperands (PATTERN (insn)) >= 0)
+ insn_size = sched_ib.filled;
+ else
+ insn_size = 0;
+
+ sched_ib.filled -= insn_size;
+
+ return can_issue_more;
+}
+
+/* Return how many instructions should scheduler lookahead to choose the
+ best one. */
+static int
+m68k_sched_first_cycle_multipass_dfa_lookahead (void)
+{
+ return m68k_sched_issue_rate () - 1;
+}
+
+/* Implementation of targetm.sched.md_init_global () hook.
+ It is invoked once per scheduling pass and is used here
+ to initialize scheduler constants. */
+static void
+m68k_sched_md_init_global (FILE *sched_dump ATTRIBUTE_UNUSED,
+ int sched_verbose ATTRIBUTE_UNUSED,
+ int n_insns ATTRIBUTE_UNUSED)
+{
+ /* Init branch types. */
+ {
+ rtx insn;
+
+ sched_branch_type = XCNEWVEC (enum attr_type, get_max_uid () + 1);
+
+ for (insn = get_insns (); insn != NULL_RTX; insn = NEXT_INSN (insn))
+ {
+ if (JUMP_P (insn))
+ /* !!! FIXME: Implement real scan here. */
+ sched_branch_type[INSN_UID (insn)] = TYPE_BCC;
+ }
+ }
+
+#ifdef ENABLE_CHECKING
+ /* Check that all instructions have DFA reservations and
+ that all instructions can be issued from a clean state. */
+ {
+ rtx insn;
+ state_t state;
+
+ state = alloca (state_size ());
+
+ for (insn = get_insns (); insn != NULL_RTX; insn = NEXT_INSN (insn))
+ {
+ if (INSN_P (insn) && recog_memoized (insn) >= 0)
+ {
+ gcc_assert (insn_has_dfa_reservation_p (insn));
+
+ state_reset (state);
+ if (state_transition (state, insn) >= 0)
+ gcc_unreachable ();
+ }
+ }
+ }
+#endif
+
+ /* Setup target cpu. */
+
+ /* ColdFire V4 has a set of features to keep its instruction buffer full
+ (e.g., a separate memory bus for instructions) and, hence, we do not model
+ buffer for this CPU. */
+ sched_ib.enabled_p = (m68k_sched_cpu != CPU_CFV4);
+
+ switch (m68k_sched_cpu)
+ {
+ case CPU_CFV4:
+ sched_ib.filled = 0;
+
+ /* FALLTHRU */
+
+ case CPU_CFV1:
+ case CPU_CFV2:
+ max_insn_size = 3;
+ sched_ib.records.n_insns = 0;
+ sched_ib.records.adjust = NULL;
+ break;
+
+ case CPU_CFV3:
+ max_insn_size = 3;
+ sched_ib.records.n_insns = 8;
+ sched_ib.records.adjust = XNEWVEC (int, sched_ib.records.n_insns);
+ break;
+
+ default:
+ gcc_unreachable ();
+ }
+
+ sched_mem_unit_code = get_cpu_unit_code ("cf_mem1");
+
+ sched_adjust_cost_state = xmalloc (state_size ());
+ state_reset (sched_adjust_cost_state);
+
+ start_sequence ();
+ emit_insn (gen_ib ());
+ sched_ib.insn = get_insns ();
+ end_sequence ();
+}
+
+/* Scheduling pass is now finished. Free/reset static variables. */
+static void
+m68k_sched_md_finish_global (FILE *dump ATTRIBUTE_UNUSED,
+ int verbose ATTRIBUTE_UNUSED)
+{
+ sched_ib.insn = NULL;
+
+ free (sched_adjust_cost_state);
+ sched_adjust_cost_state = NULL;
+
+ sched_mem_unit_code = 0;
+
+ free (sched_ib.records.adjust);
+ sched_ib.records.adjust = NULL;
+ sched_ib.records.n_insns = 0;
+ max_insn_size = 0;
+
+ free (sched_branch_type);
+ sched_branch_type = NULL;
+}
+
+/* Implementation of targetm.sched.md_init () hook.
+ It is invoked each time scheduler starts on the new block (basic block or
+ extended basic block). */
+static void
+m68k_sched_md_init (FILE *sched_dump ATTRIBUTE_UNUSED,
+ int sched_verbose ATTRIBUTE_UNUSED,
+ int n_insns ATTRIBUTE_UNUSED)
+{
+ switch (m68k_sched_cpu)
+ {
+ case CPU_CFV1:
+ case CPU_CFV2:
+ sched_ib.size = 6;
+ break;
+
+ case CPU_CFV3:
+ sched_ib.size = sched_ib.records.n_insns * max_insn_size;
+
+ memset (sched_ib.records.adjust, 0,
+ sched_ib.records.n_insns * sizeof (*sched_ib.records.adjust));
+ sched_ib.records.adjust_index = 0;
+ break;
+
+ case CPU_CFV4:
+ gcc_assert (!sched_ib.enabled_p);
+ sched_ib.size = 0;
+ break;
+
+ default:
+ gcc_unreachable ();
+ }
+
+ if (sched_ib.enabled_p)
+ /* haifa-sched.c: schedule_block () calls advance_cycle () just before
+ the first cycle. Workaround that. */
+ sched_ib.filled = -2;
+}
+
+/* Implementation of targetm.sched.dfa_pre_advance_cycle () hook.
+ It is invoked just before current cycle finishes and is used here
+ to track if instruction buffer got its two words this cycle. */
+static void
+m68k_sched_dfa_pre_advance_cycle (void)
+{
+ if (!sched_ib.enabled_p)
+ return;
+
+ if (!cpu_unit_reservation_p (curr_state, sched_mem_unit_code))
+ {
+ sched_ib.filled += 2;
+
+ if (sched_ib.filled > sched_ib.size)
+ sched_ib.filled = sched_ib.size;
+ }
+}
+
+/* Implementation of targetm.sched.dfa_post_advance_cycle () hook.
+ It is invoked just after new cycle begins and is used here
+ to setup number of filled words in the instruction buffer so that
+ instructions which won't have all their words prefetched would be
+ stalled for a cycle. */
+static void
+m68k_sched_dfa_post_advance_cycle (void)
+{
+ int i;
+
+ if (!sched_ib.enabled_p)
+ return;
+
+ /* Setup number of prefetched instruction words in the instruction
+ buffer. */
+ i = max_insn_size - sched_ib.filled;
+
+ while (--i >= 0)
+ {
+ if (state_transition (curr_state, sched_ib.insn) >= 0)
+ gcc_unreachable ();
+ }
+}
+
+/* Return X or Y (depending on OPX_P) operand of INSN,
+ if it is an integer register, or NULL overwise. */
+static rtx
+sched_get_reg_operand (rtx insn, bool opx_p)
+{
+ rtx op = NULL;
+
+ if (opx_p)
+ {
+ if (get_attr_opx_type (insn) == OPX_TYPE_RN)
+ {
+ op = sched_get_operand (insn, true);
+ gcc_assert (op != NULL);
+
+ if (!reload_completed && !REG_P (op))
+ return NULL;
+ }
+ }
+ else
+ {
+ if (get_attr_opy_type (insn) == OPY_TYPE_RN)
+ {
+ op = sched_get_operand (insn, false);
+ gcc_assert (op != NULL);
+
+ if (!reload_completed && !REG_P (op))
+ return NULL;
+ }
+ }
+
+ return op;
+}
+
+/* Return true, if X or Y (depending on OPX_P) operand of INSN
+ is a MEM. */
+static bool
+sched_mem_operand_p (rtx insn, bool opx_p)
+{
+ switch (sched_get_opxy_mem_type (insn, opx_p))
+ {
+ case OP_TYPE_MEM1:
+ case OP_TYPE_MEM6:
+ return true;
+
+ default:
+ return false;
+ }
+}
+
+/* Return X or Y (depending on OPX_P) operand of INSN,
+ if it is a MEM, or NULL overwise. */
+static rtx
+sched_get_mem_operand (rtx insn, bool must_read_p, bool must_write_p)
+{
+ bool opx_p;
+ bool opy_p;
+
+ opx_p = false;
+ opy_p = false;
+
+ if (must_read_p)
+ {
+ opx_p = true;
+ opy_p = true;
+ }
+
+ if (must_write_p)
+ {
+ opx_p = true;
+ opy_p = false;
+ }
+
+ if (opy_p && sched_mem_operand_p (insn, false))
+ return sched_get_operand (insn, false);
+
+ if (opx_p && sched_mem_operand_p (insn, true))
+ return sched_get_operand (insn, true);
+
+ gcc_unreachable ();
+ return NULL;
+}
+
+/* Return non-zero if PRO modifies register used as part of
+ address in CON. */
+int
+m68k_sched_address_bypass_p (rtx pro, rtx con)
+{
+ rtx pro_x;
+ rtx con_mem_read;
+
+ pro_x = sched_get_reg_operand (pro, true);
+ if (pro_x == NULL)
+ return 0;
+
+ con_mem_read = sched_get_mem_operand (con, true, false);
+ gcc_assert (con_mem_read != NULL);
+
+ if (reg_mentioned_p (pro_x, con_mem_read))
+ return 1;
+
+ return 0;
+}
+
+/* Helper function for m68k_sched_indexed_address_bypass_p.
+ if PRO modifies register used as index in CON,
+ return scale of indexed memory access in CON. Return zero overwise. */
+static int
+sched_get_indexed_address_scale (rtx pro, rtx con)
+{
+ rtx reg;
+ rtx mem;
+ struct m68k_address address;
+
+ reg = sched_get_reg_operand (pro, true);
+ if (reg == NULL)
+ return 0;
+
+ mem = sched_get_mem_operand (con, true, false);
+ gcc_assert (mem != NULL && MEM_P (mem));
+
+ if (!m68k_decompose_address (GET_MODE (mem), XEXP (mem, 0), reload_completed,
+ &address))
+ gcc_unreachable ();
+
+ if (REGNO (reg) == REGNO (address.index))
+ {
+ gcc_assert (address.scale != 0);
+ return address.scale;
+ }
+
+ return 0;
+}
+
+/* Return non-zero if PRO modifies register used
+ as index with scale 2 or 4 in CON. */
+int
+m68k_sched_indexed_address_bypass_p (rtx pro, rtx con)
+{
+ gcc_assert (sched_cfv4_bypass_data.pro == NULL
+ && sched_cfv4_bypass_data.con == NULL
+ && sched_cfv4_bypass_data.scale == 0);
+
+ switch (sched_get_indexed_address_scale (pro, con))
+ {
+ case 1:
+ /* We can't have a variable latency bypass, so
+ remember to adjust the insn cost in adjust_cost hook. */
+ sched_cfv4_bypass_data.pro = pro;
+ sched_cfv4_bypass_data.con = con;
+ sched_cfv4_bypass_data.scale = 1;
+ return 0;
+
+ case 2:
+ case 4:
+ return 1;
+
+ default:
+ return 0;
+ }
+}
+
+#include "gt-m68k.h"
OSDN Git Service
