/* Subroutines for insn-output.c for HPPA.
Copyright (C) 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999, 2000, 2001,
- 2002, 2003 Free Software Foundation, Inc.
+ 2002, 2003, 2004, 2005 Free Software Foundation, Inc.
Contributed by Tim Moore (moore@cs.utah.edu), based on sparc.c
This file is part of GCC.
You should have received a copy of the GNU General Public License
along with GCC; see the file COPYING. If not, write to
-the Free Software Foundation, 59 Temple Place - Suite 330,
-Boston, MA 02111-1307, USA. */
+the Free Software Foundation, 51 Franklin Street, Fifth Floor,
+Boston, MA 02110-1301, USA. */
#include "config.h"
#include "system.h"
#include "reload.h"
#include "integrate.h"
#include "function.h"
-#include "obstack.h"
#include "toplev.h"
#include "ggc.h"
#include "recog.h"
#include "target.h"
#include "target-def.h"
-static int hppa_use_dfa_pipeline_interface (void);
-
-#undef TARGET_SCHED_USE_DFA_PIPELINE_INTERFACE
-#define TARGET_SCHED_USE_DFA_PIPELINE_INTERFACE hppa_use_dfa_pipeline_interface
-
-static int
-hppa_use_dfa_pipeline_interface (void)
-{
- return 1;
-}
-
/* Return nonzero if there is a bypass for the output of
OUT_INSN and the fp store IN_INSN. */
int
#endif
#endif
+static void copy_reg_pointer (rtx, rtx);
+static void fix_range (const char *);
+static bool pa_handle_option (size_t, const char *, int);
static int hppa_address_cost (rtx);
static bool hppa_rtx_costs (rtx, int, int, int *);
static inline rtx force_mode (enum machine_mode, rtx);
static void pa_combine_instructions (void);
static int pa_can_combine_p (rtx, rtx, rtx, int, rtx, rtx, rtx);
static int forward_branch_p (rtx);
-static int shadd_constant_p (int);
static void compute_zdepwi_operands (unsigned HOST_WIDE_INT, unsigned *);
-static int compute_movstr_length (rtx);
-static int compute_clrstr_length (rtx);
+static int compute_movmem_length (rtx);
+static int compute_clrmem_length (rtx);
static bool pa_assemble_integer (rtx, unsigned int, int);
static void remove_useless_addtr_insns (int);
-static void store_reg (int, int, int);
-static void store_reg_modify (int, int, int);
-static void load_reg (int, int, int);
-static void set_reg_plus_d (int, int, int, int);
+static void store_reg (int, HOST_WIDE_INT, int);
+static void store_reg_modify (int, int, HOST_WIDE_INT);
+static void load_reg (int, HOST_WIDE_INT, int);
+static void set_reg_plus_d (int, int, HOST_WIDE_INT, int);
static void pa_output_function_prologue (FILE *, HOST_WIDE_INT);
static void update_total_code_bytes (int);
static void pa_output_function_epilogue (FILE *, HOST_WIDE_INT);
static void pa_asm_out_destructor (rtx, int);
#endif
static void pa_init_builtins (void);
+static rtx hppa_builtin_saveregs (void);
+static tree hppa_gimplify_va_arg_expr (tree, tree, tree *, tree *);
+static bool pa_scalar_mode_supported_p (enum machine_mode);
+static bool pa_commutative_p (rtx x, int outer_code);
static void copy_fp_args (rtx) ATTRIBUTE_UNUSED;
static int length_fp_args (rtx) ATTRIBUTE_UNUSED;
-static struct deferred_plabel *get_plabel (const char *)
- ATTRIBUTE_UNUSED;
static inline void pa_file_start_level (void) ATTRIBUTE_UNUSED;
static inline void pa_file_start_space (int) ATTRIBUTE_UNUSED;
static inline void pa_file_start_file (int) ATTRIBUTE_UNUSED;
static void pa_hpux64_gas_file_start (void) ATTRIBUTE_UNUSED;
static void pa_hpux64_hpas_file_start (void) ATTRIBUTE_UNUSED;
static void output_deferred_plabels (void);
+static void output_deferred_profile_counters (void) ATTRIBUTE_UNUSED;
+#ifdef ASM_OUTPUT_EXTERNAL_REAL
+static void pa_hpux_file_end (void);
+#endif
#ifdef HPUX_LONG_DOUBLE_LIBRARY
static void pa_hpux_init_libfuncs (void);
#endif
+static rtx pa_struct_value_rtx (tree, int);
+static bool pa_pass_by_reference (CUMULATIVE_ARGS *, enum machine_mode,
+ tree, bool);
+static int pa_arg_partial_bytes (CUMULATIVE_ARGS *, enum machine_mode,
+ tree, bool);
+static struct machine_function * pa_init_machine_status (void);
+static enum reg_class pa_secondary_reload (bool, rtx, enum reg_class,
+ enum machine_mode,
+ secondary_reload_info *);
+
/* Save the operands last given to a compare for use when we
generate a scc or bcc insn. */
enum cmp_type hppa_branch_type;
/* Which cpu we are scheduling for. */
-enum processor_type pa_cpu;
-
-/* String to hold which cpu we are scheduling for. */
-const char *pa_cpu_string;
-
-/* Which architecture we are generating code for. */
-enum architecture_type pa_arch;
+enum processor_type pa_cpu = TARGET_SCHED_DEFAULT;
-/* String to hold which architecture we are generating code for. */
-const char *pa_arch_string;
+/* The UNIX standard to use for predefines and linking. */
+int flag_pa_unix = TARGET_HPUX_11_11 ? 1998 : TARGET_HPUX_10_10 ? 1995 : 1993;
/* Counts for the number of callee-saved general and floating point
registers which were saved by the current function's prologue. */
struct deferred_plabel GTY(())
{
rtx internal_label;
- const char *name;
+ rtx symbol;
};
static GTY((length ("n_deferred_plabels"))) struct deferred_plabel *
deferred_plabels;
#undef TARGET_FUNCTION_OK_FOR_SIBCALL
#define TARGET_FUNCTION_OK_FOR_SIBCALL pa_function_ok_for_sibcall
+#undef TARGET_COMMUTATIVE_P
+#define TARGET_COMMUTATIVE_P pa_commutative_p
+
#undef TARGET_ASM_OUTPUT_MI_THUNK
#define TARGET_ASM_OUTPUT_MI_THUNK pa_asm_output_mi_thunk
#undef TARGET_ASM_CAN_OUTPUT_MI_THUNK
#define TARGET_ASM_CAN_OUTPUT_MI_THUNK default_can_output_mi_thunk_no_vcall
#undef TARGET_ASM_FILE_END
+#ifdef ASM_OUTPUT_EXTERNAL_REAL
+#define TARGET_ASM_FILE_END pa_hpux_file_end
+#else
#define TARGET_ASM_FILE_END output_deferred_plabels
+#endif
#if !defined(USE_COLLECT2)
#undef TARGET_ASM_CONSTRUCTOR
#define TARGET_ASM_DESTRUCTOR pa_asm_out_destructor
#endif
+#undef TARGET_DEFAULT_TARGET_FLAGS
+#define TARGET_DEFAULT_TARGET_FLAGS (TARGET_DEFAULT | TARGET_CPU_DEFAULT)
+#undef TARGET_HANDLE_OPTION
+#define TARGET_HANDLE_OPTION pa_handle_option
+
#undef TARGET_INIT_BUILTINS
#define TARGET_INIT_BUILTINS pa_init_builtins
#define TARGET_INIT_LIBFUNCS pa_hpux_init_libfuncs
#endif
+#undef TARGET_PROMOTE_FUNCTION_RETURN
+#define TARGET_PROMOTE_FUNCTION_RETURN hook_bool_tree_true
+#undef TARGET_PROMOTE_PROTOTYPES
+#define TARGET_PROMOTE_PROTOTYPES hook_bool_tree_true
+
+#undef TARGET_STRUCT_VALUE_RTX
+#define TARGET_STRUCT_VALUE_RTX pa_struct_value_rtx
+#undef TARGET_RETURN_IN_MEMORY
+#define TARGET_RETURN_IN_MEMORY pa_return_in_memory
+#undef TARGET_MUST_PASS_IN_STACK
+#define TARGET_MUST_PASS_IN_STACK must_pass_in_stack_var_size
+#undef TARGET_PASS_BY_REFERENCE
+#define TARGET_PASS_BY_REFERENCE pa_pass_by_reference
+#undef TARGET_CALLEE_COPIES
+#define TARGET_CALLEE_COPIES hook_bool_CUMULATIVE_ARGS_mode_tree_bool_true
+#undef TARGET_ARG_PARTIAL_BYTES
+#define TARGET_ARG_PARTIAL_BYTES pa_arg_partial_bytes
+
+#undef TARGET_EXPAND_BUILTIN_SAVEREGS
+#define TARGET_EXPAND_BUILTIN_SAVEREGS hppa_builtin_saveregs
+#undef TARGET_GIMPLIFY_VA_ARG_EXPR
+#define TARGET_GIMPLIFY_VA_ARG_EXPR hppa_gimplify_va_arg_expr
+
+#undef TARGET_SCALAR_MODE_SUPPORTED_P
+#define TARGET_SCALAR_MODE_SUPPORTED_P pa_scalar_mode_supported_p
+
+#undef TARGET_CANNOT_FORCE_CONST_MEM
+#define TARGET_CANNOT_FORCE_CONST_MEM pa_tls_referenced_p
+
+#undef TARGET_SECONDARY_RELOAD
+#define TARGET_SECONDARY_RELOAD pa_secondary_reload
+
struct gcc_target targetm = TARGET_INITIALIZER;
\f
-void
-override_options (void)
+/* Parse the -mfixed-range= option string. */
+
+static void
+fix_range (const char *const_str)
{
- if (pa_cpu_string == NULL)
- pa_cpu_string = TARGET_SCHED_DEFAULT;
+ int i, first, last;
+ char *str, *dash, *comma;
- if (! strcmp (pa_cpu_string, "8000"))
- {
- pa_cpu_string = "8000";
- pa_cpu = PROCESSOR_8000;
- }
- else if (! strcmp (pa_cpu_string, "7100"))
- {
- pa_cpu_string = "7100";
- pa_cpu = PROCESSOR_7100;
- }
- else if (! strcmp (pa_cpu_string, "700"))
- {
- pa_cpu_string = "700";
- pa_cpu = PROCESSOR_700;
- }
- else if (! strcmp (pa_cpu_string, "7100LC"))
- {
- pa_cpu_string = "7100LC";
- pa_cpu = PROCESSOR_7100LC;
- }
- else if (! strcmp (pa_cpu_string, "7200"))
- {
- pa_cpu_string = "7200";
- pa_cpu = PROCESSOR_7200;
- }
- else if (! strcmp (pa_cpu_string, "7300"))
- {
- pa_cpu_string = "7300";
- pa_cpu = PROCESSOR_7300;
- }
- else
+ /* str must be of the form REG1'-'REG2{,REG1'-'REG} where REG1 and
+ REG2 are either register names or register numbers. The effect
+ of this option is to mark the registers in the range from REG1 to
+ REG2 as ``fixed'' so they won't be used by the compiler. This is
+ used, e.g., to ensure that kernel mode code doesn't use fr4-fr31. */
+
+ i = strlen (const_str);
+ str = (char *) alloca (i + 1);
+ memcpy (str, const_str, i + 1);
+
+ while (1)
{
- warning ("unknown -mschedule= option (%s).\nValid options are 700, 7100, 7100LC, 7200, 7300, and 8000\n", pa_cpu_string);
+ dash = strchr (str, '-');
+ if (!dash)
+ {
+ warning (0, "value of -mfixed-range must have form REG1-REG2");
+ return;
+ }
+ *dash = '\0';
+
+ comma = strchr (dash + 1, ',');
+ if (comma)
+ *comma = '\0';
+
+ first = decode_reg_name (str);
+ if (first < 0)
+ {
+ warning (0, "unknown register name: %s", str);
+ return;
+ }
+
+ last = decode_reg_name (dash + 1);
+ if (last < 0)
+ {
+ warning (0, "unknown register name: %s", dash + 1);
+ return;
+ }
+
+ *dash = '-';
+
+ if (first > last)
+ {
+ warning (0, "%s-%s is an empty range", str, dash + 1);
+ return;
+ }
+
+ for (i = first; i <= last; ++i)
+ fixed_regs[i] = call_used_regs[i] = 1;
+
+ if (!comma)
+ break;
+
+ *comma = ',';
+ str = comma + 1;
}
- /* Set the instruction set architecture. */
- if (pa_arch_string && ! strcmp (pa_arch_string, "1.0"))
+ /* Check if all floating point registers have been fixed. */
+ for (i = FP_REG_FIRST; i <= FP_REG_LAST; i++)
+ if (!fixed_regs[i])
+ break;
+
+ if (i > FP_REG_LAST)
+ target_flags |= MASK_DISABLE_FPREGS;
+}
+
+/* Implement TARGET_HANDLE_OPTION. */
+
+static bool
+pa_handle_option (size_t code, const char *arg, int value ATTRIBUTE_UNUSED)
+{
+ switch (code)
{
- pa_arch_string = "1.0";
- pa_arch = ARCHITECTURE_10;
+ case OPT_mnosnake:
+ case OPT_mpa_risc_1_0:
+ case OPT_march_1_0:
target_flags &= ~(MASK_PA_11 | MASK_PA_20);
- }
- else if (pa_arch_string && ! strcmp (pa_arch_string, "1.1"))
- {
- pa_arch_string = "1.1";
- pa_arch = ARCHITECTURE_11;
+ return true;
+
+ case OPT_msnake:
+ case OPT_mpa_risc_1_1:
+ case OPT_march_1_1:
target_flags &= ~MASK_PA_20;
target_flags |= MASK_PA_11;
- }
- else if (pa_arch_string && ! strcmp (pa_arch_string, "2.0"))
- {
- pa_arch_string = "2.0";
- pa_arch = ARCHITECTURE_20;
+ return true;
+
+ case OPT_mpa_risc_2_0:
+ case OPT_march_2_0:
target_flags |= MASK_PA_11 | MASK_PA_20;
+ return true;
+
+ case OPT_mschedule_:
+ if (strcmp (arg, "8000") == 0)
+ pa_cpu = PROCESSOR_8000;
+ else if (strcmp (arg, "7100") == 0)
+ pa_cpu = PROCESSOR_7100;
+ else if (strcmp (arg, "700") == 0)
+ pa_cpu = PROCESSOR_700;
+ else if (strcmp (arg, "7100LC") == 0)
+ pa_cpu = PROCESSOR_7100LC;
+ else if (strcmp (arg, "7200") == 0)
+ pa_cpu = PROCESSOR_7200;
+ else if (strcmp (arg, "7300") == 0)
+ pa_cpu = PROCESSOR_7300;
+ else
+ return false;
+ return true;
+
+ case OPT_mfixed_range_:
+ fix_range (arg);
+ return true;
+
+#if TARGET_HPUX
+ case OPT_munix_93:
+ flag_pa_unix = 1993;
+ return true;
+#endif
+
+#if TARGET_HPUX_10_10
+ case OPT_munix_95:
+ flag_pa_unix = 1995;
+ return true;
+#endif
+
+#if TARGET_HPUX_11_11
+ case OPT_munix_98:
+ flag_pa_unix = 1998;
+ return true;
+#endif
+
+ default:
+ return true;
}
- else if (pa_arch_string)
- {
- warning ("unknown -march= option (%s).\nValid options are 1.0, 1.1, and 2.0\n", pa_arch_string);
- }
+}
+void
+override_options (void)
+{
/* Unconditional branches in the delay slot are not compatible with dwarf2
call frame information. There is no benefit in using this optimization
on PA8000 and later processors. */
if (flag_pic && TARGET_PORTABLE_RUNTIME)
{
- warning ("PIC code generation is not supported in the portable runtime model\n");
+ warning (0, "PIC code generation is not supported in the portable runtime model");
}
if (flag_pic && TARGET_FAST_INDIRECT_CALLS)
{
- warning ("PIC code generation is not compatible with fast indirect calls\n");
+ warning (0, "PIC code generation is not compatible with fast indirect calls");
}
if (! TARGET_GAS && write_symbols != NO_DEBUG)
{
- warning ("-g is only supported when using GAS on this processor,");
- warning ("-g option disabled");
+ warning (0, "-g is only supported when using GAS on this processor,");
+ warning (0, "-g option disabled");
write_symbols = NO_DEBUG;
}
targetm.asm_out.unaligned_op.si = NULL;
targetm.asm_out.unaligned_op.di = NULL;
}
+
+ init_machine_status = pa_init_machine_status;
}
static void
pa_init_builtins (void)
{
#ifdef DONT_HAVE_FPUTC_UNLOCKED
- built_in_decls[(int) BUILT_IN_FPUTC_UNLOCKED] = NULL_TREE;
- implicit_built_in_decls[(int) BUILT_IN_FPUTC_UNLOCKED] = NULL_TREE;
+ built_in_decls[(int) BUILT_IN_FPUTC_UNLOCKED] =
+ built_in_decls[(int) BUILT_IN_PUTC_UNLOCKED];
+ implicit_built_in_decls[(int) BUILT_IN_FPUTC_UNLOCKED]
+ = implicit_built_in_decls[(int) BUILT_IN_PUTC_UNLOCKED];
#endif
}
-/* Return nonzero only if OP is a register of mode MODE,
- or CONST0_RTX. */
-int
-reg_or_0_operand (rtx op, enum machine_mode mode)
+/* Function to init struct machine_function.
+ This will be called, via a pointer variable,
+ from push_function_context. */
+
+static struct machine_function *
+pa_init_machine_status (void)
{
- return (op == CONST0_RTX (mode) || register_operand (op, mode));
+ return ggc_alloc_cleared (sizeof (machine_function));
}
-/* Return nonzero if OP is suitable for use in a call to a named
- function.
+/* If FROM is a probable pointer register, mark TO as a probable
+ pointer register with the same pointer alignment as FROM. */
- For 2.5 try to eliminate either call_operand_address or
- function_label_operand, they perform very similar functions. */
-int
-call_operand_address (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
+static void
+copy_reg_pointer (rtx to, rtx from)
{
- return (GET_MODE (op) == word_mode
- && CONSTANT_P (op) && ! TARGET_PORTABLE_RUNTIME);
+ if (REG_POINTER (from))
+ mark_reg_pointer (to, REGNO_POINTER_ALIGN (REGNO (from)));
}
/* Return 1 if X contains a symbolic expression. We know these
return (symbolic_operand (x, VOIDmode));
}
-int
-symbolic_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
-{
- switch (GET_CODE (op))
- {
- case SYMBOL_REF:
- case LABEL_REF:
- return 1;
- case CONST:
- op = XEXP (op, 0);
- return ((GET_CODE (XEXP (op, 0)) == SYMBOL_REF
- || GET_CODE (XEXP (op, 0)) == LABEL_REF)
- && GET_CODE (XEXP (op, 1)) == CONST_INT);
- default:
- return 0;
- }
-}
-
-/* Return truth value of statement that OP is a symbolic memory
- operand of mode MODE. */
-
-int
-symbolic_memory_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
-{
- if (GET_CODE (op) == SUBREG)
- op = SUBREG_REG (op);
- if (GET_CODE (op) != MEM)
- return 0;
- op = XEXP (op, 0);
- return (GET_CODE (op) == SYMBOL_REF || GET_CODE (op) == CONST
- || GET_CODE (op) == HIGH || GET_CODE (op) == LABEL_REF);
-}
-
-/* Return 1 if the operand is either a register or a memory operand that is
- not symbolic. */
-
-int
-reg_or_nonsymb_mem_operand (rtx op, enum machine_mode mode)
-{
- if (register_operand (op, mode))
- return 1;
-
- if (memory_operand (op, mode) && ! symbolic_memory_operand (op, mode))
- return 1;
-
- return 0;
-}
-
-/* Return 1 if the operand is either a register, zero, or a memory operand
- that is not symbolic. */
-
-int
-reg_or_0_or_nonsymb_mem_operand (rtx op, enum machine_mode mode)
-{
- if (register_operand (op, mode))
- return 1;
-
- if (op == CONST0_RTX (mode))
- return 1;
-
- if (memory_operand (op, mode) && ! symbolic_memory_operand (op, mode))
- return 1;
-
- return 0;
-}
-
-/* Return 1 if the operand is a register operand or a non-symbolic memory
- operand after reload. This predicate is used for branch patterns that
- internally handle register reloading. We need to accept non-symbolic
- memory operands after reload to ensure that the pattern is still valid
- if reload didn't find a hard register for the operand. */
-
-int
-reg_before_reload_operand (rtx op, enum machine_mode mode)
-{
- /* Don't accept a SUBREG since it will need a reload. */
- if (GET_CODE (op) == SUBREG)
- return 0;
-
- if (register_operand (op, mode))
- return 1;
-
- if (reload_completed
- && memory_operand (op, mode)
- && ! symbolic_memory_operand (op, mode))
- return 1;
-
- return 0;
-}
-
/* Accept any constant that can be moved in one instruction into a
general register. */
int
|| CONST_OK_FOR_LETTER_P (intval, 'N')
|| CONST_OK_FOR_LETTER_P (intval, 'K'));
}
-
-/* Accept anything that can be moved in one instruction into a general
- register. */
-int
-move_operand (rtx op, enum machine_mode mode)
-{
- if (register_operand (op, mode))
- return 1;
-
- if (GET_CODE (op) == CONSTANT_P_RTX)
- return 1;
-
- if (GET_CODE (op) == CONST_INT)
- return cint_ok_for_move (INTVAL (op));
-
- if (GET_CODE (op) == SUBREG)
- op = SUBREG_REG (op);
- if (GET_CODE (op) != MEM)
- return 0;
-
- op = XEXP (op, 0);
-
- /* We consider a LO_SUM DLT reference a move_operand now since it has
- been merged into the normal movsi/movdi patterns. */
- if (GET_CODE (op) == LO_SUM
- && GET_CODE (XEXP (op, 0)) == REG
- && REG_OK_FOR_BASE_P (XEXP (op, 0))
- && GET_CODE (XEXP (op, 1)) == UNSPEC
- && GET_MODE (op) == Pmode)
- return 1;
-
- /* Since move_operand is only used for source operands, we can always
- allow scaled indexing! */
- if (! TARGET_DISABLE_INDEXING
- && GET_CODE (op) == PLUS
- && ((GET_CODE (XEXP (op, 0)) == MULT
- && GET_CODE (XEXP (XEXP (op, 0), 0)) == REG
- && GET_CODE (XEXP (XEXP (op, 0), 1)) == CONST_INT
- && INTVAL (XEXP (XEXP (op, 0), 1))
- == (HOST_WIDE_INT) GET_MODE_SIZE (mode)
- && GET_CODE (XEXP (op, 1)) == REG)
- || (GET_CODE (XEXP (op, 1)) == MULT
- &&GET_CODE (XEXP (XEXP (op, 1), 0)) == REG
- && GET_CODE (XEXP (XEXP (op, 1), 1)) == CONST_INT
- && INTVAL (XEXP (XEXP (op, 1), 1))
- == (HOST_WIDE_INT) GET_MODE_SIZE (mode)
- && GET_CODE (XEXP (op, 0)) == REG)))
- return 1;
-
- return memory_address_p (mode, op);
-}
-
-/* Accept REG and any CONST_INT that can be moved in one instruction into a
- general register. */
-int
-reg_or_cint_move_operand (rtx op, enum machine_mode mode)
-{
- if (register_operand (op, mode))
- return 1;
-
- if (GET_CODE (op) == CONST_INT)
- return cint_ok_for_move (INTVAL (op));
-
- return 0;
-}
-
-int
-pic_label_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
-{
- if (!flag_pic)
- return 0;
-
- switch (GET_CODE (op))
- {
- case LABEL_REF:
- return 1;
- case CONST:
- op = XEXP (op, 0);
- return (GET_CODE (XEXP (op, 0)) == LABEL_REF
- && GET_CODE (XEXP (op, 1)) == CONST_INT);
- default:
- return 0;
- }
-}
-
-int
-fp_reg_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
-{
- return reg_renumber && FP_REG_P (op);
-}
-
\f
-
-/* Return truth value of whether OP can be used as an operand in a
- three operand arithmetic insn that accepts registers of mode MODE
- or 14-bit signed integers. */
-int
-arith_operand (rtx op, enum machine_mode mode)
-{
- return (register_operand (op, mode)
- || (GET_CODE (op) == CONST_INT && INT_14_BITS (op)));
-}
-
-/* Return truth value of whether OP can be used as an operand in a
- three operand arithmetic insn that accepts registers of mode MODE
- or 11-bit signed integers. */
-int
-arith11_operand (rtx op, enum machine_mode mode)
-{
- return (register_operand (op, mode)
- || (GET_CODE (op) == CONST_INT && INT_11_BITS (op)));
-}
-
/* Return truth value of whether OP can be used as an operand in a
adddi3 insn. */
int
&& (TARGET_64BIT ? INT_14_BITS (op) : INT_11_BITS (op))));
}
-/* A constant integer suitable for use in a PRE_MODIFY memory
- reference. */
-int
-pre_cint_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
-{
- return (GET_CODE (op) == CONST_INT
- && INTVAL (op) >= -0x2000 && INTVAL (op) < 0x10);
-}
-
-/* A constant integer suitable for use in a POST_MODIFY memory
- reference. */
-int
-post_cint_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
-{
- return (GET_CODE (op) == CONST_INT
- && INTVAL (op) < 0x2000 && INTVAL (op) >= -0x10);
-}
-
-int
-arith_double_operand (rtx op, enum machine_mode mode)
-{
- return (register_operand (op, mode)
- || (GET_CODE (op) == CONST_DOUBLE
- && GET_MODE (op) == mode
- && VAL_14_BITS_P (CONST_DOUBLE_LOW (op))
- && ((CONST_DOUBLE_HIGH (op) >= 0)
- == ((CONST_DOUBLE_LOW (op) & 0x1000) == 0))));
-}
-
-/* Return truth value of whether OP is an integer which fits the
- range constraining immediate operands in three-address insns, or
- is an integer register. */
-
-int
-ireg_or_int5_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
-{
- return ((GET_CODE (op) == CONST_INT && INT_5_BITS (op))
- || (GET_CODE (op) == REG && REGNO (op) > 0 && REGNO (op) < 32));
-}
-
-/* Return nonzero if OP is an integer register, else return zero. */
-int
-ireg_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
-{
- return (GET_CODE (op) == REG && REGNO (op) > 0 && REGNO (op) < 32);
-}
-
-/* Return truth value of whether OP is an integer which fits the
- range constraining immediate operands in three-address insns. */
-
-int
-int5_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
-{
- return (GET_CODE (op) == CONST_INT && INT_5_BITS (op));
-}
-
-int
-uint5_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
-{
- return (GET_CODE (op) == CONST_INT && INT_U5_BITS (op));
-}
-
-int
-int11_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
-{
- return (GET_CODE (op) == CONST_INT && INT_11_BITS (op));
-}
-
-int
-uint32_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
-{
-#if HOST_BITS_PER_WIDE_INT > 32
- /* All allowed constants will fit a CONST_INT. */
- return (GET_CODE (op) == CONST_INT
- && (INTVAL (op) >= 0 && INTVAL (op) < (HOST_WIDE_INT) 1 << 32));
-#else
- return (GET_CODE (op) == CONST_INT
- || (GET_CODE (op) == CONST_DOUBLE
- && CONST_DOUBLE_HIGH (op) == 0));
-#endif
-}
-
-int
-arith5_operand (rtx op, enum machine_mode mode)
-{
- return register_operand (op, mode) || int5_operand (op, mode);
-}
-
/* True iff zdepi can be used to generate this CONST_INT.
zdepi first sign extends a 5 bit signed number to a given field
length, then places this field anywhere in a zero. */
return (mask & (mask - 1)) == 0;
}
-/* True iff depi or extru can be used to compute (reg & OP). */
-int
-and_operand (rtx op, enum machine_mode mode)
-{
- return (register_operand (op, mode)
- || (GET_CODE (op) == CONST_INT && and_mask_p (INTVAL (op))));
-}
-
/* True iff depi can be used to compute (reg | MASK). */
int
-ior_mask_p (unsigned HOST_WIDE_INT mask)
-{
- mask += mask & -mask;
- return (mask & (mask - 1)) == 0;
-}
-
-/* True iff depi can be used to compute (reg | OP). */
-int
-ior_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
-{
- return (GET_CODE (op) == CONST_INT && ior_mask_p (INTVAL (op)));
-}
-
-int
-lhs_lshift_operand (rtx op, enum machine_mode mode)
-{
- return register_operand (op, mode) || lhs_lshift_cint_operand (op, mode);
-}
-
-/* True iff OP is a CONST_INT of the forms 0...0xxxx or 0...01...1xxxx.
- Such values can be the left hand side x in (x << r), using the zvdepi
- instruction. */
-int
-lhs_lshift_cint_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
-{
- unsigned HOST_WIDE_INT x;
- if (GET_CODE (op) != CONST_INT)
- return 0;
- x = INTVAL (op) >> 4;
- return (x & (x + 1)) == 0;
-}
-
-int
-arith32_operand (rtx op, enum machine_mode mode)
-{
- return register_operand (op, mode) || GET_CODE (op) == CONST_INT;
-}
-
-int
-pc_or_label_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
+ior_mask_p (unsigned HOST_WIDE_INT mask)
{
- return (GET_CODE (op) == PC || GET_CODE (op) == LABEL_REF);
+ mask += mask & -mask;
+ return (mask & (mask - 1)) == 0;
}
\f
/* Legitimize PIC addresses. If the address is already
{
rtx pic_ref = orig;
+ gcc_assert (!PA_SYMBOL_REF_TLS_P (orig));
+
/* Labels need special handling. */
if (pic_label_operand (orig, mode))
{
So instead we just emit the raw set, which avoids the movXX
expanders completely. */
+ mark_reg_pointer (reg, BITS_PER_UNIT);
emit_insn (gen_rtx_SET (VOIDmode, reg, orig));
current_function_uses_pic_offset_table = 1;
return reg;
{
rtx insn, tmp_reg;
- if (reg == 0)
- abort ();
+ gcc_assert (reg);
/* Before reload, allocate a temporary register for the intermediate
result. This allows the sequence to be deleted when the final
gen_rtx_PLUS (word_mode, pic_offset_table_rtx,
gen_rtx_HIGH (word_mode, orig)));
pic_ref
- = gen_rtx_MEM (Pmode,
- gen_rtx_LO_SUM (Pmode, tmp_reg,
- gen_rtx_UNSPEC (Pmode,
- gen_rtvec (1, orig),
- 0)));
+ = gen_const_mem (Pmode,
+ gen_rtx_LO_SUM (Pmode, tmp_reg,
+ gen_rtx_UNSPEC (Pmode,
+ gen_rtvec (1, orig),
+ UNSPEC_DLTIND14R)));
current_function_uses_pic_offset_table = 1;
- MEM_NOTRAP_P (pic_ref) = 1;
- RTX_UNCHANGING_P (pic_ref) = 1;
+ mark_reg_pointer (reg, BITS_PER_UNIT);
insn = emit_move_insn (reg, pic_ref);
/* Put a REG_EQUAL note on this insn, so that it can be optimized. */
&& XEXP (XEXP (orig, 0), 0) == pic_offset_table_rtx)
return orig;
- if (reg == 0)
- abort ();
+ gcc_assert (reg);
+ gcc_assert (GET_CODE (XEXP (orig, 0)) == PLUS);
+
+ base = legitimize_pic_address (XEXP (XEXP (orig, 0), 0), Pmode, reg);
+ orig = legitimize_pic_address (XEXP (XEXP (orig, 0), 1), Pmode,
+ base == reg ? 0 : reg);
- if (GET_CODE (XEXP (orig, 0)) == PLUS)
- {
- base = legitimize_pic_address (XEXP (XEXP (orig, 0), 0), Pmode, reg);
- orig = legitimize_pic_address (XEXP (XEXP (orig, 0), 1), Pmode,
- base == reg ? 0 : reg);
- }
- else abort ();
if (GET_CODE (orig) == CONST_INT)
{
if (INT_14_BITS (orig))
pic_ref = gen_rtx_PLUS (Pmode, base, orig);
/* Likewise, should we set special REG_NOTEs here? */
}
+
return pic_ref;
}
+static GTY(()) rtx gen_tls_tga;
+
+static rtx
+gen_tls_get_addr (void)
+{
+ if (!gen_tls_tga)
+ gen_tls_tga = init_one_libfunc ("__tls_get_addr");
+ return gen_tls_tga;
+}
+
+static rtx
+hppa_tls_call (rtx arg)
+{
+ rtx ret;
+
+ ret = gen_reg_rtx (Pmode);
+ emit_library_call_value (gen_tls_get_addr (), ret,
+ LCT_CONST, Pmode, 1, arg, Pmode);
+
+ return ret;
+}
+
+static rtx
+legitimize_tls_address (rtx addr)
+{
+ rtx ret, insn, tmp, t1, t2, tp;
+ enum tls_model model = SYMBOL_REF_TLS_MODEL (addr);
+
+ switch (model)
+ {
+ case TLS_MODEL_GLOBAL_DYNAMIC:
+ tmp = gen_reg_rtx (Pmode);
+ emit_insn (gen_tgd_load (tmp, addr));
+ ret = hppa_tls_call (tmp);
+ break;
+
+ case TLS_MODEL_LOCAL_DYNAMIC:
+ ret = gen_reg_rtx (Pmode);
+ tmp = gen_reg_rtx (Pmode);
+ start_sequence ();
+ emit_insn (gen_tld_load (tmp, addr));
+ t1 = hppa_tls_call (tmp);
+ insn = get_insns ();
+ end_sequence ();
+ t2 = gen_reg_rtx (Pmode);
+ emit_libcall_block (insn, t2, t1,
+ gen_rtx_UNSPEC (Pmode, gen_rtvec (1, const0_rtx),
+ UNSPEC_TLSLDBASE));
+ emit_insn (gen_tld_offset_load (ret, addr, t2));
+ break;
+
+ case TLS_MODEL_INITIAL_EXEC:
+ tp = gen_reg_rtx (Pmode);
+ tmp = gen_reg_rtx (Pmode);
+ ret = gen_reg_rtx (Pmode);
+ emit_insn (gen_tp_load (tp));
+ emit_insn (gen_tie_load (tmp, addr));
+ emit_move_insn (ret, gen_rtx_PLUS (Pmode, tp, tmp));
+ break;
+
+ case TLS_MODEL_LOCAL_EXEC:
+ tp = gen_reg_rtx (Pmode);
+ ret = gen_reg_rtx (Pmode);
+ emit_insn (gen_tp_load (tp));
+ emit_insn (gen_tle_load (ret, addr, tp));
+ break;
+
+ default:
+ gcc_unreachable ();
+ }
+
+ return ret;
+}
+
/* Try machine-dependent ways of modifying an illegitimate address
to be legitimate. If we find one, return the new, valid address.
This macro is used in only one place: `memory_address' in explow.c.
This is for CSE to find several similar references, and only use one Z.
- X can either be a SYMBOL_REF or REG, but because combine can not
+ X can either be a SYMBOL_REF or REG, but because combine cannot
perform a 4->2 combination we do nothing for SYMBOL_REF + D where
D will not fit in 14 bits.
{
rtx orig = x;
- if (flag_pic)
+ /* We need to canonicalize the order of operands in unscaled indexed
+ addresses since the code that checks if an address is valid doesn't
+ always try both orders. */
+ if (!TARGET_NO_SPACE_REGS
+ && GET_CODE (x) == PLUS
+ && GET_MODE (x) == Pmode
+ && REG_P (XEXP (x, 0))
+ && REG_P (XEXP (x, 1))
+ && REG_POINTER (XEXP (x, 0))
+ && !REG_POINTER (XEXP (x, 1)))
+ return gen_rtx_PLUS (Pmode, XEXP (x, 1), XEXP (x, 0));
+
+ if (PA_SYMBOL_REF_TLS_P (x))
+ return legitimize_tls_address (x);
+ else if (flag_pic)
return legitimize_pic_address (x, mode, gen_reg_rtx (Pmode));
/* Strip off CONST. */
if (GET_CODE (x) == PLUS && GET_CODE (XEXP (x, 0)) == MULT
&& GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT
&& shadd_constant_p (INTVAL (XEXP (XEXP (x, 0), 1)))
- && (GET_RTX_CLASS (GET_CODE (XEXP (x, 1))) == 'o'
+ && (OBJECT_P (XEXP (x, 1))
|| GET_CODE (XEXP (x, 1)) == SUBREG)
&& GET_CODE (XEXP (x, 1)) != CONST)
{
if (mode == GET_MODE (orig))
return orig;
- if (REGNO (orig) >= FIRST_PSEUDO_REGISTER)
- abort ();
+ gcc_assert (REGNO (orig) < FIRST_PSEUDO_REGISTER);
return gen_rtx_REG (mode, REGNO (orig));
}
+/* Return 1 if *X is a thread-local symbol. */
+
+static int
+pa_tls_symbol_ref_1 (rtx *x, void *data ATTRIBUTE_UNUSED)
+{
+ return PA_SYMBOL_REF_TLS_P (*x);
+}
+
+/* Return 1 if X contains a thread-local symbol. */
+
+bool
+pa_tls_referenced_p (rtx x)
+{
+ if (!TARGET_HAVE_TLS)
+ return false;
+
+ return for_each_rtx (&x, &pa_tls_symbol_ref_1, 0);
+}
+
/* Emit insns to move operands[1] into operands[0].
Return 1 if we have written out everything that needs to be done to
register rtx operand1 = operands[1];
register rtx tem;
+ /* We can only handle indexed addresses in the destination operand
+ of floating point stores. Thus, we need to break out indexed
+ addresses from the destination operand. */
+ if (GET_CODE (operand0) == MEM && IS_INDEX_ADDR_P (XEXP (operand0, 0)))
+ {
+ /* This is only safe up to the beginning of life analysis. */
+ gcc_assert (!no_new_pseudos);
+
+ tem = copy_to_mode_reg (Pmode, XEXP (operand0, 0));
+ operand0 = replace_equiv_address (operand0, tem);
+ }
+
+ /* On targets with non-equivalent space registers, break out unscaled
+ indexed addresses from the source operand before the final CSE.
+ We have to do this because the REG_POINTER flag is not correctly
+ carried through various optimization passes and CSE may substitute
+ a pseudo without the pointer set for one with the pointer set. As
+ a result, we loose various opportunities to create insns with
+ unscaled indexed addresses. */
+ if (!TARGET_NO_SPACE_REGS
+ && !cse_not_expected
+ && GET_CODE (operand1) == MEM
+ && GET_CODE (XEXP (operand1, 0)) == PLUS
+ && REG_P (XEXP (XEXP (operand1, 0), 0))
+ && REG_P (XEXP (XEXP (operand1, 0), 1)))
+ operand1
+ = replace_equiv_address (operand1,
+ copy_to_mode_reg (Pmode, XEXP (operand1, 0)));
+
if (scratch_reg
&& reload_in_progress && GET_CODE (operand0) == REG
&& REGNO (operand0) >= FIRST_PSEUDO_REGISTER)
if (scratch_reg && reload_in_progress && GET_CODE (operand0) == MEM
&& ((tem = find_replacement (&XEXP (operand0, 0)))
!= XEXP (operand0, 0)))
- operand0 = gen_rtx_MEM (GET_MODE (operand0), tem);
+ operand0 = replace_equiv_address (operand0, tem);
+
if (scratch_reg && reload_in_progress && GET_CODE (operand1) == MEM
&& ((tem = find_replacement (&XEXP (operand1, 0)))
!= XEXP (operand1, 0)))
- operand1 = gen_rtx_MEM (GET_MODE (operand1), tem);
+ operand1 = replace_equiv_address (operand1, tem);
/* Handle secondary reloads for loads/stores of FP registers from
- REG+D addresses where D does not fit in 5 bits, including
+ REG+D addresses where D does not fit in 5 or 14 bits, including
(subreg (mem (addr))) cases. */
- if (fp_reg_operand (operand0, mode)
+ if (scratch_reg
+ && fp_reg_operand (operand0, mode)
&& ((GET_CODE (operand1) == MEM
- && ! memory_address_p (DFmode, XEXP (operand1, 0)))
+ && !memory_address_p ((GET_MODE_SIZE (mode) == 4 ? SFmode : DFmode),
+ XEXP (operand1, 0)))
|| ((GET_CODE (operand1) == SUBREG
&& GET_CODE (XEXP (operand1, 0)) == MEM
- && !memory_address_p (DFmode, XEXP (XEXP (operand1, 0), 0)))))
- && scratch_reg)
+ && !memory_address_p ((GET_MODE_SIZE (mode) == 4
+ ? SFmode : DFmode),
+ XEXP (XEXP (operand1, 0), 0))))))
{
if (GET_CODE (operand1) == SUBREG)
operand1 = XEXP (operand1, 0);
if (!memory_address_p (Pmode, XEXP (operand1, 0)))
{
emit_move_insn (scratch_reg, XEXP (XEXP (operand1, 0), 1));
- emit_move_insn (scratch_reg, gen_rtx_fmt_ee (GET_CODE (XEXP (operand1, 0)),
- Pmode,
- XEXP (XEXP (operand1, 0), 0),
- scratch_reg));
+ emit_move_insn (scratch_reg,
+ gen_rtx_fmt_ee (GET_CODE (XEXP (operand1, 0)),
+ Pmode,
+ XEXP (XEXP (operand1, 0), 0),
+ scratch_reg));
}
else
emit_move_insn (scratch_reg, XEXP (operand1, 0));
emit_insn (gen_rtx_SET (VOIDmode, operand0,
- gen_rtx_MEM (mode, scratch_reg)));
+ replace_equiv_address (operand1, scratch_reg)));
return 1;
}
- else if (fp_reg_operand (operand1, mode)
+ else if (scratch_reg
+ && fp_reg_operand (operand1, mode)
&& ((GET_CODE (operand0) == MEM
- && ! memory_address_p (DFmode, XEXP (operand0, 0)))
+ && !memory_address_p ((GET_MODE_SIZE (mode) == 4
+ ? SFmode : DFmode),
+ XEXP (operand0, 0)))
|| ((GET_CODE (operand0) == SUBREG)
&& GET_CODE (XEXP (operand0, 0)) == MEM
- && !memory_address_p (DFmode, XEXP (XEXP (operand0, 0), 0))))
- && scratch_reg)
+ && !memory_address_p ((GET_MODE_SIZE (mode) == 4
+ ? SFmode : DFmode),
+ XEXP (XEXP (operand0, 0), 0)))))
{
if (GET_CODE (operand0) == SUBREG)
operand0 = XEXP (operand0, 0);
}
else
emit_move_insn (scratch_reg, XEXP (operand0, 0));
- emit_insn (gen_rtx_SET (VOIDmode, gen_rtx_MEM (mode, scratch_reg),
+ emit_insn (gen_rtx_SET (VOIDmode,
+ replace_equiv_address (operand0, scratch_reg),
operand1));
return 1;
}
/* Handle secondary reloads for loads of FP registers from constant
expressions by forcing the constant into memory.
- use scratch_reg to hold the address of the memory location.
+ Use scratch_reg to hold the address of the memory location.
The proper fix is to change PREFERRED_RELOAD_CLASS to return
NO_REGS when presented with a const_int and a register class
containing only FP registers. Doing so unfortunately creates
more problems than it solves. Fix this for 2.5. */
- else if (fp_reg_operand (operand0, mode)
+ else if (scratch_reg
&& CONSTANT_P (operand1)
- && scratch_reg)
+ && fp_reg_operand (operand0, mode))
{
- rtx xoperands[2];
+ rtx const_mem, xoperands[2];
/* SCRATCH_REG will hold an address and maybe the actual data. We want
it in WORD_MODE regardless of what mode it was originally given
/* Force the constant into memory and put the address of the
memory location into scratch_reg. */
+ const_mem = force_const_mem (mode, operand1);
xoperands[0] = scratch_reg;
- xoperands[1] = XEXP (force_const_mem (mode, operand1), 0);
+ xoperands[1] = XEXP (const_mem, 0);
emit_move_sequence (xoperands, Pmode, 0);
/* Now load the destination register. */
emit_insn (gen_rtx_SET (mode, operand0,
- gen_rtx_MEM (mode, scratch_reg)));
+ replace_equiv_address (const_mem, scratch_reg)));
return 1;
}
/* Handle secondary reloads for SAR. These occur when trying to load
the SAR from memory, FP register, or with a constant. */
- else if (GET_CODE (operand0) == REG
+ else if (scratch_reg
+ && GET_CODE (operand0) == REG
&& REGNO (operand0) < FIRST_PSEUDO_REGISTER
&& REGNO_REG_CLASS (REGNO (operand0)) == SHIFT_REGS
&& (GET_CODE (operand1) == MEM
|| GET_CODE (operand1) == CONST_INT
|| (GET_CODE (operand1) == REG
- && FP_REG_CLASS_P (REGNO_REG_CLASS (REGNO (operand1)))))
- && scratch_reg)
+ && FP_REG_CLASS_P (REGNO_REG_CLASS (REGNO (operand1))))))
{
/* D might not fit in 14 bits either; for such cases load D into
scratch reg. */
OPERAND0. */
scratch_reg = force_mode (GET_MODE (operand0), scratch_reg);
- emit_move_insn (scratch_reg, gen_rtx_MEM (GET_MODE (operand0),
- scratch_reg));
+ emit_move_insn (scratch_reg,
+ replace_equiv_address (operand1, scratch_reg));
}
else
{
emit_move_insn (operand0, scratch_reg);
return 1;
}
- /* Handle most common case: storing into a register. */
+ /* Handle the most common case: storing into a register. */
else if (register_operand (operand0, mode))
{
if (register_operand (operand1, mode)
/* Only `general_operands' can come here, so MEM is ok. */
|| GET_CODE (operand1) == MEM)
{
- /* Run this case quickly. */
+ /* Various sets are created during RTL generation which don't
+ have the REG_POINTER flag correctly set. After the CSE pass,
+ instruction recognition can fail if we don't consistently
+ set this flag when performing register copies. This should
+ also improve the opportunities for creating insns that use
+ unscaled indexing. */
+ if (REG_P (operand0) && REG_P (operand1))
+ {
+ if (REG_POINTER (operand1)
+ && !REG_POINTER (operand0)
+ && !HARD_REGISTER_P (operand0))
+ copy_reg_pointer (operand0, operand1);
+ else if (REG_POINTER (operand0)
+ && !REG_POINTER (operand1)
+ && !HARD_REGISTER_P (operand1))
+ copy_reg_pointer (operand1, operand0);
+ }
+
+ /* When MEMs are broken out, the REG_POINTER flag doesn't
+ get set. In some cases, we can set the REG_POINTER flag
+ from the declaration for the MEM. */
+ if (REG_P (operand0)
+ && GET_CODE (operand1) == MEM
+ && !REG_POINTER (operand0))
+ {
+ tree decl = MEM_EXPR (operand1);
+
+ /* Set the register pointer flag and register alignment
+ if the declaration for this memory reference is a
+ pointer type. Fortran indirect argument references
+ are ignored. */
+ if (decl
+ && !(flag_argument_noalias > 1
+ && TREE_CODE (decl) == INDIRECT_REF
+ && TREE_CODE (TREE_OPERAND (decl, 0)) == PARM_DECL))
+ {
+ tree type;
+
+ /* If this is a COMPONENT_REF, use the FIELD_DECL from
+ tree operand 1. */
+ if (TREE_CODE (decl) == COMPONENT_REF)
+ decl = TREE_OPERAND (decl, 1);
+
+ type = TREE_TYPE (decl);
+ if (TREE_CODE (type) == ARRAY_TYPE)
+ type = get_inner_array_type (type);
+
+ if (POINTER_TYPE_P (type))
+ {
+ int align;
+
+ type = TREE_TYPE (type);
+ /* Using TYPE_ALIGN_OK is rather conservative as
+ only the ada frontend actually sets it. */
+ align = (TYPE_ALIGN_OK (type) ? TYPE_ALIGN (type)
+ : BITS_PER_UNIT);
+ mark_reg_pointer (operand0, align);
+ }
+ }
+ }
+
emit_insn (gen_rtx_SET (VOIDmode, operand0, operand1));
return 1;
}
{
/* Save away the constant part of the expression. */
const_part = XEXP (XEXP (operand1, 0), 1);
- if (GET_CODE (const_part) != CONST_INT)
- abort ();
+ gcc_assert (GET_CODE (const_part) == CONST_INT);
/* Force the function label into memory. */
temp = force_const_mem (mode, XEXP (XEXP (operand1, 0), 0));
&& (reload_completed || reload_in_progress)
&& flag_pic)
{
- operands[1] = force_const_mem (mode, operand1);
- operands[1] = legitimize_pic_address (XEXP (operands[1], 0),
+ rtx const_mem = force_const_mem (mode, operand1);
+ operands[1] = legitimize_pic_address (XEXP (const_mem, 0),
mode, temp);
+ operands[1] = replace_equiv_address (const_mem, operands[1]);
emit_move_sequence (operands, mode, temp);
}
else
{
operands[1] = legitimize_pic_address (operand1, mode, temp);
+ if (REG_P (operand0) && REG_P (operands[1]))
+ copy_reg_pointer (operand0, operands[1]);
emit_insn (gen_rtx_SET (VOIDmode, operand0, operands[1]));
}
}
Don't mark hard registers though. That loses. */
if (GET_CODE (operand0) == REG
&& REGNO (operand0) >= FIRST_PSEUDO_REGISTER)
- REG_POINTER (operand0) = 1;
+ mark_reg_pointer (operand0, BITS_PER_UNIT);
if (REGNO (temp) >= FIRST_PSEUDO_REGISTER)
- REG_POINTER (temp) = 1;
+ mark_reg_pointer (temp, BITS_PER_UNIT);
+
if (ishighonly)
set = gen_rtx_SET (mode, operand0, temp);
else
}
return 1;
}
+ else if (pa_tls_referenced_p (operand1))
+ {
+ rtx tmp = operand1;
+ rtx addend = NULL;
+
+ if (GET_CODE (tmp) == CONST && GET_CODE (XEXP (tmp, 0)) == PLUS)
+ {
+ addend = XEXP (XEXP (tmp, 0), 1);
+ tmp = XEXP (XEXP (tmp, 0), 0);
+ }
+
+ gcc_assert (GET_CODE (tmp) == SYMBOL_REF);
+ tmp = legitimize_tls_address (tmp);
+ if (addend)
+ {
+ tmp = gen_rtx_PLUS (mode, tmp, addend);
+ tmp = force_operand (tmp, operands[0]);
+ }
+ operands[1] = tmp;
+ }
else if (GET_CODE (operand1) != CONST_INT
- || ! cint_ok_for_move (INTVAL (operand1)))
+ || !cint_ok_for_move (INTVAL (operand1)))
{
- rtx extend = NULL_RTX;
- rtx temp;
+ rtx insn, temp;
+ rtx op1 = operand1;
+ HOST_WIDE_INT value = 0;
+ HOST_WIDE_INT insv = 0;
+ int insert = 0;
+
+ if (GET_CODE (operand1) == CONST_INT)
+ value = INTVAL (operand1);
- if (TARGET_64BIT && GET_CODE (operand1) == CONST_INT
+ if (TARGET_64BIT
+ && GET_CODE (operand1) == CONST_INT
&& HOST_BITS_PER_WIDE_INT > 32
&& GET_MODE_BITSIZE (GET_MODE (operand0)) > 32)
{
- HOST_WIDE_INT val = INTVAL (operand1);
HOST_WIDE_INT nval;
/* Extract the low order 32 bits of the value and sign extend.
can use the original value as-is. If the new value is
different, we use it and insert the most-significant 32-bits
of the original value into the final result. */
- nval = ((val & (((HOST_WIDE_INT) 2 << 31) - 1))
+ nval = ((value & (((HOST_WIDE_INT) 2 << 31) - 1))
^ ((HOST_WIDE_INT) 1 << 31)) - ((HOST_WIDE_INT) 1 << 31);
- if (val != nval)
+ if (value != nval)
{
#if HOST_BITS_PER_WIDE_INT > 32
- extend = GEN_INT (val >> 32);
+ insv = value >= 0 ? value >> 32 : ~(~value >> 32);
#endif
+ insert = 1;
+ value = nval;
operand1 = GEN_INT (nval);
}
}
if (reload_in_progress || reload_completed)
- temp = operand0;
+ temp = scratch_reg ? scratch_reg : operand0;
else
temp = gen_reg_rtx (mode);
because PLUS uses an 11-bit immediate and the insn sequence
generated is not as efficient as the one using HIGH/LO_SUM. */
if (GET_CODE (operand1) == CONST_INT
- && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT)
+ && GET_MODE_BITSIZE (mode) <= HOST_BITS_PER_WIDE_INT
+ && !insert)
{
/* Directly break constant into high and low parts. This
provides better optimization opportunities because various
We use a 14-bit signed low part except when the addition
of 0x4000 to the high part might change the sign of the
high part. */
- HOST_WIDE_INT value = INTVAL (operand1);
HOST_WIDE_INT low = value & 0x3fff;
HOST_WIDE_INT high = value & ~ 0x3fff;
operands[1] = gen_rtx_LO_SUM (mode, temp, operand1);
}
- emit_move_insn (operands[0], operands[1]);
+ insn = emit_move_insn (operands[0], operands[1]);
+
+ /* Now insert the most significant 32 bits of the value
+ into the register. When we don't have a second register
+ available, it could take up to nine instructions to load
+ a 64-bit integer constant. Prior to reload, we force
+ constants that would take more than three instructions
+ to load to the constant pool. During and after reload,
+ we have to handle all possible values. */
+ if (insert)
+ {
+ /* Use a HIGH/LO_SUM/INSV sequence if we have a second
+ register and the value to be inserted is outside the
+ range that can be loaded with three depdi instructions. */
+ if (temp != operand0 && (insv >= 16384 || insv < -16384))
+ {
+ operand1 = GEN_INT (insv);
+
+ emit_insn (gen_rtx_SET (VOIDmode, temp,
+ gen_rtx_HIGH (mode, operand1)));
+ emit_move_insn (temp, gen_rtx_LO_SUM (mode, temp, operand1));
+ emit_insn (gen_insv (operand0, GEN_INT (32),
+ const0_rtx, temp));
+ }
+ else
+ {
+ int len = 5, pos = 27;
+
+ /* Insert the bits using the depdi instruction. */
+ while (pos >= 0)
+ {
+ HOST_WIDE_INT v5 = ((insv & 31) ^ 16) - 16;
+ HOST_WIDE_INT sign = v5 < 0;
+
+ /* Left extend the insertion. */
+ insv = (insv >= 0 ? insv >> len : ~(~insv >> len));
+ while (pos > 0 && (insv & 1) == sign)
+ {
+ insv = (insv >= 0 ? insv >> 1 : ~(~insv >> 1));
+ len += 1;
+ pos -= 1;
+ }
+
+ emit_insn (gen_insv (operand0, GEN_INT (len),
+ GEN_INT (pos), GEN_INT (v5)));
+
+ len = pos > 0 && pos < 5 ? pos : 5;
+ pos -= len;
+ }
+ }
+ }
- if (extend != NULL_RTX)
- emit_insn (gen_insv (operands[0], GEN_INT (32), const0_rtx,
- extend));
+ REG_NOTES (insn)
+ = gen_rtx_EXPR_LIST (REG_EQUAL, op1, REG_NOTES (insn));
return 1;
}
case CONSTRUCTOR:
{
- register tree link;
- for (link = CONSTRUCTOR_ELTS (exp); link; link = TREE_CHAIN (link))
- if (TREE_VALUE (link) != 0)
- reloc |= reloc_needed (TREE_VALUE (link));
+ tree value;
+ unsigned HOST_WIDE_INT ix;
+
+ FOR_EACH_CONSTRUCTOR_VALUE (CONSTRUCTOR_ELTS (exp), ix, value)
+ if (value)
+ reloc |= reloc_needed (value);
}
break;
\f
/* Return the best assembler insn template
- for moving operands[1] into operands[0] as a fullword. */
+ for moving operands[1] into operands[0] as a fullword. */
const char *
singlemove_string (rtx *operands)
{
long i;
REAL_VALUE_TYPE d;
- if (GET_MODE (operands[1]) != SFmode)
- abort ();
+ gcc_assert (GET_MODE (operands[1]) == SFmode);
/* Translate the CONST_DOUBLE to a CONST_INT with the same target
bit pattern. */
optype1 = RNDOP;
/* Check for the cases that the operand constraints are not
- supposed to allow to happen. Abort if we get one,
- because generating code for these cases is painful. */
-
- if (optype0 != REGOP && optype1 != REGOP)
- abort ();
+ supposed to allow to happen. */
+ gcc_assert (optype0 == REGOP || optype1 == REGOP);
/* Handle auto decrementing and incrementing loads and stores
specifically, since the structure of the function doesn't work
rtx high_reg = gen_rtx_SUBREG (SImode, operands[1], 0);
operands[0] = XEXP (addr, 0);
- if (GET_CODE (operands[1]) != REG || GET_CODE (operands[0]) != REG)
- abort ();
-
- if (!reg_overlap_mentioned_p (high_reg, addr))
- {
- /* No overlap between high target register and address
- register. (We do this in a non-obvious way to
- save a register file writeback) */
- if (GET_CODE (addr) == POST_INC)
- return "{stws|stw},ma %1,8(%0)\n\tstw %R1,-4(%0)";
- return "{stws|stw},ma %1,-8(%0)\n\tstw %R1,12(%0)";
- }
- else
- abort ();
+ gcc_assert (GET_CODE (operands[1]) == REG
+ && GET_CODE (operands[0]) == REG);
+
+ gcc_assert (!reg_overlap_mentioned_p (high_reg, addr));
+
+ /* No overlap between high target register and address
+ register. (We do this in a non-obvious way to
+ save a register file writeback) */
+ if (GET_CODE (addr) == POST_INC)
+ return "{stws|stw},ma %1,8(%0)\n\tstw %R1,-4(%0)";
+ return "{stws|stw},ma %1,-8(%0)\n\tstw %R1,12(%0)";
}
else if (GET_CODE (addr) == PRE_INC || GET_CODE (addr) == PRE_DEC)
{
rtx high_reg = gen_rtx_SUBREG (SImode, operands[1], 0);
operands[0] = XEXP (addr, 0);
- if (GET_CODE (operands[1]) != REG || GET_CODE (operands[0]) != REG)
- abort ();
-
- if (!reg_overlap_mentioned_p (high_reg, addr))
- {
- /* No overlap between high target register and address
- register. (We do this in a non-obvious way to
- save a register file writeback) */
- if (GET_CODE (addr) == PRE_INC)
- return "{stws|stw},mb %1,8(%0)\n\tstw %R1,4(%0)";
- return "{stws|stw},mb %1,-8(%0)\n\tstw %R1,4(%0)";
- }
- else
- abort ();
+ gcc_assert (GET_CODE (operands[1]) == REG
+ && GET_CODE (operands[0]) == REG);
+
+ gcc_assert (!reg_overlap_mentioned_p (high_reg, addr));
+ /* No overlap between high target register and address
+ register. (We do this in a non-obvious way to save a
+ register file writeback) */
+ if (GET_CODE (addr) == PRE_INC)
+ return "{stws|stw},mb %1,8(%0)\n\tstw %R1,4(%0)";
+ return "{stws|stw},mb %1,-8(%0)\n\tstw %R1,4(%0)";
}
}
if (optype1 == MEMOP)
rtx high_reg = gen_rtx_SUBREG (SImode, operands[0], 0);
operands[1] = XEXP (addr, 0);
- if (GET_CODE (operands[0]) != REG || GET_CODE (operands[1]) != REG)
- abort ();
+ gcc_assert (GET_CODE (operands[0]) == REG
+ && GET_CODE (operands[1]) == REG);
if (!reg_overlap_mentioned_p (high_reg, addr))
{
rtx high_reg = gen_rtx_SUBREG (SImode, operands[0], 0);
operands[1] = XEXP (addr, 0);
- if (GET_CODE (operands[0]) != REG || GET_CODE (operands[1]) != REG)
- abort ();
+ gcc_assert (GET_CODE (operands[0]) == REG
+ && GET_CODE (operands[1]) == REG);
if (!reg_overlap_mentioned_p (high_reg, addr))
{
{
output_asm_insn ("fstd%F0 %1,%0", operands);
}
- else if (operands[1] == CONST0_RTX (GET_MODE (operands[0])))
+ else
{
- if (GET_CODE (operands[0]) == REG)
- {
- rtx xoperands[2];
- xoperands[1] = gen_rtx_REG (SImode, REGNO (operands[0]) + 1);
- xoperands[0] = operands[0];
- output_asm_insn ("copy %%r0,%0\n\tcopy %%r0,%1", xoperands);
- }
+ rtx xoperands[2];
+
+ gcc_assert (operands[1] == CONST0_RTX (GET_MODE (operands[0])));
+
/* This is a pain. You have to be prepared to deal with an
arbitrary address here including pre/post increment/decrement.
so avoid this in the MD. */
- else
- abort ();
+ gcc_assert (GET_CODE (operands[0]) == REG);
+
+ xoperands[1] = gen_rtx_REG (SImode, REGNO (operands[0]) + 1);
+ xoperands[0] = operands[0];
+ output_asm_insn ("copy %%r0,%0\n\tcopy %%r0,%1", xoperands);
}
- else abort ();
return "";
}
\f
else if (CONSTANT_P (XEXP (addr, 1)))
addr = XEXP (addr, 0);
else
- abort ();
+ gcc_unreachable ();
}
- if (GET_CODE (addr) == REG)
- return addr;
- abort ();
+ gcc_assert (GET_CODE (addr) == REG);
+ return addr;
}
/* Emit code to perform a block move.
OPERANDS[3] is a register for temporary storage.
OPERANDS[4] is the size as a CONST_INT
OPERANDS[5] is the alignment safe to use, as a CONST_INT.
- OPERANDS[6] is another temporary register. */
+ OPERANDS[6] is another temporary register. */
const char *
output_block_move (rtx *operands, int size_is_constant ATTRIBUTE_UNUSED)
return "";
default:
- abort ();
+ gcc_unreachable ();
}
}
count insns rather than emit them. */
static int
-compute_movstr_length (rtx insn)
+compute_movmem_length (rtx insn)
{
rtx pat = PATTERN (insn);
unsigned int align = INTVAL (XEXP (XVECEXP (pat, 0, 7), 0));
return "";
default:
- abort ();
+ gcc_unreachable ();
}
}
count insns rather than emit them. */
static int
-compute_clrstr_length (rtx insn)
+compute_clrmem_length (rtx insn)
{
rtx pat = PATTERN (insn);
unsigned int align = INTVAL (XEXP (XVECEXP (pat, 0, 4), 0));
if ((mask & (1 << ms0)) == 0)
break;
- if (ms0 != 32)
- abort ();
+ gcc_assert (ms0 == 32);
if (ls1 == 32)
{
len = ls0;
- if (len == 0)
- abort ();
+ gcc_assert (len);
operands[2] = GEN_INT (len);
return "{extru|extrw,u} %1,31,%2,%0";
if ((mask & ((unsigned HOST_WIDE_INT) 1 << ms0)) == 0)
break;
- if (ms0 != HOST_BITS_PER_WIDE_INT)
- abort ();
+ gcc_assert (ms0 == HOST_BITS_PER_WIDE_INT);
if (ls1 == HOST_BITS_PER_WIDE_INT)
{
len = ls0;
- if (len == 0)
- abort ();
+ gcc_assert (len);
operands[2] = GEN_INT (len);
return "extrd,u %1,63,%2,%0";
if ((mask & (1 << bs1)) == 0)
break;
- if (bs1 != 32 && ((unsigned HOST_WIDE_INT) 1 << bs1) <= mask)
- abort ();
+ gcc_assert (bs1 == 32 || ((unsigned HOST_WIDE_INT) 1 << bs1) > mask);
p = 31 - bs0;
len = bs1 - bs0;
if ((mask & ((unsigned HOST_WIDE_INT) 1 << bs1)) == 0)
break;
- if (bs1 != HOST_BITS_PER_WIDE_INT
- && ((unsigned HOST_WIDE_INT) 1 << bs1) <= mask)
- abort ();
+ gcc_assert (bs1 == HOST_BITS_PER_WIDE_INT
+ || ((unsigned HOST_WIDE_INT) 1 << bs1) > mask);
p = 63 - bs0;
len = bs1 - bs0;
}
\f
/* Target hook for assembling integer objects. This code handles
- aligned SI and DI integers specially, since function references must
- be preceded by P%. */
+ aligned SI and DI integers specially since function references
+ must be preceded by P%. */
static bool
pa_assemble_integer (rtx x, unsigned int size, int aligned_p)
{
- if (size == UNITS_PER_WORD && aligned_p
+ if (size == UNITS_PER_WORD
+ && aligned_p
&& function_label_operand (x, VOIDmode))
{
fputs (size == 8? "\t.dword\tP%" : "\t.word\tP%", asm_out_file);
{
int i;
int chars_output;
- unsigned char partial_output[16]; /* Max space 4 chars can occupy. */
+ unsigned char partial_output[16]; /* Max space 4 chars can occupy. */
/* The HP assembler can only take strings of 256 characters at one
time. This is a limitation on input line length, *not* the
{
rtx pattern = PATTERN (next);
- /* If it a reversed fp conditional branch (eg uses add,tr)
+ /* If it a reversed fp conditional branch (e.g. uses add,tr)
and CCFP dies, then reverse our conditional and the branch
to avoid the add,tr. */
if (GET_CODE (pattern) == SET
/* Global variables set by output_function_prologue(). */
/* Size of frame. Need to know this to emit return insns from
leaf procedures. */
-static int actual_fsize;
-static int local_fsize, save_fregs;
+static HOST_WIDE_INT actual_fsize, local_fsize;
+static int save_fregs;
/* Emit RTL to store REG at the memory location specified by BASE+DISP.
Handle case where DISP > 8k by using the add_high_const patterns.
in %r1. There is code in expand_hppa_{prologue,epilogue} that knows this.*/
static void
-store_reg (int reg, int disp, int base)
+store_reg (int reg, HOST_WIDE_INT disp, int base)
{
rtx insn, dest, src, basereg;
dest = gen_rtx_MEM (word_mode, plus_constant (basereg, disp));
insn = emit_move_insn (dest, src);
}
+ else if (TARGET_64BIT && !VAL_32_BITS_P (disp))
+ {
+ rtx delta = GEN_INT (disp);
+ rtx tmpreg = gen_rtx_REG (Pmode, 1);
+
+ emit_move_insn (tmpreg, delta);
+ insn = emit_move_insn (tmpreg, gen_rtx_PLUS (Pmode, tmpreg, basereg));
+ if (DO_FRAME_NOTES)
+ {
+ REG_NOTES (insn)
+ = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR,
+ gen_rtx_SET (VOIDmode, tmpreg,
+ gen_rtx_PLUS (Pmode, basereg, delta)),
+ REG_NOTES (insn));
+ RTX_FRAME_RELATED_P (insn) = 1;
+ }
+ dest = gen_rtx_MEM (word_mode, tmpreg);
+ insn = emit_move_insn (dest, src);
+ }
else
{
rtx delta = GEN_INT (disp);
rtx high = gen_rtx_PLUS (Pmode, basereg, gen_rtx_HIGH (Pmode, delta));
rtx tmpreg = gen_rtx_REG (Pmode, 1);
+
emit_move_insn (tmpreg, high);
dest = gen_rtx_MEM (word_mode, gen_rtx_LO_SUM (Pmode, tmpreg, delta));
insn = emit_move_insn (dest, src);
add MOD to BASE. MOD must be <= 8k. */
static void
-store_reg_modify (int base, int reg, int mod)
+store_reg_modify (int base, int reg, HOST_WIDE_INT mod)
{
rtx insn, basereg, srcreg, delta;
- if (! VAL_14_BITS_P (mod))
- abort ();
+ gcc_assert (VAL_14_BITS_P (mod));
basereg = gen_rtx_REG (Pmode, base);
srcreg = gen_rtx_REG (word_mode, reg);
RTX_FRAME_RELATED_P (insn) = 1;
/* RTX_FRAME_RELATED_P must be set on each frame related set
- in a parallel with more than one element. Don't set
- RTX_FRAME_RELATED_P in the first set if reg is temporary
- register 1. The effect of this operation is recorded in
- the initial copy. */
- if (reg != 1)
- {
- RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, 0)) = 1;
- RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, 1)) = 1;
- }
- else
- {
- /* The first element of a PARALLEL is always processed if it is
- a SET. Thus, we need an expression list for this case. */
- REG_NOTES (insn)
- = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR,
- gen_rtx_SET (VOIDmode, basereg,
- gen_rtx_PLUS (word_mode, basereg, delta)),
- REG_NOTES (insn));
- }
+ in a parallel with more than one element. */
+ RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, 0)) = 1;
+ RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, 1)) = 1;
}
}
There is code in expand_hppa_{prologue,epilogue} that knows about this. */
static void
-set_reg_plus_d (int reg, int base, int disp, int note)
+set_reg_plus_d (int reg, int base, HOST_WIDE_INT disp, int note)
{
rtx insn;
insn = emit_move_insn (gen_rtx_REG (Pmode, reg),
plus_constant (gen_rtx_REG (Pmode, base), disp));
}
+ else if (TARGET_64BIT && !VAL_32_BITS_P (disp))
+ {
+ rtx basereg = gen_rtx_REG (Pmode, base);
+ rtx delta = GEN_INT (disp);
+ rtx tmpreg = gen_rtx_REG (Pmode, 1);
+
+ emit_move_insn (tmpreg, delta);
+ insn = emit_move_insn (gen_rtx_REG (Pmode, reg),
+ gen_rtx_PLUS (Pmode, tmpreg, basereg));
+ if (DO_FRAME_NOTES)
+ REG_NOTES (insn)
+ = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR,
+ gen_rtx_SET (VOIDmode, tmpreg,
+ gen_rtx_PLUS (Pmode, basereg, delta)),
+ REG_NOTES (insn));
+ }
else
{
rtx basereg = gen_rtx_REG (Pmode, base);
rtx delta = GEN_INT (disp);
+ rtx tmpreg = gen_rtx_REG (Pmode, 1);
- emit_move_insn (gen_rtx_REG (Pmode, 1),
+ emit_move_insn (tmpreg,
gen_rtx_PLUS (Pmode, basereg,
gen_rtx_HIGH (Pmode, delta)));
insn = emit_move_insn (gen_rtx_REG (Pmode, reg),
- gen_rtx_LO_SUM (Pmode, gen_rtx_REG (Pmode, 1),
- delta));
+ gen_rtx_LO_SUM (Pmode, tmpreg, delta));
}
if (DO_FRAME_NOTES && note)
RTX_FRAME_RELATED_P (insn) = 1;
}
-int
-compute_frame_size (int size, int *fregs_live)
+HOST_WIDE_INT
+compute_frame_size (HOST_WIDE_INT size, int *fregs_live)
{
int freg_saved = 0;
int i, j;
/* hppa_expand_prologue does the dirty work now. We just need
to output the assembler directives which denote the start
of a function. */
- fprintf (file, "\t.CALLINFO FRAME=%d", actual_fsize);
+ fprintf (file, "\t.CALLINFO FRAME=" HOST_WIDE_INT_PRINT_DEC, actual_fsize);
if (regs_ever_live[2])
fputs (",CALLS,SAVE_RP", file);
else
/* The SAVE_SP flag is used to indicate that register %r3 is stored
at the beginning of the frame and that it is used as the frame
pointer for the frame. We do this because our current frame
- layout doesn't conform to that specified in the the HP runtime
+ layout doesn't conform to that specified in the HP runtime
documentation and we need a way to indicate to programs such as
GDB where %r3 is saved. The SAVE_SP flag was chosen because it
isn't used by HP compilers but is supported by the assembler.
hppa_expand_prologue (void)
{
int merge_sp_adjust_with_store = 0;
- int size = get_frame_size ();
- int i, offset;
+ HOST_WIDE_INT size = get_frame_size ();
+ HOST_WIDE_INT offset;
+ int i;
rtx insn, tmpreg;
gr_saved = 0;
frames. */
insn = emit_move_insn (tmpreg, frame_pointer_rtx);
if (DO_FRAME_NOTES)
- {
- /* We need to record the frame pointer save here since the
- new frame pointer is set in the following insn. */
- RTX_FRAME_RELATED_P (insn) = 1;
- REG_NOTES (insn)
- = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR,
- gen_rtx_SET (VOIDmode,
- gen_rtx_MEM (word_mode, stack_pointer_rtx),
- frame_pointer_rtx),
- REG_NOTES (insn));
- }
+ RTX_FRAME_RELATED_P (insn) = 1;
insn = emit_move_insn (frame_pointer_rtx, stack_pointer_rtx);
if (DO_FRAME_NOTES)
So instead use stwm to store at *sp and post-increment the
stack pointer as an atomic operation. Then increment sp to
finish allocating the new frame. */
- int adjust1 = 8192 - 64;
- int adjust2 = actual_fsize - adjust1;
+ HOST_WIDE_INT adjust1 = 8192 - 64;
+ HOST_WIDE_INT adjust2 = actual_fsize - adjust1;
store_reg_modify (STACK_POINTER_REGNUM, 1, adjust1);
set_reg_plus_d (STACK_POINTER_REGNUM, STACK_POINTER_REGNUM,
Handle case where DISP > 8k by using the add_high_const patterns. */
static void
-load_reg (int reg, int disp, int base)
+load_reg (int reg, HOST_WIDE_INT disp, int base)
{
- rtx src, dest, basereg;
+ rtx dest = gen_rtx_REG (word_mode, reg);
+ rtx basereg = gen_rtx_REG (Pmode, base);
+ rtx src;
- dest = gen_rtx_REG (word_mode, reg);
- basereg = gen_rtx_REG (Pmode, base);
if (VAL_14_BITS_P (disp))
+ src = gen_rtx_MEM (word_mode, plus_constant (basereg, disp));
+ else if (TARGET_64BIT && !VAL_32_BITS_P (disp))
{
- src = gen_rtx_MEM (word_mode, plus_constant (basereg, disp));
- emit_move_insn (dest, src);
+ rtx delta = GEN_INT (disp);
+ rtx tmpreg = gen_rtx_REG (Pmode, 1);
+
+ emit_move_insn (tmpreg, delta);
+ if (TARGET_DISABLE_INDEXING)
+ {
+ emit_move_insn (tmpreg, gen_rtx_PLUS (Pmode, tmpreg, basereg));
+ src = gen_rtx_MEM (word_mode, tmpreg);
+ }
+ else
+ src = gen_rtx_MEM (word_mode, gen_rtx_PLUS (Pmode, tmpreg, basereg));
}
else
{
rtx delta = GEN_INT (disp);
rtx high = gen_rtx_PLUS (Pmode, basereg, gen_rtx_HIGH (Pmode, delta));
rtx tmpreg = gen_rtx_REG (Pmode, 1);
+
emit_move_insn (tmpreg, high);
src = gen_rtx_MEM (word_mode, gen_rtx_LO_SUM (Pmode, tmpreg, delta));
- emit_move_insn (dest, src);
}
+
+ emit_move_insn (dest, src);
}
/* Update the total code bytes output to the text section. */
fputs ("\t.EXIT\n\t.PROCEND\n", file);
+ if (TARGET_SOM && TARGET_GAS)
+ {
+ /* We done with this subspace except possibly for some additional
+ debug information. Forget that we are in this subspace to ensure
+ that the next function is output in its own subspace. */
+ forget_section ();
+ }
+
if (INSN_ADDRESSES_SET_P ())
{
insn = get_last_nonnote_insn ();
hppa_expand_epilogue (void)
{
rtx tmpreg;
- int offset, i;
+ HOST_WIDE_INT offset;
+ HOST_WIDE_INT ret_off = 0;
+ int i;
int merge_sp_adjust_with_load = 0;
- int ret_off = 0;
/* We will use this often. */
tmpreg = gen_rtx_REG (word_mode, 1);
return get_hard_reg_initial_val (word_mode, PIC_OFFSET_TABLE_REGNUM);
}
+#ifndef NO_DEFERRED_PROFILE_COUNTERS
+#define NO_DEFERRED_PROFILE_COUNTERS 0
+#endif
+
+/* Define heap vector type for funcdef numbers. */
+DEF_VEC_I(int);
+DEF_VEC_ALLOC_I(int,heap);
+
+/* Vector of funcdef numbers. */
+static VEC(int,heap) *funcdef_nos;
+
+/* Output deferred profile counters. */
+static void
+output_deferred_profile_counters (void)
+{
+ unsigned int i;
+ int align, n;
+
+ if (VEC_empty (int, funcdef_nos))
+ return;
+
+ data_section ();
+ align = MIN (BIGGEST_ALIGNMENT, LONG_TYPE_SIZE);
+ ASM_OUTPUT_ALIGN (asm_out_file, floor_log2 (align / BITS_PER_UNIT));
+
+ for (i = 0; VEC_iterate (int, funcdef_nos, i, n); i++)
+ {
+ targetm.asm_out.internal_label (asm_out_file, "LP", n);
+ assemble_integer (const0_rtx, LONG_TYPE_SIZE / BITS_PER_UNIT, align, 1);
+ }
+
+ VEC_free (int, heap, funcdef_nos);
+}
+
void
hppa_profile_hook (int label_no)
{
+ /* We use SImode for the address of the function in both 32 and
+ 64-bit code to avoid having to provide DImode versions of the
+ lcla2 and load_offset_label_address insn patterns. */
+ rtx reg = gen_reg_rtx (SImode);
+ rtx label_rtx = gen_label_rtx ();
rtx begin_label_rtx, call_insn;
char begin_label_name[16];
ASM_GENERATE_INTERNAL_LABEL (begin_label_name, FUNC_BEGIN_PROLOG_LABEL,
label_no);
- begin_label_rtx = gen_rtx_SYMBOL_REF (Pmode, ggc_strdup (begin_label_name));
+ begin_label_rtx = gen_rtx_SYMBOL_REF (SImode, ggc_strdup (begin_label_name));
if (TARGET_64BIT)
emit_move_insn (arg_pointer_rtx,
emit_move_insn (gen_rtx_REG (word_mode, 26), gen_rtx_REG (word_mode, 2));
-#ifndef NO_PROFILE_COUNTERS
+ /* The address of the function is loaded into %r25 with a instruction-
+ relative sequence that avoids the use of relocations. The sequence
+ is split so that the load_offset_label_address instruction can
+ occupy the delay slot of the call to _mcount. */
+ if (TARGET_PA_20)
+ emit_insn (gen_lcla2 (reg, label_rtx));
+ else
+ emit_insn (gen_lcla1 (reg, label_rtx));
+
+ emit_insn (gen_load_offset_label_address (gen_rtx_REG (SImode, 25),
+ reg, begin_label_rtx, label_rtx));
+
+#if !NO_DEFERRED_PROFILE_COUNTERS
{
rtx count_label_rtx, addr, r24;
char count_label_name[16];
+ VEC_safe_push (int, heap, funcdef_nos, label_no);
ASM_GENERATE_INTERNAL_LABEL (count_label_name, "LP", label_no);
count_label_rtx = gen_rtx_SYMBOL_REF (Pmode, ggc_strdup (count_label_name));
r24 = gen_rtx_REG (Pmode, 24);
emit_move_insn (r24, addr);
- /* %r25 is set from within the output pattern. */
call_insn =
- emit_call_insn (gen_call_profiler (gen_rtx_SYMBOL_REF (Pmode, "_mcount"),
- GEN_INT (TARGET_64BIT ? 24 : 12),
- begin_label_rtx));
+ emit_call_insn (gen_call (gen_rtx_MEM (Pmode,
+ gen_rtx_SYMBOL_REF (Pmode,
+ "_mcount")),
+ GEN_INT (TARGET_64BIT ? 24 : 12)));
use_reg (&CALL_INSN_FUNCTION_USAGE (call_insn), r24);
}
#else
- /* %r25 is set from within the output pattern. */
+
call_insn =
- emit_call_insn (gen_call_profiler (gen_rtx_SYMBOL_REF (Pmode, "_mcount"),
- GEN_INT (TARGET_64BIT ? 16 : 8),
- begin_label_rtx));
+ emit_call_insn (gen_call (gen_rtx_MEM (Pmode,
+ gen_rtx_SYMBOL_REF (Pmode,
+ "_mcount")),
+ GEN_INT (TARGET_64BIT ? 16 : 8)));
+
#endif
+ use_reg (&CALL_INSN_FUNCTION_USAGE (call_insn), gen_rtx_REG (SImode, 25));
+ use_reg (&CALL_INSN_FUNCTION_USAGE (call_insn), gen_rtx_REG (SImode, 26));
+
/* Indicate the _mcount call cannot throw, nor will it execute a
non-local goto. */
REG_NOTES (call_insn)
= gen_rtx_EXPR_LIST (REG_EH_REGION, constm1_rtx, REG_NOTES (call_insn));
-
- if (flag_pic)
- {
- use_reg (&CALL_INSN_FUNCTION_USAGE (call_insn), pic_offset_table_rtx);
- if (TARGET_64BIT)
- use_reg (&CALL_INSN_FUNCTION_USAGE (call_insn), arg_pointer_rtx);
-
- emit_move_insn (pic_offset_table_rtx, hppa_pic_save_rtx ());
- }
}
/* Fetch the return address for the frame COUNT steps up from
attr_type = get_attr_type (insn);
- if (REG_NOTE_KIND (link) == REG_DEP_ANTI)
+ switch (REG_NOTE_KIND (link))
{
+ case REG_DEP_ANTI:
/* Anti dependency; DEP_INSN reads a register that INSN writes some
cycles later. */
/* For other anti dependencies, the cost is 0. */
return 0;
- }
- else if (REG_NOTE_KIND (link) == REG_DEP_OUTPUT)
- {
+
+ case REG_DEP_OUTPUT:
/* Output dependency; DEP_INSN writes a register that INSN writes some
cycles later. */
if (attr_type == TYPE_FPLOAD)
/* For other output dependencies, the cost is 0. */
return 0;
+
+ default:
+ gcc_unreachable ();
}
- else
- abort ();
}
/* Adjust scheduling priorities. We use this to try and keep addil
case PROCESSOR_8000: return 4;
default:
- abort ();
+ gcc_unreachable ();
}
}
&& GET_CODE (XEXP (XVECEXP (pat, 0, 0), 1)) == MEM
&& GET_MODE (XEXP (XVECEXP (pat, 0, 0), 0)) == BLKmode
&& GET_MODE (XEXP (XVECEXP (pat, 0, 0), 1)) == BLKmode)
- return compute_movstr_length (insn) - 4;
+ return compute_movmem_length (insn) - 4;
/* Block clear pattern. */
else if (GET_CODE (insn) == INSN
&& GET_CODE (pat) == PARALLEL
&& GET_CODE (XEXP (XVECEXP (pat, 0, 0), 0)) == MEM
&& XEXP (XVECEXP (pat, 0, 0), 1) == const0_rtx
&& GET_MODE (XEXP (XVECEXP (pat, 0, 0), 0)) == BLKmode)
- return compute_clrstr_length (insn) - 4;
+ return compute_clrmem_length (insn) - 4;
/* Conditional branch with an unfilled delay slot. */
else if (GET_CODE (insn) == JUMP_INSN && ! simplejump_p (insn))
{
case LTU:
fputs ("<<", file); break;
default:
- abort ();
+ gcc_unreachable ();
}
return;
case 'N': /* Condition, (N)egated */
case LTU:
fputs (">>=", file); break;
default:
- abort ();
+ gcc_unreachable ();
}
return;
/* For floating point comparisons. Note that the output
- predicates are the complement of the desired mode. */
+ predicates are the complement of the desired mode. The
+ conditions for GT, GE, LT, LE and LTGT cause an invalid
+ operation exception if the result is unordered and this
+ exception is enabled in the floating-point status register. */
case 'Y':
switch (GET_CODE (x))
{
case LTGT:
fputs ("!<>", file); break;
case UNLE:
- fputs (">", file); break;
+ fputs ("!?<=", file); break;
case UNLT:
- fputs (">=", file); break;
+ fputs ("!?<", file); break;
case UNGE:
- fputs ("<", file); break;
+ fputs ("!?>=", file); break;
case UNGT:
- fputs ("<=", file); break;
+ fputs ("!?>", file); break;
case UNEQ:
- fputs ("<>", file); break;
+ fputs ("!?=", file); break;
case UNORDERED:
- fputs ("<=>", file); break;
+ fputs ("!?", file); break;
case ORDERED:
- fputs ("!<=>", file); break;
+ fputs ("?", file); break;
default:
- abort ();
+ gcc_unreachable ();
}
return;
case 'S': /* Condition, operands are (S)wapped. */
case LTU:
fputs (">>", file); break;
default:
- abort ();
+ gcc_unreachable ();
}
return;
case 'B': /* Condition, (B)oth swapped and negate. */
case LTU:
fputs ("<<=", file); break;
default:
- abort ();
+ gcc_unreachable ();
}
return;
case 'k':
- if (GET_CODE (x) == CONST_INT)
- {
- fprintf (file, HOST_WIDE_INT_PRINT_DEC, ~INTVAL (x));
- return;
- }
- abort ();
+ gcc_assert (GET_CODE (x) == CONST_INT);
+ fprintf (file, HOST_WIDE_INT_PRINT_DEC, ~INTVAL (x));
+ return;
case 'Q':
- if (GET_CODE (x) == CONST_INT)
- {
- fprintf (file, HOST_WIDE_INT_PRINT_DEC, 64 - (INTVAL (x) & 63));
- return;
- }
- abort ();
+ gcc_assert (GET_CODE (x) == CONST_INT);
+ fprintf (file, HOST_WIDE_INT_PRINT_DEC, 64 - (INTVAL (x) & 63));
+ return;
case 'L':
- if (GET_CODE (x) == CONST_INT)
- {
- fprintf (file, HOST_WIDE_INT_PRINT_DEC, 32 - (INTVAL (x) & 31));
- return;
- }
- abort ();
+ gcc_assert (GET_CODE (x) == CONST_INT);
+ fprintf (file, HOST_WIDE_INT_PRINT_DEC, 32 - (INTVAL (x) & 31));
+ return;
case 'O':
- if (GET_CODE (x) == CONST_INT && exact_log2 (INTVAL (x)) >= 0)
- {
- fprintf (file, "%d", exact_log2 (INTVAL (x)));
- return;
- }
- abort ();
+ gcc_assert (GET_CODE (x) == CONST_INT && exact_log2 (INTVAL (x)) >= 0);
+ fprintf (file, "%d", exact_log2 (INTVAL (x)));
+ return;
case 'p':
- if (GET_CODE (x) == CONST_INT)
- {
- fprintf (file, HOST_WIDE_INT_PRINT_DEC, 63 - (INTVAL (x) & 63));
- return;
- }
- abort ();
+ gcc_assert (GET_CODE (x) == CONST_INT);
+ fprintf (file, HOST_WIDE_INT_PRINT_DEC, 63 - (INTVAL (x) & 63));
+ return;
case 'P':
- if (GET_CODE (x) == CONST_INT)
- {
- fprintf (file, HOST_WIDE_INT_PRINT_DEC, 31 - (INTVAL (x) & 31));
- return;
- }
- abort ();
+ gcc_assert (GET_CODE (x) == CONST_INT);
+ fprintf (file, HOST_WIDE_INT_PRINT_DEC, 31 - (INTVAL (x) & 31));
+ return;
case 'I':
if (GET_CODE (x) == CONST_INT)
fputs ("i", file);
fputs (",ma", file);
break;
case PLUS:
- if (GET_CODE (XEXP (XEXP (x, 0), 0)) == MULT
- || GET_CODE (XEXP (XEXP (x, 0), 1)) == MULT)
+ if (GET_CODE (XEXP (XEXP (x, 0), 0)) == REG
+ && GET_CODE (XEXP (XEXP (x, 0), 1)) == REG)
+ {
+ if (ASSEMBLER_DIALECT == 0)
+ fputs ("x", file);
+ }
+ else if (GET_CODE (XEXP (XEXP (x, 0), 0)) == MULT
+ || GET_CODE (XEXP (XEXP (x, 0), 1)) == MULT)
{
if (ASSEMBLER_DIALECT == 0)
fputs ("x,s", file);
addresses. */
break;
default:
- abort ();
+ gcc_unreachable ();
}
if (GET_CODE (x) == REG)
{
base = XEXP (XEXP (x, 0), 0);
fprintf (file, "%d(%s)", size, reg_names [REGNO (base)]);
break;
- default:
- if (GET_CODE (XEXP (x, 0)) == PLUS
- && GET_CODE (XEXP (XEXP (x, 0), 0)) == MULT)
+ case PLUS:
+ if (GET_CODE (XEXP (XEXP (x, 0), 0)) == MULT)
fprintf (file, "%s(%s)",
reg_names [REGNO (XEXP (XEXP (XEXP (x, 0), 0), 0))],
reg_names [REGNO (XEXP (XEXP (x, 0), 1))]);
- else if (GET_CODE (XEXP (x, 0)) == PLUS
- && GET_CODE (XEXP (XEXP (x, 0), 1)) == MULT)
+ else if (GET_CODE (XEXP (XEXP (x, 0), 1)) == MULT)
fprintf (file, "%s(%s)",
reg_names [REGNO (XEXP (XEXP (XEXP (x, 0), 1), 0))],
reg_names [REGNO (XEXP (XEXP (x, 0), 0))]);
+ else if (GET_CODE (XEXP (XEXP (x, 0), 0)) == REG
+ && GET_CODE (XEXP (XEXP (x, 0), 1)) == REG)
+ {
+ /* Because the REG_POINTER flag can get lost during reload,
+ GO_IF_LEGITIMATE_ADDRESS canonicalizes the order of the
+ index and base registers in the combined move patterns. */
+ rtx base = XEXP (XEXP (x, 0), 1);
+ rtx index = XEXP (XEXP (x, 0), 0);
+
+ fprintf (file, "%s(%s)",
+ reg_names [REGNO (index)], reg_names [REGNO (base)]);
+ }
else
output_address (XEXP (x, 0));
break;
+ default:
+ output_address (XEXP (x, 0));
+ break;
}
}
else
x = XEXP (x, 0);
if (GET_CODE (x) == SYMBOL_REF && read_only_operand (x, VOIDmode))
- assemble_name (file, XSTR (x, 0));
+ output_addr_const (file, x);
else if (GET_CODE (x) == SYMBOL_REF && !flag_pic)
{
- assemble_name (file, XSTR (x, 0));
+ output_addr_const (file, x);
fputs ("-$global$", file);
}
else if (GET_CODE (x) == CONST)
int offset = 0; /* assembler wants -$global$ at end */
rtx base = NULL_RTX;
- if (GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF)
+ switch (GET_CODE (XEXP (XEXP (x, 0), 0)))
{
+ case SYMBOL_REF:
base = XEXP (XEXP (x, 0), 0);
output_addr_const (file, base);
+ break;
+ case CONST_INT:
+ offset = INTVAL (XEXP (XEXP (x, 0), 0));
+ break;
+ default:
+ gcc_unreachable ();
}
- else if (GET_CODE (XEXP (XEXP (x, 0), 0)) == CONST_INT)
- offset = INTVAL (XEXP (XEXP (x, 0), 0));
- else abort ();
- if (GET_CODE (XEXP (XEXP (x, 0), 1)) == SYMBOL_REF)
+ switch (GET_CODE (XEXP (XEXP (x, 0), 1)))
{
+ case SYMBOL_REF:
base = XEXP (XEXP (x, 0), 1);
output_addr_const (file, base);
+ break;
+ case CONST_INT:
+ offset = INTVAL (XEXP (XEXP (x, 0), 1));
+ break;
+ default:
+ gcc_unreachable ();
}
- else if (GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT)
- offset = INTVAL (XEXP (XEXP (x, 0), 1));
- else abort ();
/* How bogus. The compiler is apparently responsible for
rounding the constant if it uses an LR field selector.
if (round_constant)
offset = ((offset + 0x1000) & ~0x1fff);
- if (GET_CODE (XEXP (x, 0)) == PLUS)
+ switch (GET_CODE (XEXP (x, 0)))
{
+ case PLUS:
if (offset < 0)
{
offset = -offset;
}
else
sep = "+";
- }
- else if (GET_CODE (XEXP (x, 0)) == MINUS
- && (GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF))
- sep = "-";
- else abort ();
+ break;
+
+ case MINUS:
+ gcc_assert (GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF);
+ sep = "-";
+ break;
+ default:
+ gcc_unreachable ();
+ }
+
if (!read_only_operand (base, VOIDmode) && !flag_pic)
fputs ("-$global$", file);
if (offset)
}
#undef aputs
-static struct deferred_plabel *
-get_plabel (const char *fname)
+/* Search the deferred plabel list for SYMBOL and return its internal
+ label. If an entry for SYMBOL is not found, a new entry is created. */
+
+rtx
+get_deferred_plabel (rtx symbol)
{
+ const char *fname = XSTR (symbol, 0);
size_t i;
/* See if we have already put this function on the list of deferred
plabels. This list is generally small, so a liner search is not
too ugly. If it proves too slow replace it with something faster. */
for (i = 0; i < n_deferred_plabels; i++)
- if (strcmp (fname, deferred_plabels[i].name) == 0)
+ if (strcmp (fname, XSTR (deferred_plabels[i].symbol, 0)) == 0)
break;
/* If the deferred plabel list is empty, or this entry was not found
on the list, create a new entry on the list. */
if (deferred_plabels == NULL || i == n_deferred_plabels)
{
- const char *real_name;
+ tree id;
if (deferred_plabels == 0)
deferred_plabels = (struct deferred_plabel *)
i = n_deferred_plabels++;
deferred_plabels[i].internal_label = gen_label_rtx ();
- deferred_plabels[i].name = ggc_strdup (fname);
+ deferred_plabels[i].symbol = symbol;
- /* Gross. We have just implicitly taken the address of this function,
- mark it as such. */
- real_name = (*targetm.strip_name_encoding) (fname);
- TREE_SYMBOL_REFERENCED (get_identifier (real_name)) = 1;
+ /* Gross. We have just implicitly taken the address of this
+ function. Mark it in the same manner as assemble_name. */
+ id = maybe_get_identifier (targetm.strip_name_encoding (fname));
+ if (id)
+ mark_referenced (id);
}
- return &deferred_plabels[i];
+ return deferred_plabels[i].internal_label;
}
static void
{
(*targetm.asm_out.internal_label) (asm_out_file, "L",
CODE_LABEL_NUMBER (deferred_plabels[i].internal_label));
- assemble_integer (gen_rtx_SYMBOL_REF (Pmode, deferred_plabels[i].name),
+ assemble_integer (deferred_plabels[i].symbol,
TARGET_64BIT ? 8 : 4, TARGET_64BIT ? 64 : 32, 1);
}
}
set_optab_libfunc (ge_optab, TFmode, "_U_Qfge");
set_optab_libfunc (lt_optab, TFmode, "_U_Qflt");
set_optab_libfunc (le_optab, TFmode, "_U_Qfle");
+ set_optab_libfunc (unord_optab, TFmode, "_U_Qfunord");
set_conv_libfunc (sext_optab, TFmode, SFmode, "_U_Qfcnvff_sgl_to_quad");
set_conv_libfunc (sext_optab, TFmode, DFmode, "_U_Qfcnvff_dbl_to_quad");
/* Emit the rtl for doing a division by a constant. */
/* Do magic division millicodes exist for this value? */
-static const int magic_milli[]= {0, 0, 0, 1, 0, 1, 1, 1, 0, 1, 1, 0, 1, 0,
- 1, 1};
+const int magic_milli[]= {0, 0, 0, 1, 0, 1, 1, 1, 0, 1, 1, 0, 1, 0, 1, 1};
/* We'll use an array to keep track of the magic millicodes and
whether or not we've used them already. [n][0] is signed, [n][1] is
static int div_milli[16][2];
int
-div_operand (rtx op, enum machine_mode mode)
-{
- return (mode == SImode
- && ((GET_CODE (op) == REG && REGNO (op) == 25)
- || (GET_CODE (op) == CONST_INT && INTVAL (op) > 0
- && INTVAL (op) < 16 && magic_milli[INTVAL (op)])));
-}
-
-int
emit_hpdiv_const (rtx *operands, int unsignedp)
{
if (GET_CODE (operands[2]) == CONST_INT
emit_move_insn (gen_rtx_REG (SImode, 26), operands[1]);
emit
- (gen_rtx
- (PARALLEL, VOIDmode,
+ (gen_rtx_PARALLEL
+ (VOIDmode,
gen_rtvec (6, gen_rtx_SET (VOIDmode, gen_rtx_REG (SImode, 29),
gen_rtx_fmt_ee (unsignedp ? UDIV : DIV,
SImode,
return;
}
- if (GET_CODE (call_insn) != CALL_INSN)
- abort ();
- for (link = CALL_INSN_FUNCTION_USAGE (call_insn); link; link = XEXP (link, 1))
+ gcc_assert (GET_CODE (call_insn) == CALL_INSN);
+ for (link = CALL_INSN_FUNCTION_USAGE (call_insn);
+ link; link = XEXP (link, 1))
{
rtx use = XEXP (link, 0);
fputc ('\n', asm_out_file);
}
\f
-/* Return the class of any secondary reload register that is needed to
- move IN into a register in class CLASS using mode MODE.
-
- Profiling has showed this routine and its descendants account for
- a significant amount of compile time (~7%). So it has been
- optimized to reduce redundant computations and eliminate useless
- function calls.
-
- It might be worthwhile to try and make this a leaf function too. */
-
-enum reg_class
-secondary_reload_class (enum reg_class class, enum machine_mode mode, rtx in)
+static enum reg_class
+pa_secondary_reload (bool in_p, rtx x, enum reg_class class,
+ enum machine_mode mode, secondary_reload_info *sri)
{
- int regno, is_symbolic;
-
- /* Trying to load a constant into a FP register during PIC code
- generation will require %r1 as a scratch register. */
- if (flag_pic
- && GET_MODE_CLASS (mode) == MODE_INT
- && FP_REG_CLASS_P (class)
- && (GET_CODE (in) == CONST_INT || GET_CODE (in) == CONST_DOUBLE))
- return R1_REGS;
+ int is_symbolic, regno;
- /* Profiling showed the PA port spends about 1.3% of its compilation
- time in true_regnum from calls inside secondary_reload_class. */
+ /* Handle the easy stuff first. */
+ if (class == R1_REGS)
+ return NO_REGS;
- if (GET_CODE (in) == REG)
+ if (REG_P (x))
{
- regno = REGNO (in);
- if (regno >= FIRST_PSEUDO_REGISTER)
- regno = true_regnum (in);
+ regno = REGNO (x);
+ if (class == BASE_REG_CLASS && regno < FIRST_PSEUDO_REGISTER)
+ return NO_REGS;
}
- else if (GET_CODE (in) == SUBREG)
- regno = true_regnum (in);
else
regno = -1;
/* If we have something like (mem (mem (...)), we can safely assume the
inner MEM will end up in a general register after reloading, so there's
no need for a secondary reload. */
- if (GET_CODE (in) == MEM
- && GET_CODE (XEXP (in, 0)) == MEM)
+ if (GET_CODE (x) == MEM && GET_CODE (XEXP (x, 0)) == MEM)
return NO_REGS;
+ /* Trying to load a constant into a FP register during PIC code
+ generation requires %r1 as a scratch register. */
+ if (flag_pic
+ && GET_MODE_CLASS (mode) == MODE_INT
+ && FP_REG_CLASS_P (class)
+ && (GET_CODE (x) == CONST_INT || GET_CODE (x) == CONST_DOUBLE))
+ {
+ gcc_assert (mode == SImode || mode == DImode);
+ sri->icode = (mode == SImode ? CODE_FOR_reload_insi_r1
+ : CODE_FOR_reload_indi_r1);
+ return NO_REGS;
+ }
+
+ /* Profiling showed the PA port spends about 1.3% of its compilation
+ time in true_regnum from calls inside pa_secondary_reload_class. */
+ if (regno >= FIRST_PSEUDO_REGISTER || GET_CODE (x) == SUBREG)
+ regno = true_regnum (x);
+
/* Handle out of range displacement for integer mode loads/stores of
FP registers. */
if (((regno >= FIRST_PSEUDO_REGISTER || regno == -1)
&& GET_MODE_CLASS (mode) == MODE_INT
&& FP_REG_CLASS_P (class))
|| (class == SHIFT_REGS && (regno <= 0 || regno >= 32)))
- return GENERAL_REGS;
+ {
+ sri->icode = in_p ? reload_in_optab[mode] : reload_out_optab[mode];
+ return NO_REGS;
+ }
/* A SAR<->FP register copy requires a secondary register (GPR) as
well as secondary memory. */
if (regno >= 0 && regno < FIRST_PSEUDO_REGISTER
&& ((REGNO_REG_CLASS (regno) == SHIFT_REGS && FP_REG_CLASS_P (class))
- || (class == SHIFT_REGS && FP_REG_CLASS_P (REGNO_REG_CLASS (regno)))))
- return GENERAL_REGS;
+ || (class == SHIFT_REGS
+ && FP_REG_CLASS_P (REGNO_REG_CLASS (regno)))))
+ {
+ sri->icode = in_p ? reload_in_optab[mode] : reload_out_optab[mode];
+ return NO_REGS;
+ }
- if (GET_CODE (in) == HIGH)
- in = XEXP (in, 0);
+ /* Secondary reloads of symbolic operands require %r1 as a scratch
+ register when we're generating PIC code and the operand isn't
+ readonly. */
+ if (GET_CODE (x) == HIGH)
+ x = XEXP (x, 0);
/* Profiling has showed GCC spends about 2.6% of its compilation
- time in symbolic_operand from calls inside secondary_reload_class.
-
- We use an inline copy and only compute its return value once to avoid
- useless work. */
- switch (GET_CODE (in))
+ time in symbolic_operand from calls inside pa_secondary_reload_class.
+ So, we use an inline copy to avoid useless work. */
+ switch (GET_CODE (x))
{
- rtx tmp;
+ rtx op;
case SYMBOL_REF:
+ is_symbolic = !SYMBOL_REF_TLS_MODEL (x);
+ break;
case LABEL_REF:
is_symbolic = 1;
break;
case CONST:
- tmp = XEXP (in, 0);
- is_symbolic = ((GET_CODE (XEXP (tmp, 0)) == SYMBOL_REF
- || GET_CODE (XEXP (tmp, 0)) == LABEL_REF)
- && GET_CODE (XEXP (tmp, 1)) == CONST_INT);
+ op = XEXP (x, 0);
+ is_symbolic = (((GET_CODE (XEXP (op, 0)) == SYMBOL_REF
+ && !SYMBOL_REF_TLS_MODEL (XEXP (op, 0)))
+ || GET_CODE (XEXP (op, 0)) == LABEL_REF)
+ && GET_CODE (XEXP (op, 1)) == CONST_INT);
break;
-
default:
is_symbolic = 0;
break;
}
- if (!flag_pic
- && is_symbolic
- && read_only_operand (in, VOIDmode))
- return NO_REGS;
-
- if (class != R1_REGS && is_symbolic)
- return R1_REGS;
+ if (is_symbolic && (flag_pic || !read_only_operand (x, VOIDmode)))
+ {
+ gcc_assert (mode == SImode || mode == DImode);
+ sri->icode = (mode == SImode ? CODE_FOR_reload_insi_r1
+ : CODE_FOR_reload_indi_r1);
+ }
return NO_REGS;
}
+/* In the 32-bit runtime, arguments larger than eight bytes are passed
+ by invisible reference. As a GCC extension, we also pass anything
+ with a zero or variable size by reference.
+
+ The 64-bit runtime does not describe passing any types by invisible
+ reference. The internals of GCC can't currently handle passing
+ empty structures, and zero or variable length arrays when they are
+ not passed entirely on the stack or by reference. Thus, as a GCC
+ extension, we pass these types by reference. The HP compiler doesn't
+ support these types, so hopefully there shouldn't be any compatibility
+ issues. This may have to be revisited when HP releases a C99 compiler
+ or updates the ABI. */
+
+static bool
+pa_pass_by_reference (CUMULATIVE_ARGS *ca ATTRIBUTE_UNUSED,
+ enum machine_mode mode, tree type,
+ bool named ATTRIBUTE_UNUSED)
+{
+ HOST_WIDE_INT size;
+
+ if (type)
+ size = int_size_in_bytes (type);
+ else
+ size = GET_MODE_SIZE (mode);
+
+ if (TARGET_64BIT)
+ return size <= 0;
+ else
+ return size <= 0 || size > 8;
+}
+
enum direction
function_arg_padding (enum machine_mode mode, tree type)
{
to determine if stdargs or varargs is used and fill in an initial
va_list. A pointer to this constructor is returned. */
-struct rtx_def *
+static rtx
hppa_builtin_saveregs (void)
{
rtx offset, dest;
std_expand_builtin_va_start (valist, nextarg);
}
-rtx
-hppa_va_arg (tree valist, tree type)
+static tree
+hppa_gimplify_va_arg_expr (tree valist, tree type, tree *pre_p, tree *post_p)
{
- HOST_WIDE_INT size = int_size_in_bytes (type);
- HOST_WIDE_INT ofs;
- tree t, ptr, pptr;
-
if (TARGET_64BIT)
{
- /* Every argument in PA64 is supposed to be passed by value
- (including large structs). However, as a GCC extension, we
- pass zero and variable sized arguments by reference. Empty
- structures are a GCC extension not supported by the HP
- compilers. Thus, passing them by reference isn't likely
- to conflict with the ABI. For variable sized arguments,
- GCC doesn't have the infrastructure to allocate these to
- registers. */
-
- /* Arguments with a size greater than 8 must be aligned 0 MOD 16. */
+ /* Args grow upward. We can use the generic routines. */
+ return std_gimplify_va_arg_expr (valist, type, pre_p, post_p);
+ }
+ else /* !TARGET_64BIT */
+ {
+ tree ptr = build_pointer_type (type);
+ tree valist_type;
+ tree t, u;
+ unsigned int size, ofs;
+ bool indirect;
- if (size > UNITS_PER_WORD)
- {
- t = build (PLUS_EXPR, TREE_TYPE (valist), valist,
- build_int_2 (2 * UNITS_PER_WORD - 1, 0));
- t = build (BIT_AND_EXPR, TREE_TYPE (t), t,
- build_int_2 (-2 * UNITS_PER_WORD, -1));
- t = build (MODIFY_EXPR, TREE_TYPE (valist), valist, t);
- TREE_SIDE_EFFECTS (t) = 1;
- expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL);
- }
-
- if (size > 0)
- return std_expand_builtin_va_arg (valist, type);
- else
+ indirect = pass_by_reference (NULL, TYPE_MODE (type), type, 0);
+ if (indirect)
{
+ type = ptr;
ptr = build_pointer_type (type);
+ }
+ size = int_size_in_bytes (type);
+ valist_type = TREE_TYPE (valist);
- /* Args grow upward. */
- t = build (POSTINCREMENT_EXPR, TREE_TYPE (valist), valist,
- build_int_2 (POINTER_SIZE / BITS_PER_UNIT, 0));
- TREE_SIDE_EFFECTS (t) = 1;
-
- pptr = build_pointer_type (ptr);
- t = build1 (NOP_EXPR, pptr, t);
- TREE_SIDE_EFFECTS (t) = 1;
+ /* Args grow down. Not handled by generic routines. */
- t = build1 (INDIRECT_REF, ptr, t);
- TREE_SIDE_EFFECTS (t) = 1;
- }
- }
- else /* !TARGET_64BIT */
- {
- ptr = build_pointer_type (type);
+ u = fold_convert (valist_type, size_in_bytes (type));
+ t = build2 (MINUS_EXPR, valist_type, valist, u);
- /* "Large" and variable sized types are passed by reference. */
- if (size > 8 || size <= 0)
- {
- /* Args grow downward. */
- t = build (PREDECREMENT_EXPR, TREE_TYPE (valist), valist,
- build_int_2 (POINTER_SIZE / BITS_PER_UNIT, 0));
- TREE_SIDE_EFFECTS (t) = 1;
+ /* Copied from va-pa.h, but we probably don't need to align to
+ word size, since we generate and preserve that invariant. */
+ u = build_int_cst (valist_type, (size > 4 ? -8 : -4));
+ t = build2 (BIT_AND_EXPR, valist_type, t, u);
- pptr = build_pointer_type (ptr);
- t = build1 (NOP_EXPR, pptr, t);
- TREE_SIDE_EFFECTS (t) = 1;
+ t = build2 (MODIFY_EXPR, valist_type, valist, t);
- t = build1 (INDIRECT_REF, ptr, t);
- TREE_SIDE_EFFECTS (t) = 1;
- }
- else
+ ofs = (8 - size) % 4;
+ if (ofs != 0)
{
- t = build (PLUS_EXPR, TREE_TYPE (valist), valist,
- build_int_2 (-size, -1));
-
- /* Copied from va-pa.h, but we probably don't need to align to
- word size, since we generate and preserve that invariant. */
- t = build (BIT_AND_EXPR, TREE_TYPE (valist), t,
- build_int_2 ((size > 4 ? -8 : -4), -1));
+ u = fold_convert (valist_type, size_int (ofs));
+ t = build2 (PLUS_EXPR, valist_type, t, u);
+ }
- t = build (MODIFY_EXPR, TREE_TYPE (valist), valist, t);
- TREE_SIDE_EFFECTS (t) = 1;
+ t = fold_convert (ptr, t);
+ t = build_va_arg_indirect_ref (t);
- ofs = (8 - size) % 4;
- if (ofs)
- {
- t = build (PLUS_EXPR, TREE_TYPE (valist), t,
- build_int_2 (ofs, 0));
- TREE_SIDE_EFFECTS (t) = 1;
- }
+ if (indirect)
+ t = build_va_arg_indirect_ref (t);
- t = build1 (NOP_EXPR, ptr, t);
- TREE_SIDE_EFFECTS (t) = 1;
- }
+ return t;
}
-
- /* Calculate! */
- return expand_expr (t, NULL_RTX, VOIDmode, EXPAND_NORMAL);
}
+/* True if MODE is valid for the target. By "valid", we mean able to
+ be manipulated in non-trivial ways. In particular, this means all
+ the arithmetic is supported.
+
+ Currently, TImode is not valid as the HP 64-bit runtime documentation
+ doesn't document the alignment and calling conventions for this type.
+ Thus, we return false when PRECISION is 2 * BITS_PER_WORD and
+ 2 * BITS_PER_WORD isn't equal LONG_LONG_TYPE_SIZE. */
+
+static bool
+pa_scalar_mode_supported_p (enum machine_mode mode)
+{
+ int precision = GET_MODE_PRECISION (mode);
+
+ switch (GET_MODE_CLASS (mode))
+ {
+ case MODE_PARTIAL_INT:
+ case MODE_INT:
+ if (precision == CHAR_TYPE_SIZE)
+ return true;
+ if (precision == SHORT_TYPE_SIZE)
+ return true;
+ if (precision == INT_TYPE_SIZE)
+ return true;
+ if (precision == LONG_TYPE_SIZE)
+ return true;
+ if (precision == LONG_LONG_TYPE_SIZE)
+ return true;
+ return false;
+
+ case MODE_FLOAT:
+ if (precision == FLOAT_TYPE_SIZE)
+ return true;
+ if (precision == DOUBLE_TYPE_SIZE)
+ return true;
+ if (precision == LONG_DOUBLE_TYPE_SIZE)
+ return true;
+ return false;
+ default:
+ gcc_unreachable ();
+ }
+}
/* This routine handles all the normal conditional branch sequences we
might need to generate. It handles compare immediate vs compare
int useskip = 0;
rtx xoperands[5];
- /* A conditional branch to the following instruction (eg the delay slot)
+ /* A conditional branch to the following instruction (e.g. the delay slot)
is asking for a disaster. This can happen when not optimizing and
when jump optimization fails.
zero for cmpb, we must ensure that we use cmpb for the comparison. */
if (GET_MODE (operands[1]) == DImode && operands[2] == const0_rtx)
operands[2] = gen_rtx_REG (DImode, 0);
+ if (GET_MODE (operands[2]) == DImode && operands[1] == const0_rtx)
+ operands[1] = gen_rtx_REG (DImode, 0);
/* If this is a long branch with its delay slot unfilled, set `nullify'
as it can nullify the delay slot and save a nop. */
return output_lbranch (operands[0], insn);
default:
- abort ();
+ gcc_unreachable ();
}
return buf;
}
if (dbr_sequence_length () != 0)
{
/* We can't handle a jump in the delay slot. */
- if (GET_CODE (NEXT_INSN (insn)) == JUMP_INSN)
- abort ();
+ gcc_assert (GET_CODE (NEXT_INSN (insn)) != JUMP_INSN);
final_scan_insn (NEXT_INSN (insn), asm_out_file,
- optimize, 0, 0);
+ optimize, 0, NULL);
/* Now delete the delay insn. */
PUT_CODE (NEXT_INSN (insn), NOTE);
static char buf[100];
int useskip = 0;
- /* A conditional branch to the following instruction (eg the delay slot) is
+ /* A conditional branch to the following instruction (e.g. the delay slot) is
asking for a disaster. I do not think this can happen as this pattern
is only used when optimizing; jump optimization should eliminate the
jump. But be prepared just in case. */
break;
default:
- abort ();
+ gcc_unreachable ();
}
return buf;
}
static char buf[100];
int useskip = 0;
- /* A conditional branch to the following instruction (eg the delay slot) is
+ /* A conditional branch to the following instruction (e.g. the delay slot) is
asking for a disaster. I do not think this can happen as this pattern
is only used when optimizing; jump optimization should eliminate the
jump. But be prepared just in case. */
break;
default:
- abort ();
+ gcc_unreachable ();
}
return buf;
}
output_dbra (rtx *operands, rtx insn, int which_alternative)
{
- /* A conditional branch to the following instruction (eg the delay slot) is
+ /* A conditional branch to the following instruction (e.g. the delay slot) is
asking for a disaster. Be prepared! */
if (next_real_insn (JUMP_LABEL (insn)) == next_real_insn (insn))
if (! nullify && length == 4 && dbr_sequence_length () == 0)
nullify = forward_branch_p (insn);
- /* Handle short versions first. */
- if (length == 4 && nullify)
- return "addib,%C2,n %1,%0,%3";
- else if (length == 4 && ! nullify)
- return "addib,%C2 %1,%0,%3";
- else if (length == 8)
+ switch (length)
{
+ case 4:
+ if (nullify)
+ return "addib,%C2,n %1,%0,%3";
+ else
+ return "addib,%C2 %1,%0,%3";
+
+ case 8:
/* Handle weird backwards branch with a fulled delay slot
which is nullified. */
if (dbr_sequence_length () != 0
return "addi,%N2 %1,%0,%0\n\tb,n %3";
else
return "addi,%N2 %1,%0,%0\n\tb %3";
+
+ default:
+ gcc_unreachable ();
}
- else
- abort ();
+
}
/* Deal with gross reload from FP register case. */
else if (which_alternative == 1)
else
{
/* Reload loop counter from memory, the store back to memory
- happens in the branch's delay slot. */
+ happens in the branch's delay slot. */
output_asm_insn ("ldw %0,%4", operands);
if (get_attr_length (insn) == 12)
return "addib,%C2 %1,%4,%3\n\tstw %4,%0";
int reverse_comparison)
{
- /* A conditional branch to the following instruction (eg the delay slot) is
+ /* A conditional branch to the following instruction (e.g. the delay slot) is
asking for a disaster. Be prepared! */
if (next_real_insn (JUMP_LABEL (insn)) == next_real_insn (insn))
if (! nullify && length == 4 && dbr_sequence_length () == 0)
nullify = forward_branch_p (insn);
- /* Handle short versions first. */
- if (length == 4 && nullify)
- return "movb,%C2,n %1,%0,%3";
- else if (length == 4 && ! nullify)
- return "movb,%C2 %1,%0,%3";
- else if (length == 8)
+ switch (length)
{
+ case 4:
+ if (nullify)
+ return "movb,%C2,n %1,%0,%3";
+ else
+ return "movb,%C2 %1,%0,%3";
+
+ case 8:
/* Handle weird backwards branch with a filled delay slot
which is nullified. */
if (dbr_sequence_length () != 0
return "or,%N2 %1,%%r0,%0\n\tb,n %3";
else
return "or,%N2 %1,%%r0,%0\n\tb %3";
+
+ default:
+ gcc_unreachable ();
}
- else
- abort ();
}
/* Deal with gross reload from FP register case. */
else if (which_alternative == 1)
else if (which_alternative == 2)
{
/* Reload loop counter from memory, the store back to memory
- happens in the branch's delay slot. */
+ happens in the branch's delay slot. */
if (get_attr_length (insn) == 8)
return "{comb|cmpb},%S2 %%r0,%1,%3\n\tstw %1,%0";
else
if (get_attr_length (insn) == 8)
return "{comb|cmpb},%S2 %%r0,%1,%3\n\tmtsar %r1";
else
- return "{comclr|cmpclr},%B2 %%r0,%1,%%r0\n\tbl %3\n\tmtsar %r1";
+ return "{comclr|cmpclr},%B2 %%r0,%1,%%r0\n\tb %3\n\tmtsar %r1";
}
}
/* ??? As far as I can tell, the HP linker doesn't support the
long pc-relative sequence described in the 64-bit runtime
architecture. So, we use a slightly longer indirect call. */
- struct deferred_plabel *p = get_plabel (XSTR (call_dest, 0));
-
- xoperands[0] = p->internal_label;
+ xoperands[0] = get_deferred_plabel (call_dest);
xoperands[1] = gen_label_rtx ();
/* If this isn't a sibcall, we put the load of %r27 into the
&& !sibcall)
{
final_scan_insn (NEXT_INSN (insn), asm_out_file,
- optimize, 0, 0);
+ optimize, 0, NULL);
/* Now delete the delay insn. */
PUT_CODE (NEXT_INSN (insn), NOTE);
/* A non-jump insn in the delay slot. By definition we can
emit this insn before the call (and in fact before argument
relocating. */
- final_scan_insn (NEXT_INSN (insn), asm_out_file, optimize, 0, 0);
+ final_scan_insn (NEXT_INSN (insn), asm_out_file, optimize, 0,
+ NULL);
/* Now delete the delay insn. */
PUT_CODE (NEXT_INSN (insn), NOTE);
essentially an inline implementation of $$dyncall.
We don't actually try to call $$dyncall as this is
as difficult as calling the function itself. */
- struct deferred_plabel *p = get_plabel (XSTR (call_dest, 0));
-
- xoperands[0] = p->internal_label;
+ xoperands[0] = get_deferred_plabel (call_dest);
xoperands[1] = gen_label_rtx ();
/* Since the call is indirect, FP arguments in registers
return "";
/* A sibcall should never have a branch in the delay slot. */
- if (sibcall)
- abort ();
+ gcc_assert (!sibcall);
/* This call has an unconditional jump in its delay slot. */
xoperands[0] = XEXP (PATTERN (NEXT_INSN (insn)), 1);
No need to check target flags as the length uniquely identifies
the remaining cases. */
if (attr_length_indirect_call (insn) == 8)
- return ".CALL\tARGW0=GR\n\t{bl|b,l} $$dyncall,%%r31\n\tcopy %%r31,%%r2";
+ {
+ /* The HP linker substitutes a BLE for millicode calls using
+ the short PIC PCREL form. Thus, we must use %r31 as the
+ link register when generating PA 1.x code. */
+ if (TARGET_PA_20)
+ return ".CALL\tARGW0=GR\n\tb,l $$dyncall,%%r2\n\tcopy %%r2,%%r31";
+ else
+ return ".CALL\tARGW0=GR\n\tbl $$dyncall,%%r31\n\tcopy %%r31,%%r2";
+ }
/* Long millicode call, but we are not generating PIC or portable runtime
code. */
static void
pa_encode_section_info (tree decl, rtx rtl, int first)
{
+ default_encode_section_info (decl, rtl, first);
+
if (first && TEXT_SPACE_P (decl))
{
SYMBOL_REF_FLAG (XEXP (rtl, 0)) = 1;
HOST_WIDE_INT vcall_offset ATTRIBUTE_UNUSED,
tree function)
{
- const char *fname = XSTR (XEXP (DECL_RTL (function), 0), 0);
- const char *tname = XSTR (XEXP (DECL_RTL (thunk_fndecl), 0), 0);
+ static unsigned int current_thunk_number;
int val_14 = VAL_14_BITS_P (delta);
int nbytes = 0;
- static unsigned int current_thunk_number;
char label[16];
+ rtx xoperands[4];
- ASM_OUTPUT_LABEL (file, tname);
- fprintf (file, "\t.PROC\n\t.CALLINFO FRAME=0,NO_CALLS\n\t.ENTRY\n");
+ xoperands[0] = XEXP (DECL_RTL (function), 0);
+ xoperands[1] = XEXP (DECL_RTL (thunk_fndecl), 0);
+ xoperands[2] = GEN_INT (delta);
- fname = (*targetm.strip_name_encoding) (fname);
- tname = (*targetm.strip_name_encoding) (tname);
+ ASM_OUTPUT_LABEL (file, XSTR (xoperands[1], 0));
+ fprintf (file, "\t.PROC\n\t.CALLINFO FRAME=0,NO_CALLS\n\t.ENTRY\n");
/* Output the thunk. We know that the function is in the same
translation unit (i.e., the same space) as the thunk, and that
&& last_address < 262132)))
|| (!targetm.have_named_sections && last_address < 262132))))
{
+ if (!val_14)
+ output_asm_insn ("addil L'%2,%%r26", xoperands);
+
+ output_asm_insn ("b %0", xoperands);
+
if (val_14)
{
- fprintf (file, "\tb %s\n\tldo " HOST_WIDE_INT_PRINT_DEC
- "(%%r26),%%r26\n", fname, delta);
+ output_asm_insn ("ldo %2(%%r26),%%r26", xoperands);
nbytes += 8;
}
else
{
- fprintf (file, "\taddil L'" HOST_WIDE_INT_PRINT_DEC
- ",%%r26\n", delta);
- fprintf (file, "\tb %s\n\tldo R'" HOST_WIDE_INT_PRINT_DEC
- "(%%r1),%%r26\n", fname, delta);
+ output_asm_insn ("ldo R'%2(%%r1),%%r26", xoperands);
nbytes += 12;
}
}
/* We only have one call-clobbered scratch register, so we can't
make use of the delay slot if delta doesn't fit in 14 bits. */
if (!val_14)
- fprintf (file, "\taddil L'" HOST_WIDE_INT_PRINT_DEC
- ",%%r26\n\tldo R'" HOST_WIDE_INT_PRINT_DEC
- "(%%r1),%%r26\n", delta, delta);
+ {
+ output_asm_insn ("addil L'%2,%%r26", xoperands);
+ output_asm_insn ("ldo R'%2(%%r1),%%r26", xoperands);
+ }
- fprintf (file, "\tb,l .+8,%%r1\n");
+ output_asm_insn ("b,l .+8,%%r1", xoperands);
if (TARGET_GAS)
{
- fprintf (file, "\taddil L'%s-$PIC_pcrel$0+4,%%r1\n", fname);
- fprintf (file, "\tldo R'%s-$PIC_pcrel$0+8(%%r1),%%r1\n", fname);
+ output_asm_insn ("addil L'%0-$PIC_pcrel$0+4,%%r1", xoperands);
+ output_asm_insn ("ldo R'%0-$PIC_pcrel$0+8(%%r1),%%r1", xoperands);
}
else
{
- int off = val_14 ? 8 : 16;
- fprintf (file, "\taddil L'%s-%s-%d,%%r1\n", fname, tname, off);
- fprintf (file, "\tldo R'%s-%s-%d(%%r1),%%r1\n", fname, tname, off);
+ xoperands[3] = GEN_INT (val_14 ? 8 : 16);
+ output_asm_insn ("addil L'%0-%1-%3,%%r1", xoperands);
}
if (val_14)
{
- fprintf (file, "\tbv %%r0(%%r1)\n\tldo ");
- fprintf (file, HOST_WIDE_INT_PRINT_DEC "(%%r26),%%r26\n", delta);
+ output_asm_insn ("bv %%r0(%%r1)", xoperands);
+ output_asm_insn ("ldo %2(%%r26),%%r26", xoperands);
nbytes += 20;
}
else
{
- fprintf (file, "\tbv,n %%r0(%%r1)\n");
+ output_asm_insn ("bv,n %%r0(%%r1)", xoperands);
nbytes += 24;
}
}
else if (TARGET_PORTABLE_RUNTIME)
{
- fprintf (file, "\tldil L'%s,%%r1\n", fname);
- fprintf (file, "\tldo R'%s(%%r1),%%r22\n", fname);
+ output_asm_insn ("ldil L'%0,%%r1", xoperands);
+ output_asm_insn ("ldo R'%0(%%r1),%%r22", xoperands);
+
+ if (!val_14)
+ output_asm_insn ("addil L'%2,%%r26", xoperands);
+
+ output_asm_insn ("bv %%r0(%%r22)", xoperands);
if (val_14)
{
- fprintf (file, "\tbv %%r0(%%r22)\n\tldo ");
- fprintf (file, HOST_WIDE_INT_PRINT_DEC "(%%r26),%%r26\n", delta);
+ output_asm_insn ("ldo %2(%%r26),%%r26", xoperands);
nbytes += 16;
}
else
{
- fprintf (file, "\taddil L'" HOST_WIDE_INT_PRINT_DEC
- ",%%r26\n", delta);
- fprintf (file, "\tbv %%r0(%%r22)\n\tldo ");
- fprintf (file, "R'" HOST_WIDE_INT_PRINT_DEC "(%%r1),%%r26\n", delta);
+ output_asm_insn ("ldo R'%2(%%r1),%%r26", xoperands);
nbytes += 20;
}
}
call the function directly with an indirect sequence similar to
that used by $$dyncall. This is possible because $$dyncall acts
as the import stub in an indirect call. */
- const char *lab;
-
ASM_GENERATE_INTERNAL_LABEL (label, "LTHN", current_thunk_number);
- lab = (*targetm.strip_name_encoding) (label);
-
- fprintf (file, "\taddil LT'%s,%%r19\n", lab);
- fprintf (file, "\tldw RT'%s(%%r1),%%r22\n", lab);
- fprintf (file, "\tldw 0(%%sr0,%%r22),%%r22\n");
- fprintf (file, "\tbb,>=,n %%r22,30,.+16\n");
- fprintf (file, "\tdepi 0,31,2,%%r22\n");
- fprintf (file, "\tldw 4(%%sr0,%%r22),%%r19\n");
- fprintf (file, "\tldw 0(%%sr0,%%r22),%%r22\n");
+ xoperands[3] = gen_rtx_SYMBOL_REF (Pmode, label);
+ output_asm_insn ("addil LT'%3,%%r19", xoperands);
+ output_asm_insn ("ldw RT'%3(%%r1),%%r22", xoperands);
+ output_asm_insn ("ldw 0(%%sr0,%%r22),%%r22", xoperands);
+ output_asm_insn ("bb,>=,n %%r22,30,.+16", xoperands);
+ output_asm_insn ("depi 0,31,2,%%r22", xoperands);
+ output_asm_insn ("ldw 4(%%sr0,%%r22),%%r19", xoperands);
+ output_asm_insn ("ldw 0(%%sr0,%%r22),%%r22", xoperands);
+
if (!val_14)
{
- fprintf (file, "\taddil L'" HOST_WIDE_INT_PRINT_DEC
- ",%%r26\n", delta);
+ output_asm_insn ("addil L'%2,%%r26", xoperands);
nbytes += 4;
}
+
if (TARGET_PA_20)
{
- fprintf (file, "\tbve (%%r22)\n\tldo ");
+ output_asm_insn ("bve (%%r22)", xoperands);
+ nbytes += 36;
+ }
+ else if (TARGET_NO_SPACE_REGS)
+ {
+ output_asm_insn ("be 0(%%sr4,%%r22)", xoperands);
nbytes += 36;
}
else
{
- if (TARGET_NO_SPACE_REGS)
- {
- fprintf (file, "\tbe 0(%%sr4,%%r22)\n\tldo ");
- nbytes += 36;
- }
- else
- {
- fprintf (file, "\tldsid (%%sr0,%%r22),%%r21\n");
- fprintf (file, "\tmtsp %%r21,%%sr0\n");
- fprintf (file, "\tbe 0(%%sr0,%%r22)\n\tldo ");
- nbytes += 44;
- }
+ output_asm_insn ("ldsid (%%sr0,%%r22),%%r21", xoperands);
+ output_asm_insn ("mtsp %%r21,%%sr0", xoperands);
+ output_asm_insn ("be 0(%%sr0,%%r22)", xoperands);
+ nbytes += 44;
}
if (val_14)
- fprintf (file, HOST_WIDE_INT_PRINT_DEC "(%%r26),%%r26\n", delta);
+ output_asm_insn ("ldo %2(%%r26),%%r26", xoperands);
else
- fprintf (file, "R'" HOST_WIDE_INT_PRINT_DEC "(%%r1),%%r26\n", delta);
+ output_asm_insn ("ldo R'%2(%%r1),%%r26", xoperands);
}
else if (flag_pic)
{
- if (TARGET_PA_20)
- fprintf (file, "\tb,l .+8,%%r1\n");
- else
- fprintf (file, "\tbl .+8,%%r1\n");
+ output_asm_insn ("{bl|b,l} .+8,%%r1", xoperands);
if (TARGET_SOM || !TARGET_GAS)
{
- fprintf (file, "\taddil L'%s-%s-8,%%r1\n", fname, tname);
- fprintf (file, "\tldo R'%s-%s-8(%%r1),%%r22\n", fname, tname);
+ output_asm_insn ("addil L'%0-%1-8,%%r1", xoperands);
+ output_asm_insn ("ldo R'%0-%1-8(%%r1),%%r22", xoperands);
}
else
{
- fprintf (file, "\taddil L'%s-$PIC_pcrel$0+4,%%r1\n", fname);
- fprintf (file, "\tldo R'%s-$PIC_pcrel$0+8(%%r1),%%r22\n", fname);
+ output_asm_insn ("addil L'%0-$PIC_pcrel$0+4,%%r1", xoperands);
+ output_asm_insn ("ldo R'%0-$PIC_pcrel$0+8(%%r1),%%r22", xoperands);
}
+ if (!val_14)
+ output_asm_insn ("addil L'%2,%%r26", xoperands);
+
+ output_asm_insn ("bv %%r0(%%r22)", xoperands);
+
if (val_14)
{
- fprintf (file, "\tbv %%r0(%%r22)\n\tldo ");
- fprintf (file, HOST_WIDE_INT_PRINT_DEC "(%%r26),%%r26\n", delta);
+ output_asm_insn ("ldo %2(%%r26),%%r26", xoperands);
nbytes += 20;
}
else
{
- fprintf (file, "\taddil L'" HOST_WIDE_INT_PRINT_DEC
- ",%%r26\n", delta);
- fprintf (file, "\tbv %%r0(%%r22)\n\tldo ");
- fprintf (file, "R'" HOST_WIDE_INT_PRINT_DEC "(%%r1),%%r26\n", delta);
+ output_asm_insn ("ldo R'%2(%%r1),%%r26", xoperands);
nbytes += 24;
}
}
else
{
if (!val_14)
- fprintf (file, "\taddil L'" HOST_WIDE_INT_PRINT_DEC ",%%r26\n", delta);
+ output_asm_insn ("addil L'%2,%%r26", xoperands);
- fprintf (file, "\tldil L'%s,%%r22\n", fname);
- fprintf (file, "\tbe R'%s(%%sr4,%%r22)\n\tldo ", fname);
+ output_asm_insn ("ldil L'%0,%%r22", xoperands);
+ output_asm_insn ("be R'%0(%%sr4,%%r22)", xoperands);
if (val_14)
{
- fprintf (file, HOST_WIDE_INT_PRINT_DEC "(%%r26),%%r26\n", delta);
+ output_asm_insn ("ldo %2(%%r26),%%r26", xoperands);
nbytes += 12;
}
else
{
- fprintf (file, "R'" HOST_WIDE_INT_PRINT_DEC "(%%r1),%%r26\n", delta);
+ output_asm_insn ("ldo R'%2(%%r1),%%r26", xoperands);
nbytes += 16;
}
}
if (TARGET_SOM && flag_pic && TREE_PUBLIC (function))
{
data_section ();
- fprintf (file, "\t.align 4\n");
+ output_asm_insn (".align 4", xoperands);
ASM_OUTPUT_LABEL (file, label);
- fprintf (file, "\t.word P'%s\n", fname);
- function_section (thunk_fndecl);
+ output_asm_insn (".word P'%0", xoperands);
}
+ else if (TARGET_SOM && TARGET_GAS)
+ forget_section ();
current_thunk_number++;
nbytes = ((nbytes + FUNCTION_BOUNDARY / BITS_PER_UNIT - 1)
static bool
pa_function_ok_for_sibcall (tree decl, tree exp ATTRIBUTE_UNUSED)
{
+ if (TARGET_PORTABLE_RUNTIME)
+ return false;
+
/* Sibcalls are ok for TARGET_ELF32 as along as the linker is used in
single subspace mode and the call is not indirect. As far as I know,
there is no operating system support for the multiple subspace mode.
if (TARGET_64BIT)
return false;
- return (decl
- && !TARGET_PORTABLE_RUNTIME
- && !TREE_PUBLIC (decl));
+ /* Sibcalls are only ok within a translation unit. */
+ return (decl && !TREE_PUBLIC (decl));
+}
+
+/* ??? Addition is not commutative on the PA due to the weird implicit
+ space register selection rules for memory addresses. Therefore, we
+ don't consider a + b == b + a, as this might be inside a MEM. */
+static bool
+pa_commutative_p (rtx x, int outer_code)
+{
+ return (COMMUTATIVE_P (x)
+ && (TARGET_NO_SPACE_REGS
+ || (outer_code != UNKNOWN && outer_code != MEM)
+ || GET_CODE (x) != PLUS));
}
/* Returns 1 if the 6 operands specified in OPERANDS are suitable for
&& ! rtx_equal_p (operands[3], operands[5]))
return 0;
- /* Inout operand of add can not conflict with any operands from multiply. */
+ /* Inout operand of add cannot conflict with any operands from multiply. */
if (rtx_equal_p (operands[3], operands[0])
|| rtx_equal_p (operands[3], operands[1])
|| rtx_equal_p (operands[3], operands[2]))
return 0;
- /* multiply can not feed into addition operands. */
+ /* multiply cannot feed into addition operands. */
if (rtx_equal_p (operands[4], operands[0])
|| rtx_equal_p (operands[5], operands[0]))
return 0;
}
#endif
+/* This function places uninitialized global data in the bss section.
+ The ASM_OUTPUT_ALIGNED_BSS macro needs to be defined to call this
+ function on the SOM port to prevent uninitialized global data from
+ being placed in the data section. */
+
+void
+pa_asm_output_aligned_bss (FILE *stream,
+ const char *name,
+ unsigned HOST_WIDE_INT size,
+ unsigned int align)
+{
+ bss_section ();
+ fprintf (stream, "\t.align %u\n", align / BITS_PER_UNIT);
+
+#ifdef ASM_OUTPUT_TYPE_DIRECTIVE
+ ASM_OUTPUT_TYPE_DIRECTIVE (stream, name, "object");
+#endif
+
+#ifdef ASM_OUTPUT_SIZE_DIRECTIVE
+ ASM_OUTPUT_SIZE_DIRECTIVE (stream, name, size);
+#endif
+
+ fprintf (stream, "\t.align %u\n", align / BITS_PER_UNIT);
+ ASM_OUTPUT_LABEL (stream, name);
+ fprintf (stream, "\t.block "HOST_WIDE_INT_PRINT_UNSIGNED"\n", size);
+}
+
+/* Both the HP and GNU assemblers under HP-UX provide a .comm directive
+ that doesn't allow the alignment of global common storage to be directly
+ specified. The SOM linker aligns common storage based on the rounded
+ value of the NUM_BYTES parameter in the .comm directive. It's not
+ possible to use the .align directive as it doesn't affect the alignment
+ of the label associated with a .comm directive. */
+
+void
+pa_asm_output_aligned_common (FILE *stream,
+ const char *name,
+ unsigned HOST_WIDE_INT size,
+ unsigned int align)
+{
+ unsigned int max_common_align;
+
+ max_common_align = TARGET_64BIT ? 128 : (size >= 4096 ? 256 : 64);
+ if (align > max_common_align)
+ {
+ warning (0, "alignment (%u) for %s exceeds maximum alignment "
+ "for global common data. Using %u",
+ align / BITS_PER_UNIT, name, max_common_align / BITS_PER_UNIT);
+ align = max_common_align;
+ }
+
+ bss_section ();
+
+ assemble_name (stream, name);
+ fprintf (stream, "\t.comm "HOST_WIDE_INT_PRINT_UNSIGNED"\n",
+ MAX (size, align / BITS_PER_UNIT));
+}
+
+/* We can't use .comm for local common storage as the SOM linker effectively
+ treats the symbol as universal and uses the same storage for local symbols
+ with the same name in different object files. The .block directive
+ reserves an uninitialized block of storage. However, it's not common
+ storage. Fortunately, GCC never requests common storage with the same
+ name in any given translation unit. */
+
+void
+pa_asm_output_aligned_local (FILE *stream,
+ const char *name,
+ unsigned HOST_WIDE_INT size,
+ unsigned int align)
+{
+ bss_section ();
+ fprintf (stream, "\t.align %u\n", align / BITS_PER_UNIT);
+
+#ifdef LOCAL_ASM_OP
+ fprintf (stream, "%s", LOCAL_ASM_OP);
+ assemble_name (stream, name);
+ fprintf (stream, "\n");
+#endif
+
+ ASM_OUTPUT_LABEL (stream, name);
+ fprintf (stream, "\t.block "HOST_WIDE_INT_PRINT_UNSIGNED"\n", size);
+}
+
/* Returns 1 if the 6 operands specified in OPERANDS are suitable for
use in fmpysub instructions. */
int
if (! rtx_equal_p (operands[3], operands[4]))
return 0;
- /* multiply can not feed into subtraction. */
+ /* multiply cannot feed into subtraction. */
if (rtx_equal_p (operands[5], operands[0]))
return 0;
- /* Inout operand of sub can not conflict with any operands from multiply. */
+ /* Inout operand of sub cannot conflict with any operands from multiply. */
if (rtx_equal_p (operands[3], operands[0])
|| rtx_equal_p (operands[3], operands[1])
|| rtx_equal_p (operands[3], operands[2]))
return 1;
}
-int
-plus_xor_ior_operator (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
-{
- return (GET_CODE (op) == PLUS || GET_CODE (op) == XOR
- || GET_CODE (op) == IOR);
-}
-
/* Return 1 if the given constant is 2, 4, or 8. These are the valid
constants for shadd instructions. */
-static int
+int
shadd_constant_p (int val)
{
if (val == 2 || val == 4 || val == 8)
return 0;
}
-/* Return 1 if OP is a CONST_INT with the value 2, 4, or 8. These are
- the valid constant for shadd instructions. */
-int
-shadd_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
-{
- return (GET_CODE (op) == CONST_INT && shadd_constant_p (INTVAL (op)));
-}
-
-/* Return 1 if OP is valid as a base register in a reg + reg address. */
+/* Return 1 if OP is valid as a base or index register in a
+ REG+REG address. */
int
-basereg_operand (rtx op, enum machine_mode mode)
+borx_reg_operand (rtx op, enum machine_mode mode)
{
- /* cse will create some unscaled indexed addresses, however; it
- generally isn't a win on the PA, so avoid creating unscaled
- indexed addresses until after cse is finished. */
- if (!cse_not_expected)
+ if (GET_CODE (op) != REG)
return 0;
- /* Allow any register when TARGET_NO_SPACE_REGS is in effect since
- we don't have to worry about the braindamaged implicit space
- register selection from the basereg. */
- if (TARGET_NO_SPACE_REGS)
- return (GET_CODE (op) == REG);
+ /* We must reject virtual registers as the only expressions that
+ can be instantiated are REG and REG+CONST. */
+ if (op == virtual_incoming_args_rtx
+ || op == virtual_stack_vars_rtx
+ || op == virtual_stack_dynamic_rtx
+ || op == virtual_outgoing_args_rtx
+ || op == virtual_cfa_rtx)
+ return 0;
/* While it's always safe to index off the frame pointer, it's not
- always profitable, particularly when the frame pointer is being
- eliminated. */
- if (! flag_omit_frame_pointer && op == frame_pointer_rtx)
- return 1;
+ profitable to do so when the frame pointer is being eliminated. */
+ if (!reload_completed
+ && flag_omit_frame_pointer
+ && !current_function_calls_alloca
+ && op == frame_pointer_rtx)
+ return 0;
- return (GET_CODE (op) == REG
- && REG_POINTER (op)
- && register_operand (op, mode));
+ return register_operand (op, mode);
}
/* Return 1 if this operand is anything other than a hard register. */
return (GET_CODE (op) == EQ || GET_CODE (op) == NE);
}
-/* Return 1 if OP is an operator suitable for use in a movb instruction. */
-int
-movb_comparison_operator (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
-{
- return (GET_CODE (op) == EQ || GET_CODE (op) == NE
- || GET_CODE (op) == LT || GET_CODE (op) == GE);
-}
-
/* Return 1 if INSN is in the delay slot of a call instruction. */
int
jump_in_call_delay (rtx insn)
will adhere to those rules.
So, late in the compilation process we find all the jump tables, and
- expand them into real code -- eg each entry in the jump table vector
+ expand them into real code -- e.g. each entry in the jump table vector
will get an appropriate label followed by a jump to the final target.
Reorg and the final jump pass can then optimize these branches and
markers disables output of the branch table to readonly memory,
and any alignment directives that might be needed. Possibly,
the begin_brtab insn should be output before the label for the
- table. This doesn matter at the moment since the tables are
+ table. This doesn't matter at the moment since the tables are
always output in the text section. */
for (insn = get_insns (); insn; insn = NEXT_INSN (insn))
{
delay slot of the millicode call -- thus they act more like traditional
CALL_INSNs.
- Note we can not consider side effects of the insn to be delayed because
+ Note we cannot consider side effects of the insn to be delayed because
the branch and link insn will clobber the return pointer. If we happened
to use the return pointer in the delay slot of the call, then we lose.
the mode is SF or DF. Then the value is returned in fr4 (32).
This must perform the same promotions as PROMOTE_MODE, else
- PROMOTE_FUNCTION_RETURN will not work correctly.
+ TARGET_PROMOTE_FUNCTION_RETURN will not work correctly.
Small structures must be returned in a PARALLEL on PA64 in order
to match the HP Compiler ABI. */
{
enum machine_mode valmode;
- /* Aggregates with a size less than or equal to 128 bits are returned
- in GR 28(-29). They are left justified. The pad bits are undefined.
- Larger aggregates are returned in memory. */
- if (TARGET_64BIT && AGGREGATE_TYPE_P (valtype))
+ if (AGGREGATE_TYPE_P (valtype))
{
- rtx loc[2];
- int i, offset = 0;
- int ub = int_size_in_bytes (valtype) <= UNITS_PER_WORD ? 1 : 2;
+ if (TARGET_64BIT)
+ {
+ /* Aggregates with a size less than or equal to 128 bits are
+ returned in GR 28(-29). They are left justified. The pad
+ bits are undefined. Larger aggregates are returned in
+ memory. */
+ rtx loc[2];
+ int i, offset = 0;
+ int ub = int_size_in_bytes (valtype) <= UNITS_PER_WORD ? 1 : 2;
+
+ for (i = 0; i < ub; i++)
+ {
+ loc[i] = gen_rtx_EXPR_LIST (VOIDmode,
+ gen_rtx_REG (DImode, 28 + i),
+ GEN_INT (offset));
+ offset += 8;
+ }
- for (i = 0; i < ub; i++)
+ return gen_rtx_PARALLEL (BLKmode, gen_rtvec_v (ub, loc));
+ }
+ else if (int_size_in_bytes (valtype) > UNITS_PER_WORD)
{
- loc[i] = gen_rtx_EXPR_LIST (VOIDmode,
- gen_rtx_REG (DImode, 28 + i),
- GEN_INT (offset));
- offset += 8;
+ /* Aggregates 5 to 8 bytes in size are returned in general
+ registers r28-r29 in the same manner as other non
+ floating-point objects. The data is right-justified and
+ zero-extended to 64 bits. This is opposite to the normal
+ justification used on big endian targets and requires
+ special treatment. */
+ rtx loc = gen_rtx_EXPR_LIST (VOIDmode,
+ gen_rtx_REG (DImode, 28), const0_rtx);
+ return gen_rtx_PARALLEL (BLKmode, gen_rtvec (1, loc));
}
-
- return gen_rtx_PARALLEL (BLKmode, gen_rtvec_v (ub, loc));
}
if ((INTEGRAL_TYPE_P (valtype)
valmode = TYPE_MODE (valtype);
if (TREE_CODE (valtype) == REAL_TYPE
+ && !AGGREGATE_TYPE_P (valtype)
&& TYPE_MODE (valtype) != TFmode
&& !TARGET_SOFT_FLOAT)
return gen_rtx_REG (valmode, 32);
arg_size = FUNCTION_ARG_SIZE (mode, type);
/* If this arg would be passed partially or totally on the stack, then
- this routine should return zero. FUNCTION_ARG_PARTIAL_NREGS will
+ this routine should return zero. pa_arg_partial_bytes will
handle arguments which are split between regs and stack slots if
the ABI mandates split arguments. */
if (! TARGET_64BIT)
/* Structures 5 to 8 bytes in size are passed in the general
registers in the same manner as other non floating-point
objects. The data is right-justified and zero-extended
- to 64 bits.
-
- This is magic. Normally, using a PARALLEL results in left
- justified data on a big-endian target. However, using a
- single double-word register provides the required right
- justification for 5 to 8 byte structures. This has nothing
- to do with the direction of padding specified for the argument.
- It has to do with how the data is widened and shifted into
- and from the register.
-
- Aside from adding load_multiple and store_multiple patterns,
- this is the only way that I have found to obtain right
- justification of BLKmode data when it has a size greater
- than one word. Splitting the operation into two SImode loads
- or returning a DImode REG results in left justified data. */
- if (mode == BLKmode)
+ to 64 bits. This is opposite to the normal justification
+ used on big endian targets and requires special treatment.
+ We now define BLOCK_REG_PADDING to pad these objects. */
+ if (mode == BLKmode || (type && AGGREGATE_TYPE_P (type)))
{
rtx loc = gen_rtx_EXPR_LIST (VOIDmode,
gen_rtx_REG (DImode, gpr_reg_base),
const0_rtx);
- return gen_rtx_PARALLEL (mode, gen_rtvec (1, loc));
+ return gen_rtx_PARALLEL (BLKmode, gen_rtvec (1, loc));
}
}
else
&& cum->indirect)
/* If the parameter is not a floating point parameter, then
it belongs in GPRs. */
- || !FLOAT_MODE_P (mode))
+ || !FLOAT_MODE_P (mode)
+ /* Structure with single SFmode field belongs in GPR. */
+ || (type && AGGREGATE_TYPE_P (type)))
retval = gen_rtx_REG (mode, gpr_reg_base);
else
retval = gen_rtx_REG (mode, fpr_reg_base);
/* If this arg would be passed totally in registers or totally on the stack,
- then this routine should return zero. It is currently called only for
- the 64-bit target. */
-int
-function_arg_partial_nregs (CUMULATIVE_ARGS *cum, enum machine_mode mode,
- tree type, int named ATTRIBUTE_UNUSED)
+ then this routine should return zero. */
+
+static int
+pa_arg_partial_bytes (CUMULATIVE_ARGS *cum, enum machine_mode mode,
+ tree type, bool named ATTRIBUTE_UNUSED)
{
unsigned int max_arg_words = 8;
unsigned int offset = 0;
+ if (!TARGET_64BIT)
+ return 0;
+
if (FUNCTION_ARG_SIZE (mode, type) > 1 && (cum->words & 1))
offset = 1;
return 0;
else
/* Arg is split. */
- return max_arg_words - cum->words - offset;
+ return (max_arg_words - cum->words - offset) * UNITS_PER_WORD;
}
-/* Return 1 if this is a comparison operator. This allows the use of
- MATCH_OPERATOR to recognize all the branch insns. */
+/* Return a string to output before text in the current function.
-int
-cmpib_comparison_operator (rtx op, enum machine_mode mode)
+ This function is only used with SOM. Because we don't support
+ named subspaces, we can only create a new subspace or switch back
+ to the default text subspace. */
+const char *
+som_text_section_asm_op (void)
{
- return ((mode == VOIDmode || GET_MODE (op) == mode)
- && (GET_CODE (op) == EQ
- || GET_CODE (op) == NE
- || GET_CODE (op) == GT
- || GET_CODE (op) == GTU
- || GET_CODE (op) == GE
- || GET_CODE (op) == LT
- || GET_CODE (op) == LE
- || GET_CODE (op) == LEU));
+ if (!TARGET_SOM)
+ return "";
+
+ if (TARGET_GAS)
+ {
+ if (cfun && !cfun->machine->in_nsubspa)
+ {
+ /* We only want to emit a .nsubspa directive once at the
+ start of the function. */
+ cfun->machine->in_nsubspa = 1;
+
+ /* Create a new subspace for the text. This provides
+ better stub placement and one-only functions. */
+ if (cfun->decl
+ && DECL_ONE_ONLY (cfun->decl)
+ && !DECL_WEAK (cfun->decl))
+ return
+ "\t.SPACE $TEXT$\n\t.NSUBSPA $CODE$,QUAD=0,ALIGN=8,ACCESS=44,SORT=24,COMDAT";
+
+ return "\t.SPACE $TEXT$\n\t.NSUBSPA $CODE$";
+ }
+ else
+ {
+ /* There isn't a current function or the body of the current
+ function has been completed. So, we are changing to the
+ text section to output debugging information. Do this in
+ the default text section. We need to forget that we are
+ in the text section so that the function text_section in
+ varasm.c will call us the next time around. */
+ forget_section ();
+ }
+ }
+
+ return "\t.SPACE $TEXT$\n\t.SUBSPA $CODE$";
}
/* On hpux10, the linker will give an error if we have a reference
&& (DECL_INITIAL (exp) == error_mark_node
|| TREE_CONSTANT (DECL_INITIAL (exp)))
&& !reloc)
+ {
+ if (TARGET_SOM
+ && DECL_ONE_ONLY (exp)
+ && !DECL_WEAK (exp))
+ som_one_only_readonly_data_section ();
+ else
+ readonly_data_section ();
+ }
+ else if (CONSTANT_CLASS_P (exp) && !reloc)
readonly_data_section ();
- else if (TREE_CODE_CLASS (TREE_CODE (exp)) == 'c'
- && !(TREE_CODE (exp) == STRING_CST && flag_writable_strings)
- && !reloc)
- readonly_data_section ();
+ else if (TARGET_SOM
+ && TREE_CODE (exp) == VAR_DECL
+ && DECL_ONE_ONLY (exp)
+ && !DECL_WEAK (exp))
+ som_one_only_data_section ();
else
data_section ();
}
fputs (",DATA\n", stream);
}
}
+
+/* Worker function for TARGET_STRUCT_VALUE_RTX. */
+
+static rtx
+pa_struct_value_rtx (tree fntype ATTRIBUTE_UNUSED,
+ int incoming ATTRIBUTE_UNUSED)
+{
+ return gen_rtx_REG (Pmode, PA_STRUCT_VALUE_REGNUM);
+}
+
+/* Worker function for TARGET_RETURN_IN_MEMORY. */
+
+bool
+pa_return_in_memory (tree type, tree fntype ATTRIBUTE_UNUSED)
+{
+ /* SOM ABI says that objects larger than 64 bits are returned in memory.
+ PA64 ABI says that objects larger than 128 bits are returned in memory.
+ Note, int_size_in_bytes can return -1 if the size of the object is
+ variable or larger than the maximum value that can be expressed as
+ a HOST_WIDE_INT. It can also return zero for an empty type. The
+ simplest way to handle variable and empty types is to pass them in
+ memory. This avoids problems in defining the boundaries of argument
+ slots, allocating registers, etc. */
+ return (int_size_in_bytes (type) > (TARGET_64BIT ? 16 : 8)
+ || int_size_in_bytes (type) <= 0);
+}
+
+/* Structure to hold declaration and name of external symbols that are
+ emitted by GCC. We generate a vector of these symbols and output them
+ at the end of the file if and only if SYMBOL_REF_REFERENCED_P is true.
+ This avoids putting out names that are never really used. */
+
+typedef struct extern_symbol GTY(())
+{
+ tree decl;
+ const char *name;
+} extern_symbol;
+
+/* Define gc'd vector type for extern_symbol. */
+DEF_VEC_O(extern_symbol);
+DEF_VEC_ALLOC_O(extern_symbol,gc);
+
+/* Vector of extern_symbol pointers. */
+static GTY(()) VEC(extern_symbol,gc) *extern_symbols;
+
+#ifdef ASM_OUTPUT_EXTERNAL_REAL
+/* Mark DECL (name NAME) as an external reference (assembler output
+ file FILE). This saves the names to output at the end of the file
+ if actually referenced. */
+
+void
+pa_hpux_asm_output_external (FILE *file, tree decl, const char *name)
+{
+ extern_symbol * p = VEC_safe_push (extern_symbol, gc, extern_symbols, NULL);
+
+ gcc_assert (file == asm_out_file);
+ p->decl = decl;
+ p->name = name;
+}
+
+/* Output text required at the end of an assembler file.
+ This includes deferred plabels and .import directives for
+ all external symbols that were actually referenced. */
+
+static void
+pa_hpux_file_end (void)
+{
+ unsigned int i;
+ extern_symbol *p;
+
+ if (!NO_DEFERRED_PROFILE_COUNTERS)
+ output_deferred_profile_counters ();
+
+ output_deferred_plabels ();
+
+ for (i = 0; VEC_iterate (extern_symbol, extern_symbols, i, p); i++)
+ {
+ tree decl = p->decl;
+
+ if (!TREE_ASM_WRITTEN (decl)
+ && SYMBOL_REF_REFERENCED_P (XEXP (DECL_RTL (decl), 0)))
+ ASM_OUTPUT_EXTERNAL_REAL (asm_out_file, decl, p->name);
+ }
+
+ VEC_free (extern_symbol, gc, extern_symbols);
+}
+#endif
+
#include "gt-pa.h"
OSDN Git Service
