/* Definitions of target machine for GNU compiler, for MMIX.
- Copyright (C) 2000, 2001, 2002 Free Software Foundation, Inc.
+ Copyright (C) 2000, 2001, 2002, 2004 Free Software Foundation, Inc.
Contributed by Hans-Peter Nilsson (hp@bitrange.com)
This file is part of GCC.
#define TARGET_OPTIONS \
{{"set-program-start=", &mmix_cc1_ignored_option, \
- N_("Set start-address of the program") }, \
+ N_("Set start-address of the program"), 0}, \
{"set-data-start=", &mmix_cc1_ignored_option, \
- N_("Set start-address of data")}}
+ N_("Set start-address of data"), 0} }
/* FIXME: There's no provision for profiling here. */
#define STARTFILE_SPEC \
address goes in a global register. When addressing, it's more like
"base address plus offset", with the offset being 0..255 from the base,
which itself can be a symbol plus an offset. The effect is like having
- a constant pool in global registers, code offseting from those
+ a constant pool in global registers, code offsetting from those
registers (automatically causing a request for a suitable constant base
address register) without having to know the specific register or the
specific offset. The setback is that there's a limited number of
registers, and you'll not find out until link time whether you
- should've compiled with -mno-base-addresses. */
+ should have compiled with -mno-base-addresses. */
#define TARGET_MASK_BASE_ADDRESSES 128
/* FIXME: Get rid of this one. */
#define FLOAT_WORDS_BIG_ENDIAN 1
#define UNITS_PER_WORD 8
-/* FIXME: This macro is correlated to MAX_FIXED_MODE_SIZE in that
- e.g. this macro must not be 8 (default, UNITS_PER_WORD) when
- MAX_FIXED_MODE_SIZE is 64 (default, DImode), or really: this must be
- set manually if MAX_FIXED_MODE_SIZE is not at least twice the register
- size. By setting it to 4, we don't have to worry about TImode things
- yet. Revisit, perhaps get TImode going or get some solution that does
- not mandate TImode or lie in other ways. */
-#define MIN_UNITS_PER_WORD 4
-
/* FIXME: Promotion of modes currently generates slow code, extending
before every operation. */
+/* I'm a little bit undecided about this one. It might be beneficial to
+ promote all operations. */
-#define PROMOTE_MODE(MODE, UNSIGNEDP, TYPE) \
+#define PROMOTE_FUNCTION_MODE(MODE, UNSIGNEDP, TYPE) \
do { \
if (GET_MODE_CLASS (MODE) == MODE_INT \
&& GET_MODE_SIZE (MODE) < 8) \
} \
} while (0)
-#define PROMOTE_FUNCTION_ARGS
-
-#if 0
-/* Apparently not doing TRT if int < register-size. FIXME: Perhaps
- FUNCTION_VALUE and LIBCALL_VALUE needs tweaking as some ports say. */
-#define PROMOTE_FUNCTION_RETURN
-#endif
-
-/* I'm a little bit undecided about this one. It might be beneficial to
- promote all operations. */
-#define PROMOTE_FOR_CALL_ONLY
-
/* We need to align everything to 64 bits that can affect the alignment
of other types. Since address N is interpreted in MMIX as (N modulo
access_size), we must align. */
/* Node: Elimination */
/* FIXME: Is this requirement built-in? Anyway, we should try to get rid
of it; we can deduce the value. */
-#define FRAME_POINTER_REQUIRED (nonlocal_goto_stack_level != NULL_RTX)
+#define FRAME_POINTER_REQUIRED current_function_has_nonlocal_label
/* The frame-pointer is stored in a location that either counts to the
offset of incoming parameters, or that counts to the offset of the
#define FUNCTION_INCOMING_ARG(CUM, MODE, TYPE, NAMED) \
mmix_function_arg (&(CUM), MODE, TYPE, NAMED, 1)
-#define FUNCTION_ARG_PASS_BY_REFERENCE(CUM, MODE, TYPE, NAMED) \
- mmix_function_arg_pass_by_reference (&(CUM), MODE, TYPE, NAMED)
-
-/* This *sounds* good, but does not seem to be implemented correctly to
- be a win; at least it wasn't in 2.7.2. FIXME: Check and perhaps
- replace with a big comment. */
-#define FUNCTION_ARG_CALLEE_COPIES(CUM, MODE, TYPE, NAMED) 1
-
typedef struct { int regs; int lib; } CUMULATIVE_ARGS;
-#define INIT_CUMULATIVE_ARGS(CUM, FNTYPE, LIBNAME, INDIRECT) \
+#define INIT_CUMULATIVE_ARGS(CUM, FNTYPE, LIBNAME, INDIRECT, N_NAMED_ARGS) \
((CUM).regs = 0, (CUM).lib = ((LIBNAME) != 0))
#define FUNCTION_ARG_ADVANCE(CUM, MODE, TYPE, NAMED) \
((CUM).regs \
- = ((MUST_PASS_IN_STACK (MODE, TYPE)) \
+ = ((targetm.calls.must_pass_in_stack (MODE, TYPE)) \
|| (MMIX_FUNCTION_ARG_SIZE (MODE, TYPE) > 8 \
&& !TARGET_LIBFUNC && !(CUM).lib)) \
? (MMIX_MAX_ARGS_IN_REGS) + 1 \
mmix_function_outgoing_value (VALTYPE, FUNC)
#define LIBCALL_VALUE(MODE) \
- gen_rtx_REG (MODE, MMIX_OUTGOING_RETURN_VALUE_REGNUM)
+ gen_rtx_REG (MODE, MMIX_RETURN_VALUE_REGNUM)
#define FUNCTION_VALUE_REGNO_P(REGNO) \
- ((REGNO) == MMIX_OUTGOING_RETURN_VALUE_REGNUM)
-
-
-/* Node: Aggregate Return */
-
-#define STRUCT_VALUE_REGNUM MMIX_STRUCT_VALUE_REGNUM
+ mmix_function_value_regno_p (REGNO)
/* Node: Caller Saves */
#define EPILOGUE_USES(REGNO) \
((REGNO) == MMIX_INCOMING_RETURN_ADDRESS_REGNUM)
-#define TARGET_ASM_OUTPUT_MI_THUNK mmix_asm_output_mi_thunk
-
-
/* Node: Profiling */
#define FUNCTION_PROFILER(FILE, LABELNO) \
mmix_function_profiler (FILE, LABELNO)
-/* Node: Varargs */
-
-/* For the moment, let's stick to pushing argument registers on the stack.
- Later, we can parse all arguments in registers, to improve
- performance. */
-#define SETUP_INCOMING_VARARGS(A, M, T, P, S) \
- mmix_setup_incoming_varargs(&(A), M, T, &(P), S)
-
-/* FIXME: This and other EXPAND_BUILTIN_VA_... target macros are not
- documented, although used by several targets. */
-#define EXPAND_BUILTIN_VA_ARG(VALIST, TYPE) \
- mmix_expand_builtin_va_arg (VALIST, TYPE)
-
/* Node: Trampolines */
#define TRAMPOLINE_TEMPLATE(FILE) \
mmix_initialize_trampoline (ADDR, FNADDR, STATIC_CHAIN)
-/* Node: Library Calls */
-
-#define TARGET_MEM_FUNCTIONS
-
-
/* Node: Addressing Modes */
#define CONSTANT_ADDRESS_P(X) \
#define REG_OK_FOR_INDEX_P(X) REG_OK_FOR_BASE_P (X)
-#define LEGITIMIZE_ADDRESS(X, OLDX, MODE, WIN)
-
#define GO_IF_MODE_DEPENDENT_ADDRESS(ADDR, LABEL)
#define LEGITIMATE_CONSTANT_P(X) \
comparisons with -1 to LT and GE respectively, and LT, LTU, GE or GEU
comparisons with 256 to 255 and LE, LEU, GT and GTU has been
ineffective; the code path for performing the changes did not trig for
- neither the GCC test-suite nor ghostscript-6.52 nor Knuth's mmix.tar.gz
+ neither the GCC testsuite nor ghostscript-6.52 nor Knuth's mmix.tar.gz
itself (core GCC functionality supposedly handling it) with sources
from 2002-06-06. */
/* Node: Costs */
-/* This one takes on both the RTX_COSTS and CONST_COSTS tasks. */
-#define DEFAULT_RTX_COSTS(X, CODE, OUTER_CODE) \
- { \
- int mmix_rtx_cost; \
- if (mmix_rtx_cost_recalculated (X, CODE, OUTER_CODE, \
- &mmix_rtx_cost)) \
- return mmix_rtx_cost; \
- }
-
-#define ADDRESS_COST(ADDRESS) mmix_address_cost (ADDRESS)
-
/* The special registers can only move to and from general regs, and we
need to check that their constraints match, so say 3 for them. */
/* WARNING: gcc-2.7.2.2 i686-pc-linux-gnulibc1 (as shipped with RH 4.2)
/* Node: File Framework */
-#define ASM_FILE_START(STREAM) \
- mmix_asm_file_start (STREAM)
-
-#define ASM_FILE_END(STREAM) \
- mmix_asm_file_end (STREAM)
-
/* While any other punctuation character but ";" would do, we prefer "%"
or "!"; "!" is an unary operator and so will not be mistakenly included
in correctly formed expressions. The hash character adds mass; catches
#define OUTPUT_QUOTED_STRING(STREAM, STRING) \
mmix_output_quoted_string (STREAM, STRING, strlen (STRING))
-#define ASM_OUTPUT_SOURCE_LINE(STREAM, LINE) \
- mmix_asm_output_source_line (STREAM, LINE)
-
#define TARGET_ASM_NAMED_SECTION default_elf_asm_named_section
-
/* Node: Data Output */
#define ASM_OUTPUT_ASCII(STREAM, PTR, LEN) \
#define ASM_OUTPUT_LABEL(STREAM, NAME) \
mmix_asm_output_label (STREAM, NAME)
+#define ASM_OUTPUT_INTERNAL_LABEL(STREAM, NAME) \
+ mmix_asm_output_internal_label (STREAM, NAME)
+
#define ASM_DECLARE_REGISTER_GLOBAL(STREAM, DECL, REGNO, NAME) \
mmix_asm_declare_register_global (STREAM, DECL, REGNO, NAME)
#define ASM_OUTPUT_LABELREF(STREAM, NAME) \
mmix_asm_output_labelref (STREAM, NAME)
-#define ASM_OUTPUT_INTERNAL_LABEL(STREAM, PREFIX, NUM) \
- mmix_asm_output_internal_label (STREAM, PREFIX, NUM)
-
/* We insert a ":" to disambiguate against user symbols like L5. */
#define ASM_GENERATE_INTERNAL_LABEL(LABEL, PREFIX, NUM) \
sprintf (LABEL, "*%s:%ld", PREFIX, (long)(NUM))
":" is seen in the object file; we don't really want that mmixal
feature visible there. We don't want the default, which uses a dot;
that'd be incompatible with mmixal. */
-#define ASM_FORMAT_PRIVATE_NAME(OUTPUT, NAME, LABELNO) \
- ((OUTPUT) = (char *) alloca (strlen ((NAME)) + 2 + 10), \
- sprintf ((OUTPUT), "%s::%d", (NAME), (LABELNO)))
+#define ASM_PN_FORMAT "%s::%lu"
#define ASM_OUTPUT_DEF(STREAM, NAME, VALUE) \
mmix_asm_output_def (STREAM, NAME, VALUE)
#define TRULY_NOOP_TRUNCATION(OUTPREC, INPREC) 1
-/* We have a choice here too. */
-#if 0
-/* FIXME: Revisit, we don't have scc expanders yet. */
-#define STORE_FLAG_VALUE 1
-#endif
+/* ??? MMIX allows a choice of STORE_FLAG_VALUE. Revisit later,
+ we don't have scc expanders yet. */
#define Pmode DImode
#define NO_IMPLICIT_EXTERN_C
-#define HANDLE_SYSV_PRAGMA
+#define HANDLE_SYSV_PRAGMA 1
/* These are checked. */
#define DOLLARS_IN_IDENTIFIERS 0
#define NO_DOLLAR_IN_LABEL
#define NO_DOT_IN_LABEL
-/* Calculate the highest used supposed saved stack register. */
-#define MACHINE_DEPENDENT_REORG(INSN) \
- mmix_machine_dependent_reorg (INSN)
-
#endif /* GCC_MMIX_H */
/*
* Local variables: