/* Output routines for GCC for ARM.
Copyright (C) 1991, 1993, 1994, 1995, 1996, 1997, 1998, 1999, 2000, 2001,
- 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009
+ 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010
Free Software Foundation, Inc.
Contributed by Pieter `Tiggr' Schoenmakers (rcpieter@win.tue.nl)
and Martin Simmons (@harleqn.co.uk).
#include "obstack.h"
#include "regs.h"
#include "hard-reg-set.h"
-#include "real.h"
#include "insn-config.h"
#include "conditions.h"
#include "output.h"
#include "optabs.h"
#include "toplev.h"
#include "recog.h"
+#include "cgraph.h"
#include "ggc.h"
#include "except.h"
#include "c-pragma.h"
#include "debug.h"
#include "langhooks.h"
#include "df.h"
+#include "intl.h"
+#include "libfuncs.h"
/* Forward definitions of types. */
typedef struct minipool_node Mnode;
typedef struct minipool_fixup Mfix;
-EXPORTED_CONST struct attribute_spec arm_attribute_table[];
-
void (*arm_lang_output_object_attributes_hook)(void);
/* Forward function declarations. */
static unsigned long arm_isr_value (tree);
static unsigned long arm_compute_func_type (void);
static tree arm_handle_fndecl_attribute (tree *, tree, tree, int, bool *);
+static tree arm_handle_pcs_attribute (tree *, tree, tree, int, bool *);
static tree arm_handle_isr_attribute (tree *, tree, tree, int, bool *);
#if TARGET_DLLIMPORT_DECL_ATTRIBUTES
static tree arm_handle_notshared_attribute (tree *, tree, tree, int, bool *);
static int count_insns_for_constant (HOST_WIDE_INT, int);
static int arm_get_strip_length (int);
static bool arm_function_ok_for_sibcall (tree, tree);
+static enum machine_mode arm_promote_function_mode (const_tree,
+ enum machine_mode, int *,
+ const_tree, int);
+static bool arm_return_in_memory (const_tree, const_tree);
+static rtx arm_function_value (const_tree, const_tree, bool);
+static rtx arm_libcall_value (enum machine_mode, const_rtx);
+
static void arm_internal_label (FILE *, const char *, unsigned long);
static void arm_output_mi_thunk (FILE *, tree, HOST_WIDE_INT, HOST_WIDE_INT,
tree);
+static bool arm_have_conditional_execution (void);
static bool arm_rtx_costs_1 (rtx, enum rtx_code, int*, bool);
static bool arm_size_rtx_costs (rtx, enum rtx_code, enum rtx_code, int *);
static bool arm_slowmul_rtx_costs (rtx, enum rtx_code, enum rtx_code, int *, bool);
static bool arm_cirrus_insn_p (rtx);
static void cirrus_reorg (rtx);
static void arm_init_builtins (void);
-static rtx arm_expand_builtin (tree, rtx, rtx, enum machine_mode, int);
static void arm_init_iwmmxt_builtins (void);
static rtx safe_vector_operand (rtx, enum machine_mode);
static rtx arm_expand_binop_builtin (enum insn_code, tree, rtx);
static rtx emit_set_insn (rtx, rtx);
static int arm_arg_partial_bytes (CUMULATIVE_ARGS *, enum machine_mode,
tree, bool);
+static rtx aapcs_allocate_return_reg (enum machine_mode, const_tree,
+ const_tree);
+static int aapcs_select_return_coproc (const_tree, const_tree);
#ifdef OBJECT_FORMAT_ELF
static void arm_elf_asm_constructor (rtx, int) ATTRIBUTE_UNUSED;
static bool arm_output_ttype (rtx);
#endif
static void arm_dwarf_handle_frame_unspec (const char *, rtx, int);
+static rtx arm_dwarf_register_span (rtx);
static tree arm_cxx_guard_type (void);
static bool arm_cxx_guard_mask_bit (void);
static int arm_issue_rate (void);
static void arm_output_dwarf_dtprel (FILE *, int, rtx) ATTRIBUTE_UNUSED;
static bool arm_allocate_stack_slots_for_args (void);
+static const char *arm_invalid_parameter_type (const_tree t);
+static const char *arm_invalid_return_type (const_tree t);
+static tree arm_promoted_type (const_tree t);
+static tree arm_convert_to_type (tree type, tree expr);
+static bool arm_scalar_mode_supported_p (enum machine_mode);
+static bool arm_frame_pointer_required (void);
+static bool arm_can_eliminate (const int, const int);
+static void arm_asm_trampoline_template (FILE *);
+static void arm_trampoline_init (rtx, tree, rtx);
+static rtx arm_trampoline_adjust_address (rtx);
+static rtx arm_pic_static_addr (rtx orig, rtx reg);
+
+\f
+/* Table of machine attributes. */
+static const struct attribute_spec arm_attribute_table[] =
+{
+ /* { name, min_len, max_len, decl_req, type_req, fn_type_req, handler } */
+ /* Function calls made to this symbol must be done indirectly, because
+ it may lie outside of the 26 bit addressing range of a normal function
+ call. */
+ { "long_call", 0, 0, false, true, true, NULL },
+ /* Whereas these functions are always known to reside within the 26 bit
+ addressing range. */
+ { "short_call", 0, 0, false, true, true, NULL },
+ /* Specify the procedure call conventions for a function. */
+ { "pcs", 1, 1, false, true, true, arm_handle_pcs_attribute },
+ /* Interrupt Service Routines have special prologue and epilogue requirements. */
+ { "isr", 0, 1, false, false, false, arm_handle_isr_attribute },
+ { "interrupt", 0, 1, false, false, false, arm_handle_isr_attribute },
+ { "naked", 0, 0, true, false, false, arm_handle_fndecl_attribute },
+#ifdef ARM_PE
+ /* ARM/PE has three new attributes:
+ interfacearm - ?
+ dllexport - for exporting a function/variable that will live in a dll
+ dllimport - for importing a function/variable from a dll
+ Microsoft allows multiple declspecs in one __declspec, separating
+ them with spaces. We do NOT support this. Instead, use __declspec
+ multiple times.
+ */
+ { "dllimport", 0, 0, true, false, false, NULL },
+ { "dllexport", 0, 0, true, false, false, NULL },
+ { "interfacearm", 0, 0, true, false, false, arm_handle_fndecl_attribute },
+#elif TARGET_DLLIMPORT_DECL_ATTRIBUTES
+ { "dllimport", 0, 0, false, false, false, handle_dll_attribute },
+ { "dllexport", 0, 0, false, false, false, handle_dll_attribute },
+ { "notshared", 0, 0, false, true, false, arm_handle_notshared_attribute },
+#endif
+ { NULL, 0, 0, false, false, false, NULL }
+};
\f
/* Initialize the GCC target structure. */
#if TARGET_DLLIMPORT_DECL_ATTRIBUTES
#undef TARGET_FUNCTION_OK_FOR_SIBCALL
#define TARGET_FUNCTION_OK_FOR_SIBCALL arm_function_ok_for_sibcall
+#undef TARGET_FUNCTION_VALUE
+#define TARGET_FUNCTION_VALUE arm_function_value
+
+#undef TARGET_LIBCALL_VALUE
+#define TARGET_LIBCALL_VALUE arm_libcall_value
+
#undef TARGET_ASM_OUTPUT_MI_THUNK
#define TARGET_ASM_OUTPUT_MI_THUNK arm_output_mi_thunk
#undef TARGET_ASM_CAN_OUTPUT_MI_THUNK
#undef TARGET_INIT_LIBFUNCS
#define TARGET_INIT_LIBFUNCS arm_init_libfuncs
-#undef TARGET_PROMOTE_FUNCTION_ARGS
-#define TARGET_PROMOTE_FUNCTION_ARGS hook_bool_const_tree_true
-#undef TARGET_PROMOTE_FUNCTION_RETURN
-#define TARGET_PROMOTE_FUNCTION_RETURN hook_bool_const_tree_true
+#undef TARGET_PROMOTE_FUNCTION_MODE
+#define TARGET_PROMOTE_FUNCTION_MODE arm_promote_function_mode
#undef TARGET_PROMOTE_PROTOTYPES
#define TARGET_PROMOTE_PROTOTYPES arm_promote_prototypes
#undef TARGET_PASS_BY_REFERENCE
#undef TARGET_ALLOCATE_STACK_SLOTS_FOR_ARGS
#define TARGET_ALLOCATE_STACK_SLOTS_FOR_ARGS arm_allocate_stack_slots_for_args
+#undef TARGET_ASM_TRAMPOLINE_TEMPLATE
+#define TARGET_ASM_TRAMPOLINE_TEMPLATE arm_asm_trampoline_template
+#undef TARGET_TRAMPOLINE_INIT
+#define TARGET_TRAMPOLINE_INIT arm_trampoline_init
+#undef TARGET_TRAMPOLINE_ADJUST_ADDRESS
+#define TARGET_TRAMPOLINE_ADJUST_ADDRESS arm_trampoline_adjust_address
+
#undef TARGET_DEFAULT_SHORT_ENUMS
#define TARGET_DEFAULT_SHORT_ENUMS arm_default_short_enums
#undef TARGET_DWARF_HANDLE_FRAME_UNSPEC
#define TARGET_DWARF_HANDLE_FRAME_UNSPEC arm_dwarf_handle_frame_unspec
+#undef TARGET_DWARF_REGISTER_SPAN
+#define TARGET_DWARF_REGISTER_SPAN arm_dwarf_register_span
+
#undef TARGET_CANNOT_COPY_INSN_P
#define TARGET_CANNOT_COPY_INSN_P arm_cannot_copy_insn_p
#define TARGET_HAVE_TLS true
#endif
+#undef TARGET_HAVE_CONDITIONAL_EXECUTION
+#define TARGET_HAVE_CONDITIONAL_EXECUTION arm_have_conditional_execution
+
#undef TARGET_CANNOT_FORCE_CONST_MEM
#define TARGET_CANNOT_FORCE_CONST_MEM arm_cannot_force_const_mem
#undef TARGET_LEGITIMATE_ADDRESS_P
#define TARGET_LEGITIMATE_ADDRESS_P arm_legitimate_address_p
+#undef TARGET_INVALID_PARAMETER_TYPE
+#define TARGET_INVALID_PARAMETER_TYPE arm_invalid_parameter_type
+
+#undef TARGET_INVALID_RETURN_TYPE
+#define TARGET_INVALID_RETURN_TYPE arm_invalid_return_type
+
+#undef TARGET_PROMOTED_TYPE
+#define TARGET_PROMOTED_TYPE arm_promoted_type
+
+#undef TARGET_CONVERT_TO_TYPE
+#define TARGET_CONVERT_TO_TYPE arm_convert_to_type
+
+#undef TARGET_SCALAR_MODE_SUPPORTED_P
+#define TARGET_SCALAR_MODE_SUPPORTED_P arm_scalar_mode_supported_p
+
+#undef TARGET_FRAME_POINTER_REQUIRED
+#define TARGET_FRAME_POINTER_REQUIRED arm_frame_pointer_required
+
+#undef TARGET_CAN_ELIMINATE
+#define TARGET_CAN_ELIMINATE arm_can_eliminate
+
struct gcc_target targetm = TARGET_INITIALIZER;
\f
/* Obstack for minipool constant handling. */
/* The processor for which instructions should be scheduled. */
enum processor_type arm_tune = arm_none;
+/* The current tuning set. */
+const struct tune_params *current_tune;
+
/* The default processor used if not overridden by commandline. */
static enum processor_type arm_default_cpu = arm_none;
-/* Which floating point model to use. */
-enum arm_fp_model arm_fp_model;
-
-/* Which floating point hardware is available. */
-enum fputype arm_fpu_arch;
-
/* Which floating point hardware to schedule for. */
-enum fputype arm_fpu_tune;
+int arm_fpu_attr;
+
+/* Which floating popint hardware to use. */
+const struct arm_fpu_desc *arm_fpu_desc;
/* Whether to use floating point hardware. */
enum float_abi_type arm_float_abi;
+/* Which __fp16 format to use. */
+enum arm_fp16_format_type arm_fp16_format;
+
/* Which ABI to use. */
enum arm_abi_type arm_abi;
#define FL_DIV (1 << 18) /* Hardware divide. */
#define FL_VFPV3 (1 << 19) /* Vector Floating Point V3. */
#define FL_NEON (1 << 20) /* Neon instructions. */
+#define FL_ARCH7EM (1 << 21) /* Instructions present in the ARMv7E-M
+ architecture. */
#define FL_IWMMXT (1 << 29) /* XScale v2 or "Intel Wireless MMX technology". */
#define FL_FOR_ARCH6T2 (FL_FOR_ARCH6 | FL_THUMB2)
#define FL_FOR_ARCH6M (FL_FOR_ARCH6 & ~FL_NOTM)
#define FL_FOR_ARCH7 (FL_FOR_ARCH6T2 &~ FL_NOTM)
-#define FL_FOR_ARCH7A (FL_FOR_ARCH7 | FL_NOTM)
+#define FL_FOR_ARCH7A (FL_FOR_ARCH7 | FL_NOTM | FL_ARCH6K)
#define FL_FOR_ARCH7R (FL_FOR_ARCH7A | FL_DIV)
#define FL_FOR_ARCH7M (FL_FOR_ARCH7 | FL_DIV)
+#define FL_FOR_ARCH7EM (FL_FOR_ARCH7M | FL_ARCH7EM)
/* The bits in this mask specify which
instructions we are allowed to generate. */
/* Nonzero if instructions not present in the 'M' profile can be used. */
int arm_arch_notm = 0;
+/* Nonzero if instructions present in ARMv7E-M can be used. */
+int arm_arch7em = 0;
+
/* Nonzero if this chip can benefit from load scheduling. */
int arm_ld_sched = 0;
the next function. */
static int after_arm_reorg = 0;
-/* The maximum number of insns to be used when loading a constant. */
-static int arm_constant_limit = 3;
+static enum arm_pcs arm_pcs_default;
/* For an explanation of these variables, see final_prescan_insn below. */
int arm_ccfsm_state;
enum processor_type core;
const char *arch;
const unsigned long flags;
- bool (* rtx_costs) (rtx, enum rtx_code, enum rtx_code, int *, bool);
+ const struct tune_params *const tune;
+};
+
+const struct tune_params arm_slowmul_tune =
+{
+ arm_slowmul_rtx_costs,
+ 3
+};
+
+const struct tune_params arm_fastmul_tune =
+{
+ arm_fastmul_rtx_costs,
+ 1
+};
+
+const struct tune_params arm_xscale_tune =
+{
+ arm_xscale_rtx_costs,
+ 2
+};
+
+const struct tune_params arm_9e_tune =
+{
+ arm_9e_rtx_costs,
+ 1
};
/* Not all of these give usefully different compilation alternatives,
{
/* ARM Cores */
#define ARM_CORE(NAME, IDENT, ARCH, FLAGS, COSTS) \
- {NAME, arm_none, #ARCH, FLAGS | FL_FOR_ARCH##ARCH, arm_##COSTS##_rtx_costs},
+ {NAME, arm_none, #ARCH, FLAGS | FL_FOR_ARCH##ARCH, &arm_##COSTS##_tune},
#include "arm-cores.def"
#undef ARM_CORE
{NULL, arm_none, NULL, 0, NULL}
static const struct processors all_architectures[] =
{
/* ARM Architectures */
- /* We don't specify rtx_costs here as it will be figured out
+ /* We don't specify tuning costs here as it will be figured out
from the core. */
{"armv2", arm2, "2", FL_CO_PROC | FL_MODE26 | FL_FOR_ARCH2, NULL},
{"armv7-a", cortexa8, "7A", FL_CO_PROC | FL_FOR_ARCH7A, NULL},
{"armv7-r", cortexr4, "7R", FL_CO_PROC | FL_FOR_ARCH7R, NULL},
{"armv7-m", cortexm3, "7M", FL_CO_PROC | FL_FOR_ARCH7M, NULL},
+ {"armv7e-m", cortexm3, "7EM", FL_CO_PROC | FL_FOR_ARCH7EM, NULL},
{"ep9312", ep9312, "4T", FL_LDSCHED | FL_CIRRUS | FL_FOR_ARCH4, NULL},
{"iwmmxt", iwmmxt, "5TE", FL_LDSCHED | FL_STRONG | FL_FOR_ARCH5TE | FL_XSCALE | FL_IWMMXT , NULL},
{"iwmmxt2", iwmmxt2, "5TE", FL_LDSCHED | FL_STRONG | FL_FOR_ARCH5TE | FL_XSCALE | FL_IWMMXT , NULL},
char arm_arch_name[] = "__ARM_ARCH_0UNK__";
-struct fpu_desc
-{
- const char * name;
- enum fputype fpu;
-};
-
-
/* Available values for -mfpu=. */
-static const struct fpu_desc all_fpus[] =
-{
- {"fpa", FPUTYPE_FPA},
- {"fpe2", FPUTYPE_FPA_EMU2},
- {"fpe3", FPUTYPE_FPA_EMU2},
- {"maverick", FPUTYPE_MAVERICK},
- {"vfp", FPUTYPE_VFP},
- {"vfp3", FPUTYPE_VFP3},
- {"vfpv3", FPUTYPE_VFP3},
- {"vfpv3-d16", FPUTYPE_VFP3D16},
- {"neon", FPUTYPE_NEON}
-};
-
-
-/* Floating point models used by the different hardware.
- See fputype in arm.h. */
-
-static const enum arm_fp_model fp_model_for_fpu[] =
-{
- /* No FP hardware. */
- ARM_FP_MODEL_UNKNOWN, /* FPUTYPE_NONE */
- ARM_FP_MODEL_FPA, /* FPUTYPE_FPA */
- ARM_FP_MODEL_FPA, /* FPUTYPE_FPA_EMU2 */
- ARM_FP_MODEL_FPA, /* FPUTYPE_FPA_EMU3 */
- ARM_FP_MODEL_MAVERICK, /* FPUTYPE_MAVERICK */
- ARM_FP_MODEL_VFP, /* FPUTYPE_VFP */
- ARM_FP_MODEL_VFP, /* FPUTYPE_VFP3D16 */
- ARM_FP_MODEL_VFP, /* FPUTYPE_VFP3 */
- ARM_FP_MODEL_VFP /* FPUTYPE_NEON */
+static const struct arm_fpu_desc all_fpus[] =
+{
+ {"fpa", ARM_FP_MODEL_FPA, 0, VFP_NONE, false, false},
+ {"fpe2", ARM_FP_MODEL_FPA, 2, VFP_NONE, false, false},
+ {"fpe3", ARM_FP_MODEL_FPA, 3, VFP_NONE, false, false},
+ {"maverick", ARM_FP_MODEL_MAVERICK, 0, VFP_NONE, false, false},
+ {"vfp", ARM_FP_MODEL_VFP, 2, VFP_REG_D16, false, false},
+ {"vfpv3", ARM_FP_MODEL_VFP, 3, VFP_REG_D32, false, false},
+ {"vfpv3-fp16", ARM_FP_MODEL_VFP, 3, VFP_REG_D32, false, true},
+ {"vfpv3-d16", ARM_FP_MODEL_VFP, 3, VFP_REG_D16, false, false},
+ {"vfpv3-d16-fp16", ARM_FP_MODEL_VFP, 3, VFP_REG_D16, false, true},
+ {"vfpv3xd", ARM_FP_MODEL_VFP, 3, VFP_REG_SINGLE, false, false},
+ {"vfpv3xd-fp16", ARM_FP_MODEL_VFP, 3, VFP_REG_SINGLE, false, true},
+ {"neon", ARM_FP_MODEL_VFP, 3, VFP_REG_D32, true , false},
+ {"neon-fp16", ARM_FP_MODEL_VFP, 3, VFP_REG_D32, true , true },
+ {"vfpv4", ARM_FP_MODEL_VFP, 4, VFP_REG_D32, false, true},
+ {"vfpv4-d16", ARM_FP_MODEL_VFP, 4, VFP_REG_D16, false, true},
+ {"fpv4-sp-d16", ARM_FP_MODEL_VFP, 4, VFP_REG_SINGLE, false, true},
+ {"neon-vfpv4", ARM_FP_MODEL_VFP, 4, VFP_REG_D32, true, true},
+ /* Compatibility aliases. */
+ {"vfp3", ARM_FP_MODEL_VFP, 3, VFP_REG_D32, false, false},
};
};
+struct fp16_format
+{
+ const char *name;
+ enum arm_fp16_format_type fp16_format_type;
+};
+
+
+/* Available values for -mfp16-format=. */
+
+static const struct fp16_format all_fp16_formats[] =
+{
+ {"none", ARM_FP16_FORMAT_NONE},
+ {"ieee", ARM_FP16_FORMAT_IEEE},
+ {"alternative", ARM_FP16_FORMAT_ALTERNATIVE}
+};
+
+
struct abi_name
{
const char *name;
TLS_LE32
};
+/* The maximum number of insns to be used when loading a constant. */
+inline static int
+arm_constant_limit (bool size_p)
+{
+ return size_p ? 1 : current_tune->constant_limit;
+}
+
/* Emit an insn that's a simple single-set. Both the operands must be known
to be valid. */
inline static rtx
set_optab_libfunc (umod_optab, DImode, NULL);
set_optab_libfunc (smod_optab, SImode, NULL);
set_optab_libfunc (umod_optab, SImode, NULL);
+
+ /* Half-precision float operations. The compiler handles all operations
+ with NULL libfuncs by converting the SFmode. */
+ switch (arm_fp16_format)
+ {
+ case ARM_FP16_FORMAT_IEEE:
+ case ARM_FP16_FORMAT_ALTERNATIVE:
+
+ /* Conversions. */
+ set_conv_libfunc (trunc_optab, HFmode, SFmode,
+ (arm_fp16_format == ARM_FP16_FORMAT_IEEE
+ ? "__gnu_f2h_ieee"
+ : "__gnu_f2h_alternative"));
+ set_conv_libfunc (sext_optab, SFmode, HFmode,
+ (arm_fp16_format == ARM_FP16_FORMAT_IEEE
+ ? "__gnu_h2f_ieee"
+ : "__gnu_h2f_alternative"));
+
+ /* Arithmetic. */
+ set_optab_libfunc (add_optab, HFmode, NULL);
+ set_optab_libfunc (sdiv_optab, HFmode, NULL);
+ set_optab_libfunc (smul_optab, HFmode, NULL);
+ set_optab_libfunc (neg_optab, HFmode, NULL);
+ set_optab_libfunc (sub_optab, HFmode, NULL);
+
+ /* Comparisons. */
+ set_optab_libfunc (eq_optab, HFmode, NULL);
+ set_optab_libfunc (ne_optab, HFmode, NULL);
+ set_optab_libfunc (lt_optab, HFmode, NULL);
+ set_optab_libfunc (le_optab, HFmode, NULL);
+ set_optab_libfunc (ge_optab, HFmode, NULL);
+ set_optab_libfunc (gt_optab, HFmode, NULL);
+ set_optab_libfunc (unord_optab, HFmode, NULL);
+ break;
+
+ default:
+ break;
+ }
+
+ if (TARGET_AAPCS_BASED)
+ synchronize_libfunc = init_one_libfunc ("__sync_synchronize");
}
/* On AAPCS systems, this is the "struct __va_list". */
gcc_assert (arm_tune != arm_none);
tune_flags = all_cores[(int)arm_tune].flags;
+ current_tune = all_cores[(int)arm_tune].tune;
+
+ if (target_fp16_format_name)
+ {
+ for (i = 0; i < ARRAY_SIZE (all_fp16_formats); i++)
+ {
+ if (streq (all_fp16_formats[i].name, target_fp16_format_name))
+ {
+ arm_fp16_format = all_fp16_formats[i].fp16_format_type;
+ break;
+ }
+ }
+ if (i == ARRAY_SIZE (all_fp16_formats))
+ error ("invalid __fp16 format option: -mfp16-format=%s",
+ target_fp16_format_name);
+ }
+ else
+ arm_fp16_format = ARM_FP16_FORMAT_NONE;
if (target_abi_name)
{
arm_arch6 = (insn_flags & FL_ARCH6) != 0;
arm_arch6k = (insn_flags & FL_ARCH6K) != 0;
arm_arch_notm = (insn_flags & FL_NOTM) != 0;
+ arm_arch7em = (insn_flags & FL_ARCH7EM) != 0;
arm_arch_thumb2 = (insn_flags & FL_THUMB2) != 0;
arm_arch_xscale = (insn_flags & FL_XSCALE) != 0;
arm_arch_cirrus = (insn_flags & FL_CIRRUS) != 0;
if (TARGET_IWMMXT_ABI && !TARGET_IWMMXT)
error ("iwmmxt abi requires an iwmmxt capable cpu");
- arm_fp_model = ARM_FP_MODEL_UNKNOWN;
if (target_fpu_name == NULL && target_fpe_name != NULL)
{
if (streq (target_fpe_name, "2"))
error ("invalid floating point emulation option: -mfpe=%s",
target_fpe_name);
}
- if (target_fpu_name != NULL)
- {
- /* The user specified a FPU. */
- for (i = 0; i < ARRAY_SIZE (all_fpus); i++)
- {
- if (streq (all_fpus[i].name, target_fpu_name))
- {
- arm_fpu_arch = all_fpus[i].fpu;
- arm_fpu_tune = arm_fpu_arch;
- arm_fp_model = fp_model_for_fpu[arm_fpu_arch];
- break;
- }
- }
- if (arm_fp_model == ARM_FP_MODEL_UNKNOWN)
- error ("invalid floating point option: -mfpu=%s", target_fpu_name);
- }
- else
+
+ if (target_fpu_name == NULL)
{
#ifdef FPUTYPE_DEFAULT
- /* Use the default if it is specified for this platform. */
- arm_fpu_arch = FPUTYPE_DEFAULT;
- arm_fpu_tune = FPUTYPE_DEFAULT;
+ target_fpu_name = FPUTYPE_DEFAULT;
#else
- /* Pick one based on CPU type. */
- /* ??? Some targets assume FPA is the default.
- if ((insn_flags & FL_VFP) != 0)
- arm_fpu_arch = FPUTYPE_VFP;
- else
- */
if (arm_arch_cirrus)
- arm_fpu_arch = FPUTYPE_MAVERICK;
+ target_fpu_name = "maverick";
else
- arm_fpu_arch = FPUTYPE_FPA_EMU2;
+ target_fpu_name = "fpe2";
#endif
- if (tune_flags & FL_CO_PROC && arm_fpu_arch == FPUTYPE_FPA_EMU2)
- arm_fpu_tune = FPUTYPE_FPA;
+ }
+
+ arm_fpu_desc = NULL;
+ for (i = 0; i < ARRAY_SIZE (all_fpus); i++)
+ {
+ if (streq (all_fpus[i].name, target_fpu_name))
+ {
+ arm_fpu_desc = &all_fpus[i];
+ break;
+ }
+ }
+
+ if (!arm_fpu_desc)
+ {
+ error ("invalid floating point option: -mfpu=%s", target_fpu_name);
+ return;
+ }
+
+ switch (arm_fpu_desc->model)
+ {
+ case ARM_FP_MODEL_FPA:
+ if (arm_fpu_desc->rev == 2)
+ arm_fpu_attr = FPU_FPE2;
+ else if (arm_fpu_desc->rev == 3)
+ arm_fpu_attr = FPU_FPE3;
else
- arm_fpu_tune = arm_fpu_arch;
- arm_fp_model = fp_model_for_fpu[arm_fpu_arch];
- gcc_assert (arm_fp_model != ARM_FP_MODEL_UNKNOWN);
+ arm_fpu_attr = FPU_FPA;
+ break;
+
+ case ARM_FP_MODEL_MAVERICK:
+ arm_fpu_attr = FPU_MAVERICK;
+ break;
+
+ case ARM_FP_MODEL_VFP:
+ arm_fpu_attr = FPU_VFP;
+ break;
+
+ default:
+ gcc_unreachable();
}
if (target_float_abi_name != NULL)
else
arm_float_abi = TARGET_DEFAULT_FLOAT_ABI;
- if (arm_float_abi == ARM_FLOAT_ABI_HARD && TARGET_VFP)
- sorry ("-mfloat-abi=hard and VFP");
-
if (TARGET_AAPCS_BASED
- && (arm_fp_model == ARM_FP_MODEL_FPA))
+ && (arm_fpu_desc->model == ARM_FP_MODEL_FPA))
error ("FPA is unsupported in the AAPCS");
+ if (TARGET_AAPCS_BASED)
+ {
+ if (TARGET_CALLER_INTERWORKING)
+ error ("AAPCS does not support -mcaller-super-interworking");
+ else
+ if (TARGET_CALLEE_INTERWORKING)
+ error ("AAPCS does not support -mcallee-super-interworking");
+ }
+
/* FPA and iWMMXt are incompatible because the insn encodings overlap.
VFP and iWMMXt can theoretically coexist, but it's unlikely such silicon
will ever exist. GCC makes no attempt to support this combination. */
if (TARGET_THUMB2 && TARGET_IWMMXT)
sorry ("Thumb-2 iWMMXt");
+ /* __fp16 support currently assumes the core has ldrh. */
+ if (!arm_arch4 && arm_fp16_format != ARM_FP16_FORMAT_NONE)
+ sorry ("__fp16 and no ldrh");
+
/* If soft-float is specified then don't use FPU. */
if (TARGET_SOFT_FLOAT)
- arm_fpu_arch = FPUTYPE_NONE;
+ arm_fpu_attr = FPU_NONE;
+
+ if (TARGET_AAPCS_BASED)
+ {
+ if (arm_abi == ARM_ABI_IWMMXT)
+ arm_pcs_default = ARM_PCS_AAPCS_IWMMXT;
+ else if (arm_float_abi == ARM_FLOAT_ABI_HARD
+ && TARGET_HARD_FLOAT
+ && TARGET_VFP)
+ arm_pcs_default = ARM_PCS_AAPCS_VFP;
+ else
+ arm_pcs_default = ARM_PCS_AAPCS;
+ }
+ else
+ {
+ if (arm_float_abi == ARM_FLOAT_ABI_HARD && TARGET_VFP)
+ sorry ("-mfloat-abi=hard and VFP");
+
+ if (arm_abi == ARM_ABI_APCS)
+ arm_pcs_default = ARM_PCS_APCS;
+ else
+ arm_pcs_default = ARM_PCS_ATPCS;
+ }
/* For arm2/3 there is no need to do any scheduling if there is only
a floating point emulator, or we are doing software floating-point. */
if ((TARGET_SOFT_FLOAT
- || arm_fpu_tune == FPUTYPE_FPA_EMU2
- || arm_fpu_tune == FPUTYPE_FPA_EMU3)
+ || (TARGET_FPA && arm_fpu_desc->rev))
&& (tune_flags & FL_MODE32) == 0)
flag_schedule_insns = flag_schedule_insns_after_reload = 0;
/* Use the cp15 method if it is available. */
if (target_thread_pointer == TP_AUTO)
{
- if (arm_arch6k && !TARGET_THUMB)
+ if (arm_arch6k && !TARGET_THUMB1)
target_thread_pointer = TP_CP15;
else
target_thread_pointer = TP_SOFT;
fix_cm3_ldrd = 0;
}
- /* ??? We might want scheduling for thumb2. */
- if (TARGET_THUMB && flag_schedule_insns)
+ if (TARGET_THUMB1 && flag_schedule_insns)
{
/* Don't warn since it's on by default in -O2. */
flag_schedule_insns = 0;
if (optimize_size)
{
- arm_constant_limit = 1;
-
/* If optimizing for size, bump the number of instructions that we
are prepared to conditionally execute (even on a StrongARM). */
max_insns_skipped = 6;
}
else
{
- /* For processors with load scheduling, it never costs more than
- 2 cycles to load a constant, and the load scheduler may well
- reduce that to 1. */
- if (arm_ld_sched)
- arm_constant_limit = 1;
-
- /* On XScale the longer latency of a load makes it more difficult
- to achieve a good schedule, so it's faster to synthesize
- constants that can be done in two insns. */
- if (arm_tune_xscale)
- arm_constant_limit = 2;
-
/* StrongARM has early execution of branches, so a sequence
that is worth skipping is shorter. */
if (arm_tune_strongarm)
max_insns_skipped = 3;
}
+ /* Hot/Cold partitioning is not currently supported, since we can't
+ handle literal pool placement in that case. */
+ if (flag_reorder_blocks_and_partition)
+ {
+ inform (input_location,
+ "-freorder-blocks-and-partition not supported on this architecture");
+ flag_reorder_blocks_and_partition = 0;
+ flag_reorder_blocks = 1;
+ }
+
/* Register global variables with the garbage collector. */
arm_add_gc_roots ();
}
}
\f
+/* Output assembler code for a block containing the constant parts
+ of a trampoline, leaving space for the variable parts.
+
+ On the ARM, (if r8 is the static chain regnum, and remembering that
+ referencing pc adds an offset of 8) the trampoline looks like:
+ ldr r8, [pc, #0]
+ ldr pc, [pc]
+ .word static chain value
+ .word function's address
+ XXX FIXME: When the trampoline returns, r8 will be clobbered. */
+
+static void
+arm_asm_trampoline_template (FILE *f)
+{
+ if (TARGET_ARM)
+ {
+ asm_fprintf (f, "\tldr\t%r, [%r, #0]\n", STATIC_CHAIN_REGNUM, PC_REGNUM);
+ asm_fprintf (f, "\tldr\t%r, [%r, #0]\n", PC_REGNUM, PC_REGNUM);
+ }
+ else if (TARGET_THUMB2)
+ {
+ /* The Thumb-2 trampoline is similar to the arm implementation.
+ Unlike 16-bit Thumb, we enter the stub in thumb mode. */
+ asm_fprintf (f, "\tldr.w\t%r, [%r, #4]\n",
+ STATIC_CHAIN_REGNUM, PC_REGNUM);
+ asm_fprintf (f, "\tldr.w\t%r, [%r, #4]\n", PC_REGNUM, PC_REGNUM);
+ }
+ else
+ {
+ ASM_OUTPUT_ALIGN (f, 2);
+ fprintf (f, "\t.code\t16\n");
+ fprintf (f, ".Ltrampoline_start:\n");
+ asm_fprintf (f, "\tpush\t{r0, r1}\n");
+ asm_fprintf (f, "\tldr\tr0, [%r, #8]\n", PC_REGNUM);
+ asm_fprintf (f, "\tmov\t%r, r0\n", STATIC_CHAIN_REGNUM);
+ asm_fprintf (f, "\tldr\tr0, [%r, #8]\n", PC_REGNUM);
+ asm_fprintf (f, "\tstr\tr0, [%r, #4]\n", SP_REGNUM);
+ asm_fprintf (f, "\tpop\t{r0, %r}\n", PC_REGNUM);
+ }
+ assemble_aligned_integer (UNITS_PER_WORD, const0_rtx);
+ assemble_aligned_integer (UNITS_PER_WORD, const0_rtx);
+}
+
+/* Emit RTL insns to initialize the variable parts of a trampoline. */
+
+static void
+arm_trampoline_init (rtx m_tramp, tree fndecl, rtx chain_value)
+{
+ rtx fnaddr, mem, a_tramp;
+
+ emit_block_move (m_tramp, assemble_trampoline_template (),
+ GEN_INT (TRAMPOLINE_SIZE), BLOCK_OP_NORMAL);
+
+ mem = adjust_address (m_tramp, SImode, TARGET_32BIT ? 8 : 12);
+ emit_move_insn (mem, chain_value);
+
+ mem = adjust_address (m_tramp, SImode, TARGET_32BIT ? 12 : 16);
+ fnaddr = XEXP (DECL_RTL (fndecl), 0);
+ emit_move_insn (mem, fnaddr);
+
+ a_tramp = XEXP (m_tramp, 0);
+ emit_library_call (gen_rtx_SYMBOL_REF (Pmode, "__clear_cache"),
+ LCT_NORMAL, VOIDmode, 2, a_tramp, Pmode,
+ plus_constant (a_tramp, TRAMPOLINE_SIZE), Pmode);
+}
+
+/* Thumb trampolines should be entered in thumb mode, so set
+ the bottom bit of the address. */
+
+static rtx
+arm_trampoline_adjust_address (rtx addr)
+{
+ if (TARGET_THUMB)
+ addr = expand_simple_binop (Pmode, IOR, addr, const1_rtx,
+ NULL, 0, OPTAB_LIB_WIDEN);
+ return addr;
+}
+\f
/* Return 1 if it is possible to return using a single instruction.
If SIBLING is non-null, this is a test for a return before a sibling
call. SIBLING is the call insn, so we can examine its register usage. */
&& !cond
&& (arm_gen_constant (code, mode, NULL_RTX, val, target, source,
1, 0)
- > arm_constant_limit + (code != SET)))
+ > (arm_constant_limit (optimize_function_for_size_p (cfun))
+ + (code != SET))))
{
if (code == SET)
{
1);
}
-/* Return the number of ARM instructions required to synthesize the given
- constant. */
+/* Return the number of instructions required to synthesize the given
+ constant, if we start emitting them from bit-position I. */
static int
count_insns_for_constant (HOST_WIDE_INT remainder, int i)
{
HOST_WIDE_INT temp1;
+ int step_size = TARGET_ARM ? 2 : 1;
int num_insns = 0;
+
+ gcc_assert (TARGET_ARM || i == 0);
+
do
{
int end;
if (i <= 0)
i += 32;
- if (remainder & (3 << (i - 2)))
+ if (remainder & (((1 << step_size) - 1) << (i - step_size)))
{
end = i - 8;
if (end < 0)
| ((i < end) ? (0xff >> (32 - end)) : 0));
remainder &= ~temp1;
num_insns++;
- i -= 6;
+ i -= 8 - step_size;
}
- i -= 2;
+ i -= step_size;
} while (remainder);
return num_insns;
}
+static int
+find_best_start (unsigned HOST_WIDE_INT remainder)
+{
+ int best_consecutive_zeros = 0;
+ int i;
+ int best_start = 0;
+
+ /* If we aren't targetting ARM, the best place to start is always at
+ the bottom. */
+ if (! TARGET_ARM)
+ return 0;
+
+ for (i = 0; i < 32; i += 2)
+ {
+ int consecutive_zeros = 0;
+
+ if (!(remainder & (3 << i)))
+ {
+ while ((i < 32) && !(remainder & (3 << i)))
+ {
+ consecutive_zeros += 2;
+ i += 2;
+ }
+ if (consecutive_zeros > best_consecutive_zeros)
+ {
+ best_consecutive_zeros = consecutive_zeros;
+ best_start = i - consecutive_zeros;
+ }
+ i -= 2;
+ }
+ }
+
+ /* So long as it won't require any more insns to do so, it's
+ desirable to emit a small constant (in bits 0...9) in the last
+ insn. This way there is more chance that it can be combined with
+ a later addressing insn to form a pre-indexed load or store
+ operation. Consider:
+
+ *((volatile int *)0xe0000100) = 1;
+ *((volatile int *)0xe0000110) = 2;
+
+ We want this to wind up as:
+
+ mov rA, #0xe0000000
+ mov rB, #1
+ str rB, [rA, #0x100]
+ mov rB, #2
+ str rB, [rA, #0x110]
+
+ rather than having to synthesize both large constants from scratch.
+
+ Therefore, we calculate how many insns would be required to emit
+ the constant starting from `best_start', and also starting from
+ zero (i.e. with bit 31 first to be output). If `best_start' doesn't
+ yield a shorter sequence, we may as well use zero. */
+ if (best_start != 0
+ && ((((unsigned HOST_WIDE_INT) 1) << best_start) < remainder)
+ && (count_insns_for_constant (remainder, 0) <=
+ count_insns_for_constant (remainder, best_start)))
+ best_start = 0;
+
+ return best_start;
+}
+
/* Emit an instruction with the indicated PATTERN. If COND is
non-NULL, conditionalize the execution of the instruction on COND
being true. */
{
int can_invert = 0;
int can_negate = 0;
+ int final_invert = 0;
int can_negate_initial = 0;
- int can_shift = 0;
int i;
int num_bits_set = 0;
int set_sign_bit_copies = 0;
int insns = 0;
unsigned HOST_WIDE_INT temp1, temp2;
unsigned HOST_WIDE_INT remainder = val & 0xffffffff;
+ int step_size = TARGET_ARM ? 2 : 1;
/* Find out which operations are safe for a given CODE. Also do a quick
check for degenerate cases; these can occur when DImode operations
{
case SET:
can_invert = 1;
- can_shift = 1;
can_negate = 1;
break;
return 1;
}
- /* We don't know how to handle other cases yet. */
- gcc_assert (remainder == 0xffffffff);
-
- if (generate)
- emit_constant_insn (cond,
- gen_rtx_SET (VOIDmode, target,
- gen_rtx_NOT (mode, source)));
- return 1;
+ if (remainder == 0xffffffff)
+ {
+ if (generate)
+ emit_constant_insn (cond,
+ gen_rtx_SET (VOIDmode, target,
+ gen_rtx_NOT (mode, source)));
+ return 1;
+ }
+ break;
case MINUS:
/* We treat MINUS as (val - source), since (source - val) is always
if ((code == AND)
|| (code != IOR && can_invert && num_bits_set > 16))
- remainder = (~remainder) & 0xffffffff;
+ remainder ^= 0xffffffff;
else if (code == PLUS && num_bits_set > 16)
remainder = (-remainder) & 0xffffffff;
+
+ /* For XOR, if more than half the bits are set and there's a sequence
+ of more than 8 consecutive ones in the pattern then we can XOR by the
+ inverted constant and then invert the final result; this may save an
+ instruction and might also lead to the final mvn being merged with
+ some other operation. */
+ else if (code == XOR && num_bits_set > 16
+ && (count_insns_for_constant (remainder ^ 0xffffffff,
+ find_best_start
+ (remainder ^ 0xffffffff))
+ < count_insns_for_constant (remainder,
+ find_best_start (remainder))))
+ {
+ remainder ^= 0xffffffff;
+ final_invert = 1;
+ }
else
{
can_invert = 0;
/* ??? Use thumb2 replicated constants when the high and low halfwords are
the same. */
{
- int best_start = 0;
- if (!TARGET_THUMB2)
- {
- int best_consecutive_zeros = 0;
-
- for (i = 0; i < 32; i += 2)
- {
- int consecutive_zeros = 0;
-
- if (!(remainder & (3 << i)))
- {
- while ((i < 32) && !(remainder & (3 << i)))
- {
- consecutive_zeros += 2;
- i += 2;
- }
- if (consecutive_zeros > best_consecutive_zeros)
- {
- best_consecutive_zeros = consecutive_zeros;
- best_start = i - consecutive_zeros;
- }
- i -= 2;
- }
- }
-
- /* So long as it won't require any more insns to do so, it's
- desirable to emit a small constant (in bits 0...9) in the last
- insn. This way there is more chance that it can be combined with
- a later addressing insn to form a pre-indexed load or store
- operation. Consider:
-
- *((volatile int *)0xe0000100) = 1;
- *((volatile int *)0xe0000110) = 2;
-
- We want this to wind up as:
-
- mov rA, #0xe0000000
- mov rB, #1
- str rB, [rA, #0x100]
- mov rB, #2
- str rB, [rA, #0x110]
-
- rather than having to synthesize both large constants from scratch.
-
- Therefore, we calculate how many insns would be required to emit
- the constant starting from `best_start', and also starting from
- zero (i.e. with bit 31 first to be output). If `best_start' doesn't
- yield a shorter sequence, we may as well use zero. */
- if (best_start != 0
- && ((((unsigned HOST_WIDE_INT) 1) << best_start) < remainder)
- && (count_insns_for_constant (remainder, 0) <=
- count_insns_for_constant (remainder, best_start)))
- best_start = 0;
- }
-
/* Now start emitting the insns. */
- i = best_start;
+ i = find_best_start (remainder);
do
{
int end;
}
else
{
- if (remainder && subtargets)
+ if ((final_invert || remainder) && subtargets)
new_src = gen_reg_rtx (mode);
else
new_src = target;
code = PLUS;
insns++;
- if (TARGET_ARM)
- i -= 6;
- else
- i -= 7;
+ i -= 8 - step_size;
}
/* Arm allows rotates by a multiple of two. Thumb-2 allows arbitrary
shifts. */
- if (TARGET_ARM)
- i -= 2;
- else
- i--;
+ i -= step_size;
}
while (remainder);
}
+ if (final_invert)
+ {
+ if (generate)
+ emit_constant_insn (cond, gen_rtx_SET (VOIDmode, target,
+ gen_rtx_NOT (mode, source)));
+ insns++;
+ }
+
return insns;
}
/* Define how to find the value returned by a function. */
-rtx
-arm_function_value(const_tree type, const_tree func ATTRIBUTE_UNUSED)
+static rtx
+arm_function_value(const_tree type, const_tree func,
+ bool outgoing ATTRIBUTE_UNUSED)
{
enum machine_mode mode;
int unsignedp ATTRIBUTE_UNUSED;
rtx r ATTRIBUTE_UNUSED;
mode = TYPE_MODE (type);
+
+ if (TARGET_AAPCS_BASED)
+ return aapcs_allocate_return_reg (mode, type, func);
+
/* Promote integer types. */
if (INTEGRAL_TYPE_P (type))
- PROMOTE_FUNCTION_MODE (mode, unsignedp, type);
+ mode = arm_promote_function_mode (type, mode, &unsignedp, func, 1);
/* Promotes small structs returned in a register to full-word size
for big-endian AAPCS. */
}
}
- return LIBCALL_VALUE(mode);
+ return LIBCALL_VALUE (mode);
+}
+
+static int
+libcall_eq (const void *p1, const void *p2)
+{
+ return rtx_equal_p ((const_rtx) p1, (const_rtx) p2);
+}
+
+static hashval_t
+libcall_hash (const void *p1)
+{
+ return hash_rtx ((const_rtx) p1, VOIDmode, NULL, NULL, FALSE);
+}
+
+static void
+add_libcall (htab_t htab, rtx libcall)
+{
+ *htab_find_slot (htab, libcall, INSERT) = libcall;
+}
+
+static bool
+arm_libcall_uses_aapcs_base (const_rtx libcall)
+{
+ static bool init_done = false;
+ static htab_t libcall_htab;
+
+ if (!init_done)
+ {
+ init_done = true;
+
+ libcall_htab = htab_create (31, libcall_hash, libcall_eq,
+ NULL);
+ add_libcall (libcall_htab,
+ convert_optab_libfunc (sfloat_optab, SFmode, SImode));
+ add_libcall (libcall_htab,
+ convert_optab_libfunc (sfloat_optab, DFmode, SImode));
+ add_libcall (libcall_htab,
+ convert_optab_libfunc (sfloat_optab, SFmode, DImode));
+ add_libcall (libcall_htab,
+ convert_optab_libfunc (sfloat_optab, DFmode, DImode));
+
+ add_libcall (libcall_htab,
+ convert_optab_libfunc (ufloat_optab, SFmode, SImode));
+ add_libcall (libcall_htab,
+ convert_optab_libfunc (ufloat_optab, DFmode, SImode));
+ add_libcall (libcall_htab,
+ convert_optab_libfunc (ufloat_optab, SFmode, DImode));
+ add_libcall (libcall_htab,
+ convert_optab_libfunc (ufloat_optab, DFmode, DImode));
+
+ add_libcall (libcall_htab,
+ convert_optab_libfunc (sext_optab, SFmode, HFmode));
+ add_libcall (libcall_htab,
+ convert_optab_libfunc (trunc_optab, HFmode, SFmode));
+ add_libcall (libcall_htab,
+ convert_optab_libfunc (sfix_optab, DImode, DFmode));
+ add_libcall (libcall_htab,
+ convert_optab_libfunc (ufix_optab, DImode, DFmode));
+ add_libcall (libcall_htab,
+ convert_optab_libfunc (sfix_optab, DImode, SFmode));
+ add_libcall (libcall_htab,
+ convert_optab_libfunc (ufix_optab, DImode, SFmode));
+ }
+
+ return libcall && htab_find (libcall_htab, libcall) != NULL;
+}
+
+rtx
+arm_libcall_value (enum machine_mode mode, const_rtx libcall)
+{
+ if (TARGET_AAPCS_BASED && arm_pcs_default != ARM_PCS_AAPCS
+ && GET_MODE_CLASS (mode) == MODE_FLOAT)
+ {
+ /* The following libcalls return their result in integer registers,
+ even though they return a floating point value. */
+ if (arm_libcall_uses_aapcs_base (libcall))
+ return gen_rtx_REG (mode, ARG_REGISTER(1));
+
+ }
+
+ return LIBCALL_VALUE (mode);
}
/* Determine the amount of memory needed to store the possible return
{
int size = 16;
- if (TARGET_ARM)
+ if (TARGET_32BIT)
{
if (TARGET_HARD_FLOAT_ABI)
{
+ if (TARGET_VFP)
+ size += 32;
if (TARGET_FPA)
size += 12;
if (TARGET_MAVERICK)
return size;
}
-/* Decide whether a type should be returned in memory (true)
- or in a register (false). This is called as the target hook
- TARGET_RETURN_IN_MEMORY. */
+/* Decide whether TYPE should be returned in memory (true)
+ or in a register (false). FNTYPE is the type of the function making
+ the call. */
static bool
-arm_return_in_memory (const_tree type, const_tree fntype ATTRIBUTE_UNUSED)
+arm_return_in_memory (const_tree type, const_tree fntype)
{
HOST_WIDE_INT size;
- size = int_size_in_bytes (type);
+ size = int_size_in_bytes (type); /* Negative if not fixed size. */
+
+ if (TARGET_AAPCS_BASED)
+ {
+ /* Simple, non-aggregate types (ie not including vectors and
+ complex) are always returned in a register (or registers).
+ We don't care about which register here, so we can short-cut
+ some of the detail. */
+ if (!AGGREGATE_TYPE_P (type)
+ && TREE_CODE (type) != VECTOR_TYPE
+ && TREE_CODE (type) != COMPLEX_TYPE)
+ return false;
+
+ /* Any return value that is no larger than one word can be
+ returned in r0. */
+ if (((unsigned HOST_WIDE_INT) size) <= UNITS_PER_WORD)
+ return false;
+
+ /* Check any available co-processors to see if they accept the
+ type as a register candidate (VFP, for example, can return
+ some aggregates in consecutive registers). These aren't
+ available if the call is variadic. */
+ if (aapcs_select_return_coproc (type, fntype) >= 0)
+ return false;
+
+ /* Vector values should be returned using ARM registers, not
+ memory (unless they're over 16 bytes, which will break since
+ we only have four call-clobbered registers to play with). */
+ if (TREE_CODE (type) == VECTOR_TYPE)
+ return (size < 0 || size > (4 * UNITS_PER_WORD));
+
+ /* The rest go in memory. */
+ return true;
+ }
- /* Vector values should be returned using ARM registers, not memory (unless
- they're over 16 bytes, which will break since we only have four
- call-clobbered registers to play with). */
if (TREE_CODE (type) == VECTOR_TYPE)
return (size < 0 || size > (4 * UNITS_PER_WORD));
if (!AGGREGATE_TYPE_P (type) &&
- !(TARGET_AAPCS_BASED && TREE_CODE (type) == COMPLEX_TYPE))
- /* All simple types are returned in registers.
- For AAPCS, complex types are treated the same as aggregates. */
- return 0;
+ (TREE_CODE (type) != VECTOR_TYPE))
+ /* All simple types are returned in registers. */
+ return false;
if (arm_abi != ARM_ABI_APCS)
{
the aggregate is either huge or of variable size, and in either case
we will want to return it via memory and not in a register. */
if (size < 0 || size > UNITS_PER_WORD)
- return 1;
+ return true;
if (TREE_CODE (type) == RECORD_TYPE)
{
continue;
if (field == NULL)
- return 0; /* An empty structure. Allowed by an extension to ANSI C. */
+ return false; /* An empty structure. Allowed by an extension to ANSI C. */
/* Check that the first field is valid for returning in a register. */
/* ... Floats are not allowed */
if (FLOAT_TYPE_P (TREE_TYPE (field)))
- return 1;
+ return true;
/* ... Aggregates that are not themselves valid for returning in
a register are not allowed. */
if (arm_return_in_memory (TREE_TYPE (field), NULL_TREE))
- return 1;
+ return true;
/* Now check the remaining fields, if any. Only bitfields are allowed,
since they are not addressable. */
continue;
if (!DECL_BIT_FIELD_TYPE (field))
- return 1;
+ return true;
}
- return 0;
+ return false;
}
if (TREE_CODE (type) == UNION_TYPE)
continue;
if (FLOAT_TYPE_P (TREE_TYPE (field)))
- return 1;
+ return true;
+
+ if (arm_return_in_memory (TREE_TYPE (field), NULL_TREE))
+ return true;
+ }
+
+ return false;
+ }
+#endif /* not ARM_WINCE */
+
+ /* Return all other types in memory. */
+ return true;
+}
+
+/* Indicate whether or not words of a double are in big-endian order. */
+
+int
+arm_float_words_big_endian (void)
+{
+ if (TARGET_MAVERICK)
+ return 0;
+
+ /* For FPA, float words are always big-endian. For VFP, floats words
+ follow the memory system mode. */
+
+ if (TARGET_FPA)
+ {
+ return 1;
+ }
+
+ if (TARGET_VFP)
+ return (TARGET_BIG_END ? 1 : 0);
+
+ return 1;
+}
+
+const struct pcs_attribute_arg
+{
+ const char *arg;
+ enum arm_pcs value;
+} pcs_attribute_args[] =
+ {
+ {"aapcs", ARM_PCS_AAPCS},
+ {"aapcs-vfp", ARM_PCS_AAPCS_VFP},
+#if 0
+ /* We could recognize these, but changes would be needed elsewhere
+ * to implement them. */
+ {"aapcs-iwmmxt", ARM_PCS_AAPCS_IWMMXT},
+ {"atpcs", ARM_PCS_ATPCS},
+ {"apcs", ARM_PCS_APCS},
+#endif
+ {NULL, ARM_PCS_UNKNOWN}
+ };
+
+static enum arm_pcs
+arm_pcs_from_attribute (tree attr)
+{
+ const struct pcs_attribute_arg *ptr;
+ const char *arg;
+
+ /* Get the value of the argument. */
+ if (TREE_VALUE (attr) == NULL_TREE
+ || TREE_CODE (TREE_VALUE (attr)) != STRING_CST)
+ return ARM_PCS_UNKNOWN;
+
+ arg = TREE_STRING_POINTER (TREE_VALUE (attr));
+
+ /* Check it against the list of known arguments. */
+ for (ptr = pcs_attribute_args; ptr->arg != NULL; ptr++)
+ if (streq (arg, ptr->arg))
+ return ptr->value;
+
+ /* An unrecognized interrupt type. */
+ return ARM_PCS_UNKNOWN;
+}
+
+/* Get the PCS variant to use for this call. TYPE is the function's type
+ specification, DECL is the specific declartion. DECL may be null if
+ the call could be indirect or if this is a library call. */
+static enum arm_pcs
+arm_get_pcs_model (const_tree type, const_tree decl)
+{
+ bool user_convention = false;
+ enum arm_pcs user_pcs = arm_pcs_default;
+ tree attr;
+
+ gcc_assert (type);
+
+ attr = lookup_attribute ("pcs", TYPE_ATTRIBUTES (type));
+ if (attr)
+ {
+ user_pcs = arm_pcs_from_attribute (TREE_VALUE (attr));
+ user_convention = true;
+ }
+
+ if (TARGET_AAPCS_BASED)
+ {
+ /* Detect varargs functions. These always use the base rules
+ (no argument is ever a candidate for a co-processor
+ register). */
+ bool base_rules = (TYPE_ARG_TYPES (type) != 0
+ && (TREE_VALUE (tree_last (TYPE_ARG_TYPES (type)))
+ != void_type_node));
+
+ if (user_convention)
+ {
+ if (user_pcs > ARM_PCS_AAPCS_LOCAL)
+ sorry ("Non-AAPCS derived PCS variant");
+ else if (base_rules && user_pcs != ARM_PCS_AAPCS)
+ error ("Variadic functions must use the base AAPCS variant");
+ }
+
+ if (base_rules)
+ return ARM_PCS_AAPCS;
+ else if (user_convention)
+ return user_pcs;
+ else if (decl && flag_unit_at_a_time)
+ {
+ /* Local functions never leak outside this compilation unit,
+ so we are free to use whatever conventions are
+ appropriate. */
+ /* FIXME: remove CONST_CAST_TREE when cgraph is constified. */
+ struct cgraph_local_info *i = cgraph_local_info (CONST_CAST_TREE(decl));
+ if (i && i->local)
+ return ARM_PCS_AAPCS_LOCAL;
+ }
+ }
+ else if (user_convention && user_pcs != arm_pcs_default)
+ sorry ("PCS variant");
+
+ /* For everything else we use the target's default. */
+ return arm_pcs_default;
+}
+
+
+static void
+aapcs_vfp_cum_init (CUMULATIVE_ARGS *pcum ATTRIBUTE_UNUSED,
+ const_tree fntype ATTRIBUTE_UNUSED,
+ rtx libcall ATTRIBUTE_UNUSED,
+ const_tree fndecl ATTRIBUTE_UNUSED)
+{
+ /* Record the unallocated VFP registers. */
+ pcum->aapcs_vfp_regs_free = (1 << NUM_VFP_ARG_REGS) - 1;
+ pcum->aapcs_vfp_reg_alloc = 0;
+}
+
+/* Walk down the type tree of TYPE counting consecutive base elements.
+ If *MODEP is VOIDmode, then set it to the first valid floating point
+ type. If a non-floating point type is found, or if a floating point
+ type that doesn't match a non-VOIDmode *MODEP is found, then return -1,
+ otherwise return the count in the sub-tree. */
+static int
+aapcs_vfp_sub_candidate (const_tree type, enum machine_mode *modep)
+{
+ enum machine_mode mode;
+ HOST_WIDE_INT size;
+
+ switch (TREE_CODE (type))
+ {
+ case REAL_TYPE:
+ mode = TYPE_MODE (type);
+ if (mode != DFmode && mode != SFmode)
+ return -1;
+
+ if (*modep == VOIDmode)
+ *modep = mode;
+
+ if (*modep == mode)
+ return 1;
+
+ break;
+
+ case COMPLEX_TYPE:
+ mode = TYPE_MODE (TREE_TYPE (type));
+ if (mode != DFmode && mode != SFmode)
+ return -1;
+
+ if (*modep == VOIDmode)
+ *modep = mode;
+
+ if (*modep == mode)
+ return 2;
+
+ break;
+
+ case VECTOR_TYPE:
+ /* Use V2SImode and V4SImode as representatives of all 64-bit
+ and 128-bit vector types, whether or not those modes are
+ supported with the present options. */
+ size = int_size_in_bytes (type);
+ switch (size)
+ {
+ case 8:
+ mode = V2SImode;
+ break;
+ case 16:
+ mode = V4SImode;
+ break;
+ default:
+ return -1;
+ }
+
+ if (*modep == VOIDmode)
+ *modep = mode;
+
+ /* Vector modes are considered to be opaque: two vectors are
+ equivalent for the purposes of being homogeneous aggregates
+ if they are the same size. */
+ if (*modep == mode)
+ return 1;
+
+ break;
+
+ case ARRAY_TYPE:
+ {
+ int count;
+ tree index = TYPE_DOMAIN (type);
+
+ /* Can't handle incomplete types. */
+ if (!COMPLETE_TYPE_P(type))
+ return -1;
+
+ count = aapcs_vfp_sub_candidate (TREE_TYPE (type), modep);
+ if (count == -1
+ || !index
+ || !TYPE_MAX_VALUE (index)
+ || !host_integerp (TYPE_MAX_VALUE (index), 1)
+ || !TYPE_MIN_VALUE (index)
+ || !host_integerp (TYPE_MIN_VALUE (index), 1)
+ || count < 0)
+ return -1;
+
+ count *= (1 + tree_low_cst (TYPE_MAX_VALUE (index), 1)
+ - tree_low_cst (TYPE_MIN_VALUE (index), 1));
+
+ /* There must be no padding. */
+ if (!host_integerp (TYPE_SIZE (type), 1)
+ || (tree_low_cst (TYPE_SIZE (type), 1)
+ != count * GET_MODE_BITSIZE (*modep)))
+ return -1;
+
+ return count;
+ }
+
+ case RECORD_TYPE:
+ {
+ int count = 0;
+ int sub_count;
+ tree field;
+
+ /* Can't handle incomplete types. */
+ if (!COMPLETE_TYPE_P(type))
+ return -1;
+
+ for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
+ {
+ if (TREE_CODE (field) != FIELD_DECL)
+ continue;
+
+ sub_count = aapcs_vfp_sub_candidate (TREE_TYPE (field), modep);
+ if (sub_count < 0)
+ return -1;
+ count += sub_count;
+ }
+
+ /* There must be no padding. */
+ if (!host_integerp (TYPE_SIZE (type), 1)
+ || (tree_low_cst (TYPE_SIZE (type), 1)
+ != count * GET_MODE_BITSIZE (*modep)))
+ return -1;
+
+ return count;
+ }
+
+ case UNION_TYPE:
+ case QUAL_UNION_TYPE:
+ {
+ /* These aren't very interesting except in a degenerate case. */
+ int count = 0;
+ int sub_count;
+ tree field;
+
+ /* Can't handle incomplete types. */
+ if (!COMPLETE_TYPE_P(type))
+ return -1;
+
+ for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
+ {
+ if (TREE_CODE (field) != FIELD_DECL)
+ continue;
+
+ sub_count = aapcs_vfp_sub_candidate (TREE_TYPE (field), modep);
+ if (sub_count < 0)
+ return -1;
+ count = count > sub_count ? count : sub_count;
+ }
+
+ /* There must be no padding. */
+ if (!host_integerp (TYPE_SIZE (type), 1)
+ || (tree_low_cst (TYPE_SIZE (type), 1)
+ != count * GET_MODE_BITSIZE (*modep)))
+ return -1;
+
+ return count;
+ }
+
+ default:
+ break;
+ }
+
+ return -1;
+}
+
+/* Return true if PCS_VARIANT should use VFP registers. */
+static bool
+use_vfp_abi (enum arm_pcs pcs_variant, bool is_double)
+{
+ if (pcs_variant == ARM_PCS_AAPCS_VFP)
+ return true;
+
+ if (pcs_variant != ARM_PCS_AAPCS_LOCAL)
+ return false;
+
+ return (TARGET_32BIT && TARGET_VFP && TARGET_HARD_FLOAT &&
+ (TARGET_VFP_DOUBLE || !is_double));
+}
+
+static bool
+aapcs_vfp_is_call_or_return_candidate (enum arm_pcs pcs_variant,
+ enum machine_mode mode, const_tree type,
+ enum machine_mode *base_mode, int *count)
+{
+ enum machine_mode new_mode = VOIDmode;
+
+ if (GET_MODE_CLASS (mode) == MODE_FLOAT
+ || GET_MODE_CLASS (mode) == MODE_VECTOR_INT
+ || GET_MODE_CLASS (mode) == MODE_VECTOR_FLOAT)
+ {
+ *count = 1;
+ new_mode = mode;
+ }
+ else if (GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT)
+ {
+ *count = 2;
+ new_mode = (mode == DCmode ? DFmode : SFmode);
+ }
+ else if (type && (mode == BLKmode || TREE_CODE (type) == VECTOR_TYPE))
+ {
+ int ag_count = aapcs_vfp_sub_candidate (type, &new_mode);
+
+ if (ag_count > 0 && ag_count <= 4)
+ *count = ag_count;
+ else
+ return false;
+ }
+ else
+ return false;
+
+
+ if (!use_vfp_abi (pcs_variant, ARM_NUM_REGS (new_mode) > 1))
+ return false;
+
+ *base_mode = new_mode;
+ return true;
+}
+
+static bool
+aapcs_vfp_is_return_candidate (enum arm_pcs pcs_variant,
+ enum machine_mode mode, const_tree type)
+{
+ int count ATTRIBUTE_UNUSED;
+ enum machine_mode ag_mode ATTRIBUTE_UNUSED;
+
+ if (!use_vfp_abi (pcs_variant, false))
+ return false;
+ return aapcs_vfp_is_call_or_return_candidate (pcs_variant, mode, type,
+ &ag_mode, &count);
+}
+
+static bool
+aapcs_vfp_is_call_candidate (CUMULATIVE_ARGS *pcum, enum machine_mode mode,
+ const_tree type)
+{
+ if (!use_vfp_abi (pcum->pcs_variant, false))
+ return false;
+
+ return aapcs_vfp_is_call_or_return_candidate (pcum->pcs_variant, mode, type,
+ &pcum->aapcs_vfp_rmode,
+ &pcum->aapcs_vfp_rcount);
+}
+
+static bool
+aapcs_vfp_allocate (CUMULATIVE_ARGS *pcum, enum machine_mode mode,
+ const_tree type ATTRIBUTE_UNUSED)
+{
+ int shift = GET_MODE_SIZE (pcum->aapcs_vfp_rmode) / GET_MODE_SIZE (SFmode);
+ unsigned mask = (1 << (shift * pcum->aapcs_vfp_rcount)) - 1;
+ int regno;
+
+ for (regno = 0; regno < NUM_VFP_ARG_REGS; regno += shift)
+ if (((pcum->aapcs_vfp_regs_free >> regno) & mask) == mask)
+ {
+ pcum->aapcs_vfp_reg_alloc = mask << regno;
+ if (mode == BLKmode || (mode == TImode && !TARGET_NEON))
+ {
+ int i;
+ int rcount = pcum->aapcs_vfp_rcount;
+ int rshift = shift;
+ enum machine_mode rmode = pcum->aapcs_vfp_rmode;
+ rtx par;
+ if (!TARGET_NEON)
+ {
+ /* Avoid using unsupported vector modes. */
+ if (rmode == V2SImode)
+ rmode = DImode;
+ else if (rmode == V4SImode)
+ {
+ rmode = DImode;
+ rcount *= 2;
+ rshift /= 2;
+ }
+ }
+ par = gen_rtx_PARALLEL (mode, rtvec_alloc (rcount));
+ for (i = 0; i < rcount; i++)
+ {
+ rtx tmp = gen_rtx_REG (rmode,
+ FIRST_VFP_REGNUM + regno + i * rshift);
+ tmp = gen_rtx_EXPR_LIST
+ (VOIDmode, tmp,
+ GEN_INT (i * GET_MODE_SIZE (rmode)));
+ XVECEXP (par, 0, i) = tmp;
+ }
+
+ pcum->aapcs_reg = par;
+ }
+ else
+ pcum->aapcs_reg = gen_rtx_REG (mode, FIRST_VFP_REGNUM + regno);
+ return true;
+ }
+ return false;
+}
+
+static rtx
+aapcs_vfp_allocate_return_reg (enum arm_pcs pcs_variant ATTRIBUTE_UNUSED,
+ enum machine_mode mode,
+ const_tree type ATTRIBUTE_UNUSED)
+{
+ if (!use_vfp_abi (pcs_variant, false))
+ return false;
+
+ if (mode == BLKmode || (mode == TImode && !TARGET_NEON))
+ {
+ int count;
+ enum machine_mode ag_mode;
+ int i;
+ rtx par;
+ int shift;
+
+ aapcs_vfp_is_call_or_return_candidate (pcs_variant, mode, type,
+ &ag_mode, &count);
+
+ if (!TARGET_NEON)
+ {
+ if (ag_mode == V2SImode)
+ ag_mode = DImode;
+ else if (ag_mode == V4SImode)
+ {
+ ag_mode = DImode;
+ count *= 2;
+ }
+ }
+ shift = GET_MODE_SIZE(ag_mode) / GET_MODE_SIZE(SFmode);
+ par = gen_rtx_PARALLEL (mode, rtvec_alloc (count));
+ for (i = 0; i < count; i++)
+ {
+ rtx tmp = gen_rtx_REG (ag_mode, FIRST_VFP_REGNUM + i * shift);
+ tmp = gen_rtx_EXPR_LIST (VOIDmode, tmp,
+ GEN_INT (i * GET_MODE_SIZE (ag_mode)));
+ XVECEXP (par, 0, i) = tmp;
+ }
+
+ return par;
+ }
+
+ return gen_rtx_REG (mode, FIRST_VFP_REGNUM);
+}
+
+static void
+aapcs_vfp_advance (CUMULATIVE_ARGS *pcum ATTRIBUTE_UNUSED,
+ enum machine_mode mode ATTRIBUTE_UNUSED,
+ const_tree type ATTRIBUTE_UNUSED)
+{
+ pcum->aapcs_vfp_regs_free &= ~pcum->aapcs_vfp_reg_alloc;
+ pcum->aapcs_vfp_reg_alloc = 0;
+ return;
+}
+
+#define AAPCS_CP(X) \
+ { \
+ aapcs_ ## X ## _cum_init, \
+ aapcs_ ## X ## _is_call_candidate, \
+ aapcs_ ## X ## _allocate, \
+ aapcs_ ## X ## _is_return_candidate, \
+ aapcs_ ## X ## _allocate_return_reg, \
+ aapcs_ ## X ## _advance \
+ }
+
+/* Table of co-processors that can be used to pass arguments in
+ registers. Idealy no arugment should be a candidate for more than
+ one co-processor table entry, but the table is processed in order
+ and stops after the first match. If that entry then fails to put
+ the argument into a co-processor register, the argument will go on
+ the stack. */
+static struct
+{
+ /* Initialize co-processor related state in CUMULATIVE_ARGS structure. */
+ void (*cum_init) (CUMULATIVE_ARGS *, const_tree, rtx, const_tree);
+
+ /* Return true if an argument of mode MODE (or type TYPE if MODE is
+ BLKmode) is a candidate for this co-processor's registers; this
+ function should ignore any position-dependent state in
+ CUMULATIVE_ARGS and only use call-type dependent information. */
+ bool (*is_call_candidate) (CUMULATIVE_ARGS *, enum machine_mode, const_tree);
+
+ /* Return true if the argument does get a co-processor register; it
+ should set aapcs_reg to an RTX of the register allocated as is
+ required for a return from FUNCTION_ARG. */
+ bool (*allocate) (CUMULATIVE_ARGS *, enum machine_mode, const_tree);
+
+ /* Return true if a result of mode MODE (or type TYPE if MODE is
+ BLKmode) is can be returned in this co-processor's registers. */
+ bool (*is_return_candidate) (enum arm_pcs, enum machine_mode, const_tree);
+
+ /* Allocate and return an RTX element to hold the return type of a
+ call, this routine must not fail and will only be called if
+ is_return_candidate returned true with the same parameters. */
+ rtx (*allocate_return_reg) (enum arm_pcs, enum machine_mode, const_tree);
+
+ /* Finish processing this argument and prepare to start processing
+ the next one. */
+ void (*advance) (CUMULATIVE_ARGS *, enum machine_mode, const_tree);
+} aapcs_cp_arg_layout[ARM_NUM_COPROC_SLOTS] =
+ {
+ AAPCS_CP(vfp)
+ };
+
+#undef AAPCS_CP
+
+static int
+aapcs_select_call_coproc (CUMULATIVE_ARGS *pcum, enum machine_mode mode,
+ tree type)
+{
+ int i;
+
+ for (i = 0; i < ARM_NUM_COPROC_SLOTS; i++)
+ if (aapcs_cp_arg_layout[i].is_call_candidate (pcum, mode, type))
+ return i;
+
+ return -1;
+}
+
+static int
+aapcs_select_return_coproc (const_tree type, const_tree fntype)
+{
+ /* We aren't passed a decl, so we can't check that a call is local.
+ However, it isn't clear that that would be a win anyway, since it
+ might limit some tail-calling opportunities. */
+ enum arm_pcs pcs_variant;
+
+ if (fntype)
+ {
+ const_tree fndecl = NULL_TREE;
+
+ if (TREE_CODE (fntype) == FUNCTION_DECL)
+ {
+ fndecl = fntype;
+ fntype = TREE_TYPE (fntype);
+ }
+
+ pcs_variant = arm_get_pcs_model (fntype, fndecl);
+ }
+ else
+ pcs_variant = arm_pcs_default;
+
+ if (pcs_variant != ARM_PCS_AAPCS)
+ {
+ int i;
+
+ for (i = 0; i < ARM_NUM_COPROC_SLOTS; i++)
+ if (aapcs_cp_arg_layout[i].is_return_candidate (pcs_variant,
+ TYPE_MODE (type),
+ type))
+ return i;
+ }
+ return -1;
+}
+
+static rtx
+aapcs_allocate_return_reg (enum machine_mode mode, const_tree type,
+ const_tree fntype)
+{
+ /* We aren't passed a decl, so we can't check that a call is local.
+ However, it isn't clear that that would be a win anyway, since it
+ might limit some tail-calling opportunities. */
+ enum arm_pcs pcs_variant;
+ int unsignedp ATTRIBUTE_UNUSED;
+
+ if (fntype)
+ {
+ const_tree fndecl = NULL_TREE;
+
+ if (TREE_CODE (fntype) == FUNCTION_DECL)
+ {
+ fndecl = fntype;
+ fntype = TREE_TYPE (fntype);
+ }
+
+ pcs_variant = arm_get_pcs_model (fntype, fndecl);
+ }
+ else
+ pcs_variant = arm_pcs_default;
+
+ /* Promote integer types. */
+ if (type && INTEGRAL_TYPE_P (type))
+ mode = arm_promote_function_mode (type, mode, &unsignedp, fntype, 1);
+
+ if (pcs_variant != ARM_PCS_AAPCS)
+ {
+ int i;
+
+ for (i = 0; i < ARM_NUM_COPROC_SLOTS; i++)
+ if (aapcs_cp_arg_layout[i].is_return_candidate (pcs_variant, mode,
+ type))
+ return aapcs_cp_arg_layout[i].allocate_return_reg (pcs_variant,
+ mode, type);
+ }
+
+ /* Promotes small structs returned in a register to full-word size
+ for big-endian AAPCS. */
+ if (type && arm_return_in_msb (type))
+ {
+ HOST_WIDE_INT size = int_size_in_bytes (type);
+ if (size % UNITS_PER_WORD != 0)
+ {
+ size += UNITS_PER_WORD - size % UNITS_PER_WORD;
+ mode = mode_for_size (size * BITS_PER_UNIT, MODE_INT, 0);
+ }
+ }
+
+ return gen_rtx_REG (mode, R0_REGNUM);
+}
+
+rtx
+aapcs_libcall_value (enum machine_mode mode)
+{
+ return aapcs_allocate_return_reg (mode, NULL_TREE, NULL_TREE);
+}
+
+/* Lay out a function argument using the AAPCS rules. The rule
+ numbers referred to here are those in the AAPCS. */
+static void
+aapcs_layout_arg (CUMULATIVE_ARGS *pcum, enum machine_mode mode,
+ tree type, int named)
+{
+ int nregs, nregs2;
+ int ncrn;
+
+ /* We only need to do this once per argument. */
+ if (pcum->aapcs_arg_processed)
+ return;
+
+ pcum->aapcs_arg_processed = true;
+
+ /* Special case: if named is false then we are handling an incoming
+ anonymous argument which is on the stack. */
+ if (!named)
+ return;
+
+ /* Is this a potential co-processor register candidate? */
+ if (pcum->pcs_variant != ARM_PCS_AAPCS)
+ {
+ int slot = aapcs_select_call_coproc (pcum, mode, type);
+ pcum->aapcs_cprc_slot = slot;
+
+ /* We don't have to apply any of the rules from part B of the
+ preparation phase, these are handled elsewhere in the
+ compiler. */
+
+ if (slot >= 0)
+ {
+ /* A Co-processor register candidate goes either in its own
+ class of registers or on the stack. */
+ if (!pcum->aapcs_cprc_failed[slot])
+ {
+ /* C1.cp - Try to allocate the argument to co-processor
+ registers. */
+ if (aapcs_cp_arg_layout[slot].allocate (pcum, mode, type))
+ return;
+
+ /* C2.cp - Put the argument on the stack and note that we
+ can't assign any more candidates in this slot. We also
+ need to note that we have allocated stack space, so that
+ we won't later try to split a non-cprc candidate between
+ core registers and the stack. */
+ pcum->aapcs_cprc_failed[slot] = true;
+ pcum->can_split = false;
+ }
- if (arm_return_in_memory (TREE_TYPE (field), NULL_TREE))
- return 1;
+ /* We didn't get a register, so this argument goes on the
+ stack. */
+ gcc_assert (pcum->can_split == false);
+ return;
}
-
- return 0;
}
-#endif /* not ARM_WINCE */
- /* Return all other types in memory. */
- return 1;
-}
+ /* C3 - For double-word aligned arguments, round the NCRN up to the
+ next even number. */
+ ncrn = pcum->aapcs_ncrn;
+ if ((ncrn & 1) && arm_needs_doubleword_align (mode, type))
+ ncrn++;
-/* Indicate whether or not words of a double are in big-endian order. */
+ nregs = ARM_NUM_REGS2(mode, type);
-int
-arm_float_words_big_endian (void)
-{
- if (TARGET_MAVERICK)
- return 0;
+ /* Sigh, this test should really assert that nregs > 0, but a GCC
+ extension allows empty structs and then gives them empty size; it
+ then allows such a structure to be passed by value. For some of
+ the code below we have to pretend that such an argument has
+ non-zero size so that we 'locate' it correctly either in
+ registers or on the stack. */
+ gcc_assert (nregs >= 0);
- /* For FPA, float words are always big-endian. For VFP, floats words
- follow the memory system mode. */
+ nregs2 = nregs ? nregs : 1;
- if (TARGET_FPA)
+ /* C4 - Argument fits entirely in core registers. */
+ if (ncrn + nregs2 <= NUM_ARG_REGS)
{
- return 1;
+ pcum->aapcs_reg = gen_rtx_REG (mode, ncrn);
+ pcum->aapcs_next_ncrn = ncrn + nregs;
+ return;
}
- if (TARGET_VFP)
- return (TARGET_BIG_END ? 1 : 0);
+ /* C5 - Some core registers left and there are no arguments already
+ on the stack: split this argument between the remaining core
+ registers and the stack. */
+ if (ncrn < NUM_ARG_REGS && pcum->can_split)
+ {
+ pcum->aapcs_reg = gen_rtx_REG (mode, ncrn);
+ pcum->aapcs_next_ncrn = NUM_ARG_REGS;
+ pcum->aapcs_partial = (NUM_ARG_REGS - ncrn) * UNITS_PER_WORD;
+ return;
+ }
- return 1;
+ /* C6 - NCRN is set to 4. */
+ pcum->aapcs_next_ncrn = NUM_ARG_REGS;
+
+ /* C7,C8 - arugment goes on the stack. We have nothing to do here. */
+ return;
}
/* Initialize a variable CUM of type CUMULATIVE_ARGS
For a library call, FNTYPE is NULL. */
void
arm_init_cumulative_args (CUMULATIVE_ARGS *pcum, tree fntype,
- rtx libname ATTRIBUTE_UNUSED,
+ rtx libname,
tree fndecl ATTRIBUTE_UNUSED)
{
+ /* Long call handling. */
+ if (fntype)
+ pcum->pcs_variant = arm_get_pcs_model (fntype, fndecl);
+ else
+ pcum->pcs_variant = arm_pcs_default;
+
+ if (pcum->pcs_variant <= ARM_PCS_AAPCS_LOCAL)
+ {
+ if (arm_libcall_uses_aapcs_base (libname))
+ pcum->pcs_variant = ARM_PCS_AAPCS;
+
+ pcum->aapcs_ncrn = pcum->aapcs_next_ncrn = 0;
+ pcum->aapcs_reg = NULL_RTX;
+ pcum->aapcs_partial = 0;
+ pcum->aapcs_arg_processed = false;
+ pcum->aapcs_cprc_slot = -1;
+ pcum->can_split = true;
+
+ if (pcum->pcs_variant != ARM_PCS_AAPCS)
+ {
+ int i;
+
+ for (i = 0; i < ARM_NUM_COPROC_SLOTS; i++)
+ {
+ pcum->aapcs_cprc_failed[i] = false;
+ aapcs_cp_arg_layout[i].cum_init (pcum, fntype, libname, fndecl);
+ }
+ }
+ return;
+ }
+
+ /* Legacy ABIs */
+
/* On the ARM, the offset starts at 0. */
pcum->nregs = 0;
pcum->iwmmxt_nregs = 0;
{
int nregs;
+ /* Handle the special case quickly. Pick an arbitrary value for op2 of
+ a call insn (op3 of a call_value insn). */
+ if (mode == VOIDmode)
+ return const0_rtx;
+
+ if (pcum->pcs_variant <= ARM_PCS_AAPCS_LOCAL)
+ {
+ aapcs_layout_arg (pcum, mode, type, named);
+ return pcum->aapcs_reg;
+ }
+
/* Varargs vectors are treated the same as long long.
named_count avoids having to change the way arm handles 'named' */
if (TARGET_IWMMXT_ABI
static int
arm_arg_partial_bytes (CUMULATIVE_ARGS *pcum, enum machine_mode mode,
- tree type, bool named ATTRIBUTE_UNUSED)
+ tree type, bool named)
{
int nregs = pcum->nregs;
+ if (pcum->pcs_variant <= ARM_PCS_AAPCS_LOCAL)
+ {
+ aapcs_layout_arg (pcum, mode, type, named);
+ return pcum->aapcs_partial;
+ }
+
if (TARGET_IWMMXT_ABI && arm_vector_mode_supported_p (mode))
return 0;
return 0;
}
+void
+arm_function_arg_advance (CUMULATIVE_ARGS *pcum, enum machine_mode mode,
+ tree type, bool named)
+{
+ if (pcum->pcs_variant <= ARM_PCS_AAPCS_LOCAL)
+ {
+ aapcs_layout_arg (pcum, mode, type, named);
+
+ if (pcum->aapcs_cprc_slot >= 0)
+ {
+ aapcs_cp_arg_layout[pcum->aapcs_cprc_slot].advance (pcum, mode,
+ type);
+ pcum->aapcs_cprc_slot = -1;
+ }
+
+ /* Generic stuff. */
+ pcum->aapcs_arg_processed = false;
+ pcum->aapcs_ncrn = pcum->aapcs_next_ncrn;
+ pcum->aapcs_reg = NULL_RTX;
+ pcum->aapcs_partial = 0;
+ }
+ else
+ {
+ pcum->nargs += 1;
+ if (arm_vector_mode_supported_p (mode)
+ && pcum->named_count > pcum->nargs
+ && TARGET_IWMMXT_ABI)
+ pcum->iwmmxt_nregs += 1;
+ else
+ pcum->nregs += ARM_NUM_REGS2 (mode, type);
+ }
+}
+
/* Variable sized types are passed by reference. This is a GCC
extension to the ARM ABI. */
arm_pragma_long_calls = OFF;
}
\f
-/* Table of machine attributes. */
-const struct attribute_spec arm_attribute_table[] =
-{
- /* { name, min_len, max_len, decl_req, type_req, fn_type_req, handler } */
- /* Function calls made to this symbol must be done indirectly, because
- it may lie outside of the 26 bit addressing range of a normal function
- call. */
- { "long_call", 0, 0, false, true, true, NULL },
- /* Whereas these functions are always known to reside within the 26 bit
- addressing range. */
- { "short_call", 0, 0, false, true, true, NULL },
- /* Interrupt Service Routines have special prologue and epilogue requirements. */
- { "isr", 0, 1, false, false, false, arm_handle_isr_attribute },
- { "interrupt", 0, 1, false, false, false, arm_handle_isr_attribute },
- { "naked", 0, 0, true, false, false, arm_handle_fndecl_attribute },
-#ifdef ARM_PE
- /* ARM/PE has three new attributes:
- interfacearm - ?
- dllexport - for exporting a function/variable that will live in a dll
- dllimport - for importing a function/variable from a dll
-
- Microsoft allows multiple declspecs in one __declspec, separating
- them with spaces. We do NOT support this. Instead, use __declspec
- multiple times.
- */
- { "dllimport", 0, 0, true, false, false, NULL },
- { "dllexport", 0, 0, true, false, false, NULL },
- { "interfacearm", 0, 0, true, false, false, arm_handle_fndecl_attribute },
-#elif TARGET_DLLIMPORT_DECL_ATTRIBUTES
- { "dllimport", 0, 0, false, false, false, handle_dll_attribute },
- { "dllexport", 0, 0, false, false, false, handle_dll_attribute },
- { "notshared", 0, 0, false, true, false, arm_handle_notshared_attribute },
-#endif
- { NULL, 0, 0, false, false, false, NULL }
-};
-
/* Handle an attribute requiring a FUNCTION_DECL;
arguments as in struct attribute_spec.handler. */
static tree
return NULL_TREE;
}
+/* Handle a "pcs" attribute; arguments as in struct
+ attribute_spec.handler. */
+static tree
+arm_handle_pcs_attribute (tree *node ATTRIBUTE_UNUSED, tree name, tree args,
+ int flags ATTRIBUTE_UNUSED, bool *no_add_attrs)
+{
+ if (arm_pcs_from_attribute (args) == ARM_PCS_UNKNOWN)
+ {
+ warning (OPT_Wattributes, "%qE attribute ignored", name);
+ *no_add_attrs = true;
+ }
+ return NULL_TREE;
+}
+
#if TARGET_DLLIMPORT_DECL_ATTRIBUTES
/* Handle the "notshared" attribute. This attribute is another way of
requesting hidden visibility. ARM's compiler supports
/* Return nonzero if it is ok to make a tail-call to DECL. */
static bool
-arm_function_ok_for_sibcall (tree decl, tree exp ATTRIBUTE_UNUSED)
+arm_function_ok_for_sibcall (tree decl, tree exp)
{
unsigned long func_type;
return false;
/* Never tailcall something for which we have no decl, or if we
- are in Thumb mode. */
- if (decl == NULL || TARGET_THUMB)
+ are generating code for Thumb-1. */
+ if (decl == NULL || TARGET_THUMB1)
return false;
/* The PIC register is live on entry to VxWorks PLT entries, so we
if (IS_INTERRUPT (func_type))
return false;
+ if (!VOID_TYPE_P (TREE_TYPE (DECL_RESULT (cfun->decl))))
+ {
+ /* Check that the return value locations are the same. For
+ example that we aren't returning a value from the sibling in
+ a VFP register but then need to transfer it to a core
+ register. */
+ rtx a, b;
+
+ a = arm_function_value (TREE_TYPE (exp), decl, false);
+ b = arm_function_value (TREE_TYPE (DECL_RESULT (cfun->decl)),
+ cfun->decl, false);
+ if (!rtx_equal_p (a, b))
+ return false;
+ }
+
/* Never tailcall if function may be called with a misaligned SP. */
if (IS_STACKALIGN (func_type))
return false;
{
rtx pic_ref, address;
rtx insn;
- int subregs = 0;
-
- /* If this function doesn't have a pic register, create one now. */
- require_pic_register ();
if (reg == 0)
{
gcc_assert (can_create_pseudo_p ());
reg = gen_reg_rtx (Pmode);
-
- subregs = 1;
+ address = gen_reg_rtx (Pmode);
}
-
- if (subregs)
- address = gen_reg_rtx (Pmode);
else
address = reg;
- if (TARGET_ARM)
- emit_insn (gen_pic_load_addr_arm (address, orig));
- else if (TARGET_THUMB2)
- emit_insn (gen_pic_load_addr_thumb2 (address, orig));
- else /* TARGET_THUMB1 */
- emit_insn (gen_pic_load_addr_thumb1 (address, orig));
-
/* VxWorks does not impose a fixed gap between segments; the run-time
gap can be different from the object-file gap. We therefore can't
use GOTOFF unless we are absolutely sure that the symbol is in the
SYMBOL_REF_LOCAL_P (orig)))
&& NEED_GOT_RELOC
&& !TARGET_VXWORKS_RTP)
- pic_ref = gen_rtx_PLUS (Pmode, cfun->machine->pic_reg, address);
+ insn = arm_pic_static_addr (orig, reg);
else
{
+ /* If this function doesn't have a pic register, create one now. */
+ require_pic_register ();
+
+ if (TARGET_32BIT)
+ emit_insn (gen_pic_load_addr_32bit (address, orig));
+ else /* TARGET_THUMB1 */
+ emit_insn (gen_pic_load_addr_thumb1 (address, orig));
+
pic_ref = gen_const_mem (Pmode,
gen_rtx_PLUS (Pmode, cfun->machine->pic_reg,
address));
+ insn = emit_move_insn (reg, pic_ref);
}
- insn = emit_move_insn (reg, pic_ref);
-
/* Put a REG_EQUAL note on this insn, so that it can be optimized
by loop. */
set_unique_reg_note (insn, REG_EQUAL, orig);
{
pic_rtx = gen_rtx_SYMBOL_REF (Pmode, VXWORKS_GOTT_BASE);
pic_rtx = gen_rtx_CONST (Pmode, pic_rtx);
- emit_insn (gen_pic_load_addr_arm (pic_reg, pic_rtx));
+ emit_insn (gen_pic_load_addr_32bit (pic_reg, pic_rtx));
emit_insn (gen_rtx_SET (Pmode, pic_reg, gen_rtx_MEM (Pmode, pic_reg)));
UNSPEC_GOTSYM_OFF);
pic_rtx = gen_rtx_CONST (Pmode, pic_rtx);
- if (TARGET_ARM)
- {
- emit_insn (gen_pic_load_addr_arm (pic_reg, pic_rtx));
- emit_insn (gen_pic_add_dot_plus_eight (pic_reg, pic_reg, labelno));
- }
- else if (TARGET_THUMB2)
+ if (TARGET_32BIT)
{
- /* Thumb-2 only allows very limited access to the PC. Calculate the
- address in a temporary register. */
- if (arm_pic_register != INVALID_REGNUM)
- {
- pic_tmp = gen_rtx_REG (SImode,
- thumb_find_work_register (saved_regs));
- }
+ emit_insn (gen_pic_load_addr_32bit (pic_reg, pic_rtx));
+ if (TARGET_ARM)
+ emit_insn (gen_pic_add_dot_plus_eight (pic_reg, pic_reg, labelno));
else
- {
- gcc_assert (can_create_pseudo_p ());
- pic_tmp = gen_reg_rtx (Pmode);
- }
-
- emit_insn (gen_pic_load_addr_thumb2 (pic_reg, pic_rtx));
- emit_insn (gen_pic_load_dot_plus_four (pic_tmp, labelno));
- emit_insn (gen_addsi3 (pic_reg, pic_reg, pic_tmp));
+ emit_insn (gen_pic_add_dot_plus_four (pic_reg, pic_reg, labelno));
}
else /* TARGET_THUMB1 */
{
emit_use (pic_reg);
}
+/* Generate code to load the address of a static var when flag_pic is set. */
+static rtx
+arm_pic_static_addr (rtx orig, rtx reg)
+{
+ rtx l1, labelno, offset_rtx, insn;
+
+ gcc_assert (flag_pic);
+
+ /* We use an UNSPEC rather than a LABEL_REF because this label
+ never appears in the code stream. */
+ labelno = GEN_INT (pic_labelno++);
+ l1 = gen_rtx_UNSPEC (Pmode, gen_rtvec (1, labelno), UNSPEC_PIC_LABEL);
+ l1 = gen_rtx_CONST (VOIDmode, l1);
+
+ /* On the ARM the PC register contains 'dot + 8' at the time of the
+ addition, on the Thumb it is 'dot + 4'. */
+ offset_rtx = plus_constant (l1, TARGET_ARM ? 8 : 4);
+ offset_rtx = gen_rtx_UNSPEC (Pmode, gen_rtvec (2, orig, offset_rtx),
+ UNSPEC_SYMBOL_OFFSET);
+ offset_rtx = gen_rtx_CONST (Pmode, offset_rtx);
+
+ if (TARGET_32BIT)
+ {
+ emit_insn (gen_pic_load_addr_32bit (reg, offset_rtx));
+ if (TARGET_ARM)
+ insn = emit_insn (gen_pic_add_dot_plus_eight (reg, reg, labelno));
+ else
+ insn = emit_insn (gen_pic_add_dot_plus_four (reg, reg, labelno));
+ }
+ else /* TARGET_THUMB1 */
+ {
+ emit_insn (gen_pic_load_addr_thumb1 (reg, offset_rtx));
+ insn = emit_insn (gen_pic_add_dot_plus_four (reg, reg, labelno));
+ }
+
+ return insn;
+}
/* Return nonzero if X is valid as an ARM state addressing register. */
static int
if (GET_MODE_SIZE (mode) <= 4
&& ! (arm_arch4
&& (mode == HImode
+ || mode == HFmode
|| (mode == QImode && outer == SIGN_EXTEND))))
{
if (code == MULT)
load. */
if (arm_arch4)
{
- if (mode == HImode || (outer == SIGN_EXTEND && mode == QImode))
+ if (mode == HImode
+ || mode == HFmode
+ || (outer == SIGN_EXTEND && mode == QImode))
range = 256;
else
range = 4096;
}
else
- range = (mode == HImode) ? 4095 : 4096;
+ range = (mode == HImode || mode == HFmode) ? 4095 : 4096;
return (code == CONST_INT
&& INTVAL (index) < range
return 1;
/* This is PC relative data after arm_reorg runs. */
- else if (GET_MODE_SIZE (mode) >= 4 && reload_completed
+ else if ((GET_MODE_SIZE (mode) >= 4 || mode == HFmode)
+ && reload_completed
&& (GET_CODE (x) == LABEL_REF
|| (GET_CODE (x) == CONST
&& GET_CODE (XEXP (x, 0)) == PLUS
if (TARGET_ARM)
emit_insn (gen_pic_add_dot_plus_eight (reg, reg, labelno));
else if (TARGET_THUMB2)
- {
- rtx tmp;
- /* Thumb-2 only allows very limited access to the PC. Calculate
- the address in a temporary register. */
- tmp = gen_reg_rtx (SImode);
- emit_insn (gen_pic_load_dot_plus_four (tmp, labelno));
- emit_insn (gen_addsi3(reg, reg, tmp));
- }
+ emit_insn (gen_pic_add_dot_plus_four (reg, reg, labelno));
else /* TARGET_THUMB1 */
emit_insn (gen_pic_add_dot_plus_four (reg, reg, labelno));
if (TARGET_ARM)
emit_insn (gen_tls_load_dot_plus_eight (reg, reg, labelno));
else if (TARGET_THUMB2)
- {
- rtx tmp;
- /* Thumb-2 only allows very limited access to the PC. Calculate
- the address in a temporary register. */
- tmp = gen_reg_rtx (SImode);
- emit_insn (gen_pic_load_dot_plus_four (tmp, labelno));
- emit_insn (gen_addsi3(reg, reg, tmp));
- emit_move_insn (reg, gen_const_mem (SImode, reg));
- }
+ emit_insn (gen_tls_load_dot_plus_four (reg, NULL, reg, labelno));
else
{
emit_insn (gen_pic_add_dot_plus_four (reg, reg, labelno));
else if ((outer == PLUS || outer == COMPARE)
&& INTVAL (x) < 256 && INTVAL (x) > -256)
return 0;
- else if (outer == AND
+ else if ((outer == IOR || outer == XOR || outer == AND)
&& INTVAL (x) < 256 && INTVAL (x) >= -256)
return COSTS_N_INSNS (1);
+ else if (outer == AND)
+ {
+ int i;
+ /* This duplicates the tests in the andsi3 expander. */
+ for (i = 9; i <= 31; i++)
+ if ((((HOST_WIDE_INT) 1) << i) - 1 == INTVAL (x)
+ || (((HOST_WIDE_INT) 1) << i) - 1 == ~INTVAL (x))
+ return COSTS_N_INSNS (2);
+ }
else if (outer == ASHIFT || outer == ASHIFTRT
|| outer == LSHIFTRT)
return 0;
enum rtx_code subcode;
rtx operand;
enum rtx_code code = GET_CODE (x);
- int extra_cost;
*total = 0;
switch (code)
case UMOD:
if (TARGET_HARD_FLOAT && mode == SFmode)
*total = COSTS_N_INSNS (2);
- else if (TARGET_HARD_FLOAT && mode == DFmode)
+ else if (TARGET_HARD_FLOAT && mode == DFmode && !TARGET_VFP_SINGLE)
*total = COSTS_N_INSNS (4);
else
*total = COSTS_N_INSNS (20);
if (GET_MODE_CLASS (mode) == MODE_FLOAT)
{
- if (TARGET_HARD_FLOAT && (mode == SFmode || mode == DFmode))
+ if (TARGET_HARD_FLOAT
+ && (mode == SFmode
+ || (mode == DFmode && !TARGET_VFP_SINGLE)))
{
*total = COSTS_N_INSNS (1);
if (GET_CODE (XEXP (x, 0)) == CONST_DOUBLE
if (GET_MODE_CLASS (mode) == MODE_FLOAT)
{
- if (TARGET_HARD_FLOAT && (mode == SFmode || mode == DFmode))
+ if (TARGET_HARD_FLOAT
+ && (mode == SFmode
+ || (mode == DFmode && !TARGET_VFP_SINGLE)))
{
*total = COSTS_N_INSNS (1);
if (GET_CODE (XEXP (x, 1)) == CONST_DOUBLE
/* Fall through */
case AND: case XOR: case IOR:
- extra_cost = 0;
/* Normally the frame registers will be spilt into reg+const during
reload, so it is a bad idea to combine them with other instructions,
case NEG:
if (GET_MODE_CLASS (mode) == MODE_FLOAT)
{
- if (TARGET_HARD_FLOAT && (mode == SFmode || mode == DFmode))
+ if (TARGET_HARD_FLOAT
+ && (mode == SFmode
+ || (mode == DFmode && !TARGET_VFP_SINGLE)))
{
*total = COSTS_N_INSNS (1);
return false;
case ABS:
if (GET_MODE_CLASS (mode) == MODE_FLOAT)
{
- if (TARGET_HARD_FLOAT && (mode == SFmode || mode == DFmode))
+ if (TARGET_HARD_FLOAT
+ && (mode == SFmode
+ || (mode == DFmode && !TARGET_VFP_SINGLE)))
{
*total = COSTS_N_INSNS (1);
return false;
*total = COSTS_N_INSNS (1) + rtx_cost (XEXP (x, 0), code, speed);
return true;
- case CONST_INT:
- if (const_ok_for_arm (INTVAL (x))
- || const_ok_for_arm (~INTVAL (x)))
- *total = COSTS_N_INSNS (1);
- else
- *total = COSTS_N_INSNS (arm_gen_constant (SET, mode, NULL_RTX,
- INTVAL (x), NULL_RTX,
- NULL_RTX, 0, 0));
- return true;
+ case CONST_INT:
+ if (const_ok_for_arm (INTVAL (x))
+ || const_ok_for_arm (~INTVAL (x)))
+ *total = COSTS_N_INSNS (1);
+ else
+ *total = COSTS_N_INSNS (arm_gen_constant (SET, mode, NULL_RTX,
+ INTVAL (x), NULL_RTX,
+ NULL_RTX, 0, 0));
+ return true;
+
+ case CONST:
+ case LABEL_REF:
+ case SYMBOL_REF:
+ *total = COSTS_N_INSNS (3);
+ return true;
+
+ case HIGH:
+ *total = COSTS_N_INSNS (1);
+ return true;
+
+ case LO_SUM:
+ *total = COSTS_N_INSNS (1);
+ *total += rtx_cost (XEXP (x, 0), code, speed);
+ return true;
+
+ case CONST_DOUBLE:
+ if (TARGET_HARD_FLOAT && vfp3_const_double_rtx (x)
+ && (mode == SFmode || !TARGET_VFP_SINGLE))
+ *total = COSTS_N_INSNS (1);
+ else
+ *total = COSTS_N_INSNS (4);
+ return true;
+
+ default:
+ *total = COSTS_N_INSNS (4);
+ return false;
+ }
+}
+
+/* Estimates the size cost of thumb1 instructions.
+ For now most of the code is copied from thumb1_rtx_costs. We need more
+ fine grain tuning when we have more related test cases. */
+static inline int
+thumb1_size_rtx_costs (rtx x, enum rtx_code code, enum rtx_code outer)
+{
+ enum machine_mode mode = GET_MODE (x);
+
+ switch (code)
+ {
+ case ASHIFT:
+ case ASHIFTRT:
+ case LSHIFTRT:
+ case ROTATERT:
+ case PLUS:
+ case MINUS:
+ case COMPARE:
+ case NEG:
+ case NOT:
+ return COSTS_N_INSNS (1);
+
+ case MULT:
+ if (GET_CODE (XEXP (x, 1)) == CONST_INT)
+ {
+ /* Thumb1 mul instruction can't operate on const. We must Load it
+ into a register first. */
+ int const_size = thumb1_size_rtx_costs (XEXP (x, 1), CONST_INT, SET);
+ return COSTS_N_INSNS (1) + const_size;
+ }
+ return COSTS_N_INSNS (1);
+
+ case SET:
+ return (COSTS_N_INSNS (1)
+ + 4 * ((GET_CODE (SET_SRC (x)) == MEM)
+ + GET_CODE (SET_DEST (x)) == MEM));
+
+ case CONST_INT:
+ if (outer == SET)
+ {
+ if ((unsigned HOST_WIDE_INT) INTVAL (x) < 256)
+ return 0;
+ if (thumb_shiftable_const (INTVAL (x)))
+ return COSTS_N_INSNS (2);
+ return COSTS_N_INSNS (3);
+ }
+ else if ((outer == PLUS || outer == COMPARE)
+ && INTVAL (x) < 256 && INTVAL (x) > -256)
+ return 0;
+ else if ((outer == IOR || outer == XOR || outer == AND)
+ && INTVAL (x) < 256 && INTVAL (x) >= -256)
+ return COSTS_N_INSNS (1);
+ else if (outer == AND)
+ {
+ int i;
+ /* This duplicates the tests in the andsi3 expander. */
+ for (i = 9; i <= 31; i++)
+ if ((((HOST_WIDE_INT) 1) << i) - 1 == INTVAL (x)
+ || (((HOST_WIDE_INT) 1) << i) - 1 == ~INTVAL (x))
+ return COSTS_N_INSNS (2);
+ }
+ else if (outer == ASHIFT || outer == ASHIFTRT
+ || outer == LSHIFTRT)
+ return 0;
+ return COSTS_N_INSNS (2);
+
+ case CONST:
+ case CONST_DOUBLE:
+ case LABEL_REF:
+ case SYMBOL_REF:
+ return COSTS_N_INSNS (3);
+
+ case UDIV:
+ case UMOD:
+ case DIV:
+ case MOD:
+ return 100;
+
+ case TRUNCATE:
+ return 99;
+
+ case AND:
+ case XOR:
+ case IOR:
+ /* XXX guess. */
+ return 8;
+
+ case MEM:
+ /* XXX another guess. */
+ /* Memory costs quite a lot for the first word, but subsequent words
+ load at the equivalent of a single insn each. */
+ return (10 + 4 * ((GET_MODE_SIZE (mode) - 1) / UNITS_PER_WORD)
+ + ((GET_CODE (x) == SYMBOL_REF && CONSTANT_POOL_ADDRESS_P (x))
+ ? 4 : 0));
+
+ case IF_THEN_ELSE:
+ /* XXX a guess. */
+ if (GET_CODE (XEXP (x, 1)) == PC || GET_CODE (XEXP (x, 2)) == PC)
+ return 14;
+ return 2;
- case CONST:
- case LABEL_REF:
- case SYMBOL_REF:
- *total = COSTS_N_INSNS (3);
- return true;
+ case ZERO_EXTEND:
+ /* XXX still guessing. */
+ switch (GET_MODE (XEXP (x, 0)))
+ {
+ case QImode:
+ return (1 + (mode == DImode ? 4 : 0)
+ + (GET_CODE (XEXP (x, 0)) == MEM ? 10 : 0));
- case HIGH:
- *total = COSTS_N_INSNS (1);
- return true;
+ case HImode:
+ return (4 + (mode == DImode ? 4 : 0)
+ + (GET_CODE (XEXP (x, 0)) == MEM ? 10 : 0));
- case LO_SUM:
- *total = COSTS_N_INSNS (1);
- *total += rtx_cost (XEXP (x, 0), code, speed);
- return true;
+ case SImode:
+ return (1 + (GET_CODE (XEXP (x, 0)) == MEM ? 10 : 0));
- case CONST_DOUBLE:
- if (TARGET_HARD_FLOAT && vfp3_const_double_rtx (x))
- *total = COSTS_N_INSNS (1);
- else
- *total = COSTS_N_INSNS (4);
- return true;
+ default:
+ return 99;
+ }
default:
- *total = COSTS_N_INSNS (4);
- return false;
+ return 99;
}
}
enum machine_mode mode = GET_MODE (x);
if (TARGET_THUMB1)
{
- /* XXX TBD. For now, use the standard costs. */
- *total = thumb1_rtx_costs (x, code, outer_code);
+ *total = thumb1_size_rtx_costs (x, code, outer_code);
return true;
}
return false;
case MINUS:
- if (TARGET_HARD_FLOAT && GET_MODE_CLASS (mode) == MODE_FLOAT)
+ if (TARGET_HARD_FLOAT && GET_MODE_CLASS (mode) == MODE_FLOAT
+ && (mode == SFmode || !TARGET_VFP_SINGLE))
{
*total = COSTS_N_INSNS (1);
return false;
return false;
case PLUS:
- if (TARGET_HARD_FLOAT && GET_MODE_CLASS (mode) == MODE_FLOAT)
+ if (TARGET_HARD_FLOAT && GET_MODE_CLASS (mode) == MODE_FLOAT
+ && (mode == SFmode || !TARGET_VFP_SINGLE))
{
*total = COSTS_N_INSNS (1);
return false;
return false;
case NEG:
- if (TARGET_HARD_FLOAT && GET_MODE_CLASS (mode) == MODE_FLOAT)
+ if (TARGET_HARD_FLOAT && GET_MODE_CLASS (mode) == MODE_FLOAT
+ && (mode == SFmode || !TARGET_VFP_SINGLE))
{
*total = COSTS_N_INSNS (1);
return false;
return false;
case ABS:
- if (TARGET_HARD_FLOAT && GET_MODE_CLASS (mode) == MODE_FLOAT)
+ if (TARGET_HARD_FLOAT && GET_MODE_CLASS (mode) == MODE_FLOAT
+ && (mode == SFmode || !TARGET_VFP_SINGLE))
*total = COSTS_N_INSNS (1);
else
*total = COSTS_N_INSNS (1 + ARM_NUM_REGS (mode));
return arm_size_rtx_costs (x, (enum rtx_code) code,
(enum rtx_code) outer_code, total);
else
- return all_cores[(int)arm_tune].rtx_costs (x, (enum rtx_code) code,
- (enum rtx_code) outer_code,
- total, speed);
+ return current_tune->rtx_costs (x, (enum rtx_code) code,
+ (enum rtx_code) outer_code,
+ total, speed);
}
/* RTX costs for cores with a slow MUL implementation. Thumb-2 is not
if (GET_MODE_CLASS (mode) == MODE_FLOAT)
{
- if (TARGET_HARD_FLOAT && (mode == SFmode || mode == DFmode))
+ if (TARGET_HARD_FLOAT
+ && (mode == SFmode
+ || (mode == DFmode && !TARGET_VFP_SINGLE)))
{
*total = COSTS_N_INSNS (1);
return false;
so it can be ignored. */
static bool
-arm_xscale_rtx_costs (rtx x, enum rtx_code code, enum rtx_code outer_code, int *total, bool speed)
+arm_xscale_rtx_costs (rtx x, enum rtx_code code, enum rtx_code outer_code,
+ int *total, bool speed)
{
enum machine_mode mode = GET_MODE (x);
if (GET_MODE_CLASS (mode) == MODE_FLOAT)
{
- if (TARGET_HARD_FLOAT && (mode == SFmode || mode == DFmode))
+ if (TARGET_HARD_FLOAT
+ && (mode == SFmode
+ || (mode == DFmode && !TARGET_VFP_SINGLE)))
{
*total = COSTS_N_INSNS (1);
return false;
if (c == MEM || c == LABEL_REF || c == SYMBOL_REF)
return 10;
- if (c == PLUS || c == MINUS)
+ if (c == PLUS)
{
- if (GET_CODE (XEXP (x, 0)) == CONST_INT)
+ if (GET_CODE (XEXP (x, 1)) == CONST_INT)
return 2;
if (ARITHMETIC_P (XEXP (x, 0)) || ARITHMETIC_P (XEXP (x, 1)))
init_fp_table ();
REAL_VALUE_FROM_CONST_DOUBLE (r, x);
- r = REAL_VALUE_NEGATE (r);
+ r = real_value_negate (&r);
if (REAL_VALUE_MINUS_ZERO (r))
return 0;
/* Extract sign, exponent and mantissa. */
sign = REAL_VALUE_NEGATIVE (r) ? 1 : 0;
- r = REAL_VALUE_ABS (r);
+ r = real_value_abs (&r);
exponent = REAL_EXP (&r);
/* For the mantissa, we expand into two HOST_WIDE_INTS, apart from the
highest (sign) bit, with a fixed binary point at bit point_pos.
}
}
-/* Initialize a vector with non-constant elements. FIXME: We can do better
- than the current implementation (building a vector on the stack and then
- loading it) in many cases. See rs6000.c. */
+/* If VALS is a vector constant that can be loaded into a register
+ using VDUP, generate instructions to do so and return an RTX to
+ assign to the register. Otherwise return NULL_RTX. */
+
+static rtx
+neon_vdup_constant (rtx vals)
+{
+ enum machine_mode mode = GET_MODE (vals);
+ enum machine_mode inner_mode = GET_MODE_INNER (mode);
+ int n_elts = GET_MODE_NUNITS (mode);
+ bool all_same = true;
+ rtx x;
+ int i;
+
+ if (GET_CODE (vals) != CONST_VECTOR || GET_MODE_SIZE (inner_mode) > 4)
+ return NULL_RTX;
+
+ for (i = 0; i < n_elts; ++i)
+ {
+ x = XVECEXP (vals, 0, i);
+ if (i > 0 && !rtx_equal_p (x, XVECEXP (vals, 0, 0)))
+ all_same = false;
+ }
+
+ if (!all_same)
+ /* The elements are not all the same. We could handle repeating
+ patterns of a mode larger than INNER_MODE here (e.g. int8x8_t
+ {0, C, 0, C, 0, C, 0, C} which can be loaded using
+ vdup.i16). */
+ return NULL_RTX;
+
+ /* We can load this constant by using VDUP and a constant in a
+ single ARM register. This will be cheaper than a vector
+ load. */
+
+ x = copy_to_mode_reg (inner_mode, XVECEXP (vals, 0, 0));
+ return gen_rtx_UNSPEC (mode, gen_rtvec (1, x),
+ UNSPEC_VDUP_N);
+}
+
+/* Generate code to load VALS, which is a PARALLEL containing only
+ constants (for vec_init) or CONST_VECTOR, efficiently into a
+ register. Returns an RTX to copy into the register, or NULL_RTX
+ for a PARALLEL that can not be converted into a CONST_VECTOR. */
+
+rtx
+neon_make_constant (rtx vals)
+{
+ enum machine_mode mode = GET_MODE (vals);
+ rtx target;
+ rtx const_vec = NULL_RTX;
+ int n_elts = GET_MODE_NUNITS (mode);
+ int n_const = 0;
+ int i;
+
+ if (GET_CODE (vals) == CONST_VECTOR)
+ const_vec = vals;
+ else if (GET_CODE (vals) == PARALLEL)
+ {
+ /* A CONST_VECTOR must contain only CONST_INTs and
+ CONST_DOUBLEs, but CONSTANT_P allows more (e.g. SYMBOL_REF).
+ Only store valid constants in a CONST_VECTOR. */
+ for (i = 0; i < n_elts; ++i)
+ {
+ rtx x = XVECEXP (vals, 0, i);
+ if (GET_CODE (x) == CONST_INT || GET_CODE (x) == CONST_DOUBLE)
+ n_const++;
+ }
+ if (n_const == n_elts)
+ const_vec = gen_rtx_CONST_VECTOR (mode, XVEC (vals, 0));
+ }
+ else
+ gcc_unreachable ();
+
+ if (const_vec != NULL
+ && neon_immediate_valid_for_move (const_vec, mode, NULL, NULL))
+ /* Load using VMOV. On Cortex-A8 this takes one cycle. */
+ return const_vec;
+ else if ((target = neon_vdup_constant (vals)) != NULL_RTX)
+ /* Loaded using VDUP. On Cortex-A8 the VDUP takes one NEON
+ pipeline cycle; creating the constant takes one or two ARM
+ pipeline cycles. */
+ return target;
+ else if (const_vec != NULL_RTX)
+ /* Load from constant pool. On Cortex-A8 this takes two cycles
+ (for either double or quad vectors). We can not take advantage
+ of single-cycle VLD1 because we need a PC-relative addressing
+ mode. */
+ return const_vec;
+ else
+ /* A PARALLEL containing something not valid inside CONST_VECTOR.
+ We can not construct an initializer. */
+ return NULL_RTX;
+}
+
+/* Initialize vector TARGET to VALS. */
void
neon_expand_vector_init (rtx target, rtx vals)
{
enum machine_mode mode = GET_MODE (target);
- enum machine_mode inner = GET_MODE_INNER (mode);
- unsigned int i, n_elts = GET_MODE_NUNITS (mode);
- rtx mem;
+ enum machine_mode inner_mode = GET_MODE_INNER (mode);
+ int n_elts = GET_MODE_NUNITS (mode);
+ int n_var = 0, one_var = -1;
+ bool all_same = true;
+ rtx x, mem;
+ int i;
+
+ for (i = 0; i < n_elts; ++i)
+ {
+ x = XVECEXP (vals, 0, i);
+ if (!CONSTANT_P (x))
+ ++n_var, one_var = i;
+
+ if (i > 0 && !rtx_equal_p (x, XVECEXP (vals, 0, 0)))
+ all_same = false;
+ }
- gcc_assert (VECTOR_MODE_P (mode));
+ if (n_var == 0)
+ {
+ rtx constant = neon_make_constant (vals);
+ if (constant != NULL_RTX)
+ {
+ emit_move_insn (target, constant);
+ return;
+ }
+ }
+
+ /* Splat a single non-constant element if we can. */
+ if (all_same && GET_MODE_SIZE (inner_mode) <= 4)
+ {
+ x = copy_to_mode_reg (inner_mode, XVECEXP (vals, 0, 0));
+ emit_insn (gen_rtx_SET (VOIDmode, target,
+ gen_rtx_UNSPEC (mode, gen_rtvec (1, x),
+ UNSPEC_VDUP_N)));
+ return;
+ }
+
+ /* One field is non-constant. Load constant then overwrite varying
+ field. This is more efficient than using the stack. */
+ if (n_var == 1)
+ {
+ rtx copy = copy_rtx (vals);
+ rtvec ops;
+
+ /* Load constant part of vector, substitute neighboring value for
+ varying element. */
+ XVECEXP (copy, 0, one_var) = XVECEXP (vals, 0, (one_var + 1) % n_elts);
+ neon_expand_vector_init (target, copy);
+ /* Insert variable. */
+ x = copy_to_mode_reg (inner_mode, XVECEXP (vals, 0, one_var));
+ ops = gen_rtvec (3, x, target, GEN_INT (one_var));
+ emit_insn (gen_rtx_SET (VOIDmode, target,
+ gen_rtx_UNSPEC (mode, ops, UNSPEC_VSET_LANE)));
+ return;
+ }
+
+ /* Construct the vector in memory one field at a time
+ and load the whole vector. */
mem = assign_stack_temp (mode, GET_MODE_SIZE (mode), 0);
for (i = 0; i < n_elts; i++)
- emit_move_insn (adjust_address_nv (mem, inner, i * GET_MODE_SIZE (inner)),
- XVECEXP (vals, 0, i));
-
+ emit_move_insn (adjust_address_nv (mem, inner_mode,
+ i * GET_MODE_SIZE (inner_mode)),
+ XVECEXP (vals, 0, i));
emit_move_insn (target, mem);
}
enum reg_class
coproc_secondary_reload_class (enum machine_mode mode, rtx x, bool wb)
{
+ if (mode == HFmode)
+ {
+ if (!TARGET_NEON_FP16)
+ return GENERAL_REGS;
+ if (s_register_operand (x, mode) || neon_vector_mem_operand (x, 2))
+ return NO_REGS;
+ return GENERAL_REGS;
+ }
+
if (TARGET_NEON
&& (GET_MODE_CLASS (mode) == MODE_VECTOR_INT
|| GET_MODE_CLASS (mode) == MODE_VECTOR_FLOAT)
}
}
-/* Must not copy a SET whose source operand is PC-relative. */
+/* Must not copy any rtx that uses a pc-relative address. */
+
+static int
+arm_note_pic_base (rtx *x, void *date ATTRIBUTE_UNUSED)
+{
+ if (GET_CODE (*x) == UNSPEC
+ && XINT (*x, 1) == UNSPEC_PIC_BASE)
+ return 1;
+ return 0;
+}
static bool
arm_cannot_copy_insn_p (rtx insn)
{
- rtx pat = PATTERN (insn);
-
- if (GET_CODE (pat) == SET)
- {
- rtx rhs = SET_SRC (pat);
-
- if (GET_CODE (rhs) == UNSPEC
- && XINT (rhs, 1) == UNSPEC_PIC_BASE)
- return TRUE;
-
- if (GET_CODE (rhs) == MEM
- && GET_CODE (XEXP (rhs, 0)) == UNSPEC
- && XINT (XEXP (rhs, 0), 1) == UNSPEC_PIC_BASE)
- return TRUE;
- }
-
- return FALSE;
+ return for_each_rtx (&PATTERN (insn), arm_note_pic_base, NULL);
}
enum rtx_code
return 0;
}
+/* Return true iff it would be profitable to turn a sequence of NOPS loads
+ or stores (depending on IS_STORE) into a load-multiple or store-multiple
+ instruction. ADD_OFFSET is nonzero if the base address register needs
+ to be modified with an add instruction before we can use it. */
+
+static bool
+multiple_operation_profitable_p (bool is_store ATTRIBUTE_UNUSED,
+ int nops, HOST_WIDE_INT add_offset)
+ {
+ /* For ARM8,9 & StrongARM, 2 ldr instructions are faster than an ldm
+ if the offset isn't small enough. The reason 2 ldrs are faster
+ is because these ARMs are able to do more than one cache access
+ in a single cycle. The ARM9 and StrongARM have Harvard caches,
+ whilst the ARM8 has a double bandwidth cache. This means that
+ these cores can do both an instruction fetch and a data fetch in
+ a single cycle, so the trick of calculating the address into a
+ scratch register (one of the result regs) and then doing a load
+ multiple actually becomes slower (and no smaller in code size).
+ That is the transformation
+
+ ldr rd1, [rbase + offset]
+ ldr rd2, [rbase + offset + 4]
+
+ to
+
+ add rd1, rbase, offset
+ ldmia rd1, {rd1, rd2}
+
+ produces worse code -- '3 cycles + any stalls on rd2' instead of
+ '2 cycles + any stalls on rd2'. On ARMs with only one cache
+ access per cycle, the first sequence could never complete in less
+ than 6 cycles, whereas the ldm sequence would only take 5 and
+ would make better use of sequential accesses if not hitting the
+ cache.
+
+ We cheat here and test 'arm_ld_sched' which we currently know to
+ only be true for the ARM8, ARM9 and StrongARM. If this ever
+ changes, then the test below needs to be reworked. */
+ if (nops == 2 && arm_ld_sched && add_offset != 0)
+ return false;
+
+ return true;
+}
+
+/* Subroutine of load_multiple_sequence and store_multiple_sequence.
+ Given an array of UNSORTED_OFFSETS, of which there are NOPS, compute
+ an array ORDER which describes the sequence to use when accessing the
+ offsets that produces an ascending order. In this sequence, each
+ offset must be larger by exactly 4 than the previous one. ORDER[0]
+ must have been filled in with the lowest offset by the caller.
+ If UNSORTED_REGS is nonnull, it is an array of register numbers that
+ we use to verify that ORDER produces an ascending order of registers.
+ Return true if it was possible to construct such an order, false if
+ not. */
+
+static bool
+compute_offset_order (int nops, HOST_WIDE_INT *unsorted_offsets, int *order,
+ int *unsorted_regs)
+{
+ int i;
+ for (i = 1; i < nops; i++)
+ {
+ int j;
+
+ order[i] = order[i - 1];
+ for (j = 0; j < nops; j++)
+ if (unsorted_offsets[j] == unsorted_offsets[order[i - 1]] + 4)
+ {
+ /* We must find exactly one offset that is higher than the
+ previous one by 4. */
+ if (order[i] != order[i - 1])
+ return false;
+ order[i] = j;
+ }
+ if (order[i] == order[i - 1])
+ return false;
+ /* The register numbers must be ascending. */
+ if (unsorted_regs != NULL
+ && unsorted_regs[order[i]] <= unsorted_regs[order[i - 1]])
+ return false;
+ }
+ return true;
+}
+
int
load_multiple_sequence (rtx *operands, int nops, int *regs, int *base,
HOST_WIDE_INT *load_offset)
{
- int unsorted_regs[4];
- HOST_WIDE_INT unsorted_offsets[4];
- int order[4];
+ int unsorted_regs[MAX_LDM_STM_OPS];
+ HOST_WIDE_INT unsorted_offsets[MAX_LDM_STM_OPS];
+ int order[MAX_LDM_STM_OPS];
int base_reg = -1;
- int i;
+ int i, ldm_case;
- /* Can only handle 2, 3, or 4 insns at present,
- though could be easily extended if required. */
- gcc_assert (nops >= 2 && nops <= 4);
+ /* Can only handle up to MAX_LDM_STM_OPS insns at present, though could be
+ easily extended if required. */
+ gcc_assert (nops >= 2 && nops <= MAX_LDM_STM_OPS);
- memset (order, 0, 4 * sizeof (int));
+ memset (order, 0, MAX_LDM_STM_OPS * sizeof (int));
/* Loop over the operands and check that the memory references are
suitable (i.e. immediate offsets from the same base register). At
== CONST_INT)))
{
if (i == 0)
- {
- base_reg = REGNO (reg);
- unsorted_regs[0] = (GET_CODE (operands[i]) == REG
- ? REGNO (operands[i])
- : REGNO (SUBREG_REG (operands[i])));
- order[0] = 0;
- }
+ base_reg = REGNO (reg);
else
{
if (base_reg != (int) REGNO (reg))
/* Not addressed from the same base register. */
return 0;
-
- unsorted_regs[i] = (GET_CODE (operands[i]) == REG
- ? REGNO (operands[i])
- : REGNO (SUBREG_REG (operands[i])));
- if (unsorted_regs[i] < unsorted_regs[order[0]])
- order[0] = i;
}
+ unsorted_regs[i] = (GET_CODE (operands[i]) == REG
+ ? REGNO (operands[i])
+ : REGNO (SUBREG_REG (operands[i])));
/* If it isn't an integer register, or if it overwrites the
base register but isn't the last insn in the list, then
return 0;
unsorted_offsets[i] = INTVAL (offset);
+ if (i == 0 || unsorted_offsets[i] < unsorted_offsets[order[0]])
+ order[0] = i;
}
else
/* Not a suitable memory address. */
/* All the useful information has now been extracted from the
operands into unsorted_regs and unsorted_offsets; additionally,
- order[0] has been set to the lowest numbered register in the
- list. Sort the registers into order, and check that the memory
- offsets are ascending and adjacent. */
-
- for (i = 1; i < nops; i++)
- {
- int j;
-
- order[i] = order[i - 1];
- for (j = 0; j < nops; j++)
- if (unsorted_regs[j] > unsorted_regs[order[i - 1]]
- && (order[i] == order[i - 1]
- || unsorted_regs[j] < unsorted_regs[order[i]]))
- order[i] = j;
-
- /* Have we found a suitable register? if not, one must be used more
- than once. */
- if (order[i] == order[i - 1])
- return 0;
-
- /* Is the memory address adjacent and ascending? */
- if (unsorted_offsets[order[i]] != unsorted_offsets[order[i - 1]] + 4)
- return 0;
- }
+ order[0] has been set to the lowest offset in the list. Sort
+ the offsets into order, verifying that they are adjacent, and
+ check that the register numbers are ascending. */
+ if (!compute_offset_order (nops, unsorted_offsets, order, unsorted_regs))
+ return 0;
if (base)
{
}
if (unsorted_offsets[order[0]] == 0)
- return 1; /* ldmia */
-
- if (TARGET_ARM && unsorted_offsets[order[0]] == 4)
- return 2; /* ldmib */
-
- if (TARGET_ARM && unsorted_offsets[order[nops - 1]] == 0)
- return 3; /* ldmda */
-
- if (unsorted_offsets[order[nops - 1]] == -4)
- return 4; /* ldmdb */
-
- /* For ARM8,9 & StrongARM, 2 ldr instructions are faster than an ldm
- if the offset isn't small enough. The reason 2 ldrs are faster
- is because these ARMs are able to do more than one cache access
- in a single cycle. The ARM9 and StrongARM have Harvard caches,
- whilst the ARM8 has a double bandwidth cache. This means that
- these cores can do both an instruction fetch and a data fetch in
- a single cycle, so the trick of calculating the address into a
- scratch register (one of the result regs) and then doing a load
- multiple actually becomes slower (and no smaller in code size).
- That is the transformation
-
- ldr rd1, [rbase + offset]
- ldr rd2, [rbase + offset + 4]
-
- to
-
- add rd1, rbase, offset
- ldmia rd1, {rd1, rd2}
-
- produces worse code -- '3 cycles + any stalls on rd2' instead of
- '2 cycles + any stalls on rd2'. On ARMs with only one cache
- access per cycle, the first sequence could never complete in less
- than 6 cycles, whereas the ldm sequence would only take 5 and
- would make better use of sequential accesses if not hitting the
- cache.
+ ldm_case = 1; /* ldmia */
+ else if (TARGET_ARM && unsorted_offsets[order[0]] == 4)
+ ldm_case = 2; /* ldmib */
+ else if (TARGET_ARM && unsorted_offsets[order[nops - 1]] == 0)
+ ldm_case = 3; /* ldmda */
+ else if (unsorted_offsets[order[nops - 1]] == -4)
+ ldm_case = 4; /* ldmdb */
+ else if (const_ok_for_arm (unsorted_offsets[order[0]])
+ || const_ok_for_arm (-unsorted_offsets[order[0]]))
+ ldm_case = 5;
+ else
+ return 0;
- We cheat here and test 'arm_ld_sched' which we currently know to
- only be true for the ARM8, ARM9 and StrongARM. If this ever
- changes, then the test below needs to be reworked. */
- if (nops == 2 && arm_ld_sched)
+ if (!multiple_operation_profitable_p (false, nops,
+ ldm_case == 5
+ ? unsorted_offsets[order[0]] : 0))
return 0;
- /* Can't do it without setting up the offset, only do this if it takes
- no more than one insn. */
- return (const_ok_for_arm (unsorted_offsets[order[0]])
- || const_ok_for_arm (-unsorted_offsets[order[0]])) ? 5 : 0;
+ return ldm_case;
}
const char *
emit_ldm_seq (rtx *operands, int nops)
{
- int regs[4];
+ int regs[MAX_LDM_STM_OPS];
int base_reg;
HOST_WIDE_INT offset;
char buf[100];
store_multiple_sequence (rtx *operands, int nops, int *regs, int *base,
HOST_WIDE_INT * load_offset)
{
- int unsorted_regs[4];
- HOST_WIDE_INT unsorted_offsets[4];
- int order[4];
+ int unsorted_regs[MAX_LDM_STM_OPS];
+ HOST_WIDE_INT unsorted_offsets[MAX_LDM_STM_OPS];
+ int order[MAX_LDM_STM_OPS];
int base_reg = -1;
- int i;
+ int i, stm_case;
- /* Can only handle 2, 3, or 4 insns at present, though could be easily
- extended if required. */
- gcc_assert (nops >= 2 && nops <= 4);
+ /* Can only handle up to MAX_LDM_STM_OPS insns at present, though could be
+ easily extended if required. */
+ gcc_assert (nops >= 2 && nops <= MAX_LDM_STM_OPS);
- memset (order, 0, 4 * sizeof (int));
+ memset (order, 0, MAX_LDM_STM_OPS * sizeof (int));
/* Loop over the operands and check that the memory references are
suitable (i.e. immediate offsets from the same base register). At
&& (GET_CODE (offset = XEXP (XEXP (operands[nops + i], 0), 1))
== CONST_INT)))
{
+ unsorted_regs[i] = (GET_CODE (operands[i]) == REG
+ ? REGNO (operands[i])
+ : REGNO (SUBREG_REG (operands[i])));
if (i == 0)
- {
- base_reg = REGNO (reg);
- unsorted_regs[0] = (GET_CODE (operands[i]) == REG
- ? REGNO (operands[i])
- : REGNO (SUBREG_REG (operands[i])));
- order[0] = 0;
- }
- else
- {
- if (base_reg != (int) REGNO (reg))
- /* Not addressed from the same base register. */
- return 0;
-
- unsorted_regs[i] = (GET_CODE (operands[i]) == REG
- ? REGNO (operands[i])
- : REGNO (SUBREG_REG (operands[i])));
- if (unsorted_regs[i] < unsorted_regs[order[0]])
- order[0] = i;
- }
+ base_reg = REGNO (reg);
+ else if (base_reg != (int) REGNO (reg))
+ /* Not addressed from the same base register. */
+ return 0;
/* If it isn't an integer register, then we can't do this. */
if (unsorted_regs[i] < 0 || unsorted_regs[i] > 14)
return 0;
unsorted_offsets[i] = INTVAL (offset);
+ if (i == 0 || unsorted_offsets[i] < unsorted_offsets[order[0]])
+ order[0] = i;
}
else
/* Not a suitable memory address. */
/* All the useful information has now been extracted from the
operands into unsorted_regs and unsorted_offsets; additionally,
- order[0] has been set to the lowest numbered register in the
- list. Sort the registers into order, and check that the memory
- offsets are ascending and adjacent. */
-
- for (i = 1; i < nops; i++)
- {
- int j;
-
- order[i] = order[i - 1];
- for (j = 0; j < nops; j++)
- if (unsorted_regs[j] > unsorted_regs[order[i - 1]]
- && (order[i] == order[i - 1]
- || unsorted_regs[j] < unsorted_regs[order[i]]))
- order[i] = j;
-
- /* Have we found a suitable register? if not, one must be used more
- than once. */
- if (order[i] == order[i - 1])
- return 0;
-
- /* Is the memory address adjacent and ascending? */
- if (unsorted_offsets[order[i]] != unsorted_offsets[order[i - 1]] + 4)
- return 0;
- }
+ order[0] has been set to the lowest offset in the list. Sort
+ the offsets into order, verifying that they are adjacent, and
+ check that the register numbers are ascending. */
+ if (!compute_offset_order (nops, unsorted_offsets, order, unsorted_regs))
+ return 0;
if (base)
{
}
if (unsorted_offsets[order[0]] == 0)
- return 1; /* stmia */
-
- if (unsorted_offsets[order[0]] == 4)
- return 2; /* stmib */
-
- if (unsorted_offsets[order[nops - 1]] == 0)
- return 3; /* stmda */
+ stm_case = 1; /* stmia */
+ else if (TARGET_ARM && unsorted_offsets[order[0]] == 4)
+ stm_case = 2; /* stmib */
+ else if (TARGET_ARM && unsorted_offsets[order[nops - 1]] == 0)
+ stm_case = 3; /* stmda */
+ else if (unsorted_offsets[order[nops - 1]] == -4)
+ stm_case = 4; /* stmdb */
+ else
+ return 0;
- if (unsorted_offsets[order[nops - 1]] == -4)
- return 4; /* stmdb */
+ if (!multiple_operation_profitable_p (false, nops, 0))
+ return 0;
- return 0;
+ return stm_case;
}
const char *
emit_stm_seq (rtx *operands, int nops)
{
- int regs[4];
+ int regs[MAX_LDM_STM_OPS];
int base_reg;
HOST_WIDE_INT offset;
char buf[100];
/* A compare with a shifted operand. Because of canonicalization, the
comparison will have to be swapped when we emit the assembler. */
- if (GET_MODE (y) == SImode && GET_CODE (y) == REG
+ if (GET_MODE (y) == SImode
+ && (REG_P (y) || (GET_CODE (y) == SUBREG))
&& (GET_CODE (x) == ASHIFT || GET_CODE (x) == ASHIFTRT
|| GET_CODE (x) == LSHIFTRT || GET_CODE (x) == ROTATE
|| GET_CODE (x) == ROTATERT))
/* This operation is performed swapped, but since we only rely on the Z
flag we don't need an additional mode. */
- if (GET_MODE (y) == SImode && REG_P (y)
+ if (GET_MODE (y) == SImode
+ && (REG_P (y) || (GET_CODE (y) == SUBREG))
&& GET_CODE (x) == NEG
&& (op == EQ || op == NE))
return CC_Zmode;
XVECEXP (par, 0, 0)
= gen_rtx_SET (VOIDmode,
- gen_frame_mem (BLKmode,
- gen_rtx_PRE_DEC (BLKmode,
- stack_pointer_rtx)),
+ gen_frame_mem
+ (BLKmode,
+ gen_rtx_PRE_MODIFY (Pmode,
+ stack_pointer_rtx,
+ plus_constant
+ (stack_pointer_rtx,
+ - (count * 8)))
+ ),
gen_rtx_UNSPEC (BLKmode,
gen_rtvec (1, reg),
UNSPEC_PUSH_MULT));
return "";
}
-/* Output a 'call' insn that is a reference in memory. */
+/* Output a 'call' insn that is a reference in memory. This is
+ disabled for ARMv5 and we prefer a blx instead because otherwise
+ there's a significant performance overhead. */
const char *
output_call_mem (rtx *operands)
{
- if (TARGET_INTERWORK && !arm_arch5)
+ gcc_assert (!arm_arch5);
+ if (TARGET_INTERWORK)
{
output_asm_insn ("ldr%?\t%|ip, %0", operands);
output_asm_insn ("mov%?\t%|lr, %|pc", operands);
first instruction. It's safe to use IP as the target of the
load since the call will kill it anyway. */
output_asm_insn ("ldr%?\t%|ip, %0", operands);
- if (arm_arch5)
- output_asm_insn ("blx%?\t%|ip", operands);
+ output_asm_insn ("mov%?\t%|lr, %|pc", operands);
+ if (arm_arch4t)
+ output_asm_insn ("bx%?\t%|ip", operands);
else
- {
- output_asm_insn ("mov%?\t%|lr, %|pc", operands);
- if (arm_arch4t)
- output_asm_insn ("bx%?\t%|ip", operands);
- else
- output_asm_insn ("mov%?\t%|pc, %|ip", operands);
- }
+ output_asm_insn ("mov%?\t%|pc, %|ip", operands);
}
else
{
return "";
}
-
-/* Emit a MOVW/MOVT pair. */
-void arm_emit_movpair (rtx dest, rtx src)
-{
- emit_set_insn (dest, gen_rtx_HIGH (SImode, src));
- emit_set_insn (dest, gen_rtx_LO_SUM (SImode, dest, src));
-}
-
+void
+arm_emit_movpair (rtx dest, rtx src)
+ {
+ /* If the src is an immediate, simplify it. */
+ if (CONST_INT_P (src))
+ {
+ HOST_WIDE_INT val = INTVAL (src);
+ emit_set_insn (dest, GEN_INT (val & 0x0000ffff));
+ if ((val >> 16) & 0x0000ffff)
+ emit_set_insn (gen_rtx_ZERO_EXTRACT (SImode, dest, GEN_INT (16),
+ GEN_INT (16)),
+ GEN_INT ((val >> 16) & 0x0000ffff));
+ return;
+ }
+ emit_set_insn (dest, gen_rtx_HIGH (SImode, src));
+ emit_set_insn (dest, gen_rtx_LO_SUM (SImode, dest, src));
+ }
/* Output a move from arm registers to an fpa registers.
OPERANDS[0] is an fpa register.
{
/* We're only using DImode here because it's a convenient size. */
ops[0] = gen_rtx_REG (DImode, REGNO (reg) + 2 * i);
- ops[1] = adjust_address (mem, SImode, 8 * i);
+ ops[1] = adjust_address (mem, DImode, 8 * i);
if (reg_overlap_mentioned_p (ops[0], mem))
{
gcc_assert (overlap == -1);
&& crtl->uses_pic_offset_table)
save_reg_mask |= 1 << PIC_OFFSET_TABLE_REGNUM;
}
+ else if (IS_VOLATILE(func_type))
+ {
+ /* For noreturn functions we historically omitted register saves
+ altogether. However this really messes up debugging. As a
+ compromise save just the frame pointers. Combined with the link
+ register saved elsewhere this should be sufficient to get
+ a backtrace. */
+ if (frame_pointer_needed)
+ save_reg_mask |= 1 << HARD_FRAME_POINTER_REGNUM;
+ if (df_regs_ever_live_p (ARM_HARD_FRAME_POINTER_REGNUM))
+ save_reg_mask |= 1 << ARM_HARD_FRAME_POINTER_REGNUM;
+ if (df_regs_ever_live_p (THUMB_HARD_FRAME_POINTER_REGNUM))
+ save_reg_mask |= 1 << THUMB_HARD_FRAME_POINTER_REGNUM;
+ }
else
{
/* In the normal case we only need to save those registers
| (1 << LR_REGNUM)
| (1 << PC_REGNUM);
- /* Volatile functions do not return, so there
- is no need to save any other registers. */
- if (IS_VOLATILE (func_type))
- return save_reg_mask;
-
save_reg_mask |= arm_compute_save_reg0_reg12_mask ();
/* Decide if we need to save the link register.
gcc_assert (stack_adjust == 0 || stack_adjust == 4);
if (stack_adjust && arm_arch5 && TARGET_ARM)
- sprintf (instr, "ldm%sib\t%%|sp, {", conditional);
+ if (TARGET_UNIFIED_ASM)
+ sprintf (instr, "ldmib%s\t%%|sp, {", conditional);
+ else
+ sprintf (instr, "ldm%sib\t%%|sp, {", conditional);
else
{
/* If we can't use ldmib (SA110 bug),
then try to pop r3 instead. */
if (stack_adjust)
live_regs_mask |= 1 << 3;
- sprintf (instr, "ldm%sfd\t%%|sp, {", conditional);
+
+ if (TARGET_UNIFIED_ASM)
+ sprintf (instr, "ldmfd%s\t%%|sp, {", conditional);
+ else
+ sprintf (instr, "ldm%sfd\t%%|sp, {", conditional);
}
}
else
- sprintf (instr, "ldm%sfd\t%%|sp!, {", conditional);
+ if (TARGET_UNIFIED_ASM)
+ sprintf (instr, "pop%s\t{", conditional);
+ else
+ sprintf (instr, "ldm%sfd\t%%|sp!, {", conditional);
p = instr + strlen (instr);
/* This variable is for the Virtual Frame Pointer, not VFP regs. */
int vfp_offset = offsets->frame;
- if (arm_fpu_arch == FPUTYPE_FPA_EMU2)
+ if (TARGET_FPA_EMU2)
{
for (reg = LAST_FPA_REGNUM; reg >= FIRST_FPA_REGNUM; reg--)
if (df_regs_ever_live_p (reg) && !call_used_regs[reg])
SP_REGNUM, HARD_FRAME_POINTER_REGNUM);
}
- if (arm_fpu_arch == FPUTYPE_FPA_EMU2)
+ if (TARGET_FPA_EMU2)
{
for (reg = FIRST_FPA_REGNUM; reg <= LAST_FPA_REGNUM; reg++)
if (df_regs_ever_live_p (reg) && !call_used_regs[reg])
if (TARGET_HARD_FLOAT && TARGET_VFP)
{
- start_reg = FIRST_VFP_REGNUM;
- for (reg = FIRST_VFP_REGNUM; reg < LAST_VFP_REGNUM; reg += 2)
+ int end_reg = LAST_VFP_REGNUM + 1;
+
+ /* Scan the registers in reverse order. We need to match
+ any groupings made in the prologue and generate matching
+ pop operations. */
+ for (reg = LAST_VFP_REGNUM - 1; reg >= FIRST_VFP_REGNUM; reg -= 2)
{
if ((!df_regs_ever_live_p (reg) || call_used_regs[reg])
- && (!df_regs_ever_live_p (reg + 1) || call_used_regs[reg + 1]))
+ && (!df_regs_ever_live_p (reg + 1)
+ || call_used_regs[reg + 1]))
{
- if (start_reg != reg)
+ if (end_reg > reg + 2)
vfp_output_fldmd (f, SP_REGNUM,
- (start_reg - FIRST_VFP_REGNUM) / 2,
- (reg - start_reg) / 2);
- start_reg = reg + 2;
+ (reg + 2 - FIRST_VFP_REGNUM) / 2,
+ (end_reg - (reg + 2)) / 2);
+ end_reg = reg;
}
}
- if (start_reg != reg)
- vfp_output_fldmd (f, SP_REGNUM,
- (start_reg - FIRST_VFP_REGNUM) / 2,
- (reg - start_reg) / 2);
+ if (end_reg > reg + 2)
+ vfp_output_fldmd (f, SP_REGNUM, 0,
+ (end_reg - (reg + 2)) / 2);
}
+
if (TARGET_IWMMXT)
for (reg = FIRST_IWMMXT_REGNUM; reg <= LAST_IWMMXT_REGNUM; reg++)
if (df_regs_ever_live_p (reg) && !call_used_regs[reg])
/* For the body of the insn we are going to generate an UNSPEC in
parallel with several USEs. This allows the insn to be recognized
- by the push_multi pattern in the arm.md file. The insn looks
- something like this:
+ by the push_multi pattern in the arm.md file.
+
+ The body of the insn looks something like this:
(parallel [
- (set (mem:BLK (pre_dec:BLK (reg:SI sp)))
+ (set (mem:BLK (pre_modify:SI (reg:SI sp)
+ (const_int:SI <num>)))
(unspec:BLK [(reg:SI r4)] UNSPEC_PUSH_MULT))
- (use (reg:SI 11 fp))
- (use (reg:SI 12 ip))
- (use (reg:SI 14 lr))
- (use (reg:SI 15 pc))
+ (use (reg:SI XX))
+ (use (reg:SI YY))
+ ...
])
For the frame note however, we try to be more explicit and actually
(sequence [
(set (reg:SI sp) (plus:SI (reg:SI sp) (const_int -20)))
(set (mem:SI (reg:SI sp)) (reg:SI r4))
- (set (mem:SI (plus:SI (reg:SI sp) (const_int 4))) (reg:SI fp))
- (set (mem:SI (plus:SI (reg:SI sp) (const_int 8))) (reg:SI ip))
- (set (mem:SI (plus:SI (reg:SI sp) (const_int 12))) (reg:SI lr))
+ (set (mem:SI (plus:SI (reg:SI sp) (const_int 4))) (reg:SI XX))
+ (set (mem:SI (plus:SI (reg:SI sp) (const_int 8))) (reg:SI YY))
+ ...
])
- This sequence is used both by the code to support stack unwinding for
- exceptions handlers and the code to generate dwarf2 frame debugging. */
+ FIXME:: In an ideal world the PRE_MODIFY would not exist and
+ instead we'd have a parallel expression detailing all
+ the stores to the various memory addresses so that debug
+ information is more up-to-date. Remember however while writing
+ this to take care of the constraints with the push instruction.
+
+ Note also that this has to be taken care of for the VFP registers.
+
+ For more see PR43399. */
par = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (num_regs));
dwarf = gen_rtx_SEQUENCE (VOIDmode, rtvec_alloc (num_dwarf_regs + 1));
XVECEXP (par, 0, 0)
= gen_rtx_SET (VOIDmode,
- gen_frame_mem (BLKmode,
- gen_rtx_PRE_DEC (BLKmode,
- stack_pointer_rtx)),
+ gen_frame_mem
+ (BLKmode,
+ gen_rtx_PRE_MODIFY (Pmode,
+ stack_pointer_rtx,
+ plus_constant
+ (stack_pointer_rtx,
+ -4 * num_regs))
+ ),
gen_rtx_UNSPEC (BLKmode,
gen_rtvec (1, reg),
UNSPEC_PUSH_MULT));
{
tmp
= gen_rtx_SET (VOIDmode,
- gen_frame_mem (SImode,
- plus_constant (stack_pointer_rtx,
- 4 * j)),
+ gen_frame_mem
+ (SImode,
+ plus_constant (stack_pointer_rtx,
+ 4 * j)),
reg);
RTX_FRAME_RELATED_P (tmp) = 1;
XVECEXP (dwarf, 0, dwarf_par_index++) = tmp;
XVECEXP (par, 0, 0)
= gen_rtx_SET (VOIDmode,
- gen_frame_mem (BLKmode,
- gen_rtx_PRE_DEC (BLKmode,
- stack_pointer_rtx)),
+ gen_frame_mem
+ (BLKmode,
+ gen_rtx_PRE_MODIFY (Pmode,
+ stack_pointer_rtx,
+ plus_constant
+ (stack_pointer_rtx,
+ -12 * count))
+ ),
gen_rtx_UNSPEC (BLKmode,
gen_rtvec (1, reg),
UNSPEC_PUSH_MULT));
}
}
+/* Given FROM and TO register numbers, say whether this elimination is
+ allowed. Frame pointer elimination is automatically handled.
+
+ All eliminations are permissible. Note that ARG_POINTER_REGNUM and
+ HARD_FRAME_POINTER_REGNUM are in fact the same thing. If we need a frame
+ pointer, we must eliminate FRAME_POINTER_REGNUM into
+ HARD_FRAME_POINTER_REGNUM and not into STACK_POINTER_REGNUM or
+ ARG_POINTER_REGNUM. */
+
+bool
+arm_can_eliminate (const int from, const int to)
+{
+ return ((to == FRAME_POINTER_REGNUM && from == ARG_POINTER_REGNUM) ? false :
+ (to == STACK_POINTER_REGNUM && frame_pointer_needed) ? false :
+ (to == ARM_HARD_FRAME_POINTER_REGNUM && TARGET_THUMB) ? false :
+ (to == THUMB_HARD_FRAME_POINTER_REGNUM && TARGET_ARM) ? false :
+ true);
+}
/* Emit RTL to save coprocessor registers on function entry. Returns the
number of bytes pushed. */
for (reg = LAST_IWMMXT_REGNUM; reg >= FIRST_IWMMXT_REGNUM; reg--)
if (df_regs_ever_live_p (reg) && ! call_used_regs[reg])
{
- insn = gen_rtx_PRE_DEC (V2SImode, stack_pointer_rtx);
+ insn = gen_rtx_PRE_DEC (Pmode, stack_pointer_rtx);
insn = gen_rtx_MEM (V2SImode, insn);
insn = emit_set_insn (insn, gen_rtx_REG (V2SImode, reg));
RTX_FRAME_RELATED_P (insn) = 1;
/* Save any floating point call-saved registers used by this
function. */
- if (arm_fpu_arch == FPUTYPE_FPA_EMU2)
+ if (TARGET_FPA_EMU2)
{
for (reg = LAST_FPA_REGNUM; reg >= FIRST_FPA_REGNUM; reg--)
if (df_regs_ever_live_p (reg) && !call_used_regs[reg])
{
- insn = gen_rtx_PRE_DEC (XFmode, stack_pointer_rtx);
+ insn = gen_rtx_PRE_DEC (Pmode, stack_pointer_rtx);
insn = gen_rtx_MEM (XFmode, insn);
insn = emit_set_insn (insn, gen_rtx_REG (XFmode, reg));
RTX_FRAME_RELATED_P (insn) = 1;
{
REAL_VALUE_TYPE r;
REAL_VALUE_FROM_CONST_DOUBLE (r, x);
- r = REAL_VALUE_NEGATE (r);
+ r = real_value_negate (&r);
fprintf (stream, "%s", fp_const_from_val (&r));
}
return;
}
return;
+ /* Print the high single-precision register of a VFP double-precision
+ register. */
+ case 'p':
+ {
+ int mode = GET_MODE (x);
+ int regno;
+
+ if (GET_MODE_SIZE (mode) != 8 || GET_CODE (x) != REG)
+ {
+ output_operand_lossage ("invalid operand for code '%c'", code);
+ return;
+ }
+
+ regno = REGNO (x);
+ if (!VFP_REGNO_OK_FOR_DOUBLE (regno))
+ {
+ output_operand_lossage ("invalid operand for code '%c'", code);
+ return;
+ }
+
+ fprintf (stream, "s%d", regno - FIRST_VFP_REGNUM + 1);
+ }
+ return;
+
/* Print a VFP/Neon double precision or quad precision register name. */
case 'P':
case 'q':
}
return;
+ /* Translate an S register number into a D register number and element index. */
+ case 'y':
+ {
+ int mode = GET_MODE (x);
+ int regno;
+
+ if (GET_MODE_SIZE (mode) != 4 || GET_CODE (x) != REG)
+ {
+ output_operand_lossage ("invalid operand for code '%c'", code);
+ return;
+ }
+
+ regno = REGNO (x);
+ if (!VFP_REGNO_OK_FOR_SINGLE (regno))
+ {
+ output_operand_lossage ("invalid operand for code '%c'", code);
+ return;
+ }
+
+ regno = regno - FIRST_VFP_REGNUM;
+ fprintf (stream, "d%d[%d]", regno / 2, regno % 2);
+ }
+ return;
+
+ /* Register specifier for vld1.16/vst1.16. Translate the S register
+ number into a D register number and element index. */
+ case 'z':
+ {
+ int mode = GET_MODE (x);
+ int regno;
+
+ if (GET_MODE_SIZE (mode) != 2 || GET_CODE (x) != REG)
+ {
+ output_operand_lossage ("invalid operand for code '%c'", code);
+ return;
+ }
+
+ regno = REGNO (x);
+ if (!VFP_REGNO_OK_FOR_SINGLE (regno))
+ {
+ output_operand_lossage ("invalid operand for code '%c'", code);
+ return;
+ }
+
+ regno = regno - FIRST_VFP_REGNUM;
+ fprintf (stream, "d%d[%d]", regno/2, ((regno % 2) ? 2 : 0));
+ }
+ return;
+
default:
if (x == 0)
{
default:
gcc_assert (GET_CODE (x) != NEG);
fputc ('#', stream);
+ if (GET_CODE (x) == HIGH)
+ {
+ fputs (":lower16:", stream);
+ x = XEXP (x, 0);
+ }
+
output_addr_const (stream, x);
break;
}
reversed if it appears to fail. */
int reverse = 0;
- /* JUMP_CLOBBERS will be one implies that the conditions if a branch is
- taken are clobbered, even if the rtl suggests otherwise. It also
- means that we have to grub around within the jump expression to find
- out what the conditions are when the jump isn't taken. */
- int jump_clobbers = 0;
-
/* If we start with a return insn, we only succeed if we find another one. */
int seeking_return = 0;
int then_not_else = TRUE;
rtx this_insn = start_insn, label = 0;
- /* If the jump cannot be done with one instruction, we cannot
- conditionally execute the instruction in the inverse case. */
- if (get_attr_conds (insn) == CONDS_JUMP_CLOB)
- {
- jump_clobbers = 1;
- return;
- }
-
/* Register the insn jumped to. */
if (reverse)
{
control falls in from somewhere else. */
if (this_insn == label)
{
- if (jump_clobbers)
- {
- arm_ccfsm_state = 2;
- this_insn = next_nonnote_insn (this_insn);
- }
- else
- arm_ccfsm_state = 1;
+ arm_ccfsm_state = 1;
succeed = TRUE;
}
else
this_insn = next_nonnote_insn (this_insn);
if (this_insn && this_insn == label)
{
- if (jump_clobbers)
- {
- arm_ccfsm_state = 2;
- this_insn = next_nonnote_insn (this_insn);
- }
- else
- arm_ccfsm_state = 1;
+ arm_ccfsm_state = 1;
succeed = TRUE;
}
else
if (this_insn && this_insn == label
&& insns_skipped < max_insns_skipped)
{
- if (jump_clobbers)
- {
- arm_ccfsm_state = 2;
- this_insn = next_nonnote_insn (this_insn);
- }
- else
- arm_ccfsm_state = 1;
+ arm_ccfsm_state = 1;
succeed = TRUE;
}
else
}
arm_target_insn = this_insn;
}
- if (jump_clobbers)
- {
- gcc_assert (!reverse);
- arm_current_cc =
- get_arm_condition_code (XEXP (XEXP (XEXP (SET_SRC (body),
- 0), 0), 1));
- if (GET_CODE (XEXP (XEXP (SET_SRC (body), 0), 0)) == AND)
- arm_current_cc = ARM_INVERSE_CONDITION_CODE (arm_current_cc);
- if (GET_CODE (XEXP (SET_SRC (body), 0)) == NE)
- arm_current_cc = ARM_INVERSE_CONDITION_CODE (arm_current_cc);
- }
- else
- {
- /* If REVERSE is true, ARM_CURRENT_CC needs to be inverted from
- what it was. */
- if (!reverse)
- arm_current_cc = get_arm_condition_code (XEXP (SET_SRC (body),
- 0));
- }
+
+ /* If REVERSE is true, ARM_CURRENT_CC needs to be inverted from
+ what it was. */
+ if (!reverse)
+ arm_current_cc = get_arm_condition_code (XEXP (SET_SRC (body), 0));
if (reverse || then_not_else)
arm_current_cc = ARM_INVERSE_CONDITION_CODE (arm_current_cc);
if (mode == DFmode)
return VFP_REGNO_OK_FOR_DOUBLE (regno);
+ /* VFP registers can hold HFmode values, but there is no point in
+ putting them there unless we have hardware conversion insns. */
+ if (mode == HFmode)
+ return TARGET_FP16 && VFP_REGNO_OK_FOR_SINGLE (regno);
+
if (TARGET_NEON)
return (VALID_NEON_DREG_MODE (mode) && VFP_REGNO_OK_FOR_DOUBLE (regno))
|| (VALID_NEON_QREG_MODE (mode)
}
static void
+arm_init_fp16_builtins (void)
+{
+ tree fp16_type = make_node (REAL_TYPE);
+ TYPE_PRECISION (fp16_type) = 16;
+ layout_type (fp16_type);
+ (*lang_hooks.types.register_builtin_type) (fp16_type, "__fp16");
+}
+
+static void
arm_init_builtins (void)
{
arm_init_tls_builtins ();
if (TARGET_NEON)
arm_init_neon_builtins ();
+
+ if (arm_fp16_format)
+ arm_init_fp16_builtins ();
+}
+
+/* Implement TARGET_INVALID_PARAMETER_TYPE. */
+
+static const char *
+arm_invalid_parameter_type (const_tree t)
+{
+ if (SCALAR_FLOAT_TYPE_P (t) && TYPE_PRECISION (t) == 16)
+ return N_("function parameters cannot have __fp16 type");
+ return NULL;
+}
+
+/* Implement TARGET_INVALID_PARAMETER_TYPE. */
+
+static const char *
+arm_invalid_return_type (const_tree t)
+{
+ if (SCALAR_FLOAT_TYPE_P (t) && TYPE_PRECISION (t) == 16)
+ return N_("functions cannot return __fp16 type");
+ return NULL;
+}
+
+/* Implement TARGET_PROMOTED_TYPE. */
+
+static tree
+arm_promoted_type (const_tree t)
+{
+ if (SCALAR_FLOAT_TYPE_P (t) && TYPE_PRECISION (t) == 16)
+ return float_type_node;
+ return NULL_TREE;
+}
+
+/* Implement TARGET_CONVERT_TO_TYPE.
+ Specifically, this hook implements the peculiarity of the ARM
+ half-precision floating-point C semantics that requires conversions between
+ __fp16 to or from double to do an intermediate conversion to float. */
+
+static tree
+arm_convert_to_type (tree type, tree expr)
+{
+ tree fromtype = TREE_TYPE (expr);
+ if (!SCALAR_FLOAT_TYPE_P (fromtype) || !SCALAR_FLOAT_TYPE_P (type))
+ return NULL_TREE;
+ if ((TYPE_PRECISION (fromtype) == 16 && TYPE_PRECISION (type) > 32)
+ || (TYPE_PRECISION (type) == 16 && TYPE_PRECISION (fromtype) > 32))
+ return convert (type, convert (float_type_node, expr));
+ return NULL_TREE;
+}
+
+/* Implement TARGET_SCALAR_MODE_SUPPORTED_P.
+ This simply adds HFmode as a supported mode; even though we don't
+ implement arithmetic on this type directly, it's supported by
+ optabs conversions, much the way the double-word arithmetic is
+ special-cased in the default hook. */
+
+static bool
+arm_scalar_mode_supported_p (enum machine_mode mode)
+{
+ if (mode == HFmode)
+ return (arm_fp16_format != ARM_FP16_FORMAT_NONE);
+ else
+ return default_scalar_mode_supported_p (mode);
}
/* Errors in the source file can cause expand_expr to return const0_rtx
unsigned HOST_WIDE_INT mask = 0xff;
int i;
+ val = val & (unsigned HOST_WIDE_INT)0xffffffffu;
if (val == 0) /* XXX */
return 0;
}
}
+/* Given the stack offsets and register mask in OFFSETS, decide
+ how many additional registers to push instead of subtracting
+ a constant from SP. */
+static int
+thumb1_extra_regs_pushed (arm_stack_offsets *offsets)
+{
+ HOST_WIDE_INT amount = offsets->outgoing_args - offsets->saved_regs;
+ unsigned long live_regs_mask = offsets->saved_regs_mask;
+ /* Extract a mask of the ones we can give to the Thumb's push instruction. */
+ unsigned long l_mask = live_regs_mask & 0x40ff;
+ /* Then count how many other high registers will need to be pushed. */
+ unsigned long high_regs_pushed = bit_count (live_regs_mask & 0x0f00);
+ int n_free;
+
+ /* If the stack frame size is 512 exactly, we can save one load
+ instruction, which should make this a win even when optimizing
+ for speed. */
+ if (!optimize_size && amount != 512)
+ return 0;
+
+ /* Can't do this if there are high registers to push, or if we
+ are not going to do a push at all. */
+ if (high_regs_pushed != 0 || l_mask == 0)
+ return 0;
+
+ /* Don't do this if thumb1_expand_prologue wants to emit instructions
+ between the push and the stack frame allocation. */
+ if ((flag_pic && arm_pic_register != INVALID_REGNUM)
+ || (!frame_pointer_needed && CALLER_INTERWORKING_SLOT_SIZE > 0))
+ return 0;
+
+ for (n_free = 0; n_free < 8 && !(live_regs_mask & 1); live_regs_mask >>= 1)
+ n_free++;
+
+ if (n_free == 0)
+ return 0;
+ gcc_assert (amount / 4 * 4 == amount);
+
+ if (amount >= 512 && (amount - n_free * 4) < 512)
+ return (amount - 508) / 4;
+ if (amount <= n_free * 4)
+ return amount / 4;
+ return 0;
+}
+
/* Generate the rest of a function's prologue. */
void
thumb1_expand_prologue (void)
stack_pointer_rtx);
amount = offsets->outgoing_args - offsets->saved_regs;
+ amount -= 4 * thumb1_extra_regs_pushed (offsets);
if (amount)
{
if (amount < 512)
register. */
else if ((l_mask & 0xff) != 0
|| (high_regs_pushed == 0 && l_mask))
- thumb_pushpop (f, l_mask, 1, &cfa_offset, l_mask);
+ {
+ unsigned long mask = l_mask;
+ mask |= (1 << thumb1_extra_regs_pushed (offsets)) - 1;
+ thumb_pushpop (f, mask, 1, &cfa_offset, mask);
+ }
if (high_regs_pushed)
{
}
else
{
- int set_float_abi_attributes = 0;
- switch (arm_fpu_arch)
- {
- case FPUTYPE_FPA:
- fpu_name = "fpa";
- break;
- case FPUTYPE_FPA_EMU2:
- fpu_name = "fpe2";
- break;
- case FPUTYPE_FPA_EMU3:
- fpu_name = "fpe3";
- break;
- case FPUTYPE_MAVERICK:
- fpu_name = "maverick";
- break;
- case FPUTYPE_VFP:
- fpu_name = "vfp";
- set_float_abi_attributes = 1;
- break;
- case FPUTYPE_VFP3D16:
- fpu_name = "vfpv3-d16";
- set_float_abi_attributes = 1;
- break;
- case FPUTYPE_VFP3:
- fpu_name = "vfpv3";
- set_float_abi_attributes = 1;
- break;
- case FPUTYPE_NEON:
- fpu_name = "neon";
- set_float_abi_attributes = 1;
- break;
- default:
- abort();
- }
- if (set_float_abi_attributes)
+ fpu_name = arm_fpu_desc->name;
+ if (arm_fpu_desc->model == ARM_FP_MODEL_VFP)
{
if (TARGET_HARD_FLOAT)
asm_fprintf (asm_out_file, "\t.eabi_attribute 27, 3\n");
val = 6;
asm_fprintf (asm_out_file, "\t.eabi_attribute 30, %d\n", val);
+ /* Tag_ABI_FP_16bit_format. */
+ if (arm_fp16_format)
+ asm_fprintf (asm_out_file, "\t.eabi_attribute 38, %d\n",
+ (int)arm_fp16_format);
+
if (arm_lang_output_object_attributes_hook)
arm_lang_output_object_attributes_hook();
}
return 1;
}
+/* Emit a fp16 constant appropriately padded to occupy a 4-byte word.
+ HFmode constant pool entries are actually loaded with ldr. */
+void
+arm_emit_fp16_const (rtx c)
+{
+ REAL_VALUE_TYPE r;
+ long bits;
+
+ REAL_VALUE_FROM_CONST_DOUBLE (r, c);
+ bits = real_to_target (NULL, &r, HFmode);
+ if (WORDS_BIG_ENDIAN)
+ assemble_zeros (2);
+ assemble_integer (GEN_INT (bits), 2, BITS_PER_WORD, 1);
+ if (!WORDS_BIG_ENDIAN)
+ assemble_zeros (2);
+}
+
const char *
arm_output_load_gr (rtx *operands)
{
that way. */
static void
-arm_setup_incoming_varargs (CUMULATIVE_ARGS *cum,
+arm_setup_incoming_varargs (CUMULATIVE_ARGS *pcum,
enum machine_mode mode,
tree type,
int *pretend_size,
int second_time ATTRIBUTE_UNUSED)
{
- int nregs = cum->nregs;
- if (nregs & 1
- && ARM_DOUBLEWORD_ALIGN
- && arm_needs_doubleword_align (mode, type))
- nregs++;
-
+ int nregs;
+
cfun->machine->uses_anonymous_args = 1;
+ if (pcum->pcs_variant <= ARM_PCS_AAPCS_LOCAL)
+ {
+ nregs = pcum->aapcs_ncrn;
+ if ((nregs & 1) && arm_needs_doubleword_align (mode, type))
+ nregs++;
+ }
+ else
+ nregs = pcum->nregs;
+
if (nregs < NUM_ARG_REGS)
*pretend_size = (NUM_ARG_REGS - nregs) * UNITS_PER_WORD;
}
return !TARGET_AAPCS_BASED;
}
+static enum machine_mode
+arm_promote_function_mode (const_tree type ATTRIBUTE_UNUSED,
+ enum machine_mode mode,
+ int *punsignedp ATTRIBUTE_UNUSED,
+ const_tree fntype ATTRIBUTE_UNUSED,
+ int for_return ATTRIBUTE_UNUSED)
+{
+ if (GET_MODE_CLASS (mode) == MODE_INT
+ && GET_MODE_SIZE (mode) < 4)
+ return SImode;
+
+ return mode;
+}
/* AAPCS based ABIs use short enums by default. */
|| mode == V16QImode || mode == V4SFmode || mode == V2DImode))
return true;
- if ((mode == V2SImode)
- || (mode == V4HImode)
- || (mode == V8QImode))
+ if ((TARGET_NEON || TARGET_IWMMXT)
+ && ((mode == V2SImode)
+ || (mode == V4HImode)
+ || (mode == V8QImode)))
return true;
return false;
}
+/* Implements target hook small_register_classes_for_mode_p. */
+bool
+arm_small_register_classes_for_mode_p (enum machine_mode mode ATTRIBUTE_UNUSED)
+{
+ return TARGET_THUMB1;
+}
+
/* Implement TARGET_SHIFT_TRUNCATION_MASK. SImode shifts use normal
ARM insns and therefore guarantee that the shift count is modulo 256.
DImode shifts (those implemented by lib1funcs.asm or by optabs.c)
if (IS_FPA_REGNUM (regno))
return (TARGET_AAPCS_BASED ? 96 : 16) + regno - FIRST_FPA_REGNUM;
- /* FIXME: VFPv3 register numbering. */
if (IS_VFP_REGNUM (regno))
- return 64 + regno - FIRST_VFP_REGNUM;
+ {
+ /* See comment in arm_dwarf_register_span. */
+ if (VFP_REGNO_OK_FOR_SINGLE (regno))
+ return 64 + regno - FIRST_VFP_REGNUM;
+ else
+ return 256 + (regno - FIRST_VFP_REGNUM) / 2;
+ }
if (IS_IWMMXT_GR_REGNUM (regno))
return 104 + regno - FIRST_IWMMXT_GR_REGNUM;
gcc_unreachable ();
}
+/* Dwarf models VFPv3 registers as 32 64-bit registers.
+ GCC models tham as 64 32-bit registers, so we need to describe this to
+ the DWARF generation code. Other registers can use the default. */
+static rtx
+arm_dwarf_register_span (rtx rtl)
+{
+ unsigned regno;
+ int nregs;
+ int i;
+ rtx p;
+
+ regno = REGNO (rtl);
+ if (!IS_VFP_REGNUM (regno))
+ return NULL_RTX;
+
+ /* XXX FIXME: The EABI defines two VFP register ranges:
+ 64-95: Legacy VFPv2 numbering for S0-S31 (obsolescent)
+ 256-287: D0-D31
+ The recommended encoding for S0-S31 is a DW_OP_bit_piece of the
+ corresponding D register. Until GDB supports this, we shall use the
+ legacy encodings. We also use these encodings for D0-D15 for
+ compatibility with older debuggers. */
+ if (VFP_REGNO_OK_FOR_SINGLE (regno))
+ return NULL_RTX;
+
+ nregs = GET_MODE_SIZE (GET_MODE (rtl)) / 8;
+ p = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (nregs));
+ regno = (regno - FIRST_VFP_REGNUM) / 2;
+ for (i = 0; i < nregs; i++)
+ XVECEXP (p, 0, i) = gen_rtx_REG (DImode, 256 + regno + i);
+
+ return p;
+}
#ifdef TARGET_UNWIND_INFO
/* Emit unwind directives for a store-multiple instruction or stack pointer
offset = INTVAL (XEXP (e1, 1));
asm_fprintf (asm_out_file, "\t.setfp %r, %r, #%wd\n",
HARD_FRAME_POINTER_REGNUM, reg,
- INTVAL (XEXP (e1, 1)));
+ offset);
}
else if (GET_CODE (e1) == REG)
{
fputc (')', fp);
return TRUE;
}
+ else if (GET_CODE (x) == UNSPEC && XINT (x, 1) == UNSPEC_SYMBOL_OFFSET)
+ {
+ output_addr_const (fp, XVECEXP (x, 0, 0));
+ if (GOT_PCREL)
+ fputs ("+.", fp);
+ fputs ("-(", fp);
+ output_addr_const (fp, XVECEXP (x, 0, 1));
+ fputc (')', fp);
+ return TRUE;
+ }
else if (GET_CODE (x) == CONST_VECTOR)
return arm_emit_vector_const (fp, x);
return "";
}
+/* Output a Thumb-1 casesi dispatch sequence. */
+const char *
+thumb1_output_casesi (rtx *operands)
+{
+ rtx diff_vec = PATTERN (next_real_insn (operands[0]));
+
+ gcc_assert (GET_CODE (diff_vec) == ADDR_DIFF_VEC);
+
+ switch (GET_MODE(diff_vec))
+ {
+ case QImode:
+ return (ADDR_DIFF_VEC_FLAGS (diff_vec).offset_unsigned ?
+ "bl\t%___gnu_thumb1_case_uqi" : "bl\t%___gnu_thumb1_case_sqi");
+ case HImode:
+ return (ADDR_DIFF_VEC_FLAGS (diff_vec).offset_unsigned ?
+ "bl\t%___gnu_thumb1_case_uhi" : "bl\t%___gnu_thumb1_case_shi");
+ case SImode:
+ return "bl\t%___gnu_thumb1_case_si";
+ default:
+ gcc_unreachable ();
+ }
+}
+
/* Output a Thumb-2 casesi instruction. */
const char *
thumb2_output_casesi (rtx *operands)
&& lang_hooks.types_compatible_p (CONST_CAST_TREE (type), va_list_type))
{
static bool warned;
- if (!warned && warn_psabi)
+ if (!warned && warn_psabi && !in_system_header)
{
warned = true;
inform (input_location,
return "St9__va_list";
}
+ /* Half-precision float. */
+ if (TREE_CODE (type) == REAL_TYPE && TYPE_PRECISION (type) == 16)
+ return "Dh";
+
if (TREE_CODE (type) != VECTOR_TYPE)
return NULL;
flag_section_anchors = 2;
}
+/* Implement TARGET_FRAME_POINTER_REQUIRED. */
+
+bool
+arm_frame_pointer_required (void)
+{
+ return (cfun->has_nonlocal_label
+ || SUBTARGET_FRAME_POINTER_REQUIRED
+ || (TARGET_ARM && TARGET_APCS_FRAME && ! leaf_function_p ()));
+}
+
+/* Only thumb1 can't support conditional execution, so return true if
+ the target is not thumb1. */
+static bool
+arm_have_conditional_execution (void)
+{
+ return !TARGET_THUMB1;
+}
+
#include "gt-arm.h"
OSDN Git Service
