/* tc-aarch64.c -- Assemble for the AArch64 ISA Copyright (C) 2009-2018 Free Software Foundation, Inc. Contributed by ARM Ltd. This file is part of GAS. GAS is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation; either version 3 of the license, or (at your option) any later version. GAS is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details. You should have received a copy of the GNU General Public License along with this program; see the file COPYING3. If not, see . */ #include "as.h" #include #include #include "bfd_stdint.h" #define NO_RELOC 0 #include "safe-ctype.h" #include "subsegs.h" #include "obstack.h" #ifdef OBJ_ELF #include "elf/aarch64.h" #include "dw2gencfi.h" #endif #include "dwarf2dbg.h" /* Types of processor to assemble for. */ #ifndef CPU_DEFAULT #define CPU_DEFAULT AARCH64_ARCH_V8 #endif #define streq(a, b) (strcmp (a, b) == 0) #define END_OF_INSN '\0' static aarch64_feature_set cpu_variant; /* Variables that we set while parsing command-line options. Once all options have been read we re-process these values to set the real assembly flags. */ static const aarch64_feature_set *mcpu_cpu_opt = NULL; static const aarch64_feature_set *march_cpu_opt = NULL; /* Constants for known architecture features. */ static const aarch64_feature_set cpu_default = CPU_DEFAULT; #ifdef OBJ_ELF /* Pre-defined "_GLOBAL_OFFSET_TABLE_" */ static symbolS *GOT_symbol; /* Which ABI to use. */ enum aarch64_abi_type { AARCH64_ABI_NONE = 0, AARCH64_ABI_LP64 = 1, AARCH64_ABI_ILP32 = 2 }; #ifndef DEFAULT_ARCH #define DEFAULT_ARCH "aarch64" #endif /* DEFAULT_ARCH is initialized in gas/configure.tgt. */ static const char *default_arch = DEFAULT_ARCH; /* AArch64 ABI for the output file. */ static enum aarch64_abi_type aarch64_abi = AARCH64_ABI_NONE; /* When non-zero, program to a 32-bit model, in which the C data types int, long and all pointer types are 32-bit objects (ILP32); or to a 64-bit model, in which the C int type is 32-bits but the C long type and all pointer types are 64-bit objects (LP64). */ #define ilp32_p (aarch64_abi == AARCH64_ABI_ILP32) #endif enum vector_el_type { NT_invtype = -1, NT_b, NT_h, NT_s, NT_d, NT_q, NT_zero, NT_merge }; /* Bits for DEFINED field in vector_type_el. */ #define NTA_HASTYPE 1 #define NTA_HASINDEX 2 #define NTA_HASVARWIDTH 4 struct vector_type_el { enum vector_el_type type; unsigned char defined; unsigned width; int64_t index; }; #define FIXUP_F_HAS_EXPLICIT_SHIFT 0x00000001 struct reloc { bfd_reloc_code_real_type type; expressionS exp; int pc_rel; enum aarch64_opnd opnd; uint32_t flags; unsigned need_libopcodes_p : 1; }; struct aarch64_instruction { /* libopcodes structure for instruction intermediate representation. */ aarch64_inst base; /* Record assembly errors found during the parsing. */ struct { enum aarch64_operand_error_kind kind; const char *error; } parsing_error; /* The condition that appears in the assembly line. */ int cond; /* Relocation information (including the GAS internal fixup). */ struct reloc reloc; /* Need to generate an immediate in the literal pool. */ unsigned gen_lit_pool : 1; }; typedef struct aarch64_instruction aarch64_instruction; static aarch64_instruction inst; static bfd_boolean parse_operands (char *, const aarch64_opcode *); static bfd_boolean programmer_friendly_fixup (aarch64_instruction *); /* Diagnostics inline function utilities. These are lightweight utilities which should only be called by parse_operands and other parsers. GAS processes each assembly line by parsing it against instruction template(s), in the case of multiple templates (for the same mnemonic name), those templates are tried one by one until one succeeds or all fail. An assembly line may fail a few templates before being successfully parsed; an error saved here in most cases is not a user error but an error indicating the current template is not the right template. Therefore it is very important that errors can be saved at a low cost during the parsing; we don't want to slow down the whole parsing by recording non-user errors in detail. Remember that the objective is to help GAS pick up the most appropriate error message in the case of multiple templates, e.g. FMOV which has 8 templates. */ static inline void clear_error (void) { inst.parsing_error.kind = AARCH64_OPDE_NIL; inst.parsing_error.error = NULL; } static inline bfd_boolean error_p (void) { return inst.parsing_error.kind != AARCH64_OPDE_NIL; } static inline const char * get_error_message (void) { return inst.parsing_error.error; } static inline enum aarch64_operand_error_kind get_error_kind (void) { return inst.parsing_error.kind; } static inline void set_error (enum aarch64_operand_error_kind kind, const char *error) { inst.parsing_error.kind = kind; inst.parsing_error.error = error; } static inline void set_recoverable_error (const char *error) { set_error (AARCH64_OPDE_RECOVERABLE, error); } /* Use the DESC field of the corresponding aarch64_operand entry to compose the error message. */ static inline void set_default_error (void) { set_error (AARCH64_OPDE_SYNTAX_ERROR, NULL); } static inline void set_syntax_error (const char *error) { set_error (AARCH64_OPDE_SYNTAX_ERROR, error); } static inline void set_first_syntax_error (const char *error) { if (! error_p ()) set_error (AARCH64_OPDE_SYNTAX_ERROR, error); } static inline void set_fatal_syntax_error (const char *error) { set_error (AARCH64_OPDE_FATAL_SYNTAX_ERROR, error); } /* Number of littlenums required to hold an extended precision number. */ #define MAX_LITTLENUMS 6 /* Return value for certain parsers when the parsing fails; those parsers return the information of the parsed result, e.g. register number, on success. */ #define PARSE_FAIL -1 /* This is an invalid condition code that means no conditional field is present. */ #define COND_ALWAYS 0x10 typedef struct { const char *template; unsigned long value; } asm_barrier_opt; typedef struct { const char *template; uint32_t value; } asm_nzcv; struct reloc_entry { char *name; bfd_reloc_code_real_type reloc; }; /* Macros to define the register types and masks for the purpose of parsing. */ #undef AARCH64_REG_TYPES #define AARCH64_REG_TYPES \ BASIC_REG_TYPE(R_32) /* w[0-30] */ \ BASIC_REG_TYPE(R_64) /* x[0-30] */ \ BASIC_REG_TYPE(SP_32) /* wsp */ \ BASIC_REG_TYPE(SP_64) /* sp */ \ BASIC_REG_TYPE(Z_32) /* wzr */ \ BASIC_REG_TYPE(Z_64) /* xzr */ \ BASIC_REG_TYPE(FP_B) /* b[0-31] *//* NOTE: keep FP_[BHSDQ] consecutive! */\ BASIC_REG_TYPE(FP_H) /* h[0-31] */ \ BASIC_REG_TYPE(FP_S) /* s[0-31] */ \ BASIC_REG_TYPE(FP_D) /* d[0-31] */ \ BASIC_REG_TYPE(FP_Q) /* q[0-31] */ \ BASIC_REG_TYPE(VN) /* v[0-31] */ \ BASIC_REG_TYPE(ZN) /* z[0-31] */ \ BASIC_REG_TYPE(PN) /* p[0-15] */ \ /* Typecheck: any 64-bit int reg (inc SP exc XZR). */ \ MULTI_REG_TYPE(R64_SP, REG_TYPE(R_64) | REG_TYPE(SP_64)) \ /* Typecheck: same, plus SVE registers. */ \ MULTI_REG_TYPE(SVE_BASE, REG_TYPE(R_64) | REG_TYPE(SP_64) \ | REG_TYPE(ZN)) \ /* Typecheck: x[0-30], w[0-30] or [xw]zr. */ \ MULTI_REG_TYPE(R_Z, REG_TYPE(R_32) | REG_TYPE(R_64) \ | REG_TYPE(Z_32) | REG_TYPE(Z_64)) \ /* Typecheck: same, plus SVE registers. */ \ MULTI_REG_TYPE(SVE_OFFSET, REG_TYPE(R_32) | REG_TYPE(R_64) \ | REG_TYPE(Z_32) | REG_TYPE(Z_64) \ | REG_TYPE(ZN)) \ /* Typecheck: x[0-30], w[0-30] or {w}sp. */ \ MULTI_REG_TYPE(R_SP, REG_TYPE(R_32) | REG_TYPE(R_64) \ | REG_TYPE(SP_32) | REG_TYPE(SP_64)) \ /* Typecheck: any int (inc {W}SP inc [WX]ZR). */ \ MULTI_REG_TYPE(R_Z_SP, REG_TYPE(R_32) | REG_TYPE(R_64) \ | REG_TYPE(SP_32) | REG_TYPE(SP_64) \ | REG_TYPE(Z_32) | REG_TYPE(Z_64)) \ /* Typecheck: any [BHSDQ]P FP. */ \ MULTI_REG_TYPE(BHSDQ, REG_TYPE(FP_B) | REG_TYPE(FP_H) \ | REG_TYPE(FP_S) | REG_TYPE(FP_D) | REG_TYPE(FP_Q)) \ /* Typecheck: any int or [BHSDQ]P FP or V reg (exc SP inc [WX]ZR). */ \ MULTI_REG_TYPE(R_Z_BHSDQ_V, REG_TYPE(R_32) | REG_TYPE(R_64) \ | REG_TYPE(Z_32) | REG_TYPE(Z_64) | REG_TYPE(VN) \ | REG_TYPE(FP_B) | REG_TYPE(FP_H) \ | REG_TYPE(FP_S) | REG_TYPE(FP_D) | REG_TYPE(FP_Q)) \ /* Typecheck: as above, but also Zn, Pn, and {W}SP. This should only \ be used for SVE instructions, since Zn and Pn are valid symbols \ in other contexts. */ \ MULTI_REG_TYPE(R_Z_SP_BHSDQ_VZP, REG_TYPE(R_32) | REG_TYPE(R_64) \ | REG_TYPE(SP_32) | REG_TYPE(SP_64) \ | REG_TYPE(Z_32) | REG_TYPE(Z_64) | REG_TYPE(VN) \ | REG_TYPE(FP_B) | REG_TYPE(FP_H) \ | REG_TYPE(FP_S) | REG_TYPE(FP_D) | REG_TYPE(FP_Q) \ | REG_TYPE(ZN) | REG_TYPE(PN)) \ /* Any integer register; used for error messages only. */ \ MULTI_REG_TYPE(R_N, REG_TYPE(R_32) | REG_TYPE(R_64) \ | REG_TYPE(SP_32) | REG_TYPE(SP_64) \ | REG_TYPE(Z_32) | REG_TYPE(Z_64)) \ /* Pseudo type to mark the end of the enumerator sequence. */ \ BASIC_REG_TYPE(MAX) #undef BASIC_REG_TYPE #define BASIC_REG_TYPE(T) REG_TYPE_##T, #undef MULTI_REG_TYPE #define MULTI_REG_TYPE(T,V) BASIC_REG_TYPE(T) /* Register type enumerators. */ typedef enum aarch64_reg_type_ { /* A list of REG_TYPE_*. */ AARCH64_REG_TYPES } aarch64_reg_type; #undef BASIC_REG_TYPE #define BASIC_REG_TYPE(T) 1 << REG_TYPE_##T, #undef REG_TYPE #define REG_TYPE(T) (1 << REG_TYPE_##T) #undef MULTI_REG_TYPE #define MULTI_REG_TYPE(T,V) V, /* Structure for a hash table entry for a register. */ typedef struct { const char *name; unsigned char number; ENUM_BITFIELD (aarch64_reg_type_) type : 8; unsigned char builtin; } reg_entry; /* Values indexed by aarch64_reg_type to assist the type checking. */ static const unsigned reg_type_masks[] = { AARCH64_REG_TYPES }; #undef BASIC_REG_TYPE #undef REG_TYPE #undef MULTI_REG_TYPE #undef AARCH64_REG_TYPES /* Diagnostics used when we don't get a register of the expected type. Note: this has to synchronized with aarch64_reg_type definitions above. */ static const char * get_reg_expected_msg (aarch64_reg_type reg_type) { const char *msg; switch (reg_type) { case REG_TYPE_R_32: msg = N_("integer 32-bit register expected"); break; case REG_TYPE_R_64: msg = N_("integer 64-bit register expected"); break; case REG_TYPE_R_N: msg = N_("integer register expected"); break; case REG_TYPE_R64_SP: msg = N_("64-bit integer or SP register expected"); break; case REG_TYPE_SVE_BASE: msg = N_("base register expected"); break; case REG_TYPE_R_Z: msg = N_("integer or zero register expected"); break; case REG_TYPE_SVE_OFFSET: msg = N_("offset register expected"); break; case REG_TYPE_R_SP: msg = N_("integer or SP register expected"); break; case REG_TYPE_R_Z_SP: msg = N_("integer, zero or SP register expected"); break; case REG_TYPE_FP_B: msg = N_("8-bit SIMD scalar register expected"); break; case REG_TYPE_FP_H: msg = N_("16-bit SIMD scalar or floating-point half precision " "register expected"); break; case REG_TYPE_FP_S: msg = N_("32-bit SIMD scalar or floating-point single precision " "register expected"); break; case REG_TYPE_FP_D: msg = N_("64-bit SIMD scalar or floating-point double precision " "register expected"); break; case REG_TYPE_FP_Q: msg = N_("128-bit SIMD scalar or floating-point quad precision " "register expected"); break; case REG_TYPE_R_Z_BHSDQ_V: case REG_TYPE_R_Z_SP_BHSDQ_VZP: msg = N_("register expected"); break; case REG_TYPE_BHSDQ: /* any [BHSDQ]P FP */ msg = N_("SIMD scalar or floating-point register expected"); break; case REG_TYPE_VN: /* any V reg */ msg = N_("vector register expected"); break; case REG_TYPE_ZN: msg = N_("SVE vector register expected"); break; case REG_TYPE_PN: msg = N_("SVE predicate register expected"); break; default: as_fatal (_("invalid register type %d"), reg_type); } return msg; } /* Some well known registers that we refer to directly elsewhere. */ #define REG_SP 31 /* Instructions take 4 bytes in the object file. */ #define INSN_SIZE 4 static struct hash_control *aarch64_ops_hsh; static struct hash_control *aarch64_cond_hsh; static struct hash_control *aarch64_shift_hsh; static struct hash_control *aarch64_sys_regs_hsh; static struct hash_control *aarch64_pstatefield_hsh; static struct hash_control *aarch64_sys_regs_ic_hsh; static struct hash_control *aarch64_sys_regs_dc_hsh; static struct hash_control *aarch64_sys_regs_at_hsh; static struct hash_control *aarch64_sys_regs_tlbi_hsh; static struct hash_control *aarch64_reg_hsh; static struct hash_control *aarch64_barrier_opt_hsh; static struct hash_control *aarch64_nzcv_hsh; static struct hash_control *aarch64_pldop_hsh; static struct hash_control *aarch64_hint_opt_hsh; /* Stuff needed to resolve the label ambiguity As: ... label: may differ from: ... label: */ static symbolS *last_label_seen; /* Literal pool structure. Held on a per-section and per-sub-section basis. */ #define MAX_LITERAL_POOL_SIZE 1024 typedef struct literal_expression { expressionS exp; /* If exp.op == O_big then this bignum holds a copy of the global bignum value. */ LITTLENUM_TYPE * bignum; } literal_expression; typedef struct literal_pool { literal_expression literals[MAX_LITERAL_POOL_SIZE]; unsigned int next_free_entry; unsigned int id; symbolS *symbol; segT section; subsegT sub_section; int size; struct literal_pool *next; } literal_pool; /* Pointer to a linked list of literal pools. */ static literal_pool *list_of_pools = NULL; /* Pure syntax. */ /* This array holds the chars that always start a comment. If the pre-processor is disabled, these aren't very useful. */ const char comment_chars[] = ""; /* This array holds the chars that only start a comment at the beginning of a line. If the line seems to have the form '# 123 filename' .line and .file directives will appear in the pre-processed output. */ /* Note that input_file.c hand checks for '#' at the beginning of the first line of the input file. This is because the compiler outputs #NO_APP at the beginning of its output. */ /* Also note that comments like this one will always work. */ const char line_comment_chars[] = "#"; const char line_separator_chars[] = ";"; /* Chars that can be used to separate mant from exp in floating point numbers. */ const char EXP_CHARS[] = "eE"; /* Chars that mean this number is a floating point constant. */ /* As in 0f12.456 */ /* or 0d1.2345e12 */ const char FLT_CHARS[] = "rRsSfFdDxXeEpP"; /* Prefix character that indicates the start of an immediate value. */ #define is_immediate_prefix(C) ((C) == '#') /* Separator character handling. */ #define skip_whitespace(str) do { if (*(str) == ' ') ++(str); } while (0) static inline bfd_boolean skip_past_char (char **str, char c) { if (**str == c) { (*str)++; return TRUE; } else return FALSE; } #define skip_past_comma(str) skip_past_char (str, ',') /* Arithmetic expressions (possibly involving symbols). */ static bfd_boolean in_my_get_expression_p = FALSE; /* Third argument to my_get_expression. */ #define GE_NO_PREFIX 0 #define GE_OPT_PREFIX 1 /* Return TRUE if the string pointed by *STR is successfully parsed as an valid expression; *EP will be filled with the information of such an expression. Otherwise return FALSE. */ static bfd_boolean my_get_expression (expressionS * ep, char **str, int prefix_mode, int reject_absent) { char *save_in; segT seg; int prefix_present_p = 0; switch (prefix_mode) { case GE_NO_PREFIX: break; case GE_OPT_PREFIX: if (is_immediate_prefix (**str)) { (*str)++; prefix_present_p = 1; } break; default: abort (); } memset (ep, 0, sizeof (expressionS)); save_in = input_line_pointer; input_line_pointer = *str; in_my_get_expression_p = TRUE; seg = expression (ep); in_my_get_expression_p = FALSE; if (ep->X_op == O_illegal || (reject_absent && ep->X_op == O_absent)) { /* We found a bad expression in md_operand(). */ *str = input_line_pointer; input_line_pointer = save_in; if (prefix_present_p && ! error_p ()) set_fatal_syntax_error (_("bad expression")); else set_first_syntax_error (_("bad expression")); return FALSE; } #ifdef OBJ_AOUT if (seg != absolute_section && seg != text_section && seg != data_section && seg != bss_section && seg != undefined_section) { set_syntax_error (_("bad segment")); *str = input_line_pointer; input_line_pointer = save_in; return FALSE; } #else (void) seg; #endif *str = input_line_pointer; input_line_pointer = save_in; return TRUE; } /* Turn a string in input_line_pointer into a floating point constant of type TYPE, and store the appropriate bytes in *LITP. The number of LITTLENUMS emitted is stored in *SIZEP. An error message is returned, or NULL on OK. */ const char * md_atof (int type, char *litP, int *sizeP) { return ieee_md_atof (type, litP, sizeP, target_big_endian); } /* We handle all bad expressions here, so that we can report the faulty instruction in the error message. */ void md_operand (expressionS * exp) { if (in_my_get_expression_p) exp->X_op = O_illegal; } /* Immediate values. */ /* Errors may be set multiple times during parsing or bit encoding (particularly in the Neon bits), but usually the earliest error which is set will be the most meaningful. Avoid overwriting it with later (cascading) errors by calling this function. */ static void first_error (const char *error) { if (! error_p ()) set_syntax_error (error); } /* Similar to first_error, but this function accepts formatted error message. */ static void first_error_fmt (const char *format, ...) { va_list args; enum { size = 100 }; /* N.B. this single buffer will not cause error messages for different instructions to pollute each other; this is because at the end of processing of each assembly line, error message if any will be collected by as_bad. */ static char buffer[size]; if (! error_p ()) { int ret ATTRIBUTE_UNUSED; va_start (args, format); ret = vsnprintf (buffer, size, format, args); know (ret <= size - 1 && ret >= 0); va_end (args); set_syntax_error (buffer); } } /* Register parsing. */ /* Generic register parser which is called by other specialized register parsers. CCP points to what should be the beginning of a register name. If it is indeed a valid register name, advance CCP over it and return the reg_entry structure; otherwise return NULL. It does not issue diagnostics. */ static reg_entry * parse_reg (char **ccp) { char *start = *ccp; char *p; reg_entry *reg; #ifdef REGISTER_PREFIX if (*start != REGISTER_PREFIX) return NULL; start++; #endif p = start; if (!ISALPHA (*p) || !is_name_beginner (*p)) return NULL; do p++; while (ISALPHA (*p) || ISDIGIT (*p) || *p == '_'); reg = (reg_entry *) hash_find_n (aarch64_reg_hsh, start, p - start); if (!reg) return NULL; *ccp = p; return reg; } /* Return TRUE if REG->TYPE is a valid type of TYPE; otherwise return FALSE. */ static bfd_boolean aarch64_check_reg_type (const reg_entry *reg, aarch64_reg_type type) { return (reg_type_masks[type] & (1 << reg->type)) != 0; } /* Try to parse a base or offset register. Allow SVE base and offset registers if REG_TYPE includes SVE registers. Return the register entry on success, setting *QUALIFIER to the register qualifier. Return null otherwise. Note that this function does not issue any diagnostics. */ static const reg_entry * aarch64_addr_reg_parse (char **ccp, aarch64_reg_type reg_type, aarch64_opnd_qualifier_t *qualifier) { char *str = *ccp; const reg_entry *reg = parse_reg (&str); if (reg == NULL) return NULL; switch (reg->type) { case REG_TYPE_R_32: case REG_TYPE_SP_32: case REG_TYPE_Z_32: *qualifier = AARCH64_OPND_QLF_W; break; case REG_TYPE_R_64: case REG_TYPE_SP_64: case REG_TYPE_Z_64: *qualifier = AARCH64_OPND_QLF_X; break; case REG_TYPE_ZN: if ((reg_type_masks[reg_type] & (1 << REG_TYPE_ZN)) == 0 || str[0] != '.') return NULL; switch (TOLOWER (str[1])) { case 's': *qualifier = AARCH64_OPND_QLF_S_S; break; case 'd': *qualifier = AARCH64_OPND_QLF_S_D; break; default: return NULL; } str += 2; break; default: return NULL; } *ccp = str; return reg; } /* Try to parse a base or offset register. Return the register entry on success, setting *QUALIFIER to the register qualifier. Return null otherwise. Note that this function does not issue any diagnostics. */ static const reg_entry * aarch64_reg_parse_32_64 (char **ccp, aarch64_opnd_qualifier_t *qualifier) { return aarch64_addr_reg_parse (ccp, REG_TYPE_R_Z_SP, qualifier); } /* Parse the qualifier of a vector register or vector element of type REG_TYPE. Fill in *PARSED_TYPE and return TRUE if the parsing succeeds; otherwise return FALSE. Accept only one occurrence of: 4b 8b 16b 2h 4h 8h 2s 4s 1d 2d b h s d q */ static bfd_boolean parse_vector_type_for_operand (aarch64_reg_type reg_type, struct vector_type_el *parsed_type, char **str) { char *ptr = *str; unsigned width; unsigned element_size; enum vector_el_type type; /* skip '.' */ gas_assert (*ptr == '.'); ptr++; if (reg_type == REG_TYPE_ZN || reg_type == REG_TYPE_PN || !ISDIGIT (*ptr)) { width = 0; goto elt_size; } width = strtoul (ptr, &ptr, 10); if (width != 1 && width != 2 && width != 4 && width != 8 && width != 16) { first_error_fmt (_("bad size %d in vector width specifier"), width); return FALSE; } elt_size: switch (TOLOWER (*ptr)) { case 'b': type = NT_b; element_size = 8; break; case 'h': type = NT_h; element_size = 16; break; case 's': type = NT_s; element_size = 32; break; case 'd': type = NT_d; element_size = 64; break; case 'q': if (reg_type == REG_TYPE_ZN || width == 1) { type = NT_q; element_size = 128; break; } /* fall through. */ default: if (*ptr != '\0') first_error_fmt (_("unexpected character `%c' in element size"), *ptr); else first_error (_("missing element size")); return FALSE; } if (width != 0 && width * element_size != 64 && width * element_size != 128 && !(width == 2 && element_size == 16) && !(width == 4 && element_size == 8)) { first_error_fmt (_ ("invalid element size %d and vector size combination %c"), width, *ptr); return FALSE; } ptr++; parsed_type->type = type; parsed_type->width = width; *str = ptr; return TRUE; } /* *STR contains an SVE zero/merge predication suffix. Parse it into *PARSED_TYPE and point *STR at the end of the suffix. */ static bfd_boolean parse_predication_for_operand (struct vector_type_el *parsed_type, char **str) { char *ptr = *str; /* Skip '/'. */ gas_assert (*ptr == '/'); ptr++; switch (TOLOWER (*ptr)) { case 'z': parsed_type->type = NT_zero; break; case 'm': parsed_type->type = NT_merge; break; default: if (*ptr != '\0' && *ptr != ',') first_error_fmt (_("unexpected character `%c' in predication type"), *ptr); else first_error (_("missing predication type")); return FALSE; } parsed_type->width = 0; *str = ptr + 1; return TRUE; } /* Parse a register of the type TYPE. Return PARSE_FAIL if the string pointed by *CCP is not a valid register name or the parsed register is not of TYPE. Otherwise return the register number, and optionally fill in the actual type of the register in *RTYPE when multiple alternatives were given, and return the register shape and element index information in *TYPEINFO. IN_REG_LIST should be set with TRUE if the caller is parsing a register list. */ static int parse_typed_reg (char **ccp, aarch64_reg_type type, aarch64_reg_type *rtype, struct vector_type_el *typeinfo, bfd_boolean in_reg_list) { char *str = *ccp; const reg_entry *reg = parse_reg (&str); struct vector_type_el atype; struct vector_type_el parsetype; bfd_boolean is_typed_vecreg = FALSE; atype.defined = 0; atype.type = NT_invtype; atype.width = -1; atype.index = 0; if (reg == NULL) { if (typeinfo) *typeinfo = atype; set_default_error (); return PARSE_FAIL; } if (! aarch64_check_reg_type (reg, type)) { DEBUG_TRACE ("reg type check failed"); set_default_error (); return PARSE_FAIL; } type = reg->type; if ((type == REG_TYPE_VN || type == REG_TYPE_ZN || type == REG_TYPE_PN) && (*str == '.' || (type == REG_TYPE_PN && *str == '/'))) { if (*str == '.') { if (!parse_vector_type_for_operand (type, &parsetype, &str)) return PARSE_FAIL; } else { if (!parse_predication_for_operand (&parsetype, &str)) return PARSE_FAIL; } /* Register if of the form Vn.[bhsdq]. */ is_typed_vecreg = TRUE; if (type == REG_TYPE_ZN || type == REG_TYPE_PN) { /* The width is always variable; we don't allow an integer width to be specified. */ gas_assert (parsetype.width == 0); atype.defined |= NTA_HASVARWIDTH | NTA_HASTYPE; } else if (parsetype.width == 0) /* Expect index. In the new scheme we cannot have Vn.[bhsdq] represent a scalar. Therefore any Vn.[bhsdq] should have an index following it. Except in reglists of course. */ atype.defined |= NTA_HASINDEX; else atype.defined |= NTA_HASTYPE; atype.type = parsetype.type; atype.width = parsetype.width; } if (skip_past_char (&str, '[')) { expressionS exp; /* Reject Sn[index] syntax. */ if (!is_typed_vecreg) { first_error (_("this type of register can't be indexed")); return PARSE_FAIL; } if (in_reg_list) { first_error (_("index not allowed inside register list")); return PARSE_FAIL; } atype.defined |= NTA_HASINDEX; my_get_expression (&exp, &str, GE_NO_PREFIX, 1); if (exp.X_op != O_constant) { first_error (_("constant expression required")); return PARSE_FAIL; } if (! skip_past_char (&str, ']')) return PARSE_FAIL; atype.index = exp.X_add_number; } else if (!in_reg_list && (atype.defined & NTA_HASINDEX) != 0) { /* Indexed vector register expected. */ first_error (_("indexed vector register expected")); return PARSE_FAIL; } /* A vector reg Vn should be typed or indexed. */ if (type == REG_TYPE_VN && atype.defined == 0) { first_error (_("invalid use of vector register")); } if (typeinfo) *typeinfo = atype; if (rtype) *rtype = type; *ccp = str; return reg->number; } /* Parse register. Return the register number on success; return PARSE_FAIL otherwise. If RTYPE is not NULL, return in *RTYPE the (possibly restricted) type of the register (e.g. NEON double or quad reg when either has been requested). If this is a NEON vector register with additional type information, fill in the struct pointed to by VECTYPE (if non-NULL). This parser does not handle register list. */ static int aarch64_reg_parse (char **ccp, aarch64_reg_type type, aarch64_reg_type *rtype, struct vector_type_el *vectype) { struct vector_type_el atype; char *str = *ccp; int reg = parse_typed_reg (&str, type, rtype, &atype, /*in_reg_list= */ FALSE); if (reg == PARSE_FAIL) return PARSE_FAIL; if (vectype) *vectype = atype; *ccp = str; return reg; } static inline bfd_boolean eq_vector_type_el (struct vector_type_el e1, struct vector_type_el e2) { return e1.type == e2.type && e1.defined == e2.defined && e1.width == e2.width && e1.index == e2.index; } /* This function parses a list of vector registers of type TYPE. On success, it returns the parsed register list information in the following encoded format: bit 18-22 | 13-17 | 7-11 | 2-6 | 0-1 4th regno | 3rd regno | 2nd regno | 1st regno | num_of_reg The information of the register shape and/or index is returned in *VECTYPE. It returns PARSE_FAIL if the register list is invalid. The list contains one to four registers. Each register can be one of: .[] . All should be identical. All should be identical. There are restrictions on numbers which are checked later (by reg_list_valid_p). */ static int parse_vector_reg_list (char **ccp, aarch64_reg_type type, struct vector_type_el *vectype) { char *str = *ccp; int nb_regs; struct vector_type_el typeinfo, typeinfo_first; int val, val_range; int in_range; int ret_val; int i; bfd_boolean error = FALSE; bfd_boolean expect_index = FALSE; if (*str != '{') { set_syntax_error (_("expecting {")); return PARSE_FAIL; } str++; nb_regs = 0; typeinfo_first.defined = 0; typeinfo_first.type = NT_invtype; typeinfo_first.width = -1; typeinfo_first.index = 0; ret_val = 0; val = -1; val_range = -1; in_range = 0; do { if (in_range) { str++; /* skip over '-' */ val_range = val; } val = parse_typed_reg (&str, type, NULL, &typeinfo, /*in_reg_list= */ TRUE); if (val == PARSE_FAIL) { set_first_syntax_error (_("invalid vector register in list")); error = TRUE; continue; } /* reject [bhsd]n */ if (type == REG_TYPE_VN && typeinfo.defined == 0) { set_first_syntax_error (_("invalid scalar register in list")); error = TRUE; continue; } if (typeinfo.defined & NTA_HASINDEX) expect_index = TRUE; if (in_range) { if (val < val_range) { set_first_syntax_error (_("invalid range in vector register list")); error = TRUE; } val_range++; } else { val_range = val; if (nb_regs == 0) typeinfo_first = typeinfo; else if (! eq_vector_type_el (typeinfo_first, typeinfo)) { set_first_syntax_error (_("type mismatch in vector register list")); error = TRUE; } } if (! error) for (i = val_range; i <= val; i++) { ret_val |= i << (5 * nb_regs); nb_regs++; } in_range = 0; } while (skip_past_comma (&str) || (in_range = 1, *str == '-')); skip_whitespace (str); if (*str != '}') { set_first_syntax_error (_("end of vector register list not found")); error = TRUE; } str++; skip_whitespace (str); if (expect_index) { if (skip_past_char (&str, '[')) { expressionS exp; my_get_expression (&exp, &str, GE_NO_PREFIX, 1); if (exp.X_op != O_constant) { set_first_syntax_error (_("constant expression required.")); error = TRUE; } if (! skip_past_char (&str, ']')) error = TRUE; else typeinfo_first.index = exp.X_add_number; } else { set_first_syntax_error (_("expected index")); error = TRUE; } } if (nb_regs > 4) { set_first_syntax_error (_("too many registers in vector register list")); error = TRUE; } else if (nb_regs == 0) { set_first_syntax_error (_("empty vector register list")); error = TRUE; } *ccp = str; if (! error) *vectype = typeinfo_first; return error ? PARSE_FAIL : (ret_val << 2) | (nb_regs - 1); } /* Directives: register aliases. */ static reg_entry * insert_reg_alias (char *str, int number, aarch64_reg_type type) { reg_entry *new; const char *name; if ((new = hash_find (aarch64_reg_hsh, str)) != 0) { if (new->builtin) as_warn (_("ignoring attempt to redefine built-in register '%s'"), str); /* Only warn about a redefinition if it's not defined as the same register. */ else if (new->number != number || new->type != type) as_warn (_("ignoring redefinition of register alias '%s'"), str); return NULL; } name = xstrdup (str); new = XNEW (reg_entry); new->name = name; new->number = number; new->type = type; new->builtin = FALSE; if (hash_insert (aarch64_reg_hsh, name, (void *) new)) abort (); return new; } /* Look for the .req directive. This is of the form: new_register_name .req existing_register_name If we find one, or if it looks sufficiently like one that we want to handle any error here, return TRUE. Otherwise return FALSE. */ static bfd_boolean create_register_alias (char *newname, char *p) { const reg_entry *old; char *oldname, *nbuf; size_t nlen; /* The input scrubber ensures that whitespace after the mnemonic is collapsed to single spaces. */ oldname = p; if (strncmp (oldname, " .req ", 6) != 0) return FALSE; oldname += 6; if (*oldname == '\0') return FALSE; old = hash_find (aarch64_reg_hsh, oldname); if (!old) { as_warn (_("unknown register '%s' -- .req ignored"), oldname); return TRUE; } /* If TC_CASE_SENSITIVE is defined, then newname already points to the desired alias name, and p points to its end. If not, then the desired alias name is in the global original_case_string. */ #ifdef TC_CASE_SENSITIVE nlen = p - newname; #else newname = original_case_string; nlen = strlen (newname); #endif nbuf = xmemdup0 (newname, nlen); /* Create aliases under the new name as stated; an all-lowercase version of the new name; and an all-uppercase version of the new name. */ if (insert_reg_alias (nbuf, old->number, old->type) != NULL) { for (p = nbuf; *p; p++) *p = TOUPPER (*p); if (strncmp (nbuf, newname, nlen)) { /* If this attempt to create an additional alias fails, do not bother trying to create the all-lower case alias. We will fail and issue a second, duplicate error message. This situation arises when the programmer does something like: foo .req r0 Foo .req r1 The second .req creates the "Foo" alias but then fails to create the artificial FOO alias because it has already been created by the first .req. */ if (insert_reg_alias (nbuf, old->number, old->type) == NULL) { free (nbuf); return TRUE; } } for (p = nbuf; *p; p++) *p = TOLOWER (*p); if (strncmp (nbuf, newname, nlen)) insert_reg_alias (nbuf, old->number, old->type); } free (nbuf); return TRUE; } /* Should never be called, as .req goes between the alias and the register name, not at the beginning of the line. */ static void s_req (int a ATTRIBUTE_UNUSED) { as_bad (_("invalid syntax for .req directive")); } /* The .unreq directive deletes an alias which was previously defined by .req. For example: my_alias .req r11 .unreq my_alias */ static void s_unreq (int a ATTRIBUTE_UNUSED) { char *name; char saved_char; name = input_line_pointer; while (*input_line_pointer != 0 && *input_line_pointer != ' ' && *input_line_pointer != '\n') ++input_line_pointer; saved_char = *input_line_pointer; *input_line_pointer = 0; if (!*name) as_bad (_("invalid syntax for .unreq directive")); else { reg_entry *reg = hash_find (aarch64_reg_hsh, name); if (!reg) as_bad (_("unknown register alias '%s'"), name); else if (reg->builtin) as_warn (_("ignoring attempt to undefine built-in register '%s'"), name); else { char *p; char *nbuf; hash_delete (aarch64_reg_hsh, name, FALSE); free ((char *) reg->name); free (reg); /* Also locate the all upper case and all lower case versions. Do not complain if we cannot find one or the other as it was probably deleted above. */ nbuf = strdup (name); for (p = nbuf; *p; p++) *p = TOUPPER (*p); reg = hash_find (aarch64_reg_hsh, nbuf); if (reg) { hash_delete (aarch64_reg_hsh, nbuf, FALSE); free ((char *) reg->name); free (reg); } for (p = nbuf; *p; p++) *p = TOLOWER (*p); reg = hash_find (aarch64_reg_hsh, nbuf); if (reg) { hash_delete (aarch64_reg_hsh, nbuf, FALSE); free ((char *) reg->name); free (reg); } free (nbuf); } } *input_line_pointer = saved_char; demand_empty_rest_of_line (); } /* Directives: Instruction set selection. */ #ifdef OBJ_ELF /* This code is to handle mapping symbols as defined in the ARM AArch64 ELF spec. (See "Mapping symbols", section 4.5.4, ARM AAELF64 version 0.05). Note that previously, $a and $t has type STT_FUNC (BSF_OBJECT flag), and $d has type STT_OBJECT (BSF_OBJECT flag). Now all three are untyped. */ /* Create a new mapping symbol for the transition to STATE. */ static void make_mapping_symbol (enum mstate state, valueT value, fragS * frag) { symbolS *symbolP; const char *symname; int type; switch (state) { case MAP_DATA: symname = "$d"; type = BSF_NO_FLAGS; break; case MAP_INSN: symname = "$x"; type = BSF_NO_FLAGS; break; default: abort (); } symbolP = symbol_new (symname, now_seg, value, frag); symbol_get_bfdsym (symbolP)->flags |= type | BSF_LOCAL; /* Save the mapping symbols for future reference. Also check that we do not place two mapping symbols at the same offset within a frag. We'll handle overlap between frags in check_mapping_symbols. If .fill or other data filling directive generates zero sized data, the mapping symbol for the following code will have the same value as the one generated for the data filling directive. In this case, we replace the old symbol with the new one at the same address. */ if (value == 0) { if (frag->tc_frag_data.first_map != NULL) { know (S_GET_VALUE (frag->tc_frag_data.first_map) == 0); symbol_remove (frag->tc_frag_data.first_map, &symbol_rootP, &symbol_lastP); } frag->tc_frag_data.first_map = symbolP; } if (frag->tc_frag_data.last_map != NULL) { know (S_GET_VALUE (frag->tc_frag_data.last_map) <= S_GET_VALUE (symbolP)); if (S_GET_VALUE (frag->tc_frag_data.last_map) == S_GET_VALUE (symbolP)) symbol_remove (frag->tc_frag_data.last_map, &symbol_rootP, &symbol_lastP); } frag->tc_frag_data.last_map = symbolP; } /* We must sometimes convert a region marked as code to data during code alignment, if an odd number of bytes have to be padded. The code mapping symbol is pushed to an aligned address. */ static void insert_data_mapping_symbol (enum mstate state, valueT value, fragS * frag, offsetT bytes) { /* If there was already a mapping symbol, remove it. */ if (frag->tc_frag_data.last_map != NULL && S_GET_VALUE (frag->tc_frag_data.last_map) == frag->fr_address + value) { symbolS *symp = frag->tc_frag_data.last_map; if (value == 0) { know (frag->tc_frag_data.first_map == symp); frag->tc_frag_data.first_map = NULL; } frag->tc_frag_data.last_map = NULL; symbol_remove (symp, &symbol_rootP, &symbol_lastP); } make_mapping_symbol (MAP_DATA, value, frag); make_mapping_symbol (state, value + bytes, frag); } static void mapping_state_2 (enum mstate state, int max_chars); /* Set the mapping state to STATE. Only call this when about to emit some STATE bytes to the file. */ void mapping_state (enum mstate state) { enum mstate mapstate = seg_info (now_seg)->tc_segment_info_data.mapstate; if (state == MAP_INSN) /* AArch64 instructions require 4-byte alignment. When emitting instructions into any section, record the appropriate section alignment. */ record_alignment (now_seg, 2); if (mapstate == state) /* The mapping symbol has already been emitted. There is nothing else to do. */ return; #define TRANSITION(from, to) (mapstate == (from) && state == (to)) if (TRANSITION (MAP_UNDEFINED, MAP_DATA) && !subseg_text_p (now_seg)) /* Emit MAP_DATA within executable section in order. Otherwise, it will be evaluated later in the next else. */ return; else if (TRANSITION (MAP_UNDEFINED, MAP_INSN)) { /* Only add the symbol if the offset is > 0: if we're at the first frag, check it's size > 0; if we're not at the first frag, then for sure the offset is > 0. */ struct frag *const frag_first = seg_info (now_seg)->frchainP->frch_root; const int add_symbol = (frag_now != frag_first) || (frag_now_fix () > 0); if (add_symbol) make_mapping_symbol (MAP_DATA, (valueT) 0, frag_first); } #undef TRANSITION mapping_state_2 (state, 0); } /* Same as mapping_state, but MAX_CHARS bytes have already been allocated. Put the mapping symbol that far back. */ static void mapping_state_2 (enum mstate state, int max_chars) { enum mstate mapstate = seg_info (now_seg)->tc_segment_info_data.mapstate; if (!SEG_NORMAL (now_seg)) return; if (mapstate == state) /* The mapping symbol has already been emitted. There is nothing else to do. */ return; seg_info (now_seg)->tc_segment_info_data.mapstate = state; make_mapping_symbol (state, (valueT) frag_now_fix () - max_chars, frag_now); } #else #define mapping_state(x) /* nothing */ #define mapping_state_2(x, y) /* nothing */ #endif /* Directives: sectioning and alignment. */ static void s_bss (int ignore ATTRIBUTE_UNUSED) { /* We don't support putting frags in the BSS segment, we fake it by marking in_bss, then looking at s_skip for clues. */ subseg_set (bss_section, 0); demand_empty_rest_of_line (); mapping_state (MAP_DATA); } static void s_even (int ignore ATTRIBUTE_UNUSED) { /* Never make frag if expect extra pass. */ if (!need_pass_2) frag_align (1, 0, 0); record_alignment (now_seg, 1); demand_empty_rest_of_line (); } /* Directives: Literal pools. */ static literal_pool * find_literal_pool (int size) { literal_pool *pool; for (pool = list_of_pools; pool != NULL; pool = pool->next) { if (pool->section == now_seg && pool->sub_section == now_subseg && pool->size == size) break; } return pool; } static literal_pool * find_or_make_literal_pool (int size) { /* Next literal pool ID number. */ static unsigned int latest_pool_num = 1; literal_pool *pool; pool = find_literal_pool (size); if (pool == NULL) { /* Create a new pool. */ pool = XNEW (literal_pool); if (!pool) return NULL; /* Currently we always put the literal pool in the current text section. If we were generating "small" model code where we knew that all code and initialised data was within 1MB then we could output literals to mergeable, read-only data sections. */ pool->next_free_entry = 0; pool->section = now_seg; pool->sub_section = now_subseg; pool->size = size; pool->next = list_of_pools; pool->symbol = NULL; /* Add it to the list. */ list_of_pools = pool; } /* New pools, and emptied pools, will have a NULL symbol. */ if (pool->symbol == NULL) { pool->symbol = symbol_create (FAKE_LABEL_NAME, undefined_section, (valueT) 0, &zero_address_frag); pool->id = latest_pool_num++; } /* Done. */ return pool; } /* Add the literal of size SIZE in *EXP to the relevant literal pool. Return TRUE on success, otherwise return FALSE. */ static bfd_boolean add_to_lit_pool (expressionS *exp, int size) { literal_pool *pool; unsigned int entry; pool = find_or_make_literal_pool (size); /* Check if this literal value is already in the pool. */ for (entry = 0; entry < pool->next_free_entry; entry++) { expressionS * litexp = & pool->literals[entry].exp; if ((litexp->X_op == exp->X_op) && (exp->X_op == O_constant) && (litexp->X_add_number == exp->X_add_number) && (litexp->X_unsigned == exp->X_unsigned)) break; if ((litexp->X_op == exp->X_op) && (exp->X_op == O_symbol) && (litexp->X_add_number == exp->X_add_number) && (litexp->X_add_symbol == exp->X_add_symbol) && (litexp->X_op_symbol == exp->X_op_symbol)) break; } /* Do we need to create a new entry? */ if (entry == pool->next_free_entry) { if (entry >= MAX_LITERAL_POOL_SIZE) { set_syntax_error (_("literal pool overflow")); return FALSE; } pool->literals[entry].exp = *exp; pool->next_free_entry += 1; if (exp->X_op == O_big) { /* PR 16688: Bignums are held in a single global array. We must copy and preserve that value now, before it is overwritten. */ pool->literals[entry].bignum = XNEWVEC (LITTLENUM_TYPE, exp->X_add_number); memcpy (pool->literals[entry].bignum, generic_bignum, CHARS_PER_LITTLENUM * exp->X_add_number); } else pool->literals[entry].bignum = NULL; } exp->X_op = O_symbol; exp->X_add_number = ((int) entry) * size; exp->X_add_symbol = pool->symbol; return TRUE; } /* Can't use symbol_new here, so have to create a symbol and then at a later date assign it a value. That's what these functions do. */ static void symbol_locate (symbolS * symbolP, const char *name,/* It is copied, the caller can modify. */ segT segment, /* Segment identifier (SEG_). */ valueT valu, /* Symbol value. */ fragS * frag) /* Associated fragment. */ { size_t name_length; char *preserved_copy_of_name; name_length = strlen (name) + 1; /* +1 for \0. */ obstack_grow (¬es, name, name_length); preserved_copy_of_name = obstack_finish (¬es); #ifdef tc_canonicalize_symbol_name preserved_copy_of_name = tc_canonicalize_symbol_name (preserved_copy_of_name); #endif S_SET_NAME (symbolP, preserved_copy_of_name); S_SET_SEGMENT (symbolP, segment); S_SET_VALUE (symbolP, valu); symbol_clear_list_pointers (symbolP); symbol_set_frag (symbolP, frag); /* Link to end of symbol chain. */ { extern int symbol_table_frozen; if (symbol_table_frozen) abort (); } symbol_append (symbolP, symbol_lastP, &symbol_rootP, &symbol_lastP); obj_symbol_new_hook (symbolP); #ifdef tc_symbol_new_hook tc_symbol_new_hook (symbolP); #endif #ifdef DEBUG_SYMS verify_symbol_chain (symbol_rootP, symbol_lastP); #endif /* DEBUG_SYMS */ } static void s_ltorg (int ignored ATTRIBUTE_UNUSED) { unsigned int entry; literal_pool *pool; char sym_name[20]; int align; for (align = 2; align <= 4; align++) { int size = 1 << align; pool = find_literal_pool (size); if (pool == NULL || pool->symbol == NULL || pool->next_free_entry == 0) continue; /* Align pool as you have word accesses. Only make a frag if we have to. */ if (!need_pass_2) frag_align (align, 0, 0); mapping_state (MAP_DATA); record_alignment (now_seg, align); sprintf (sym_name, "$$lit_\002%x", pool->id); symbol_locate (pool->symbol, sym_name, now_seg, (valueT) frag_now_fix (), frag_now); symbol_table_insert (pool->symbol); for (entry = 0; entry < pool->next_free_entry; entry++) { expressionS * exp = & pool->literals[entry].exp; if (exp->X_op == O_big) { /* PR 16688: Restore the global bignum value. */ gas_assert (pool->literals[entry].bignum != NULL); memcpy (generic_bignum, pool->literals[entry].bignum, CHARS_PER_LITTLENUM * exp->X_add_number); } /* First output the expression in the instruction to the pool. */ emit_expr (exp, size); /* .word|.xword */ if (exp->X_op == O_big) { free (pool->literals[entry].bignum); pool->literals[entry].bignum = NULL; } } /* Mark the pool as empty. */ pool->next_free_entry = 0; pool->symbol = NULL; } } #ifdef OBJ_ELF /* Forward declarations for functions below, in the MD interface section. */ static fixS *fix_new_aarch64 (fragS *, int, short, expressionS *, int, int); static struct reloc_table_entry * find_reloc_table_entry (char **); /* Directives: Data. */ /* N.B. the support for relocation suffix in this directive needs to be implemented properly. */ static void s_aarch64_elf_cons (int nbytes) { expressionS exp; #ifdef md_flush_pending_output md_flush_pending_output (); #endif if (is_it_end_of_statement ()) { demand_empty_rest_of_line (); return; } #ifdef md_cons_align md_cons_align (nbytes); #endif mapping_state (MAP_DATA); do { struct reloc_table_entry *reloc; expression (&exp); if (exp.X_op != O_symbol) emit_expr (&exp, (unsigned int) nbytes); else { skip_past_char (&input_line_pointer, '#'); if (skip_past_char (&input_line_pointer, ':')) { reloc = find_reloc_table_entry (&input_line_pointer); if (reloc == NULL) as_bad (_("unrecognized relocation suffix")); else as_bad (_("unimplemented relocation suffix")); ignore_rest_of_line (); return; } else emit_expr (&exp, (unsigned int) nbytes); } } while (*input_line_pointer++ == ','); /* Put terminator back into stream. */ input_line_pointer--; demand_empty_rest_of_line (); } #endif /* OBJ_ELF */ /* Output a 32-bit word, but mark as an instruction. */ static void s_aarch64_inst (int ignored ATTRIBUTE_UNUSED) { expressionS exp; #ifdef md_flush_pending_output md_flush_pending_output (); #endif if (is_it_end_of_statement ()) { demand_empty_rest_of_line (); return; } /* Sections are assumed to start aligned. In executable section, there is no MAP_DATA symbol pending. So we only align the address during MAP_DATA --> MAP_INSN transition. For other sections, this is not guaranteed. */ enum mstate mapstate = seg_info (now_seg)->tc_segment_info_data.mapstate; if (!need_pass_2 && subseg_text_p (now_seg) && mapstate == MAP_DATA) frag_align_code (2, 0); #ifdef OBJ_ELF mapping_state (MAP_INSN); #endif do { expression (&exp); if (exp.X_op != O_constant) { as_bad (_("constant expression required")); ignore_rest_of_line (); return; } if (target_big_endian) { unsigned int val = exp.X_add_number; exp.X_add_number = SWAP_32 (val); } emit_expr (&exp, 4); } while (*input_line_pointer++ == ','); /* Put terminator back into stream. */ input_line_pointer--; demand_empty_rest_of_line (); } #ifdef OBJ_ELF /* Emit BFD_RELOC_AARCH64_TLSDESC_ADD on the next ADD instruction. */ static void s_tlsdescadd (int ignored ATTRIBUTE_UNUSED) { expressionS exp; expression (&exp); frag_grow (4); fix_new_aarch64 (frag_now, frag_more (0) - frag_now->fr_literal, 4, &exp, 0, BFD_RELOC_AARCH64_TLSDESC_ADD); demand_empty_rest_of_line (); } /* Emit BFD_RELOC_AARCH64_TLSDESC_CALL on the next BLR instruction. */ static void s_tlsdesccall (int ignored ATTRIBUTE_UNUSED) { expressionS exp; /* Since we're just labelling the code, there's no need to define a mapping symbol. */ expression (&exp); /* Make sure there is enough room in this frag for the following blr. This trick only works if the blr follows immediately after the .tlsdesc directive. */ frag_grow (4); fix_new_aarch64 (frag_now, frag_more (0) - frag_now->fr_literal, 4, &exp, 0, BFD_RELOC_AARCH64_TLSDESC_CALL); demand_empty_rest_of_line (); } /* Emit BFD_RELOC_AARCH64_TLSDESC_LDR on the next LDR instruction. */ static void s_tlsdescldr (int ignored ATTRIBUTE_UNUSED) { expressionS exp; expression (&exp); frag_grow (4); fix_new_aarch64 (frag_now, frag_more (0) - frag_now->fr_literal, 4, &exp, 0, BFD_RELOC_AARCH64_TLSDESC_LDR); demand_empty_rest_of_line (); } #endif /* OBJ_ELF */ static void s_aarch64_arch (int); static void s_aarch64_cpu (int); static void s_aarch64_arch_extension (int); /* This table describes all the machine specific pseudo-ops the assembler has to support. The fields are: pseudo-op name without dot function to call to execute this pseudo-op Integer arg to pass to the function. */ const pseudo_typeS md_pseudo_table[] = { /* Never called because '.req' does not start a line. */ {"req", s_req, 0}, {"unreq", s_unreq, 0}, {"bss", s_bss, 0}, {"even", s_even, 0}, {"ltorg", s_ltorg, 0}, {"pool", s_ltorg, 0}, {"cpu", s_aarch64_cpu, 0}, {"arch", s_aarch64_arch, 0}, {"arch_extension", s_aarch64_arch_extension, 0}, {"inst", s_aarch64_inst, 0}, #ifdef OBJ_ELF {"tlsdescadd", s_tlsdescadd, 0}, {"tlsdesccall", s_tlsdesccall, 0}, {"tlsdescldr", s_tlsdescldr, 0}, {"word", s_aarch64_elf_cons, 4}, {"long", s_aarch64_elf_cons, 4}, {"xword", s_aarch64_elf_cons, 8}, {"dword", s_aarch64_elf_cons, 8}, #endif {0, 0, 0} }; /* Check whether STR points to a register name followed by a comma or the end of line; REG_TYPE indicates which register types are checked against. Return TRUE if STR is such a register name; otherwise return FALSE. The function does not intend to produce any diagnostics, but since the register parser aarch64_reg_parse, which is called by this function, does produce diagnostics, we call clear_error to clear any diagnostics that may be generated by aarch64_reg_parse. Also, the function returns FALSE directly if there is any user error present at the function entry. This prevents the existing diagnostics state from being spoiled. The function currently serves parse_constant_immediate and parse_big_immediate only. */ static bfd_boolean reg_name_p (char *str, aarch64_reg_type reg_type) { int reg; /* Prevent the diagnostics state from being spoiled. */ if (error_p ()) return FALSE; reg = aarch64_reg_parse (&str, reg_type, NULL, NULL); /* Clear the parsing error that may be set by the reg parser. */ clear_error (); if (reg == PARSE_FAIL) return FALSE; skip_whitespace (str); if (*str == ',' || is_end_of_line[(unsigned int) *str]) return TRUE; return FALSE; } /* Parser functions used exclusively in instruction operands. */ /* Parse an immediate expression which may not be constant. To prevent the expression parser from pushing a register name into the symbol table as an undefined symbol, firstly a check is done to find out whether STR is a register of type REG_TYPE followed by a comma or the end of line. Return FALSE if STR is such a string. */ static bfd_boolean parse_immediate_expression (char **str, expressionS *exp, aarch64_reg_type reg_type) { if (reg_name_p (*str, reg_type)) { set_recoverable_error (_("immediate operand required")); return FALSE; } my_get_expression (exp, str, GE_OPT_PREFIX, 1); if (exp->X_op == O_absent) { set_fatal_syntax_error (_("missing immediate expression")); return FALSE; } return TRUE; } /* Constant immediate-value read function for use in insn parsing. STR points to the beginning of the immediate (with the optional leading #); *VAL receives the value. REG_TYPE says which register names should be treated as registers rather than as symbolic immediates. Return TRUE on success; otherwise return FALSE. */ static bfd_boolean parse_constant_immediate (char **str, int64_t *val, aarch64_reg_type reg_type) { expressionS exp; if (! parse_immediate_expression (str, &exp, reg_type)) return FALSE; if (exp.X_op != O_constant) { set_syntax_error (_("constant expression required")); return FALSE; } *val = exp.X_add_number; return TRUE; } static uint32_t encode_imm_float_bits (uint32_t imm) { return ((imm >> 19) & 0x7f) /* b[25:19] -> b[6:0] */ | ((imm >> (31 - 7)) & 0x80); /* b[31] -> b[7] */ } /* Return TRUE if the single-precision floating-point value encoded in IMM can be expressed in the AArch64 8-bit signed floating-point format with 3-bit exponent and normalized 4 bits of precision; in other words, the floating-point value must be expressable as (+/-) n / 16 * power (2, r) where n and r are integers such that 16 <= n <=31 and -3 <= r <= 4. */ static bfd_boolean aarch64_imm_float_p (uint32_t imm) { /* If a single-precision floating-point value has the following bit pattern, it can be expressed in the AArch64 8-bit floating-point format: 3 32222222 2221111111111 1 09876543 21098765432109876543210 n Eeeeeexx xxxx0000000000000000000 where n, e and each x are either 0 or 1 independently, with E == ~ e. */ uint32_t pattern; /* Prepare the pattern for 'Eeeeee'. */ if (((imm >> 30) & 0x1) == 0) pattern = 0x3e000000; else pattern = 0x40000000; return (imm & 0x7ffff) == 0 /* lower 19 bits are 0. */ && ((imm & 0x7e000000) == pattern); /* bits 25 - 29 == ~ bit 30. */ } /* Return TRUE if the IEEE double value encoded in IMM can be expressed as an IEEE float without any loss of precision. Store the value in *FPWORD if so. */ static bfd_boolean can_convert_double_to_float (uint64_t imm, uint32_t *fpword) { /* If a double-precision floating-point value has the following bit pattern, it can be expressed in a float: 6 66655555555 5544 44444444 33333333 33222222 22221111 111111 3 21098765432 1098 76543210 98765432 10987654 32109876 54321098 76543210 n E~~~eeeeeee ssss ssssssss ssssssss SSS00000 00000000 00000000 00000000 -----------------------------> nEeeeeee esssssss ssssssss sssssSSS if Eeee_eeee != 1111_1111 where n, e, s and S are either 0 or 1 independently and where ~ is the inverse of E. */ uint32_t pattern; uint32_t high32 = imm >> 32; uint32_t low32 = imm; /* Lower 29 bits need to be 0s. */ if ((imm & 0x1fffffff) != 0) return FALSE; /* Prepare the pattern for 'Eeeeeeeee'. */ if (((high32 >> 30) & 0x1) == 0) pattern = 0x38000000; else pattern = 0x40000000; /* Check E~~~. */ if ((high32 & 0x78000000) != pattern) return FALSE; /* Check Eeee_eeee != 1111_1111. */ if ((high32 & 0x7ff00000) == 0x47f00000) return FALSE; *fpword = ((high32 & 0xc0000000) /* 1 n bit and 1 E bit. */ | ((high32 << 3) & 0x3ffffff8) /* 7 e and 20 s bits. */ | (low32 >> 29)); /* 3 S bits. */ return TRUE; } /* Return true if we should treat OPERAND as a double-precision floating-point operand rather than a single-precision one. */ static bfd_boolean double_precision_operand_p (const aarch64_opnd_info *operand) { /* Check for unsuffixed SVE registers, which are allowed for LDR and STR but not in instructions that require an immediate. We get better error messages if we arbitrarily pick one size, parse the immediate normally, and then report the match failure in the normal way. */ return (operand->qualifier == AARCH64_OPND_QLF_NIL || aarch64_get_qualifier_esize (operand->qualifier) == 8); } /* Parse a floating-point immediate. Return TRUE on success and return the value in *IMMED in the format of IEEE754 single-precision encoding. *CCP points to the start of the string; DP_P is TRUE when the immediate is expected to be in double-precision (N.B. this only matters when hexadecimal representation is involved). REG_TYPE says which register names should be treated as registers rather than as symbolic immediates. This routine accepts any IEEE float; it is up to the callers to reject invalid ones. */ static bfd_boolean parse_aarch64_imm_float (char **ccp, int *immed, bfd_boolean dp_p, aarch64_reg_type reg_type) { char *str = *ccp; char *fpnum; LITTLENUM_TYPE words[MAX_LITTLENUMS]; int64_t val = 0; unsigned fpword = 0; bfd_boolean hex_p = FALSE; skip_past_char (&str, '#'); fpnum = str; skip_whitespace (fpnum); if (strncmp (fpnum, "0x", 2) == 0) { /* Support the hexadecimal representation of the IEEE754 encoding. Double-precision is expected when DP_P is TRUE, otherwise the representation should be in single-precision. */ if (! parse_constant_immediate (&str, &val, reg_type)) goto invalid_fp; if (dp_p) { if (!can_convert_double_to_float (val, &fpword)) goto invalid_fp; } else if ((uint64_t) val > 0xffffffff) goto invalid_fp; else fpword = val; hex_p = TRUE; } else if (reg_name_p (str, reg_type)) { set_recoverable_error (_("immediate operand required")); return FALSE; } if (! hex_p) { int i; if ((str = atof_ieee (str, 's', words)) == NULL) goto invalid_fp; /* Our FP word must be 32 bits (single-precision FP). */ for (i = 0; i < 32 / LITTLENUM_NUMBER_OF_BITS; i++) { fpword <<= LITTLENUM_NUMBER_OF_BITS; fpword |= words[i]; } } *immed = fpword; *ccp = str; return TRUE; invalid_fp: set_fatal_syntax_error (_("invalid floating-point constant")); return FALSE; } /* Less-generic immediate-value read function with the possibility of loading a big (64-bit) immediate, as required by AdvSIMD Modified immediate instructions. To prevent the expression parser from pushing a register name into the symbol table as an undefined symbol, a check is firstly done to find out whether STR is a register of type REG_TYPE followed by a comma or the end of line. Return FALSE if STR is such a register. */ static bfd_boolean parse_big_immediate (char **str, int64_t *imm, aarch64_reg_type reg_type) { char *ptr = *str; if (reg_name_p (ptr, reg_type)) { set_syntax_error (_("immediate operand required")); return FALSE; } my_get_expression (&inst.reloc.exp, &ptr, GE_OPT_PREFIX, 1); if (inst.reloc.exp.X_op == O_constant) *imm = inst.reloc.exp.X_add_number; *str = ptr; return TRUE; } /* Set operand IDX of the *INSTR that needs a GAS internal fixup. if NEED_LIBOPCODES is non-zero, the fixup will need assistance from the libopcodes. */ static inline void aarch64_set_gas_internal_fixup (struct reloc *reloc, const aarch64_opnd_info *operand, int need_libopcodes_p) { reloc->type = BFD_RELOC_AARCH64_GAS_INTERNAL_FIXUP; reloc->opnd = operand->type; if (need_libopcodes_p) reloc->need_libopcodes_p = 1; }; /* Return TRUE if the instruction needs to be fixed up later internally by the GAS; otherwise return FALSE. */ static inline bfd_boolean aarch64_gas_internal_fixup_p (void) { return inst.reloc.type == BFD_RELOC_AARCH64_GAS_INTERNAL_FIXUP; } /* Assign the immediate value to the relevant field in *OPERAND if RELOC->EXP is a constant expression; otherwise, flag that *OPERAND needs an internal fixup in a later stage. ADDR_OFF_P determines whether it is the field ADDR.OFFSET.IMM or IMM.VALUE that may get assigned with the constant. */ static inline void assign_imm_if_const_or_fixup_later (struct reloc *reloc, aarch64_opnd_info *operand, int addr_off_p, int need_libopcodes_p, int skip_p) { if (reloc->exp.X_op == O_constant) { if (addr_off_p) operand->addr.offset.imm = reloc->exp.X_add_number; else operand->imm.value = reloc->exp.X_add_number; reloc->type = BFD_RELOC_UNUSED; } else { aarch64_set_gas_internal_fixup (reloc, operand, need_libopcodes_p); /* Tell libopcodes to ignore this operand or not. This is helpful when one of the operands needs to be fixed up later but we need libopcodes to check the other operands. */ operand->skip = skip_p; } } /* Relocation modifiers. Each entry in the table contains the textual name for the relocation which may be placed before a symbol used as a load/store offset, or add immediate. It must be surrounded by a leading and trailing colon, for example: ldr x0, [x1, #:rello:varsym] add x0, x1, #:rello:varsym */ struct reloc_table_entry { const char *name; int pc_rel; bfd_reloc_code_real_type adr_type; bfd_reloc_code_real_type adrp_type; bfd_reloc_code_real_type movw_type; bfd_reloc_code_real_type add_type; bfd_reloc_code_real_type ldst_type; bfd_reloc_code_real_type ld_literal_type; }; static struct reloc_table_entry reloc_table[] = { /* Low 12 bits of absolute address: ADD/i and LDR/STR */ {"lo12", 0, 0, /* adr_type */ 0, 0, BFD_RELOC_AARCH64_ADD_LO12, BFD_RELOC_AARCH64_LDST_LO12, 0}, /* Higher 21 bits of pc-relative page offset: ADRP */ {"pg_hi21", 1, 0, /* adr_type */ BFD_RELOC_AARCH64_ADR_HI21_PCREL, 0, 0, 0, 0}, /* Higher 21 bits of pc-relative page offset: ADRP, no check */ {"pg_hi21_nc", 1, 0, /* adr_type */ BFD_RELOC_AARCH64_ADR_HI21_NC_PCREL, 0, 0, 0, 0}, /* Most significant bits 0-15 of unsigned address/value: MOVZ */ {"abs_g0", 0, 0, /* adr_type */ 0, BFD_RELOC_AARCH64_MOVW_G0, 0, 0, 0}, /* Most significant bits 0-15 of signed address/value: MOVN/Z */ {"abs_g0_s", 0, 0, /* adr_type */ 0, BFD_RELOC_AARCH64_MOVW_G0_S, 0, 0, 0}, /* Less significant bits 0-15 of address/value: MOVK, no check */ {"abs_g0_nc", 0, 0, /* adr_type */ 0, BFD_RELOC_AARCH64_MOVW_G0_NC, 0, 0, 0}, /* Most significant bits 16-31 of unsigned address/value: MOVZ */ {"abs_g1", 0, 0, /* adr_type */ 0, BFD_RELOC_AARCH64_MOVW_G1, 0, 0, 0}, /* Most significant bits 16-31 of signed address/value: MOVN/Z */ {"abs_g1_s", 0, 0, /* adr_type */ 0, BFD_RELOC_AARCH64_MOVW_G1_S, 0, 0, 0}, /* Less significant bits 16-31 of address/value: MOVK, no check */ {"abs_g1_nc", 0, 0, /* adr_type */ 0, BFD_RELOC_AARCH64_MOVW_G1_NC, 0, 0, 0}, /* Most significant bits 32-47 of unsigned address/value: MOVZ */ {"abs_g2", 0, 0, /* adr_type */ 0, BFD_RELOC_AARCH64_MOVW_G2, 0, 0, 0}, /* Most significant bits 32-47 of signed address/value: MOVN/Z */ {"abs_g2_s", 0, 0, /* adr_type */ 0, BFD_RELOC_AARCH64_MOVW_G2_S, 0, 0, 0}, /* Less significant bits 32-47 of address/value: MOVK, no check */ {"abs_g2_nc", 0, 0, /* adr_type */ 0, BFD_RELOC_AARCH64_MOVW_G2_NC, 0, 0, 0}, /* Most significant bits 48-63 of signed/unsigned address/value: MOVZ */ {"abs_g3", 0, 0, /* adr_type */ 0, BFD_RELOC_AARCH64_MOVW_G3, 0, 0, 0}, /* Most significant bits 0-15 of signed/unsigned address/value: MOVZ */ {"prel_g0", 1, 0, /* adr_type */ 0, BFD_RELOC_AARCH64_MOVW_PREL_G0, 0, 0, 0}, /* Most significant bits 0-15 of signed/unsigned address/value: MOVK */ {"prel_g0_nc", 1, 0, /* adr_type */ 0, BFD_RELOC_AARCH64_MOVW_PREL_G0_NC, 0, 0, 0}, /* Most significant bits 16-31 of signed/unsigned address/value: MOVZ */ {"prel_g1", 1, 0, /* adr_type */ 0, BFD_RELOC_AARCH64_MOVW_PREL_G1, 0, 0, 0}, /* Most significant bits 16-31 of signed/unsigned address/value: MOVK */ {"prel_g1_nc", 1, 0, /* adr_type */ 0, BFD_RELOC_AARCH64_MOVW_PREL_G1_NC, 0, 0, 0}, /* Most significant bits 32-47 of signed/unsigned address/value: MOVZ */ {"prel_g2", 1, 0, /* adr_type */ 0, BFD_RELOC_AARCH64_MOVW_PREL_G2, 0, 0, 0}, /* Most significant bits 32-47 of signed/unsigned address/value: MOVK */ {"prel_g2_nc", 1, 0, /* adr_type */ 0, BFD_RELOC_AARCH64_MOVW_PREL_G2_NC, 0, 0, 0}, /* Most significant bits 48-63 of signed/unsigned address/value: MOVZ */ {"prel_g3", 1, 0, /* adr_type */ 0, BFD_RELOC_AARCH64_MOVW_PREL_G3, 0, 0, 0}, /* Get to the page containing GOT entry for a symbol. */ {"got", 1, 0, /* adr_type */ BFD_RELOC_AARCH64_ADR_GOT_PAGE, 0, 0, 0, BFD_RELOC_AARCH64_GOT_LD_PREL19}, /* 12 bit offset into the page containing GOT entry for that symbol. */ {"got_lo12", 0, 0, /* adr_type */ 0, 0, 0, BFD_RELOC_AARCH64_LD_GOT_LO12_NC, 0}, /* 0-15 bits of address/value: MOVk, no check. */ {"gotoff_g0_nc", 0, 0, /* adr_type */ 0, BFD_RELOC_AARCH64_MOVW_GOTOFF_G0_NC, 0, 0, 0}, /* Most significant bits 16-31 of address/value: MOVZ. */ {"gotoff_g1", 0, 0, /* adr_type */ 0, BFD_RELOC_AARCH64_MOVW_GOTOFF_G1, 0, 0, 0}, /* 15 bit offset into the page containing GOT entry for that symbol. */ {"gotoff_lo15", 0, 0, /* adr_type */ 0, 0, 0, BFD_RELOC_AARCH64_LD64_GOTOFF_LO15, 0}, /* Get to the page containing GOT TLS entry for a symbol */ {"gottprel_g0_nc", 0, 0, /* adr_type */ 0, BFD_RELOC_AARCH64_TLSIE_MOVW_GOTTPREL_G0_NC, 0, 0, 0}, /* Get to the page containing GOT TLS entry for a symbol */ {"gottprel_g1", 0, 0, /* adr_type */ 0, BFD_RELOC_AARCH64_TLSIE_MOVW_GOTTPREL_G1, 0, 0, 0}, /* Get to the page containing GOT TLS entry for a symbol */ {"tlsgd", 0, BFD_RELOC_AARCH64_TLSGD_ADR_PREL21, /* adr_type */ BFD_RELOC_AARCH64_TLSGD_ADR_PAGE21, 0, 0, 0, 0}, /* 12 bit offset into the page containing GOT TLS entry for a symbol */ {"tlsgd_lo12", 0, 0, /* adr_type */ 0, 0, BFD_RELOC_AARCH64_TLSGD_ADD_LO12_NC, 0, 0}, /* Lower 16 bits address/value: MOVk. */ {"tlsgd_g0_nc", 0, 0, /* adr_type */ 0, BFD_RELOC_AARCH64_TLSGD_MOVW_G0_NC, 0, 0, 0}, /* Most significant bits 16-31 of address/value: MOVZ. */ {"tlsgd_g1", 0, 0, /* adr_type */ 0, BFD_RELOC_AARCH64_TLSGD_MOVW_G1, 0, 0, 0}, /* Get to the page containing GOT TLS entry for a symbol */ {"tlsdesc", 0, BFD_RELOC_AARCH64_TLSDESC_ADR_PREL21, /* adr_type */ BFD_RELOC_AARCH64_TLSDESC_ADR_PAGE21, 0, 0, 0, BFD_RELOC_AARCH64_TLSDESC_LD_PREL19}, /* 12 bit offset into the page containing GOT TLS entry for a symbol */ {"tlsdesc_lo12", 0, 0, /* adr_type */ 0, 0, BFD_RELOC_AARCH64_TLSDESC_ADD_LO12, BFD_RELOC_AARCH64_TLSDESC_LD_LO12_NC, 0}, /* Get to the page containing GOT TLS entry for a symbol. The same as GD, we allocate two consecutive GOT slots for module index and module offset, the only difference with GD is the module offset should be initialized to zero without any outstanding runtime relocation. */ {"tlsldm", 0, BFD_RELOC_AARCH64_TLSLD_ADR_PREL21, /* adr_type */ BFD_RELOC_AARCH64_TLSLD_ADR_PAGE21, 0, 0, 0, 0}, /* 12 bit offset into the page containing GOT TLS entry for a symbol */ {"tlsldm_lo12_nc", 0, 0, /* adr_type */ 0, 0, BFD_RELOC_AARCH64_TLSLD_ADD_LO12_NC, 0, 0}, /* 12 bit offset into the module TLS base address. */ {"dtprel_lo12", 0, 0, /* adr_type */ 0, 0, BFD_RELOC_AARCH64_TLSLD_ADD_DTPREL_LO12, BFD_RELOC_AARCH64_TLSLD_LDST_DTPREL_LO12, 0}, /* Same as dtprel_lo12, no overflow check. */ {"dtprel_lo12_nc", 0, 0, /* adr_type */ 0, 0, BFD_RELOC_AARCH64_TLSLD_ADD_DTPREL_LO12_NC, BFD_RELOC_AARCH64_TLSLD_LDST_DTPREL_LO12_NC, 0}, /* bits[23:12] of offset to the module TLS base address. */ {"dtprel_hi12", 0, 0, /* adr_type */ 0, 0, BFD_RELOC_AARCH64_TLSLD_ADD_DTPREL_HI12, 0, 0}, /* bits[15:0] of offset to the module TLS base address. */ {"dtprel_g0", 0, 0, /* adr_type */ 0, BFD_RELOC_AARCH64_TLSLD_MOVW_DTPREL_G0, 0, 0, 0}, /* No overflow check version of BFD_RELOC_AARCH64_TLSLD_MOVW_DTPREL_G0. */ {"dtprel_g0_nc", 0, 0, /* adr_type */ 0, BFD_RELOC_AARCH64_TLSLD_MOVW_DTPREL_G0_NC, 0, 0, 0}, /* bits[31:16] of offset to the module TLS base address. */ {"dtprel_g1", 0, 0, /* adr_type */ 0, BFD_RELOC_AARCH64_TLSLD_MOVW_DTPREL_G1, 0, 0, 0}, /* No overflow check version of BFD_RELOC_AARCH64_TLSLD_MOVW_DTPREL_G1. */ {"dtprel_g1_nc", 0, 0, /* adr_type */ 0, BFD_RELOC_AARCH64_TLSLD_MOVW_DTPREL_G1_NC, 0, 0, 0}, /* bits[47:32] of offset to the module TLS base address. */ {"dtprel_g2", 0, 0, /* adr_type */ 0, BFD_RELOC_AARCH64_TLSLD_MOVW_DTPREL_G2, 0, 0, 0}, /* Lower 16 bit offset into GOT entry for a symbol */ {"tlsdesc_off_g0_nc", 0, 0, /* adr_type */ 0, BFD_RELOC_AARCH64_TLSDESC_OFF_G0_NC, 0, 0, 0}, /* Higher 16 bit offset into GOT entry for a symbol */ {"tlsdesc_off_g1", 0, 0, /* adr_type */ 0, BFD_RELOC_AARCH64_TLSDESC_OFF_G1, 0, 0, 0}, /* Get to the page containing GOT TLS entry for a symbol */ {"gottprel", 0, 0, /* adr_type */ BFD_RELOC_AARCH64_TLSIE_ADR_GOTTPREL_PAGE21, 0, 0, 0, BFD_RELOC_AARCH64_TLSIE_LD_GOTTPREL_PREL19}, /* 12 bit offset into the page containing GOT TLS entry for a symbol */ {"gottprel_lo12", 0, 0, /* adr_type */ 0, 0, 0, BFD_RELOC_AARCH64_TLSIE_LD_GOTTPREL_LO12_NC, 0}, /* Get tp offset for a symbol. */ {"tprel", 0, 0, /* adr_type */ 0, 0, BFD_RELOC_AARCH64_TLSLE_ADD_TPREL_LO12, 0, 0}, /* Get tp offset for a symbol. */ {"tprel_lo12", 0, 0, /* adr_type */ 0, 0, BFD_RELOC_AARCH64_TLSLE_ADD_TPREL_LO12, BFD_RELOC_AARCH64_TLSLE_LDST_TPREL_LO12, 0}, /* Get tp offset for a symbol. */ {"tprel_hi12", 0, 0, /* adr_type */ 0, 0, BFD_RELOC_AARCH64_TLSLE_ADD_TPREL_HI12, 0, 0}, /* Get tp offset for a symbol. */ {"tprel_lo12_nc", 0, 0, /* adr_type */ 0, 0, BFD_RELOC_AARCH64_TLSLE_ADD_TPREL_LO12_NC, BFD_RELOC_AARCH64_TLSLE_LDST_TPREL_LO12_NC, 0}, /* Most significant bits 32-47 of address/value: MOVZ. */ {"tprel_g2", 0, 0, /* adr_type */ 0, BFD_RELOC_AARCH64_TLSLE_MOVW_TPREL_G2, 0, 0, 0}, /* Most significant bits 16-31 of address/value: MOVZ. */ {"tprel_g1", 0, 0, /* adr_type */ 0, BFD_RELOC_AARCH64_TLSLE_MOVW_TPREL_G1, 0, 0, 0}, /* Most significant bits 16-31 of address/value: MOVZ, no check. */ {"tprel_g1_nc", 0, 0, /* adr_type */ 0, BFD_RELOC_AARCH64_TLSLE_MOVW_TPREL_G1_NC, 0, 0, 0}, /* Most significant bits 0-15 of address/value: MOVZ. */ {"tprel_g0", 0, 0, /* adr_type */ 0, BFD_RELOC_AARCH64_TLSLE_MOVW_TPREL_G0, 0, 0, 0}, /* Most significant bits 0-15 of address/value: MOVZ, no check. */ {"tprel_g0_nc", 0, 0, /* adr_type */ 0, BFD_RELOC_AARCH64_TLSLE_MOVW_TPREL_G0_NC, 0, 0, 0}, /* 15bit offset from got entry to base address of GOT table. */ {"gotpage_lo15", 0, 0, 0, 0, 0, BFD_RELOC_AARCH64_LD64_GOTPAGE_LO15, 0}, /* 14bit offset from got entry to base address of GOT table. */ {"gotpage_lo14", 0, 0, 0, 0, 0, BFD_RELOC_AARCH64_LD32_GOTPAGE_LO14, 0}, }; /* Given the address of a pointer pointing to the textual name of a relocation as may appear in assembler source, attempt to find its details in reloc_table. The pointer will be updated to the character after the trailing colon. On failure, NULL will be returned; otherwise return the reloc_table_entry. */ static struct reloc_table_entry * find_reloc_table_entry (char **str) { unsigned int i; for (i = 0; i < ARRAY_SIZE (reloc_table); i++) { int length = strlen (reloc_table[i].name); if (strncasecmp (reloc_table[i].name, *str, length) == 0 && (*str)[length] == ':') { *str += (length + 1); return &reloc_table[i]; } } return NULL; } /* Mode argument to parse_shift and parser_shifter_operand. */ enum parse_shift_mode { SHIFTED_NONE, /* no shifter allowed */ SHIFTED_ARITH_IMM, /* "rn{,lsl|lsr|asl|asr|uxt|sxt #n}" or "#imm{,lsl #n}" */ SHIFTED_LOGIC_IMM, /* "rn{,lsl|lsr|asl|asr|ror #n}" or "#imm" */ SHIFTED_LSL, /* bare "lsl #n" */ SHIFTED_MUL, /* bare "mul #n" */ SHIFTED_LSL_MSL, /* "lsl|msl #n" */ SHIFTED_MUL_VL, /* "mul vl" */ SHIFTED_REG_OFFSET /* [su]xtw|sxtx {#n} or lsl #n */ }; /* Parse a operator on an AArch64 data processing instruction. Return TRUE on success; otherwise return FALSE. */ static bfd_boolean parse_shift (char **str, aarch64_opnd_info *operand, enum parse_shift_mode mode) { const struct aarch64_name_value_pair *shift_op; enum aarch64_modifier_kind kind; expressionS exp; int exp_has_prefix; char *s = *str; char *p = s; for (p = *str; ISALPHA (*p); p++) ; if (p == *str) { set_syntax_error (_("shift expression expected")); return FALSE; } shift_op = hash_find_n (aarch64_shift_hsh, *str, p - *str); if (shift_op == NULL) { set_syntax_error (_("shift operator expected")); return FALSE; } kind = aarch64_get_operand_modifier (shift_op); if (kind == AARCH64_MOD_MSL && mode != SHIFTED_LSL_MSL) { set_syntax_error (_("invalid use of 'MSL'")); return FALSE; } if (kind == AARCH64_MOD_MUL && mode != SHIFTED_MUL && mode != SHIFTED_MUL_VL) { set_syntax_error (_("invalid use of 'MUL'")); return FALSE; } switch (mode) { case SHIFTED_LOGIC_IMM: if (aarch64_extend_operator_p (kind)) { set_syntax_error (_("extending shift is not permitted")); return FALSE; } break; case SHIFTED_ARITH_IMM: if (kind == AARCH64_MOD_ROR) { set_syntax_error (_("'ROR' shift is not permitted")); return FALSE; } break; case SHIFTED_LSL: if (kind != AARCH64_MOD_LSL) { set_syntax_error (_("only 'LSL' shift is permitted")); return FALSE; } break; case SHIFTED_MUL: if (kind != AARCH64_MOD_MUL) { set_syntax_error (_("only 'MUL' is permitted")); return FALSE; } break; case SHIFTED_MUL_VL: /* "MUL VL" consists of two separate tokens. Require the first token to be "MUL" and look for a following "VL". */ if (kind == AARCH64_MOD_MUL) { skip_whitespace (p); if (strncasecmp (p, "vl", 2) == 0 && !ISALPHA (p[2])) { p += 2; kind = AARCH64_MOD_MUL_VL; break; } } set_syntax_error (_("only 'MUL VL' is permitted")); return FALSE; case SHIFTED_REG_OFFSET: if (kind != AARCH64_MOD_UXTW && kind != AARCH64_MOD_LSL && kind != AARCH64_MOD_SXTW && kind != AARCH64_MOD_SXTX) { set_fatal_syntax_error (_("invalid shift for the register offset addressing mode")); return FALSE; } break; case SHIFTED_LSL_MSL: if (kind != AARCH64_MOD_LSL && kind != AARCH64_MOD_MSL) { set_syntax_error (_("invalid shift operator")); return FALSE; } break; default: abort (); } /* Whitespace can appear here if the next thing is a bare digit. */ skip_whitespace (p); /* Parse shift amount. */ exp_has_prefix = 0; if ((mode == SHIFTED_REG_OFFSET && *p == ']') || kind == AARCH64_MOD_MUL_VL) exp.X_op = O_absent; else { if (is_immediate_prefix (*p)) { p++; exp_has_prefix = 1; } my_get_expression (&exp, &p, GE_NO_PREFIX, 0); } if (kind == AARCH64_MOD_MUL_VL) /* For consistency, give MUL VL the same shift amount as an implicit MUL #1. */ operand->shifter.amount = 1; else if (exp.X_op == O_absent) { if (!aarch64_extend_operator_p (kind) || exp_has_prefix) { set_syntax_error (_("missing shift amount")); return FALSE; } operand->shifter.amount = 0; } else if (exp.X_op != O_constant) { set_syntax_error (_("constant shift amount required")); return FALSE; } /* For parsing purposes, MUL #n has no inherent range. The range depends on the operand and will be checked by operand-specific routines. */ else if (kind != AARCH64_MOD_MUL && (exp.X_add_number < 0 || exp.X_add_number > 63)) { set_fatal_syntax_error (_("shift amount out of range 0 to 63")); return FALSE; } else { operand->shifter.amount = exp.X_add_number; operand->shifter.amount_present = 1; } operand->shifter.operator_present = 1; operand->shifter.kind = kind; *str = p; return TRUE; } /* Parse a for a data processing instruction: # #, LSL #imm Validation of immediate operands is deferred to md_apply_fix. Return TRUE on success; otherwise return FALSE. */ static bfd_boolean parse_shifter_operand_imm (char **str, aarch64_opnd_info *operand, enum parse_shift_mode mode) { char *p; if (mode != SHIFTED_ARITH_IMM && mode != SHIFTED_LOGIC_IMM) return FALSE; p = *str; /* Accept an immediate expression. */ if (! my_get_expression (&inst.reloc.exp, &p, GE_OPT_PREFIX, 1)) return FALSE; /* Accept optional LSL for arithmetic immediate values. */ if (mode == SHIFTED_ARITH_IMM && skip_past_comma (&p)) if (! parse_shift (&p, operand, SHIFTED_LSL)) return FALSE; /* Not accept any shifter for logical immediate values. */ if (mode == SHIFTED_LOGIC_IMM && skip_past_comma (&p) && parse_shift (&p, operand, mode)) { set_syntax_error (_("unexpected shift operator")); return FALSE; } *str = p; return TRUE; } /* Parse a for a data processing instruction: , # #, LSL #imm where is handled by parse_shift above, and the last two cases are handled by the function above. Validation of immediate operands is deferred to md_apply_fix. Return TRUE on success; otherwise return FALSE. */ static bfd_boolean parse_shifter_operand (char **str, aarch64_opnd_info *operand, enum parse_shift_mode mode) { const reg_entry *reg; aarch64_opnd_qualifier_t qualifier; enum aarch64_operand_class opd_class = aarch64_get_operand_class (operand->type); reg = aarch64_reg_parse_32_64 (str, &qualifier); if (reg) { if (opd_class == AARCH64_OPND_CLASS_IMMEDIATE) { set_syntax_error (_("unexpected register in the immediate operand")); return FALSE; } if (!aarch64_check_reg_type (reg, REG_TYPE_R_Z)) { set_syntax_error (_(get_reg_expected_msg (REG_TYPE_R_Z))); return FALSE; } operand->reg.regno = reg->number; operand->qualifier = qualifier; /* Accept optional shift operation on register. */ if (! skip_past_comma (str)) return TRUE; if (! parse_shift (str, operand, mode)) return FALSE; return TRUE; } else if (opd_class == AARCH64_OPND_CLASS_MODIFIED_REG) { set_syntax_error (_("integer register expected in the extended/shifted operand " "register")); return FALSE; } /* We have a shifted immediate variable. */ return parse_shifter_operand_imm (str, operand, mode); } /* Return TRUE on success; return FALSE otherwise. */ static bfd_boolean parse_shifter_operand_reloc (char **str, aarch64_opnd_info *operand, enum parse_shift_mode mode) { char *p = *str; /* Determine if we have the sequence of characters #: or just : coming next. If we do, then we check for a :rello: relocation modifier. If we don't, punt the whole lot to parse_shifter_operand. */ if ((p[0] == '#' && p[1] == ':') || p[0] == ':') { struct reloc_table_entry *entry; if (p[0] == '#') p += 2; else p++; *str = p; /* Try to parse a relocation. Anything else is an error. */ if (!(entry = find_reloc_table_entry (str))) { set_syntax_error (_("unknown relocation modifier")); return FALSE; } if (entry->add_type == 0) { set_syntax_error (_("this relocation modifier is not allowed on this instruction")); return FALSE; } /* Save str before we decompose it. */ p = *str; /* Next, we parse the expression. */ if (! my_get_expression (&inst.reloc.exp, str, GE_NO_PREFIX, 1)) return FALSE; /* Record the relocation type (use the ADD variant here). */ inst.reloc.type = entry->add_type; inst.reloc.pc_rel = entry->pc_rel; /* If str is empty, we've reached the end, stop here. */ if (**str == '\0') return TRUE; /* Otherwise, we have a shifted reloc modifier, so rewind to recover the variable name and continue parsing for the shifter. */ *str = p; return parse_shifter_operand_imm (str, operand, mode); } return parse_shifter_operand (str, operand, mode); } /* Parse all forms of an address expression. Information is written to *OPERAND and/or inst.reloc. The A64 instruction set has the following addressing modes: Offset [base] // in SIMD ld/st structure [base{,#0}] // in ld/st exclusive [base{,#imm}] [base,Xm{,LSL #imm}] [base,Xm,SXTX {#imm}] [base,Wm,(S|U)XTW {#imm}] Pre-indexed [base,#imm]! Post-indexed [base],#imm [base],Xm // in SIMD ld/st structure PC-relative (literal) label SVE: [base,#imm,MUL VL] [base,Zm.D{,LSL #imm}] [base,Zm.S,(S|U)XTW {#imm}] [base,Zm.D,(S|U)XTW {#imm}] // ignores top 32 bits of Zm.D elements [Zn.S,#imm] [Zn.D,#imm] [Zn.S,Zm.S{,LSL #imm}] // in ADR [Zn.D,Zm.D{,LSL #imm}] // in ADR [Zn.D,Zm.D,(S|U)XTW {#imm}] // in ADR (As a convenience, the notation "=immediate" is permitted in conjunction with the pc-relative literal load instructions to automatically place an immediate value or symbolic address in a nearby literal pool and generate a hidden label which references it.) Upon a successful parsing, the address structure in *OPERAND will be filled in the following way: .base_regno = .offset.is_reg // 1 if the offset is a register .offset.imm = .offset.regno = For different addressing modes defined in the A64 ISA: Offset .pcrel=0; .preind=1; .postind=0; .writeback=0 Pre-indexed .pcrel=0; .preind=1; .postind=0; .writeback=1 Post-indexed .pcrel=0; .preind=0; .postind=1; .writeback=1 PC-relative (literal) .pcrel=1; .preind=1; .postind=0; .writeback=0 The shift/extension information, if any, will be stored in .shifter. The base and offset qualifiers will be stored in *BASE_QUALIFIER and *OFFSET_QUALIFIER respectively, with NIL being used if there's no corresponding register. BASE_TYPE says which types of base register should be accepted and OFFSET_TYPE says the same for offset registers. IMM_SHIFT_MODE is the type of shifter that is allowed for immediate offsets, or SHIFTED_NONE if none. In all other respects, it is the caller's responsibility to check for addressing modes not supported by the instruction, and to set inst.reloc.type. */ static bfd_boolean parse_address_main (char **str, aarch64_opnd_info *operand, aarch64_opnd_qualifier_t *base_qualifier, aarch64_opnd_qualifier_t *offset_qualifier, aarch64_reg_type base_type, aarch64_reg_type offset_type, enum parse_shift_mode imm_shift_mode) { char *p = *str; const reg_entry *reg; expressionS *exp = &inst.reloc.exp; *base_qualifier = AARCH64_OPND_QLF_NIL; *offset_qualifier = AARCH64_OPND_QLF_NIL; if (! skip_past_char (&p, '[')) { /* =immediate or label. */ operand->addr.pcrel = 1; operand->addr.preind = 1; /* #:: */ skip_past_char (&p, '#'); if (skip_past_char (&p, ':')) { bfd_reloc_code_real_type ty; struct reloc_table_entry *entry; /* Try to parse a relocation modifier. Anything else is an error. */ entry = find_reloc_table_entry (&p); if (! entry) { set_syntax_error (_("unknown relocation modifier")); return FALSE; } switch (operand->type) { case AARCH64_OPND_ADDR_PCREL21: /* adr */ ty = entry->adr_type; break; default: ty = entry->ld_literal_type; break; } if (ty == 0) { set_syntax_error (_("this relocation modifier is not allowed on this " "instruction")); return FALSE; } /* #:: */ if (! my_get_expression (exp, &p, GE_NO_PREFIX, 1)) { set_syntax_error (_("invalid relocation expression")); return FALSE; } /* #:: */ /* Record the relocation type. */ inst.reloc.type = ty; inst.reloc.pc_rel = entry->pc_rel; } else { if (skip_past_char (&p, '=')) /* =immediate; need to generate the literal in the literal pool. */ inst.gen_lit_pool = 1; if (!my_get_expression (exp, &p, GE_NO_PREFIX, 1)) { set_syntax_error (_("invalid address")); return FALSE; } } *str = p; return TRUE; } /* [ */ reg = aarch64_addr_reg_parse (&p, base_type, base_qualifier); if (!reg || !aarch64_check_reg_type (reg, base_type)) { set_syntax_error (_(get_reg_expected_msg (base_type))); return FALSE; } operand->addr.base_regno = reg->number; /* [Xn */ if (skip_past_comma (&p)) { /* [Xn, */ operand->addr.preind = 1; reg = aarch64_addr_reg_parse (&p, offset_type, offset_qualifier); if (reg) { if (!aarch64_check_reg_type (reg, offset_type)) { set_syntax_error (_(get_reg_expected_msg (offset_type))); return FALSE; } /* [Xn,Rm */ operand->addr.offset.regno = reg->number; operand->addr.offset.is_reg = 1; /* Shifted index. */ if (skip_past_comma (&p)) { /* [Xn,Rm, */ if (! parse_shift (&p, operand, SHIFTED_REG_OFFSET)) /* Use the diagnostics set in parse_shift, so not set new error message here. */ return FALSE; } /* We only accept: [base,Xm{,LSL #imm}] [base,Xm,SXTX {#imm}] [base,Wm,(S|U)XTW {#imm}] */ if (operand->shifter.kind == AARCH64_MOD_NONE || operand->shifter.kind == AARCH64_MOD_LSL || operand->shifter.kind == AARCH64_MOD_SXTX) { if (*offset_qualifier == AARCH64_OPND_QLF_W) { set_syntax_error (_("invalid use of 32-bit register offset")); return FALSE; } if (aarch64_get_qualifier_esize (*base_qualifier) != aarch64_get_qualifier_esize (*offset_qualifier)) { set_syntax_error (_("offset has different size from base")); return FALSE; } } else if (*offset_qualifier == AARCH64_OPND_QLF_X) { set_syntax_error (_("invalid use of 64-bit register offset")); return FALSE; } } else { /* [Xn,#:: */ skip_past_char (&p, '#'); if (skip_past_char (&p, ':')) { struct reloc_table_entry *entry; /* Try to parse a relocation modifier. Anything else is an error. */ if (!(entry = find_reloc_table_entry (&p))) { set_syntax_error (_("unknown relocation modifier")); return FALSE; } if (entry->ldst_type == 0) { set_syntax_error (_("this relocation modifier is not allowed on this " "instruction")); return FALSE; } /* [Xn,#:: */ /* We now have the group relocation table entry corresponding to the name in the assembler source. Next, we parse the expression. */ if (! my_get_expression (exp, &p, GE_NO_PREFIX, 1)) { set_syntax_error (_("invalid relocation expression")); return FALSE; } /* [Xn,#:: */ /* Record the load/store relocation type. */ inst.reloc.type = entry->ldst_type; inst.reloc.pc_rel = entry->pc_rel; } else { if (! my_get_expression (exp, &p, GE_OPT_PREFIX, 1)) { set_syntax_error (_("invalid expression in the address")); return FALSE; } /* [Xn, */ if (imm_shift_mode != SHIFTED_NONE && skip_past_comma (&p)) /* [Xn,, */ if (! parse_shift (&p, operand, imm_shift_mode)) return FALSE; } } } if (! skip_past_char (&p, ']')) { set_syntax_error (_("']' expected")); return FALSE; } if (skip_past_char (&p, '!')) { if (operand->addr.preind && operand->addr.offset.is_reg) { set_syntax_error (_("register offset not allowed in pre-indexed " "addressing mode")); return FALSE; } /* [Xn]! */ operand->addr.writeback = 1; } else if (skip_past_comma (&p)) { /* [Xn], */ operand->addr.postind = 1; operand->addr.writeback = 1; if (operand->addr.preind) { set_syntax_error (_("cannot combine pre- and post-indexing")); return FALSE; } reg = aarch64_reg_parse_32_64 (&p, offset_qualifier); if (reg) { /* [Xn],Xm */ if (!aarch64_check_reg_type (reg, REG_TYPE_R_64)) { set_syntax_error (_(get_reg_expected_msg (REG_TYPE_R_64))); return FALSE; } operand->addr.offset.regno = reg->number; operand->addr.offset.is_reg = 1; } else if (! my_get_expression (exp, &p, GE_OPT_PREFIX, 1)) { /* [Xn],#expr */ set_syntax_error (_("invalid expression in the address")); return FALSE; } } /* If at this point neither .preind nor .postind is set, we have a bare [Rn]{!}; reject [Rn]! but accept [Rn] as a shorthand for [Rn,#0]. */ if (operand->addr.preind == 0 && operand->addr.postind == 0) { if (operand->addr.writeback) { /* Reject [Rn]! */ set_syntax_error (_("missing offset in the pre-indexed address")); return FALSE; } operand->addr.preind = 1; inst.reloc.exp.X_op = O_constant; inst.reloc.exp.X_add_number = 0; } *str = p; return TRUE; } /* Parse a base AArch64 address (as opposed to an SVE one). Return TRUE on success. */ static bfd_boolean parse_address (char **str, aarch64_opnd_info *operand) { aarch64_opnd_qualifier_t base_qualifier, offset_qualifier; return parse_address_main (str, operand, &base_qualifier, &offset_qualifier, REG_TYPE_R64_SP, REG_TYPE_R_Z, SHIFTED_NONE); } /* Parse an address in which SVE vector registers and MUL VL are allowed. The arguments have the same meaning as for parse_address_main. Return TRUE on success. */ static bfd_boolean parse_sve_address (char **str, aarch64_opnd_info *operand, aarch64_opnd_qualifier_t *base_qualifier, aarch64_opnd_qualifier_t *offset_qualifier) { return parse_address_main (str, operand, base_qualifier, offset_qualifier, REG_TYPE_SVE_BASE, REG_TYPE_SVE_OFFSET, SHIFTED_MUL_VL); } /* Parse an operand for a MOVZ, MOVN or MOVK instruction. Return TRUE on success; otherwise return FALSE. */ static bfd_boolean parse_half (char **str, int *internal_fixup_p) { char *p = *str; skip_past_char (&p, '#'); gas_assert (internal_fixup_p); *internal_fixup_p = 0; if (*p == ':') { struct reloc_table_entry *entry; /* Try to parse a relocation. Anything else is an error. */ ++p; if (!(entry = find_reloc_table_entry (&p))) { set_syntax_error (_("unknown relocation modifier")); return FALSE; } if (entry->movw_type == 0) { set_syntax_error (_("this relocation modifier is not allowed on this instruction")); return FALSE; } inst.reloc.type = entry->movw_type; } else *internal_fixup_p = 1; if (! my_get_expression (&inst.reloc.exp, &p, GE_NO_PREFIX, 1)) return FALSE; *str = p; return TRUE; } /* Parse an operand for an ADRP instruction: ADRP ,