NeXTSTEP 3.2 (Developer)

home *** CD-ROM | disk | FTP | other *** search

/ NeXTSTEP 3.2 (Developer) / NS_dev_3.2.iso / NextDeveloper / Source / GNU / cctools / as / notes < prev next >

Wrap

Text File | 1993-02-05 | 65.0 KB | 1,235 lines

Known bugs: - The m88k instruction "tb0 0,r0,undef" trashes the instruction because of the undefined. - "0: jmp 0b" does not work, 0 is the problem in the 0b (1-9 work). - 68k does not handle packed immediates (the tables have been changed to dis- allow this) because there is no routine to convert a flonum to a 68k packed form. - The logical operators && and || are not implemented. To do: - Mark the instructions that are 601 specific (book IV) and registers with /* 601 only */ comments on them and flag as an error unless -force_cpusubtype_ALL is specified. Changes for the 3.1 release (the cctools-20 release): - Fixed a bug for the m98k that did not correctly check for too few parameters. Two bugs here one in calcomp that was testing != NONE which should have been == NONE and a bug in md_assemble in advanceing past the '+' or '-' when it did not exist and there was nothing but a '\0' after the op it advanced past it. Changes for the 3.1 release (the cctools-16 release): - Fixed a bug with the m98k opcodes for stwcx. stdcx. where bit 0 was not set to a 1. - Changed the following instructions so that for the SH field the value 32 (or 64) assembles as 0: rldicl rldicl. rldicr rldicr. rldic rldic. rldimi rldimi. rlwinm rlwinm. rlwimi rlwimi. - Fixed a bug in the m98k assembler where the value of exprressions which was exactly (1 << width) was not treated as an error (4 places > was changed to >= in m98k.c). Changes for the 3.1 release (the cctools-15 release): - Moved the m88k and m98k to be install in /usr/local/lib not /lib. - Fixed a bug in the m98k assembler that did not detect instructions with too many parameters. - Added macros and register names for batX[ul] the same as ibatX[ul] since the 601 does not have split i and d for these. - Changed the m98k instruction's "icbi" first paramenter to G0REG from GREG. - Back out the below fix and added new code in try_to_make_absolute() that walked the frags between the the symbols L0 and L1 to calculate the absolute value. - Fixed a bug where the expression L1-L0 was not coming up right when it had a .align between L0 and L1. A hack was removed from try_to_make_absolute() in expr.c that had the code ifdef out that was trying to say the expresion could change due to relaxation. The the routine s_align() in read.c was ifdef RISC to do the alignment instead of creating an align frag. Changes for the 3.1 release (the cctools-14 release): - Added a form of fcmpu and fcmpo that takes a crX as it's first argument. - Added the opcodes for tlbiex (31,338) and tlbia (31,370). - Fixed a bug in the m68k assembler where the code to handle implementation specific instructions had || in two places where && was supposed to be. The change was on lines 2693 and 2706 in md_assemble() in m68k.c . - Changed the m98k instructions: lwarx, ldarx, stwcx. and stdcx. second arg from GREG to G0REG. - Fixed the Makefile to install the m98k assembler it built. Changes for the 3.1 release (the cctools-13 release): - Added the m98k (PowerPC) architecture. Changes for the 3.1 release (the cctools-10 release): - Changed the default .include directories to /NextDeveloper/Headers and /LocalDeveloper/Headers in as.c and made them work (the default never worked). - Corrected the following table entries for i386 floating point instructions which had a FloatD or'ed into them which was wrong: faddp, fmulp. - Fixed a bug that caused an error message that a segment override was specified more than once for string instructions for the i386. The fix was in i386.c where i.seg was need to be set to zero after the string instruction operands were parsed (bug #29867). - Fixed the assembler driver /bin/as to take machine specific architecture flags and run the family architecture named assembler. So "as -arch m68040" will run /lib/m68k/as and not /lib/m68040/as. - Changed the handling of the cpusubtypes to support this design: The design allows each architecture family (m68k, m88k, i386) to have a cpusubtype that will run on all implementations of that architecture and have cpusubtypes that will run only on specific implementations of an architecture. The design also gives the programmer the option to force the cpusubtype indicating all implementations (at build time with a flag) when using instructions that are implementation specific leaving it up to the program to runtime select the correct code to execute. The assemblers implementation of this design is: By default the assembler will produce the cpusubtype ALL for the object file it is assembling if it finds no implementation specific instructions. Again by default the assembler will allow implementation specific instructions for implementations and combine the cpusubtype for those specific implementations. The combining of specific implementations is architecture dependent and may not be allowed for some architectures and are flagged as an error. With the optional -force_cpusubtype_ALL flag all instructions are allowed and the object file's cpusubtype will be ALL. If an -arch flag for a specific implementation (ie. -arch m68040 or -arch i586) is used the assembler will flag as errors instructions that are not supported on that architecture and produce the cpusubtype for that specific implementation in the object file (even if no specific instructions are used). This effected as.c, as.h, Mach-O.c, m68k.c, m68k-opcode.h, i386.c, i386.h, and i386-opcode.h. The m88k and i860 assemblers had no machine specific modifications. Changes for the 3.1 release (the cctools-9 release): - Fixed a bug that caused the .include feature to fail in some cases. The value of the stuff saved by save_scrub_context() in app.c was not reset which caused the app preprecessor to start parsing the included file and think it was in the case of a string. Changes for the 3.1 release (the cctools-8 release): - Fixed a bug the did not cause m68k floating point branchs to undefined symbols to have relocation entries that make it to the object file to work with scattered loading. On line 3299 in m68k.c was: fix_new(fragP,(int)(fragP->fr_fix),4,fragP->fr_symbol, (symbolS *)0,fragP->fr_offset+4,1); changed to: fix_new(fragP,(int)(fragP->fr_fix),4,fragP->fr_symbol, (symbolS *)0,fragP->fr_offset+4,1+2); - Fixed a bug in the i386 assembler for these two instructions where the segment override did not get picked up from: mov %eax,%gs:sr_mem_offset jmp %gs:sr_mem_offset The first is bug #29555 the second is just another form of the same logic bug in another place. There maybe more of this same logic bug. The fixes are in i386.c in md_assemble() when putting out the opcode. - Fixed a bug in the string instructions where segment overrides in the operand fields were not picked up. To do this the kludge that ignored the operands of string instructions had to be removed as special case table entries and matching checking had to be added (bug #26409). - Fixed a bug in the i386 assembler where the invlpg instruction did not take a Mem operand (it was Mem32). The fix was in the table entry for invlpg in i386-opcode.h (change requested by the Lono group). The manual is confusing on this instruction. - Fixed a bug in the i386 assembler where a call or jmp instruction to an absolute address was not getting put out pc relitive and no relocation entry was produced (line 1050 and line 1155 in i386.c). - Fixed the problem of getting alignment correct for .align directives that are greater than the default alignment. This effected the struct frchain in subsegs.h, the routine set_section_align() in Mach-O.c and the routine write_object_file() in write.c and the initialization of the new field in subsegs.c (bug #29432). - Changed the I386 bound instruction such that the parameters are consistant with gas (reversed them). Also fixed the boundw so it only put out one '0x66' data16 prefix byte. - Fixed a bug for the I386 that padded the text section with zeros (changed to pad with nops) in write.c. - Added the wait (0x9b) prefix to the following instructions: "finit", "fstcw", "fstsw", "fclex", "fstenv" and "fsave" the "XnXXX" form does not have the wait prefix. - Added "fucom" as an alias for "fucom %st(1)" - Added "fucomp" as an alias for "fucomp %st(1)" Changes for the 3.1 release (the cctools-7 release): - Added the i486 and i586 specific stuff to i386-opcode.h (bug #27475). The changes are ifdef'ed i486 and i586 and these are turned in in the Makefile. Also the define STRICT_i586 is ifdef'ed (but not defined) where the i586 does not allow certian things (test register instructions). - Fixed a bug in md_assemble() in i386.c where the instruction "mov %ah,%al" would assemble wrong. The problem was when the suffix was selected based on the register type the "i.types[o]" needed to and'ed with `Reg' because %al and %ax have `Acc' in their types and they were coming up with a 'l' suffix. This is ifdefed NeXT. - Fixed a bug in m68k_reg_parse() in m68k.c where the registers "ic", dc" and "bc" were incorrectly parsed because of an "if (c2 = 'c')" (bug #27954). - Added an ifdef CHECK_WORD_IMMEDIATES in m68k.c and code to make checking of 16-bit immediates consistant in the m68k assembler (bug #26863). Also to make this work the "link" (no suffix) for an immediate word entry in m68k-opcode.h had a new "place" character created for it (#z for #w) and code was added to m68k.c to handle it. The define CHECK_WORD_IMMEDIATES is left off to cause truncation of too large immediate words. - Fixed a bug that did not allow -arch and -arch_multiple (as.c). This was put in to the NRW cctools-6 but not into lono's. Changes for the 3.1 release (the cctools-6 release): - Added the -arch_multiple flag that will cause one line to printed before all error messages which will contain the architecture name. - Fixed the m88k pmul and punpk instructions where the last register is not required to be an even register. - Fixed a bug in atof-ieee.c in gen_to_words() that did not round IEEE denorms correctly and caused the smallest denorm to become zero (bug #23178). Changes for the 3.1 release (the cctools-5 release): - Picked up the lono team's cctools-4_2 i386-opcode.h . - Added the pseudo op for the m88k assembler ".dot symbol" that sets the value of the location counter into the value of the symbol. - Removed the trnc.x mnemonic as it is not legal "trnc.sx" is the correct form which remains in m88k-opcode.h { 0x8400d900, "trnc.x", { {21,5,REG}, {0,5,XREG}, {0,0,NIL} } }, Changes for the 3.1 release (the cctools-4 release): - Fixed a bug in parse_cst() in m88k.c that did not allow expressions for constant operands. This bug was found with "tb0 0,r0,(129)" where the ()'s caused the problem. (bug #21052) - Changed installing the i386 assembler into /lib/i386/as from ix86 (and changed the -arch name to i386). - Changed CPU_TYPE_I80x86 to CPU_TYPE_I386 in Mach-O.c - Picked up the changes for the i386 assembler to allow scattered loading from the lono team. Changes for the 3.0 release (the -57 compiler release) - Removed the following opcodes from m88k-opcode.h as siff(1) and the newest 110 manual says they are not valid. { 0X840008A0, "fcvt.dd", { {21,5,REG}, {0,5,REG}, {0,0,NIL} } }, { 0X840088A0, "fcvt.dd", { {21,5,XREG}, {0,5,XREG}, {0,0,NIL} } }, { 0X84000800, "fcvt.ss", { {21,5,REG}, {0,5,REG}, {0,0,NIL} } }, { 0X84008800, "fcvt.ss", { {21,5,XREG}, {0,5,XREG}, {0,0,NIL} } }, { 0X84000800, "fcvt.ss", { {21,5,REG}, {0,5,REG}, {0,0,NIL} } }, { 0X84008800, "fcvt.ss", { {21,5,XREG}, {0,5,XREG}, {0,0,NIL} } }, (bug #20021) - Fixed a bug introduced with the change of the SUB_SEGMENT_ALIGN. It turns out it broke the objective-C runtime that assumed that the __protocol section (amoung others) can be indexed like an array of structs which are not multiples of 8 bytes. The fix was to align all objective-C sections to 4 bytes. Again the change was in write_object_file() in write.c (bug #20022). Changes for the 3.0 release (the -56 compiler release) (performance month): - Changed the order of the objective-C sections. The message_refs and the cls_refs section were switched. The meta_cls_refs section was removed. This change effected Mach-O.c, write.c and read.c. - Changed write_object_file() in write.c to used the normal subsegment alignment: #ifdef RISC #define SUB_SEGMENT_ALIGN (3) #else #define SUB_SEGMENT_ALIGN (2) #endif and handle the literal pointer sections special (by knowing their subsegment values). This fixes a problem on the m88k where the const section had a .align 3 directive but started on a align of 2 boundary. This still has the problem if a section has an align greater 3 the data in the output file will end up aligned correctly but the section start will not resulting in the link edited object to having the data not aligned correctly. (bug #19492) Changes for the 3.0 release (the -55 compiler release) (performance month): - Changed the Makefile to install the driver in /usr/local/bin and the 68k assembler in /bin/as and all other assemblers in /usr/local/lib/* . - Changed the as driver (driver.c) to look in /lib and in /usr/local/lib for assemblers. - Changed the order of the objective-C setions to: 2 (__OBJC,__class) 3 (__OBJC,__meta_class) 4 (__OBJC,__string_object) 5 (__OBJC,__protocol) 6 (__OBJC,__cat_cls_meth) 7 (__OBJC,__cat_inst_meth) 8 (__OBJC,__cls_meth) 9 (__OBJC,__inst_meth) 10 (__OBJC,__cls_refs) 11 (__OBJC,__meta_cls_refs) 12 (__OBJC,__message_refs) 13 (__OBJC,__class_names) 14 (__OBJC,__module_info) 15 (__OBJC,__symbols) 16 (__OBJC,__category) 17 (__OBJC,__meth_var_types) 18 (__OBJC,__class_vars) 19 (__OBJC,__instance_vars) 20 (__OBJC,__meth_var_names) 21 (__OBJC,__selector_strs) Also the special casing of the objective-C section in determing to created a scatter or non-scattered relocation entry was removed for all but the (__OBJC,__selector_strs) section. The directive ".objc_selector_refs" is still there and the cc-55 compiler will be changed to use the correct directive ".objc_message_refs" and then this can be removed. These changes effected read.c and Mach-O.c Changes for the 3.0 release (the -54 compiler release) (performance fortnight): - Added three string sections to the Objective-C segment: .objc_class_names, __OBJC __class_names .objc_meth_var_names, __OBJC __meth_var_names .objc_meth_var_types, __OBJC __meth_var_types This effected read.c and Mach-O.c. - Added the following lines to i386-opcode.h at the request of the lono guys: {"repz", 0, 0xf3, _, NoModrm, 0, 0, 0}, { "repnz", 0, 0xf2, _, NoModrm, 0, 0, 0}, Plus allow .word on the ix86 (ifdef added in read.c). - Added const to lex_type, is_end_of_line, potable in read.c to make read-only. - Added const to op_encoding in expr.c to make read-only. - Added const to m68k_opcodes and endop in m68k-opcode.h to make read-only. - Changed in_buf in input-file.c to allways be malloc()'ed. - Added const to the Makefile for the next_version.c echo line. Changes for the 3.0 release (the -51 compiler release): - Changed Mach-O.c to pad out the string table to a multiple of 4 and set the padded bytes to '\0'. Changes for the 3.0 release (the -50 compiler release): - Changed the instructions f[sd]sqrtx for operands FPm,FPn and FPm to f[sd]sqrt (without the trailing 'x'). This was just wrong (but the moto opcode is inconsistant with all others in this case). - Fixed a bug in the fsincos instruction where the FPc and FPs registers were switched in the instruction. - Changed the order of the objective-C classes to: __class - always written on __meta_class - always written on __string_object - always written on __protocol - always written on __cls_meth - sometimes written on __inst_meth - sometimes written on __cat_cls_meth - sometimes written on <these will not be used soon> __cat_inst_meth - sometimes written on <these will not be used soon> __cls_refs - sometimes written on (uniqued) <these are not used now> __meta_cls_refs - sometimes written on (uniqued) <these are not used now> __message_refs - sometimes written on (uniqued) __symbols - never written on __category - never written on __class_vars - never written on __instance_vars - never written on __module_info - never written on __selector_strs - never written on (uniqued) The six sections starting from the __string_object section were effected. The change was made in read.c and Mach-O.c . Changes for the 3.0 release (the -49 compiler release): - Fixed a bug where the assembler was padding literal pointer sections with a zero for RISC machines and causing the link editor to complain. The fix was to change the macro SUB_SEGMENT_ALIGN from 3 to 2 in write.c and to set the alignment of S_LITERAL_POINTER sections in Mach_O.c to 2. - Fixed the passing and using of RC_CFLAGS correctly in the Makefile. Changes for the 3.0 release (the -49 compiler release): - Changed the Makefile to meet the RC api. Changes for the 3.0 release (the -47 compiler release): - Added the missing 040 "ptest[rw] An@" instructions. - Changed the constant CPU_TYPE_I386 to CPU_TYPE_I80x86 to match header file. - Changed the behavior so if warning message is produced (with as_warn()) that an object is not produced. The change was in as_warn() in messages.c and is ifdef'ed NeXT which sets bad_error to 1 just like as_bad(). (bug #16137 and #16044) - Added the (__OBJC,__string_object) section with the directive .objc_string_object (read.c and Mach-O.c where changed). Changes for the 3.0 release (the -44 compiler release): - Created an assembler driver to be placed in /bin/as and the assemblers are then moved to /lib/<arch_flag>/as . The Makefile was updated to build and install this way. as.c was changed to take "-arch <arch_flag>" and check it against the type of assembler it is. - Switch over to the new header file organization. Changes for the 3.0 release (the -43 compiler release): - Changed the Makefile to install the i860 assembler in /usr/local/bin/as860 . - Picked up md.h from 1.38.1 which added const to md_pseudo_table and md_relax_table so i860.c, m68k.c and m88k.c were all updated as were the uses in read.c and write.c. - Picked up the files: i386.c, i386.h and i386-opcode.h from the 1.38.1 release. Changes for the 3.0 release (the -39 compiler release): - Fixed so that strings can have characters with the 8th bit set in them. This involved adding the lines in next_char_of_string() in read.c: #ifdef NeXT /* allow 8 bit chars in strings */ c = (c & MASK_CHAR); /* to make sure the 0xff char is not returned as -1*/ #endif NeXT so that the high bit does not get sign extened and the -1 return code that is tested for at the call sites as >= 0 is not tripped over. Second changed all 8th bit set chars in lex_type[] in read.c to be allowed in names. Also had to change the macros in read.h #define is_name_beginner(c) ( lex_type[(c) & 0xff] & LEX_BEGIN_NAME ) #define is_part_of_name(c) ( lex_type[(c) & 0xff] & LEX_NAME ) to add the "& 0xff" because of the sign extension of chars (bug #15597). Changes for the 3.0 release (the -37 compiler release): - Fixed the relocation entries for the 88k so that 88k objects can be scatter loaded by the link editor. This involves adding a PAIR relocation entry for the LO16 type (as well as having one for the HI16 type) and moving the place in the relocation entry where the other half of the relocatable expression is store from the r_symbolnum field to the r_address field (so that a scattered relocation can be used, since it does not have an r_symbolnum field). Also removed all support for signed immediates on the 88110 since NeXT will not use this feature. Also to be consistant the i860's PAIR's will also use the r_address field even though they will not use scattered relocation entries. These changes were made in Mach-O.c . Also required forcing relocation entries for non-local labels for 88k branch instructions which was done with the same kludge as the 68k by setting the 0x2 bit in the fx_pcrel fix structure when it is created in m88k.c in md_assemble(). This also required an extra ifdef M68k in Mach-O.c in fix_to_relocation_info() when choosing to put out a scattered relocation entry because of the way 68 branch instructions work. Changes for the 3.0 release (the -36f compiler release): - Fixed a bug that did not catch a bit field expression error between the <>'s. A check for the trailing '<' was missing. This was added in parse_bf() in m88k.c . - Fixed the .abs directive for 88k registers. The fix was for it to handle scaled register expressions and also not generate an undefined symbol for the register name "U r1". The changes were to m88k.c adding s_m88k_abs() for the .abs pseudo-op and to read.c to leave s_abs() but to ifdef it out of the table for the 88k as s_m88k_abs() uses s_abs(). - Corrected the lex_type table in read.c to not allow the character '[' as part of a name. - Added '@' as a statement separator for the 88k (change the "# <line> <file> <level>" stuff to use it when generating ".line <line> @ .file <file>" ). Changed app.c and read.c ifdef'ed M88K. Also s_reg(), s_scaled() and no_delay() in m88k.c also need this because they can't use the macros in read.c . - Added the .no_delay 88k pseudo-op. Changed m88k.c and m88k-opcode.h to add the delay_slot field to the instruction table and a static variable, in_delay_slot, that gets set each time an instruction is assembled. - Fixed a bug not allowing macro names to start with '.' . The fix was in read.c in parse_a_buffer() right before it detects an unknown pseudo-op. Also change it so the unknown pseudo-op is printed when an error happens. Also changed s_macro() in read.c to print a warning if a known pseudo-op name is used a macro name. Changes for the 3.0 release (the -36e compiler release): - Fixed a bug where the operand "pc@" did not assemble correctly. The fix was in m68_ip() in m68k.c on linr 1604 for the "case AINDR" which can be set with PC in the opP->reg. In this case the mode pc@(d16) is used. This is ifdef'ed NeXT. - Fixed a bug where "foo :" did not recognize "foo" as a label. The fix was in app.c in do_scrub_begin() where the line: lex [':'] |= LEX_IS_LINE_SEPERATOR; was ifdef'ed out since ':' did not work. But ... This DOES not cause ':' to be a LINE SEPERATOR but does make the second if logic after flushchar: in do_scrub_next_char() to handle "foo :" and strip the blanks. This is the way has always been and must be this way to work. Changes for the 3.0 release (the -36d compiler release): - Fixed a bug in the 88k assember that did not handle "# <line> <file> <level>" comments correctly because it uses ";" which is a comment and the .file gets ignored. The fix was ugly. The change was to app.c and read.c ifdef'ed M88K to allow '\001' as a statement seporator (CHANGED IN -36f see above). - Changed the marking of literal sections from not marking them for RISC to not marking them for only the I860. This change is since the 88k compiler will ALLWAYS make a 32 bit reference to an item and leave it to the link editor to find ways to make 16 bit references these sections can marked for uniqueing for the 88k. - Added the following directives for the following new sections: .constructor for __TEXT, __constructor .destructor for __TEXT, __destructor .objc_protocol for __OBJC, __protocol .objc_cls_refs for __OBJC, __cls_refs (S_LITERAL_POINTERS) .objc_meta_cls_refs for __OBJC, __meta_cls_refs (S_LITERAL_POINTERS) Changes for the 3.0 release (the -36c compiler release): - Fixed a bug involving expressions with unknown symbols for operators other than '+' or '-'. The problem is that in expr() in expr.c if an expression operator is something other than '+' or '-' then it sets the need_pass_2 flag and no other "frags" (bytes of output) are generated. You would think it would want to run another pass but the code doesn't do that (major bug)! So now it just does what it would in the case the symbol is known which is report the expression can't be relocated. - Fixed m88k.c to use LO16 relocation types not LO16EXT types. Changes for the 3.0 release (the -36b compiler release): - Added the m88k directive .scaled as requested by the OS group. - Allow expressions for the bit number, parse_cmp() like in bb0, condition match, parse_cnd() like in bcnd, and even 4 bit rotation, parse_e4rot like in prot. Changes for the 3.0 release (the -36a compiler release): - Added the opcodes "illop[123]" as per the version 3.0 88110 spec. - Removed the "lda.b rD,rS1[rS2]" instruction and replaced that opcode with "lda.x rD,rS1[rS2]" as per the version 3.0 88110 spec. - Corrected "nint.[sdx]" to be "nint.s[sdx]" and "int.[sdx]" to be "int.s[sdx]" which was just wrong in the GNU assembler (trnc was previous corrected but flaged as an 88110 versrion 3 change but that was incorrect as the assembler was just wrong (even for the 88100)). - Corrected "mov.s xD,xS2" to be "mov xD,xS2" as per the version 3.0 88110 spec. - Removed the old (version 2.0 of the 88110) opcodes: "mov.t xrD,rD2" and "mov.t rD,xrS2" which used blank instead of .s for single. - Removed the old (version 2.0 of the 88110) opcodes: "trnc.t rD,xrs2" and "trnc.t rD,rS2" (where t is the type of the result) which used only the type of the result and implied the .s for single. - Removed the "ppack.8.b", "ppack.8.h", and "ppack.16.b" opcodes from the m88k opcode table. These operations are undefined. Changes for the 3.0 release (the -35 compiler release): - Fixed a bug in parse_bf() when expressions were added expressions that did not start with a digit (for example a '(', '+', '-', '~' or '!') were not recognized. - Changed the action for .abort to print the remaining line as part of the error message. Feature request by the OS group. - Added an option [,alignment] expression to the end of the .lcomm directive which aligns the symbol to the power of 2 alignment expression (same as the .align directive). This is ifdef'ed NeXT in s_lcomm() in read.c . - Changed which directives are allowed on which machines .word 68k and i860 only (machine specific) NOT 88k .long 68k and i860 only NOT 88k .quad 68k only .octa 68k only .float 68k and i860 only NOT 88k These changes are in read.c, m68k.c and i860.c . Feature request by the OS group and removal of .quad and .octa for the i860 approved by the NeXT Dimension group. - Added the directive .elseif . This involed a bit of reworking the .if, .else and .endif stuff in read.c . Feature request by the OS group. - Fixed a bug that would allow you to use the macro in a macro definition and cause the assember to core dump. A limit, MAX_MACRO_DEPTH of 20, is used to avoid this. - Added the directives .macros_on .macros_off. This is to allow macros to be turned off, which allows macros to override a machine instruction and still use the machine instruction. This is in read.c and toggles the variable macros_on in s_macros_on() and s_macros_off() which is tested in parse_a_buffer(). Feature request by the OS group. - Added s_abs() in read.c to implement ".abs symbol,exp" which set symbol to 1 or 0 depending on if the expression is an absolute expression. This is intended for use in macros. Feature request by the OS group. - Added s_reg() to m88k.c to implement ".greg symbol,exp" and ".xreg symbol,exp" which set symbol to 1 or 0 depending on if the expression is a general register or extended register respectfully. These are intended for use in macros. Feature request by the OS group. - Added $n in expand_macro() in read.c to substitute the number of actual arguments for the macro. Feature request by the OS group. - Changed the code for setting the line separator's (character for multiple statements on a line) in do_scrub_begin() in app.c . The character '/' tried to be a separator for the 88k but code down stream prevented it from working so it was removed and the 88k does not allow multiple statements on a line. Also removed the NeXT ifdef for the ':' character which also did not work. Changes for the 3.0 release (the -34 compiler release): - Fixed a bug that for all floating-point branches it did not generate a relocation entry. The fix is in md_estimate_size_before_relax() in m68k.c where the case of 'TAB(FBRANCH,SZ_UNDEF)' was not handled in the first switch and code to generate the relocation entry (a call to add_fix) was not done. This is ifdef'ed NeXT. - Fixed a bug for branches that use displacements to absolute addresses which produced a displacement off by -4. This is ifdef'ed NeXT in m68_ip() in m68k.c in the second main switch statement that sets the bits in the instruction for the 'B' case. There are two ifdef's, one for 'g' sub case (normal branches) and one for the 'c' sub case (floating-point branches). - Disallow all floating-point packed immediates for the 68k assembler because the gen_to_words() routine in atof-ieee.c does not produce the correct 68k packed decimal format. This simply disallows this but does not fix it. So "fmovep #0r1.0,fp0" will no longer assemble instead of assemble wrong. This is ifdef'ed PACKED_IMMEDIATE in m68k-opcode.h and m68k.c in m68_ip(). (internal bug #5) - Fixed a bug in the assembler which matched the "fmoveml #1,fpc" where the immediate #1 cause an internal FATAL error because it can't decode the mode "*s". The fix is in m68_ip() in m68k.c where the case for 's' was ifdef'ed NeXT in just like the long case. This is legal but the instruction "fmoveml #1,fpc/fpi" is not and the assembler STILL accepts it. (internal bug #4). - Fixed a bug in the assembler which matched the "movec sfc,#1" where the immediate #1 cause an internal FATAL error because it can't decode the mode "#j". The fix in m68_ip() in m68k.c in the loop to install the operand bits for the '#' case was missing the second sub case for 'j' that check the range of the value and installed the operand. If the immediate is a variable my guess this will still fail but in a different way. (internal bug #3). - Fixed a bug that caused the assembler to call abort with no error message when assembling "andiw #0x8001,fpir/fpcr/fpsr". In get_num() in m68k.c the case for a SEG_PASS1 was missing from the switch statement from the type of the expression() call. It was ifdef'ed NeXT in and handled like SEG_UNKNOWN and a bunch of others that just print out it can't handle the expression. STILL BROKEN! (internal bug #2). - Fixed a bug that the operand of the form "zpc@(0)" was trying to use the pc@(16) (AOFF) form which does not allow the base register to be suppressed which is what zpc is. So this now uses the pc@(bd,Xn) form (AINDX). The bug caused "zpc@(0)" to generate garbage, namely "d1". The change is in m68k_ip_op() in m68k.c and ifdef'ed NeXT with a comment like above. (internal bug #1). - Ifdef'ed out the turning operands into PC relative in m68_ip() in m68k.c (this is a 1.36 feature) because it breaks scattered loading. - Fixed a bug in the 1.36 version of GAS where the table of fmovem instructions were reordered. See the comment in m68k-opcode.h with the header "REGISTER LIST BUG:". The fix was to put the list back in the previous order. There is a design bug here that needs to be fixed. - Fixed a bug where the .align directives were not propagated into the section headers of the object file. A new routine, set_section_align() in Mach_O.c, is called from s_align() in read.c . - Put the change in atof-ieee() in atof-ieee.c that creates infinities on overflows. This fixes bug #13409. - Picked up a change in i860_ip() in i860.c from the NDTools-6 version. Having to do with constant offset alignments. - Added expressions to the width and <offset> bit field instructions. Since the parameter syntax is width<offset> and offset may be a two character 'cmp' bit designator, the width expression may not contain the character '<' and the offset expression must start with a digit. - Changed "mov.t xrD,rD2" and "mov.t rD,xrS2" to use .s for single instead of blank. (version 3.0 of 88110 spec). - Changed "trnc.t rD,xrs2" and "trnc.t rD,rS2" (where t is the type of the result) to use .st where the s is for single and t is the type of the result. (version 3.0 of 88110 spec). - Changed the pflusha instruction to pflusha030 and pflusha040 because there is no way to tell them apart. - Added automatic creation of stabs for assembly code debugging with -g. The comment that explains in detail how this is done is in read_a_source_file() in read.c, The other changes are in make_stab_for_symbol() in symbols.c, s_include(), s_line() in read.c, and md_assemble() in m68k.c and m88k.c also two static declarations were removed from input-scrub.c. These changes are marked with pairs of: #ifdef NeXT /* generate stabs for debugging assembly code */ ... #endif NeXT /* generate stabs for debugging assembly code */ - Added the MMU instructions for the 030 and 040 (ifdef'ed BUILTIN_MMUS) and turned off the m68851 define for that set of MMU instructions. The reason to turn it off is because of the register names it must recognize (see bug #7525 why we don't want to do this). This change is not ifdef'ed NeXT because it is very intertwined with the 68851 stuff. Also with this change the "MMU status register" correct name of "mmusr" was added but the old name of "psr" was retained for compatiblity because of assembler code that might use it. - Added installsrc, installIBMsrc and installGNUsrc targets to the Makefile. - Bug #8331, feature request for hex immediate bit-patterns for floating-point immediates. Added the constant radix 0b... like 0x... except that it would be assumed to be a "bignum" and then a binary bit pattern for a hex immediate. This effected the routines operand() in expr.c, get_num() in m68k.c and m68_ip() in m68k.c . All of these are ifdef'ed NeXT with the comment /* fix for bug #8331 */ . - Bug #13017, this is where ".fill x,3,x" would cause the assembler to call abort because the repeat size is 3 bytes. This is now dissallowed in s_fill() in read.c and only repeat sizes of 0,1,2 and 4 are allowed. - Bug #11209, this is where if the file system fills up or something and the file can't be closed the object file was left and would confuse later make(1)'s because the object file would be present but would then just hand this off to the link editor and it would complain about a bad object file. The fix in output_file_close() in output-file.c was to remove the file in this case because it might be bad. - Bug #8920, where s file containing just "bra L1" would produce a bad object file because the undefined local lable L1 was not defined is fixed. The fix is in write_object_file() in write.c (and one line in write_Mach_O() in Mach-O.c to test bad_error). The undefined local symbols are printed with an error message in this case and then the object file is not written. - Bug #8918, where a line of the form "# 40 MP1 = M + 1" gets confused with a line of the form "# 1 "hello.c" 1" and causes a bug that ignores the rest of the file. This was fixed in app.c when in state 4 (after putting out a .line, put out digits) and then not finding an expected '"' for the name of the file it ignores the rest of the line but forgets to set the state to 0 (begining of a line). This is ifdef'ed NeXT. - Bug #7525 (second part), where "bfffo d0{#2,#32},d1" would not work with the field width of 32 is now fixed. (I'm not sure exactly what the fix was it probably came from the 1.36 version of GNU). - Bug #5384, where if a ".globl foo" precedes "foo=1" foo does not end up global has been verfied to be fixed (I'm not sure exactly what the fix was it probably came from the 1.36 version of GNU). - Changed the default alignment of sections to 3 (8) for RISC machines from 2 (4) in both write.c and MachO.c. - Print a warning for -R (make data text) to used .const and not put the data in the text. - Cleaned up Mach-O.c and read.c by changing/adding message_refs where selector_refs was used. --- Changes to merge in John Anderson's (DJA) version of GAS --- - added relational binary operators (<, ==, >, <= and >=) and modified the precedence to conform to 'C'. The code is marked with pairs of: #ifdef NeXT /* added relational ops, changed precedence to be same as C */ ... #endif NeXT /* added relational ops, changed precedence to be same as C */ and is contained in the file expr.c and is the DJA version with a few bug fixes to make it work. Found a logic bug when "<>" was used as an operator it was recognized as a "<". This "operator" appears in the WriteNow source so I added "<>" as a form of "!=". - added logical negation unary operator (!). The code is marked with pairs of: #ifdef NeXT /* logical negation feature */ ... #endif NeXT /* logical negation feature */ and is contained in the file expr.c and is exactly the DJA version. - added code to try to make expresions absolute. The code is marked with pairs of: #ifdef NeXT /* feature to try to make expressions absolute */ ... #endif NeXT /* feature to try to make expressions absolute */ and is contained in the files expr.c and m68k.c (the code is exactly the DJA version). - added the .dump/.load feature (this is based on top of the .include and .macro features). The code is marked with pairs of: #ifdef NeXT /* the .dump/.load feature */ ... #endif NeXT /* the .dump/.load feature */ and is in read.c (and one line in symbols.c) and is the DJA version. Fixed a bug in write_symbol() in read.c where the symbol's n_type field needed to be and'ed with the N_TYPE macro before checking for equal to N_ABS. not checked - added the conditional assembly feature (pseudo ops .if .else .endif) and the macro feature (pseudo ops .macro and .endmacro). This is all contined read.c and required a major rewrite of the main parsing routine read_a_source_file(). This was replaced by three routines read_a_source_file(), parse_a_buffer() and parse_line_comment(). Since the their was no way to ifdef the old code it was removed. Where possible the conditional assembly feature code is marked with pairs of: #ifdef NeXT /* the conditional assembly feature (.if, .else, and .endif) */ ... #endif NeXT /* the conditional assembly feature (.if, .else, and .endif) */ and the macro feature code is marked with pairs of: #ifdef NeXT /* the .macro feature */ ... #endif NeXT /* the .macro feature */ All of these changes are in read.c and except for the rewrite read_a_source_file() the changes are the DJA version. - added the .include "filename" pseudo op. This is marked with pairs of: #ifdef NeXT /* .include feature */ ... #endif NeXT /* .include feature */ the code in in read.c, as.c, app.c, as.h and input-scrub.c. Except for the code in app.c and the typedef scrub_context_data in as.h (related to the major changes in the app.c code from the DJA version) it is exactly what was in the DJA version. Fixed a bug in input_file_open() in input-file.c where it was doing a setbuffer() call using a staticly allocated buffer for all the file's in read. This was changed to use a dynamicly allocated buffer when processing an include file so the buffer does not get reused by include files. Changes for the 3.0 release (the -33 compiler release): - Fixed trap*.w and trap*.l to take one immediate operand of word or long (this was just wrong in GAS). --- Changes to merged in the 1.36 version of GAS --- app.c: (1.36 version picked up) - This deals with the "# <number> <filename> <garbage>" in the state machine (the NeXT fix in s_line() was much cleaner). - Picked up the 1.36 version. The only odd difference is that ':' was ifdef'ed OUT in the 1.36 version and IN the the NeXT 1.28 version. #ifdef DONTDEF <- 1.36 #ifndef DONTDEF <- NeXT 1.28 lex [':'] |= LEX_IS_LINE_SEPERATOR; #endif I did the NeXT thing in fear of breaking something. Done with: #if defined(DONTDEF) || defined(NeXT) append.c: (1.36 version picked up) - Only Copyright comment changed as.c: (1.36 version picked up) - The machine specific command line options have been moved to routines named md_parse_option() in the machine specific files. - The handling of assembly errors has changed from using as_warn() to the new routine as_bad() which if called sets bad_error and will not produce an output file if that gets set (see the file messages.c for definitions). - Handling of signals has changed to an array of signal numbers and a routine that catches them and prints out the signal number. messages.c: (1.36 version picked up) - The addition of the routine as_bad() and the variable bad_error. If as_bad() is called then bad_error gets set and the output file does not get produced (see main() in as.c). as.h: (1.36 version picked up) - The following macros had ()'s added around their parameters: #define bzero(s,n) memset((s),0,(n)) #define bcopy(from,to,n) memcpy((to),(from),(n)) atof-generic.c: (1.36 version picked up) - Macro for alloca ifdef'ed __GNUC__ added: #ifdef __GNUC__ #define alloca __builtin_alloca #else #ifdef sparc #include <alloca.h> #endif #endif - Macros for bzero and index ifdef'ed USG added: #ifdef USG #define bzero(s,n) memset(s,0,n) #define index strchr #endif - The strings "nan", "inf" or "infinity" (in either case) are recognized first and NaN's get the sign set to 0, +infinity gets the sign set to 'P' and -infinity gets the sign set to 'N' (see flonum.h). They used to be caught at the end and the strings "Infinity", "infinity", "NaN", "nan", "SNan", or "snan" had been recognized and some note about see atof-m68k.c was there (this file was removed and atof-ieee.c was added). - A loop was added to strip leading '0' characters: while(number_of_digits_after_decimal && first_digit[number_of_digits_before_decimal + number_of_digits_after_decimal] == '0') --number_of_digits_after_decimal; After they were picked up. - Looks like the extra precision was move into two extra littlenums worth in the implementation of converting digit strings into flonum's. flonum-const.c: (1.36 version picked up) - Comment changes. flonum-copy.c: (1.36 version picked up) - Copyright comment changed. flonum-mult.c: (1.36 version picked up) - Added a check if the signs of the two numbers are one of '+' or '-' it is an error and returns zero. This happens with infinities as the sign is set to 'P' or 'M' or NaNs and the sign is set to zero ('\0' or 0). - Also some extra term in an if statement: 146c141 < if (significant || P<0) --- > if (significant) I did figure out what it was. flonum.h: (1.36 version picked up) - Comment about NaN and infinities was added: /* JF: A sign value of 0 means we have been asked to assemble NaN A sign value of 'P' means we've been asked to assemble +Inf A sign value of 'N' means we've been asked to assemble -Inf */ atof-ieee.c: (1.36 version picked up) - Replaces atof-m68k.c bignum-copy.c: (1.36 version picked up) - The addtion of the explit return type of 'int' was added to the routine bignum_copy(). - Copyright comment changed bignum.h: (1.36 version picked up) - The commented out extra digits of LOG_TO_BASE_2_OF_10 were uncommented. the comment above this was that this was done to get around a problem in GCC (I'm assuming that has been fixed). < #define LOG_TO_BASE_2_OF_10 (3.3219280948873623478703194294893901758651) --- > #define LOG_TO_BASE_2_OF_10 (3.321928 /* 0948873623478703194294893901758651 */) - Copyright comment changed. expr.c: (1.36 version picked up with Mach_O and NeXT ifdef's merged in) - Copyright comment changed and top comment removed. - A hack was changed with respect to the variable generic_bignum[]. The comment explains: /* Seems atof_machine can backscan through generic_bignum and hit whatever happens to be loaded before it in memory. And its way too complicated for me to fix right. Thus a hack. JF: Just make generic_bignum bigger, and never write into the early words, thus they'll always be zero. I hate Dean's floating-point code. Bleh. */ - This varable and comment was added but no one uses it. See flonum.h for how NaNs and infinities are handled. /* If nonzero, we've been asked to assemble nan, +inf or -inf */ int generic_floating_point_magic; - Changes to allow d$ where d is a digit has been added ifdef'ed SUN_ASM_SYNTAX. But according to the Sun assembler manual, page 10 section 2.4, the local labels are n$ where n is any integer (I wounder if it really supports negitive integers). There is also code in expr.c, symbols.c and read.c to support this. expr.h: (1.36 version picked up) - Copyright comment changed. frags.h: (1.36 version picked up) - Copyright comment changed. hash.c: (1.36 version picked up with error() calls ifdef NeXT to as_fatal) - Copyright comment changed and two /* in comments changed to / * - A change from: newsize = handle->hash_stat[STAT_SIZE] <<= 1; to handle->hash_stat[STAT_SIZE] <<= 1; newsize = handle->hash_stat[STAT_SIZE]; in hash_grow(); hash.h: (1.36 version picked up) - Copyright comment changed. - The following line removed: static char * hash_grow(); /* error text (internal) */ hex-value.c: (1.36 version picked up) - Copyright comment changed. - The following routine was added: #ifdef VMS dummy2() { } #endif input-file.c: (1.36 version picked up) - Copyright comment changed. - The commented out declaration was removed (but not the comment out code) /* static int file_handle; /* <0 === not open */ - The explict declaration of the pre prameter was added to the routine input_file_open(). - The explict declaration of the routine do_scrub_next_char() was added inside the routine input_file_give_next_buffer() in a local scope. input-file.h: (1.36 version picked up) - Copyright comment changed. input-scrub.c: (1.36 version picked up) - Copyright comment changed. - The macro AFTER_STRING was changed from: #define AFTER_STRING (" ") /* bcopy of 0 chars might choke. */ to: #define AFTER_STRING ("\0") /* bcopy of 0 chars might choke. */ - The varables used by the ifdef'ed DONTDEF code was removed (why not just also ifdef'ed?): char *p; char *out_string; int out_length; static char *save_buffer = 0; extern int preprocess; m68k-opcode.h: (1.36 version merged in) - Copyright comment changed. - The bras and bsrs were ifdef'ed NeXT to not use word displacements. - some reordering of the movem and fmovem type instructions. - all m68851 stuff pulled in (comments and opcodes), pmmu.h was removed. m68k.c: (1.36 version merged in) - Copyright comment changed - Lots of changes related to the DBCC and DBCC68000 with jumps to jumps (see GAS 1.36 version change log). - The characters 'e' and 'E' were added to FLT_CHARS[] - In the md_relax_table the long branches (BRANCH,FBRANCH & PCREL) had their forward and backward reach changed by 2 where (the 2 was removed from the expression). - Constants for the BCC68000 and DBCC branch types were added as well as entries in the md_relax_table. - The .proc pseudo op was added - The register defines for m68851 were added to m68k.c and pmmu.h was removed. - Fixed a bunch of the macros like add_fix which did NOT have ()'s around the parameters which was the source of a nasty bug NeXT tracked down. - The routine m68k_reg_parse() takes something of the form fp0: and turns the ':' into a ',' . - A fix to handling big numbers (greater than 32 bits) as a floating-point bit pattern was made to put the bits out in the correct order. The loop was changed from: for(wordp=generic_bignum;offs(opP->con1)--;wordp++) to: for(wordp=generic_bignum+offs(opP->con1)-1;offs(opP->con1)--;--wordp) - The the routine md_atof() was changed to use atof_ieee() from atof_m68k(). - Picked up the md_parse_option() routine. - The NeXT made change to allow hex immediates for floating-point (which broke decimal immediates like #1 and did not work for doubles) was removed. Also see bug #8331 in bug tracker. This change is in the routine m68_ip() (which converts a string into a 68k instruction) in the code for handling immediates which are some type of floating point number that is not a SEG_BIG. This next #if 0 #endif pair comments out these two lines: int_to_gen(nextword); gen_to_words(words,baseo,(long int)outro); and replaces it with this line: /* modified by mself to support hex immediates. */ *(int *)words = nextword; The effect is that the non SEG_BIG expression (which is just an integer, not a floating point value) is not converted to a float but just used as a bit pattern for the floating point number. This fails for doubles since some random bits left in the local array words[] get stuffed into the 64 bit double value and of course breaks the common case of #1 for decimal numbers. - The NeXT use of atof_m68k was removed in the case of getting a floating point immediate and the code to call gen_to_words() was put back. - The NeXT change of #if 0'ing out the line: (I don't know why): gen_to_words(words,2,8L);/* These numbers are magic! */ was removed the the #if removed and the code left in. obstack.c: (1.36 version picked up) - Lots of changes but diffed with the same file in the cc directory (which is based on 1.36) it looks very close to the same. Since the NeXT 2.0 compiler uses it it is picked up here on faith. obstack.h: (1.36 version picked up) - Lots of changes but diffed with the same file in the cc directory (which is based on 1.36) it looks very close to the same. Since the NeXT 2.0 compiler uses it it is picked up here on faith. output-file.c: (1.36 version picked up with NeXT and Mach_O ifdef's put in) - Copyright comment changed. - The NeXT ifdef is to unlink the file before doing a create on it. - The Mach_O ifdef is for the routine output_write(). pmmu.h: removed (1.36 has this stuff moved into m68k-opcode.h and m68k.c) read.c: (1.36 version picked up with NeXT, Mach_O and I860 ifdefs added) - Copyright comment changed. - There is a differing set of changes related to the bumping of the line counters with respect to #NO_APP and #APP. One in the 1.28 version ifdef'ed NeXT and the other in the 1.36 version. The 1.36 set of changes were picked up. - A bunch of changes to the s_set routine (not use in the NeXT compiler suite). read.h: (1.36 version picked up) - Copyright comment changed. strstr.c: (1.36 version picked up with NeXT ifdef code added) - Only Copyright comment changed - The routine strstrn() apperently was added by NeXT and is used in read.c for searching for "#NO_APP\n". struc-symbol.h: (1.36 version picked up with NeXT ifdef code added) - Only Copyright comment changed - The ifdef NeXT code is to the sy_other macro to refer to the n_sect field instead of the n_other field. subsegs.c: (1.36 version picked up) - Only Copyright comment changed subsegs.h: (1.36 version picked up) - Only Copyright comment changed symbols.c: (1.36 version picked up with Mach_O ifdef code added) - Only Copyright comment changed - Changes to allow d$ where d is a digit has been added ifdef'ed SUN_ASM_SYNTAX. But according to the Sun assembler manual, page 10 section 2.4, the local labels are n$ where n is any integer (I wounder if it really supports negitive integers). There is also code in expr.c, symbols.c and read.c to support this. - The ifdef Mach_O code is to set the n_sect field. symbols.h: (1.36 version picked up) - Only Copyright comment changed version.c: (1.36 version picked up) - The comments were removed and place in a file ChangeLog write.c: (1.36 version picked up with NeXT, M68K, Mach_O and I860 ifdefs added) write.h: (1.36 version picked up with the NeXT ifdef added) xmalloc.c: (1.36 version picked up with NeXT ifdef code added) - Only Copyright comment changed - The NeXT ifdefs are changing the call to error() to as_fatal() so the macro -Derror=as_fatal does not have to be used (since it could substitue in places where it shouldn't. xrealloc.c: (1.36 version picked up with NeXT ifdef code added) - Only Copyright comment changed - The NeXT ifdefs are changing the call to error() to as_fatal() so the macro -Derror=as_fatal does not have to be used (since it could substitue in places where it shouldn't. --- Changes to merged in the i860 version of GAS by NeXT Dimension team --- (NDTools-4) - i860.h: This contained the i860 relocation stuff. This was moved into reloc.h Also there was a bug in the GNU version of ld that relocated the RELOC_HIGHADJ wrong. The adjustment was always done out of the assembler and should have been taken out and put back everytime. This is now the case in the NeXT Mach-O link editor in i860_reloc.c . - I860 changes to read.c: big_cons(), get_known_segmented_expression() and stringer() no longer static Mike changed s_file() and s_line() to handle the cpp line directive nesting level by adding discard_rest_of_line() to it. The complier group's version just recognized the extra digits in s_file(). The compiler group's version was retained and Mike's changes were left out. The i860 has it's own align syntax and the "align" pseudo-op is ifdef'ed out for the i860 (what is this symtax?). The i860 has the "org" and "quad" pseudo-op's ifdef'ed out. The as_fatal() call in pobegin() has "... (%s)", errtxt ); added to it. An Intel "lable::" crock, which also makes the symbol global The fix_new() call in cons() has an extra RELOC_VANILLA argument added to it that is ifdef'ed I860. This also requires i860.h to be included which defines RELOC_VANILLA to be added at line 37: #if defined(I860) #include <i860.h> #endif - I860 changes to write.c: Added at line 50 (for the NO_RELOC relocation r_type) #if defined(I860) #include "i860.h" #endif The variable next_object_file_charP is not static for the i860 (ifdef'ed I860). fix_new has an extra prameter r_type (ifdef'ed I860) and it is set in to the fixP struct via: fixP->fx_r_type = r_type; also ifdef'ed I860. In write_object_file() after the relax segment calls the text alignment is forced to 32 byte alignment, the data and bss to 16 byte alignment. The code for text at line 316 is: /* Check/force alignment here! */ #if defined(I860) text_siz = (text_siz + 0x1F) & (~0x1F);/* Keep 32 byte alignment (most restrictive) */ text_last_frag->fr_address = text_siz; /* and pad the last fragment.*/ #endif for data at line 388 is: #if defined(I860) data_siz += (16 - (data_siz % 16)) % 16; /* Pad data seg to preserve alignment */ data_last_frag->fr_address = data_siz; /* to quad-word boundries */ #endif and for bss at line 361 is: #if defined(I860) local_bss_counter=(local_bss_counter+0xF)&(~0xF); /* Pad BSS to preserve alignment */ #endif The call to fix_new() in write_object_file() has an extra parameter added to it, NO_RELOC, which is ifdef'ed I860. At line 522: #if defined(I860) fix_new(lie->frag,lie->word_goes_here - lie->frag->fr_literal,2, lie->add,lie->sub,lie->addnum,0,NO_RELOC); #else fix_new(lie->frag,lie->word_goes_here - lie->frag->fr_literal,2, lie->add,lie->sub,lie->addnum,0); #endif In write_object_file() a bunch of checks were added. Just before emitting relocations at line 675: know(next_object_file_charP== (the_object_file+(N_TXTOFF(the_exec)+the_exec.a_text+the_exec.a_data))); Just before emiting the symbols at line 684: know(next_object_file_charP == (the_object_file + N_SYMOFF(the_exec)) ); Just before emiting the strings at line 710: know(next_object_file_charP == (the_object_file + N_STROFF(the_exec)) ); In fixup_segment() the switch statement for immediate displacement types for case 0 is ifdef'ed I860 with this change (at line 1209): #if defined(I860) fixP->fx_addnumber = add_number; /* * fixup_segment is expected to return a count of the number of * relocation_info structures needed for an object module. * Two specific relocation types encode only the high half * of an address, and so are followed by a second relocation_info * structure which encodes the low half. We allow for this * by bumping seg_reloc_count an extra time here. * * The extra item is generated in emit_relocations(). */ if ( fixP->fx_addsy && (fixP->fx_r_type==RELOC_HIGH || fixP->fx_r_type==RELOC_HIGHADJ)) { ++seg_reloc_count; } md_number_to_imm (place, add_number, size,fixP,this_segment_type); #else md_number_to_imm (place, add_number, size); #endif and for case 1 the comment was added (at line 1232): case 1: /* Not used in i860 version */ In emit_relocations() the following line was ifdef'ed in the other two were else'ed out (at line 1276): #if defined(I860) ri . r_type = fixP -> fx_r_type; #else /* I860 */ /* These two 'cuz of NS32K */ ri . r_bsr = fixP -> fx_bsr; ri . r_disp = fixP -> fx_im_disp; #endif /* I860 */ In emit_relocations() at the end of the loop processing the fixS structures the following lines were added to handle split relocations (at line 1425): #if defined(I860) /* Whenever we have a relocation item using the high half of an * address, we also emit a relocation item describing the low * half of the address, so the linker can reconstruct the address * being relocated in a reasonable manner. * * We set r_extern to 0, so other apps won't try to use r_symbolnum * as a symbol table indice. We OR in some bits in bits 16-23 of * r_symbolnum so it is guaranteed to be outside the range we use * for non-external types to denote what segment the relocation is in. */ if ( fixP->fx_r_type == RELOC_HIGH || fixP->fx_r_type == RELOC_HIGHADJ ) { ri.r_length = nbytes_r_length [fixP->fx_size]; ri.r_pcrel = fixP->fx_pcrel; ri.r_address = fixP -> fx_frag->fr_address + fixP->fx_where - segment_address_in_file; ri.r_extern = 0; ri.r_type = RELOC_PAIR; /* Hide the low half of the addr in r_symbolnum. More overloading...*/ ri.r_symbolnum = (fixP->fx_addnumber & 0xFFFF) | 0x7F0000; md_ri_to_chars((char *) &ri, ri); append(&next_object_file_charP, (char *)&ri, (unsigned long)sizeof(ri)); } #endif --- Changes made to do the merges of 1.36 and i860 versions --- - Removed the cpp macro "error" which was set on the compiler line to -Derror=as_fatal and changed the 4 uses in hash.c, xmalloc.c and xrealloc.c to just use as_fatal. - Added the cpp macro M68K for 68k specific ifdef that are needed (like in Mach-O.c). This is instead of the "default case" without a target processor macro meaning that it is the 68k case. This is set in the Makefile as the target processor that the assembler is for in the make macro COPTS. - Changed the only use of the cpp macro CROSS in output-file.c to use NeXT to get rid of this macro. The line of code that is ifdef'ed is is the unlink of "name" in output_file_create(). - Removed a.out.h and letting the one in ../include get used which is a merge of the original and includes NeXT's files (nlist.h and reloc.h). - Removed the file atom.c since Mach-O.c replaces it (also removed all the code in write.c that used it). - Removed all machine specific files except for the target processors that NeXT uses. The remaining code that used this stuff has been ifdef'ed where needed to preserved the code in the files we use. - Removed the files gdb.c, gdb-file.c, gdb-symbols.c, gdb-blocks.c and gdb-lines.c and ifdef'ed DONTDEF the code in as.c, read.c and write.c that used this stuff since the GNU 1.36 version of GAS did the same. - Removed the files m-68k.h, m-sun3.h, m-hpux and m-generic and ifndef'ed the include of m-68k.h out of m-68k.c. - Removed the files atof-m68k.c atof-m68k-assist.s since they are no longer used (see the change below for the -27 compiler release). And replaced the the file atof-m68k.c with the 1.36 atof-ieee.c . The 2.0 Release (the -32 compiler release) Changes for the Warp ?? release (the -27 compiler release): - Fixed m68_ip() to handle hex immediate operands to floating point insn's. Now fadds #0xffffffff,fp0 works correctly. The fix only works for .s, not for .d or .x. This orignally worked, but was broken by NeXT's mods to atof-m68k.c. (mself) - Added new 68040 floating-point instructions to m68k-opcode.h (mself) - Changed the name of the the section generated by the .objc_selector_refs directive from __selector_refs to __message_refs and set the flags field of this section to S_LITERAL_POINTERS. This change requires a link editor that knows how to handle a S_LITERAL_POINTERS section. - Changed m68k.c to use the reguar atof (actually strtod) instead of using atof-m68k.c and atof-m68k-assist.s, since these instructions will be emulated on th '040. (mself) Changes for the Warp ?? release (the -26 compiler release): - Added the file Mach-O.c and the ablity to have a subset of a fixed number of sections. All changes ifdef'ed MACH_O. This removes atom.c (ifdef'ed out). New sections include const, literal4, literal8, 11 new objc sections, etc. Basicly a lot of changes. Changes for the Warp 3 (OS update) release (the -25 compiler release): - Added scattered relocation entries to the assembler in emit_relocation() in write.c (see extensive comments in there and in <reloc.h>). - Changed fixup_segment() in write.c and md_estimate_size_before_relax() in m68k.c to make branches to lables that persist on output to be long in length and have a relocation entry (to make scattered loading in the link editor work). This was done by using the value of 3 in fx_pcrel (see the comment in write.h) for force this to happen. Changes for the Warp ?? release (the -24 compiler release): - Fixed the bug that would not assemble the following instruction: L2:movl #1,@(_down:l,d7:l:4) The fix was a bug in the macro use in m68k.c for add_fix() which the macro did not put ()'s around it's arguments (bugs 5207 and 5270). - Fixed the bug with cpp processed assembler files (bug 4280). The new syntax of the cpp output now includes an optional number after the file name, for example: `# 1 "x.c" 2' the 2 is the new number. This was done by changing the routine s_file() in read.c which recognizes the .file directive (that the assembler's preprecessor inserts) to optionally recognize this new number. - Changed the section alignment of the text section to 2 byte alignment so that scattered loading will work (the branch table of the shlibs will not move). Changes for the 2.0 impulse X.X release (the -23 compiler release): - Now is linked with libsys. Changes for the 2.0 impulse X.X release (the -22 compiler release): - Allow symbol names to appear in ""'s . This is so that the symbol names for methods can be "+[Class(category) method:name:]" and tools will not have to go through the objective-C section to get these names. Changes how get_symbol_end() works and how the callers of it use it. Changes for the 2.0 impulse X.X release (the -19 compiler release): - as is no longer installed as as-<version_number> to match the rest of the project. - Updated atom.c to the changes to CPU_TYPE and CPU_SUBTYPE with the changes to <sys/machine.h> Changes for the 0.91 release (the -10 compiler release): * s.stone fixed a bug in `#APP', `#NO_APP' that affected read.c & strstr.c. + Fixed a bug in converting to Mach-O object files with the new sections for the objective-C runtime. The bug was if a local relocation item refered to a symbol plus an offset the incorrect section number could be assigned if the value of the symbol plus offset was in a different section than the value of the symbol. This is an un fixable bug in atom(1) but fixed in here by moving the assignment of the section number into the r_symbolnum field into the assembler and using just the symbol's value (not plus the offset) to pick the section number. The fix is in write.c in emit_relocation() (plus a call to a new function get_objc_section_bounds() before calling emit_relocation). + Fixed a bug where a file had no symbols and the result was a Mach-O file. What would happen was a 4 (the size of the string table) was written at offset 0 in the output file (overwriting the magic number). Also did some major clean up of atom.c and removed all the garbage that did not apply (about half of what was there). + Added the .reference pseudo op to read.c. This was added for the new objective-C runtime to use so that archive semantaic could be maintained but no globals (that 'C' could use) are created. + Fixed the exponent overflow handling in atof-m68k.c to not print a warning (ifdef NeXT) and to get the right answer (a bzero of the 'words' as added, and corrected the reversed sign for infinities). New notes go at the TOP of this file.