TODO for coalesced symbols:
- Should external relocation entries for defined coalesced symbols only be
created with -dynamic and not -static?
TODO:
- Add MacOS line termination \r .
Known bugs:
- The assembly line:
bl ""foo""
causes the symbol name "" to be used.
- cmpwi seems to be the same as cmpi for PowerPC.
- Can't optimize because of compiler bug #50416 prevents line 235 in symbols.c
from working (currently has #pragma CC_OPT_OFF and #pragma CC_OPT_RESUME
around that routine).
- The m88k instruction "tb0 0,r0,undef" trashes the instruction because of
the undefined.
- 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.
Bugs to be fixed:
- The m68k "jmp @(_foo)" is not legal and needs to be flaged.
- The PowerPC extended branch mnemonic like beqlrl need to take a CRFONLY or
a number shifted over by 2 (like the fcmpu instruction).
Changes for the 5.31 release (the cctools-828 release):
- Back out the change below in relax_immediate() in arm.c that was done as
an attempt to fix Radar bug #11215603 as it is causing problems. Radar bug
#11403124.
Changes for the 5.31 release (the cctools-827 release):
- Fixed relax_immediate() in arm.c to correctly handle a negative immediate
to see that it fits in the instruction. Radar bug #11215603.
Changes for the 5.31 release (the cctools-825 release):
- Added the -Q flag and tweaked handling the running of clang as the assembler
to the driver. Cleaned up a bit of the code in there. Radar bug #11139995.
- Added a missing extern to subsections_via_symbols in as.h, and to
dwarf2_loc_mark_labels in dwarf2dbg.h and to now_subseg in arm.c . Radar bug
#11136207.
Changes for the 5.31 release (the cctools-824 release):
- Added a check in do_neon_ld_st_interleave() in arm.c to check that the size
of 64 is only used with VLD1/VST1. Radar bug #11024563.
Changes for the 5.31 release (the cctools-823 release):
- Added the -q option to the assembler driver to run clang(1) for x86_64.
Basis for Radar bug #9801003.
Changes for the 5.30 release (the cctools-820 release):
- Removed creating the apple_version.c file from the Makefile and now use the
new one created in libstuff. Radar bug #10596469.
- Another change to out_debug_info() in dwarf2dbg.c for Radar bug #10466662.
For the AT_high_pc, needed to the next symbol with different address or
the end of the section.
Changes for the 5.30 release (the cctools-819 release):
- Changed out_debug_info() in dwarf2dbg.c to check if the previous symbol in
the list of subprogram's info has the same address and to not emit it so
there aren't subprograms with zero lengths. Radar bug #10466662.
- Added AT_prototyped with a value of 0 to the TAG_subprogram when producing
dwarf for assembly source and added the TAG_unspecified_parameters after the
TAG_subprogram. The changes are in out_debug_abbrev() and out_debug_info() in
dwarf2dbg.c . Radar bug #10537250.
Changes for the 5.30 release (the cctools-816 release):
- Changed the arm assembler so that for the thumb multiply instructions it
checks that either source register is the same as the destination register
and uses the 16-bit T1 encoding. The fix was in arm.c in do_t_mul to change
this:
|| inst.operands[0].reg != inst.operands[2].reg))
to this:
|| (inst.operands[0].reg != inst.operands[2].reg &&
inst.operands[0].reg != inst.operands[1].reg)))
Radar bug #10433396.
- Fixed a problem with dwarf generated for assembly source with -g for arm
assembly code. Symbols marked with .thumb_func would have the low bit set
for their AT_low_pc value. The fix was to not use the symbol directly for the
expression but to create a temporary symbol without the N_ARM_THUMB_DEF bit
set in make_subprogram_for_symbol() in symbols.c . Radar bug #10409052.
Changes for the 5.30 release (the cctools-815 release):
- Changed the i386 popcnt instructions to not have NoSuf on their entries in
i386-opcode.h. Radar bug #10385539.
- Added a check in dwarf2_directive_file() in dwarf2dbg.c to see if the -g flag
is being used to generate dwarf debug info for assembly code and print
an error because the input has .file dwarf directives. Radar bug #9976587.
Changes for the 5.30 release (the cctools-814 release):
- Fixed the .incbin directive so it doesn't require the file to be writable.
Radar bug #10288224.
Changes for the 5.30 release (the cctools-812 release):
- Fix the encoding of the arm sev instruction. Radar bug #10133308.
Changes for the 5.30 release (the cctools-809 release):
- Changed the assembler driver to no longer look in ../libexec/gcc/darwin paths
and to no longer install target assemblers in those paths. Radar bug
#9908131.
Changes for the 5.30 release (the cctools-807 release):
- Changed cons() in read.c to call dwarf2_emit_insn() to allow instructions
that are being assembled with things like .long to trigger and outstanding
.loc line number creation. Radar bug #9811389.
Changes for the 5.29 release (the cctools-806.2 release):
- Changed so that the target assembler get installed both in ../libexec/as and
in ../libexec/gcc/darwin and changed the driver to look in the first path then
fall back to the second path. Radar bug #9661352.
Changes for the 5.29 release (the cctools-800.7 release):
- Added -gdwarf-2 to be the same as -g and --gwarf2 in main() in as.c .
Radar bug #9249853.
Changes for the 5.29 release (the cctools-800.4 release):
- Added missing support for -arch armv7[fk] for md_begin() in arm.c . Radar
bug #9055988.
Changes for the 5.29 release (the cctools-800.3 release):
- Added support for -arch armv7[fk] for main() in as.c and in s_align() in
read.c . Radar bug #8987684.
Changes for the 5.29 release (the cctools-800 release):
- Fixed the arm VCVT (between half-precision and single-precision) instructions
encodings. Had the op bit inverted in the table. It should have been:
VCVT<c><q>.F32.F16 <Qd>, <Dm> Encoded as op = 1 (aka fcvtshp)
VCVT<c><q>.F16.F32 <Dd>, <Qm> Encoded as op = 0 (aka fcvthps)
Radar bug #8951023.
- Added the .incbin directive. Radar bug #8945167.
- Changed the Makefile to install the ppc and ppc64 assembler in LOCLIBDIR.
Radar bug #8828970.
Changes for the 5.28 release (the cctools-799 release):
- Found another path that -g was not checking to produce dwarf line table
entries only for the text section. That was in dwarf2_emit_insn() in
dwarf2dbg.c. Radar bug #8792738.
- Fixed the mishanding of the thumb2 tst instruction. Changed do_t_mvn_tst()
in arm.c. Normally, the versions of the instructions that set condition flags
aren't allowed inside IT blocks, so it's checking that. TST<c>, however, is
allowed and isn't accounted for. Radar bug #8783763.
- Changed the arm assembler for ADD/SUB (immediate, Thumb) to correctly follow
the manual's note about these instructions:
encoding T1 is to be preferred if Rd is specified, and T2 is preferred if
it is not.
To do this do_t_add_sub() sets the unused 0x0100 in inst.instruction when
using BFD_RELOC_ARM_THUMB_ADD to communicate to md_apply_fix() that two
operands were present in assembly code. Then md_apply_fix() uses this bit
and checks the size of the immediate to see if the 3-bit immediate form can
be uses as it should be preferred in that case. Radar bug #8756030.
Changes for the 5.28 release (the cctools-798 release):
- Changed to error messages in arm.c which were:
"invalid constant (%x) after fixup"
to have 0x in front of the hex value. Radar bug #8772178.
- Changed do_t_mul() in arm.c . The problem is that the assembler didn't take
into account that the "mul" mnemonic (w/o the 's' suffix) is valid for the
16-bit wide encoding when in an IT block. Radar bug #8755919.
- Added support for make aarm_test .
- Added a arm-check.c that currently only does arm ALU instructions can set
into account that the "mul" mnemonic (w/o the 's' suffix) is valid for the
16-bit wide encoding when in an IT block. Radar bug #8755919.
- Added support for make aarm_test .
- Added a arm-check.c that currently only does arm ALU instructions can set
the condition codes.
- Added #ifndef INSNS_TABLE_ONLY to arm.c so that can be #defined and included
in arm-check.c .
- Made the needed changes in the Makefile to add support for aarm_test .
Changes for the 5.28 release (the cctools-797 release):
- Changed the code that was adding the NULL DIE terminating the Compile Unit
DIE's when producing dwarf debuging info for assembly files to not be
conditional on if dwarf2_subprograms_info != NULL. The change is at the end
of out_debug_info() in dwarf2dbg.c . Radar bug #8441990 and #8443177.
Changes for the 5.28 release (the cctools-796 release):
- Changed the x86-64 assembler to allow immediate expressions of the form:
movq $(xxx-L$y),%rax
by changing output_imm() to look for 4 byte signed 32-bit expression and use
the same trick in output_disp() to call fix_new() so it will use two
relocation entries pointing at the same 32-bits. Radar bug #7081262.
- Added the OpenSourceURL key to the cctools.plist file as:
<key>OpenSourceURL</key>
<string>http://ftp.gnu.org/gnu/binutils/binutils-1.38.1.tar.gz</string>
Radar bug #8569655.
Changes for the 5.28 release (the cctools-795 release):
- Added checks or overlowing the 24-bits of the r_address relocation entry in
fix_to_relocation_entries() in write_object.c . Radar bug #8348673.
Changes for the 5.28 release (the cctools-794 release):
- Added the NULL DIE terminating the Compile Unit DIE's when producing dwarf
debuging info for assembly files. The change is at the end of
out_debug_info() in dwarf2dbg.c . Radar bug #8453745.
Changes for the 5.28 release (the cctools-793 release):
- Fixed more problems with support for generating dwarf debugging information
for assembly source, with symbols in other sections besides .text .
- The routine make_subprogram_for_symbol() in symbols.c was not ignoring
symbols not in the .text section. Also the calls to
- Generation of the .loc directive done in md_assemble() in i386.c and in
arm.c by calling dwarf2_loc() also needed to check that the section was
the .text section.
Radar bug #8425965.
- Fixed problems with support for generating dwarf debugging information for
assembly source. The routine make_subprogram_for_symbol() in symbols.c was
not updating last_dwarf2_subprogram_info when dwarf2_subprograms_info != NULL
and it was not ignoring symbols that stared with 'L'. Radar bug #8387908.
Changes for the 5.28 release (the cctools-790 release):
- Added the arm Media and VFP Feature Registers 1 and 0 to arm.c as mvfr0,
MVFR0, mvfr1 and MVFR1. Radar bug #8246341.
Changes for the 5.28 release (the cctools-788 release):
- Remove the use of -mdynamic-no-pic from the Makefile. Radar bug #8207834.
Changes for the 5.28 release (the cctools-787 release):
- Added support for the .symbol_resolver directive. Added s_symbol_resolver()
in read.c and an entry for it in pseudo_table[]. Also added N_SYMBOL_RESOLVER
in the conditional test in colon() in symbols.c to allow it to be set first
then the symbol defined. Radar bug #7901954.
Changes for the 5.28 release (the cctools-786 release):
- Fixed the incorrect trigger for outputing dwarf info in dwarf2_finish() in
dwarf2dbg.c . The change was to make sure that all_segs was not NULL and the
either all_segs->next was not NULL or debug_type == DEBUG_DWARF2.
Radar bug #8154276.
Changes for the 5.28 release (the cctools-785 release):
- Fixed a bug with support for thread local variables in fixup_section() in
layout.c that did not include the address of the fragment when calculating
the "pc" value. Radar bug #8113809.
- Added support for generating dwarf debugging information for assembly source
files when the --gdwarf2 or -g flag is specified. Radar bug #7580966.
- main() in as.c was changed to make -g the same as --gdwarf2.
- The declaration of apple_version[] was moved to as.h out of as.c.
- The variable version_string[] in as.c was made global and the declaration
was added to as.h.
- The macro APPLE_INC_VERSION was added to as.h for print the version string
in main() in as.c
- The global varaible apple_flags was added to as.[ch] and set in main() when
the environment variable RC_DEBUG_OPTIONS is set.
- The global variable dwarf_file_number was added to dwarf2dbg.[ch] .
- Generation of the .file directive is done in read_a_source_file() in read.c
by calling dwarf2_file() which was added in dwarf2dbg.c . The call uses
global variable dwarf_file_number and increments it.
- Generation of the .loc directive is done in md_assemble() in i386.c and in
arm.c by calling dwarf2_loc() which was added in dwarf2dbg.c .
- The struct dwarf2_subprogram_info and global variable
dwarf2_subprograms_info was added to dwarf2dbg.h to collect info and use it
for the subprogram entires.
- Collecting information for the symbols to make dwarf2 subprogram entires
is done by calling the new routine, make_subprogram_for_symbol(), added
to symbols.c called from colon() in symbols.c .
- out_debug_abbrev() in dwarf2dbg.c was change to add the DW_TAG_subprogram
DIE as #2 and changed to DW_CHILDREN_yes for the DW_TAG_compile_unit DIE.
Also the AT_APPLE_flags is added when apple_flags is set.
- In out_debug_info() in dwarf2dbg.c it was change to output the subprograms
and the DW_AT_producer was changed to use the version information so it
matches what as(1) -v would print. Also the AT_APPLE_flags added to the
output when apple_flags is set.
- The trigger for calling the out_debug_*() routines at the end of
dwarf2_finish() in dwarf2dbg.c was change to include a test of
debug_type == DEBUG_DWARF2 (when --gdwarf2 is specified).
Changes for the 5.28 release (the cctools-784 release):
- Changed the arm assembler to not used the thumb2 opcode when targetting armv6
and unified syntax. The fix was in do_t_nop() in arm.c to add a check in
additiona to unified syntax to make sure the cpu_variant core specified also
supports thumb2. Radar bug #7876935.
- Add a check for too many sections being created in sections.c . Radar bug
#7920267.
Changes for the 5.28 release (the cctools-783 release):
- Added control register names as %ecr0 and %rcr0 in i386-opcode.h that is the
names llvm-mc uses. And added $(LLVM_MC) to the otool command lines in the
Makefile.
- Added the -llvm-mc option to the driver to run llvm-mc. Radar bug #7837868.
- Added the support for thread local variables.
- Changed the _i386_insn in i386.c to include a reloc[] array like for ARCH64
so it can hold a GENERIC_RELOC_TLV.
- Changed md_assemble() in i386.c to zero out reloc[] without the #ifdef
ARCH64.
- Changed the call for fix_new() in output_disp() in i386.c to pass i.reloc[n]
as the last argument in the #if !ARCH64 case.
- Changed the #if expression around lex_got() and removed the !defined(ARCH64)
part of the condition.
- Inside lex_got() in i386 changed gotrel[] initialization on the #ifdef
NeXT_MOD side. Wrapped that in #ifdef ARCH64 and added the TLVP to the table
and added an #else side for 32-bit with the entry for the TLVP.
- In layout_addresses() in layout.c made changes to test for
S_THREAD_LOCAL_ZEROFILL when testing for S_ZEROFILL.
- In fixup_section() in layout.c added code that is
#if defined(I386) && !defined(ARCH64)
- To test GENERIC_RELOC_TLV relocation type when sub_symbolP is not NULL
and adjust the value to have to be the difference of the end of the
relocated item and the symbol. Then clear the fx_subsy field and set the
fx_pcrel to get the proper TVL relocation entry. This is done in two
places, one where both symbols are in a section and one when the
add symbol is undefined.
- Added a check of fixP->fx_r_type != GENERIC_RELOC_TLV where it would
add in the symbol value so it does not do that (it will be an external
reloc in this case even if the symbol is defined).
- In read.c added the new section type names thread_local_regular,
thread_local_variables and thread_local_init_function_pointers to
type_names[].
- In read.c added the new built in section directives tdata, tlv and
thread_init_func to builtin_sections[].
- In read.c changed s_zerofill() and passed the section type as the value
parameter and added the tbss directive to use s_zerofill. And changed the
zerofill directive to pass S_ZEROFILL as the value. In s_zerofill() it
checks for the value passes as S_THREAD_LOCAL_ZEROFILL and creates the
(__DATA, __thread_bss) section with that type.
- In fix_to_relocation_entries() in write_object.c added a test of
((symbolP->sy_type & N_TYPE) == N_SECT &&
fixP->fx_r_type == GENERIC_RELOC_TLV)
that is
#if defined(I386) && !defined(ARCH64)
to where it sets the r_extern bit and the r_symbolnum to the symbol index
as @TLVP that are defined get external relocation entries.
- In sections.c changed section_new() and in write_object.c in write_object()
to also test for S_THREAD_LOCAL_ZEROFILL when it is testing for S_ZEROFILL.
Radar bug #7796734.
Changes for the 5.28 release (the cctools-782 release):
- Added the .weak_def_can_be_hidden to read.c and changed colon() in symbols.c
to also preserve N_WEAK_DEF bit. Radar bug #5021217.
- Fixed the arm assembler to correctly detect errors with floating point to
integer conversion instruction that use the rounding mode (vcvtr.*) when they
are not valid and generate an error. The fix was in do_vfp_nsyn_cvtr() in
arm.c to cause an error if one has not already be caught by checking that
the flavour index into the static arrary enc[] does not have a null value.
Radar bug #7719236.
- Added the Intel aes instructions. Radar bug #7709572.
Changes for the 5.28 release (the cctools-780 release):
- Changed the Makefile to use -Os for OFLAG in building the driver.
- Fixed a bug in parse_neon_mov() in arm.c that caused assembling:
vmov.f32 s2, s1
to end up with an undefined symbol named "s1" in the output file. This was
caused by the call to parse_big_immediate(). Which was incorrectly called
before trying arm_typed_reg_parse(). The call to parse_big_immediate()
failed but as a side effect created the undefined symbol and then the later
call to arm_typed_reg_parse() did not fail. Reversing the two removes the
problem and then in the case when parse_big_immediate() fails it causes an
error. Radar bug #7567713.
- Fixed arm.c that incorrectly had an ARM_VARIANT set for the Thumb only
instructions cbnz and cbz. Also fixed do_neon_mov() in arm.c when it was
called for something like "vshr q0,d0,#0" to not abort() in the switch(rs)
case for default if a inst.error and just return and let the error get
reported. Radar bug #7559323.
Changes for the 5.28 release (the cctools-779 release):
- Changed the test in encode_thumb2_ldmstm() for the warning of the value
stored is UNPREDICTABLE to check that writeback is set before the rest of the
checks such as the base is in the list and is also not the lowest register
in the mask. Radar bug #7528057.
Changes for the 5.28 release (the cctools-777 release):
- Changed round() to rnd() and stuff/round.h to stuff/rnd.h in all files and
removed -fno-builtin-round from the Makefile. Radar bug #7475139.
- Changed the Makefile to use -Os for OFLAG. Radar bug #7475093.
Changes for the 5.28 release (the cctools-775 release):
- Added support for the new arm ARM_RELOC_HALF and ARM_RELOC_HALF_SECTDIFF
relocation entries used with the movw and movt instruction immediates.
- Added a new file arm_reloc.h to have internal #defines for
ARM_{,THUMB_}RELOC_{LO16,HI16} with notes about the encodings for the
reloc entries in <mach-o/arm/reloc.h> .
- In arm.c assigned the enum BFD_RELOC_{ARM,THUMB}_MOV{W,T} values
ARM_{,THUMB_}RELOC_{LO16,HI16} from arm_reloc.h .
- Fixed the code in md_apply_fix() for the BFD_RELOC_{ARM,THUMB}_MOV{W,T}
case values so the *_MOVT values would have high 16 bits of the value used.
- Fixed the code in write_object.c to correctly deal with the ARM relocs.
- In nrelocs_for_fix() added the #ifdef ARM code for the paired relocs.
- In fix_to_relocation_entries() added the #ifdef ARM code for the handling
of the LO16,HI16 and SECTDIFF ARM relocs.
Radar bug #7421496.
Changes for the 5.28 release (the cctools-774 release):
- Added the invvpid and invept instructions to i386-opcode.h .
Radar bug #7383743.
- Fixed x86_64 instructions that have a 0xf2, 0xf3, etc. "mandatory" prefixes to
follow optional prefixes. This is done by adding a entry in the i.prefix[]
array, increasing MAX_PREFIXES in i386.h by one and adding MAN_PREFIX just
before REX_PREFIX. Then add_prefix() in i386.c had the parameter mandatory
added to it which was passed MAN_PREFIX when called with a mandatory prefix or
0 otherwise. Then add_prefix() had code added to it to place the passed
prefix into i.prefix[mandatory] when mandatory was not zero and return 1
to indicate a prefix was added but was not a REP prefix. Radar bug #7403143.
- Fixed the i386 bound instruction to take WordMem, WordReg not WordReg,
WordMem . Radar bug #7394709.
- Update the i386-check.c program to work for both i386 and x86_64 and added in
the Makefile an ax86_64_test target.
Changes for the 5.27 release (the cctools-772 release):
- Fixed the arm's assembler's vmls.f32 and vmls.f64 encodings, the changes were
in do_vfp_nsyn_mla_mls() in arm.c to use "fnmacs" instead of "fmscs" and
"fnmacd" instead of "fmscd". Fixed the arm's assembler's vnmla.f32, vnmls.f32
and vnmla.f64, vnmls.f64 encoding in their respective table entries.
Radar bug #7331301.
- Changed the arm assembler to make the vcvt.* forms match the manual the same
as what was the vcvtz.* forms. And removed the vcvtz.* forms. Then added
the vcvtr.* forms as in the manual as what was the old vcvt.* forms. Changes
were in do_vfp_nsyn_cvt() in arm.c in the last enc[] array entries and a
change to the "if (rs == NS_FF" to add "&& (flavour == 18 || flavour == 19))".
Then changed do_vfp_nsyn_cvtz() to do_vfp_nsyn_cvtr() and changed the entries
in the enc[] array in there. And also changed the vcvt entry to use
vfp_nsyn_cvtr . Radar bug #7329319 & #7329326.
- Changed the arm assembler to allow "vmrs APSR_nzcv, FPSCR" by adding the
routine do_vmrs() in arm.c and changing the table entry for "vmrs" to allow
ASPR_RR for the first argument and use the new routine. Radar bug #7327031.
- Changed i386_displacement() in i386.c and turned on the GCC_ASM_O_HACK to
allow displacements like "4+(%esp)". Radar bug #5828463.
Changes for the 5.27 release (the cctools-771 release):
- Changed fixup_section() in layout.c for the first #if ARM case and removed the
check for the add symbol not being weak to set the low bit of the value.
Radar bug #7262421.
- Changed the test in encode_thumb2_ldmstm() for the warning of the value
stored is UNPREDICTABLE to check that the base is also not the lowest register
in the mask. Radar bug #7268941.
Changes for the 5.27 release (the cctools-769 release):
- Fixed a bug in s_zerofill() in read.c that did not allow zerofill sizes
greater than 32-bits. In s_zerofill() in read.c changed the local variable
"size" to a uint64_t type. Changed the type of the field "fr_address"
in the frag struct in frags.h to a uint64_t type. And fixed types for
relax_addressT in relax.h which were reversed in the #ifdef ARCH64.
Radar bug #7245879.
Changes for the 5.27 release (the cctools-767 release):
- Added an error check for the call to realpath(3) in driver.c in main().
Radar bug #7201696.
Changes for the 5.27 release (the cctools-766 release):
- Picked up the change that never made it in the sources from cctools-685 but
made in in the notes below. Changed relax_section() in layout.c in two plaves
#ifdef'ed ARM. In the first look it now sets frag's relax_marker field to
zero. In the "Do relax" loop it flips the relax_marker field with an "^= 1".
This is to fix a problem in relaxed_symbol_addr() so it knows to add the
stretch amount to the addr it is returning. Radar bug #7141908.
Changes for the 5.27 release (the cctools-765 release):
- Fixed a problem with the mask used for checking overflow of 2 byte values
in fixup_section() in layout.c that was 0xffff8000 which should have been
0xffff0000. Also fixed the resulting error message so it would print the
source line number. Radar bug #7114412.
- Added the assembly for the orn (immediate) instructions to arm.c in
t_do_orn(). Radar bug #7096522.
Changes for the 5.27 release (the cctools-763 release):
- Added the assembly for the orn instructions to arm.c . Radar bug #7096522.
- Added the assembly for the rrx instructions to arm.c . Radar bug #7098328.
Changes for the 5.27 release (the cctools-762 release):
- Changed the Makefile to set CC when when SDKROOT is set. Radar bug #7078805.
- Changed output_relax_insn() in arm.c to allow symbols that are just an
absolute symbol. As this occurs when using a symbol from .set for an
immediate value. Radar bug #7069304.
Changes for the 5.27 release (the cctools-760 release):
- Changed the sources so the armv7 stuff is a no longer #ifndef __OPEN_SOURCE__
Radar bug #7016121.
- Changed the Makefile and removed the #ifndef __OPEN_SOURCE__ for the arm
build.
- Removed the #ifndef __OPEN_SOURCE__ from as.c that was bracketing armv7
the reference.
Changes for the 5.27 release (the cctools-759 release):
- Added the assembly for the pldw, vcvt.f32.f16, vcvt.f16.f32, vcvtt and
vcvtb instructions to arm.c . Radar bug #6959129.
- Added the memory barrior options ish, ishst, osh, oshst, nsh and nshst
to arm.c . Radar bug #6982242.
- Changed the Makefile to add -isysroot $(SDKROOT) to compile and link lines
when SDKROOT is set. Radar bug #6965029.
Changes for the 5.27 release (the cctools-758 release):
- Backed out the change in Radar bug #6511482 in the arm assembler to put out a
relocation entry for 32-bit thumb branches using the ARM_THUMB_32BIT_BRANCH
relocation type.
- Changed arm.c to no longer assign the enum type
BFD_RELOC_THUMB_PCREL_BRANCH20 to the ARM_THUMB_32BIT_BRANCH relocation
type's value.
- Changed nrelocs_for_fix() and fix_to_relocation_entries() in write_object.c
and removed the #ifdef ARM and tests for the ARM_THUMB_32BIT_BRANCH
relocation type when archflag_cpusubtype == CPU_SUBTYPE_ARM_V7. That code
use to count 1 for the fixes in nrelocs_for_fix() and to put out a
relocation entry in fix_to_relocation_entries() for the type
ARM_THUMB_32BIT_BRANCH. The relocation entry put out was a local relocation
entry that was pcrel, length of long and an symbolnum of R_ABS (0).
Also in fix_to_relocation_entries() it was setting fr_type to
ARM_THUMB_32BIT_BRANCH so that BFD_RELOC_THUMB_PCREL_BRANCH20 versions with
no symbol were geting set correctly.
Radar bug #6945428.
Changes for the 5.27 release (the cctools-757 release):
- Fixed the arm assembler to correctly place Rm in the rbit instruction in
both halfs of the 16 bit opcode. The fix was is do_t_rbit() in arm.c .
Radar bug #6915130.
Changes for the 5.27 release (the cctools-754 release):
- Changed the arm assembler to allow -arch armv7 to accept ARMv7M special
registers. The change was in md_begin() in arm.c for the case statement
CPU_SUBTYPE_ARM_V7. Radar bug #6782055.
Changes for the 5.27 release (the cctools-752 release):
- Ported to build with clang.
- Changed Makefile to work around bugs 6761248, 6756531 and 6756295.
- Added a (void) cast in the SKIP_WHITESPACE macro in read.h .
- Added (void) casts in the obstack_finish macro in obstack.h .
- Added (void) cast before obstack_finish() in:
frag_new() in frags.c
two places add_last_frags_to_sections() and two places in
layout_addresses() in layout.c
section_new() in sections.c
- Added (void) case before obstack_1grow() in:
demand_copy_string(), s_macro(), s_endmacro(), add_to_macro_definition(),
expand_macro() and s_load() in read.c .
- Changed decode_local_label_name() in symbols.c so to use a literal string
instead of a pointer to a string to get rid of a "potentially insecure"
warning.
- Changed ppc.c and the use and setting of error_param_message so it is just
a string and not a format string to get rid of a "potentially insecure"
warning. Workaround for 6765807.
- Fixed warnings about extra tokens at end of #endif directive in ofile.c
by putting them in a comment.
- Changed input_file_give_next_buffer() in input-file.[ch] to also return the
size of the buffer it will give. Then changed input_scrub_next_buffer() in
input-scrub.c to use this to know when it didn't need to reallocate the
buffer to fix it from needlessly running out of memory. Radar bug #6720994.
- Changed md_number_to_imm() in arm.c to have a case for 8 so .quad with a non-
relocatable expression would work. Radar bug #6731807.
Changes for the 5.27 release (the cctools-751 release):
- Fixed a problem with the arm assembler not accepting the sdiv instruction with
armv7. Radar bug #6697528.
- Fixed an initialization problem in fix_new_internal() in fixes.c that was not
initializing the fx_sectdiff_divide_by_two field. Radar bug #5930362.
Changes for the 5.26 release (the cctools-747 release):
- Changed the assembler to put out a line table if no .loc directives are
used but a .file directive to be used. This was causing a bad dwarf for data
only files that used .file and had no .loc directives. The fix is in
dwarf2_finish() in dwarf2dbg.c to add to the test to see we have seen any
line number segs or .file was used before calling out_debug_line(). Radar
bug #6593344.
Changes for the 5.26 release (the cctools-746 release):
- Added the support for the .sleb128 and .uleb128 directives. Radar bug
#4545225.
- Added rs_leb128 to the relax_stateT enum in relax.h .
- Added the fields has_rs_leb128s and layout_pass to the frchain struct in
sections.h .
- Added entry for sleb128 and uleb128 to pseudo_table[] in read.c .
- Added s_leb128(), sizeof_uleb128, emit_leb128_expr() to read.c .
Where emit_leb128_expr() is heavily hacked from the original due to
differences in expressions. In emit_leb128_expr() after the call to
frag_var() for a rs_leb128 frchain_now->has_rs_leb128s is set to TRUE to
mark the section as containing an rs_leb128 frag.
- Added a case label for rs_leb128 in the first loop in relax_section()
in layout.c with a line of code:
address += fragP->fr_offset = 1;
- Also added a case label for rs_leb128 in the second loop in relax_section()
in layout.c to get the value of the symbol's expression and call
sizeof_leb128() to get the current size, adjust the growth and save the new
size in the fr_offset field of the frag.
- Added code in layout_addresses() in layout.c to make three layout passes
over the frags. Doing the debug sections last, ones with containing
rs_leb128 frags before that and all other sections first.
- Added a loop in the layout pass if relax_section() changes things. To do
that the new field last_fr_address to the frag struct in frags.h, then
changed relax_section() in layout.c to return non-zero if the things change
by setting and testing the last_fr_address feild in each of the frags.
- Added a case label for rs_leb128 in layout_addresses() in layout.c and
code to get the value of the symbol's expression, then call output_leb128()
to place that value in the frag and finally convert the frag to an rs_fill
of the size of the bytes put in the frag.
Changes for the 5.26 release (the cctools-745 release):
- Changed the assembler to not put out a line table if no .loc directives are
used. This was causing a second line table in the output with older compilers
causing bad dSYM files to be created. The fix is in dwarf2_finish() in
dwarf2dbg.c to test to see we have seen any line number segs before calling
out_debug_line(). Radar bug #6589056.
Changes for the 5.26 release (the cctools-744 release):
- Fixed a problem with the code to support .file/.loc which was incorrectly
putting out a Compile Unit into the debug_info section when it shouldn't.
The code in dwarf2_finish() in dwarf2dbg.c was not ported correctly for
checking if the debug_info section was empty. It was testing the section
size for zero but that is not valid at that point in the code and needed
to look for fragments in the section. So a version of the routine
seg_not_empty_p() was ported and added to sections.[ch] . Radar bug #6573528.
Changes for the 5.26 release (the cctools-741 release):
- Fixed the encoding of the armv7 thumb2 instructions for where it places Rd in
qadd, qdadd, qsub and qdsub. Added the do_t_rd_rm_rn() routine and changed
the opcode entries for these instructions to use them. Radar bug #6522202.
Changes for the 5.26 release (the cctools-740 release):
- Removed the last abort() call in the reloc() routine in i386.c so user errors
do not appear like assembler crashes. Radar bug #6516397.
Changes for the 5.26 release (the cctools-739 release):
- Changed the arm assembler to use 0x46c0 as the thumb1 nop pad for sections
with pure instructions. The change was in s_align() in read.c. Radar bug
#6517544.
Changes for the 5.26 release (the cctools-738 release):
- Removed the code and files for the old ProjectBuilder interface including the
use of mig(1) in the Makefile. The files make.defs make_defs.h were removed
and the code in as.c and messages.[ch] with the OLD_PROJECTBUILDER_INTERFACE
#ifdef was removed. Radar bug #6457206.
- Changed the arm assembler to put out a relocation entry for 32-bit thumb
branches using the new ARM_THUMB_32BIT_BRANCH relocation type. Radar bug
#6511482.
- Changed arm.c to assign the enum type BFD_RELOC_THUMB_PCREL_BRANCH20 to
the new ARM_THUMB_32BIT_BRANCH relocation type's value.
- Changed nrelocs_for_fix() and fix_to_relocation_entries() in write_object.c
so there is an #ifdef ARM and tests for the new ARM_THUMB_32BIT_BRANCH
relocation type when archflag_cpusubtype == CPU_SUBTYPE_ARM_V7. To count 1
for the fixes in nrelocs_for_fix() and to put out a relocation entry in
fix_to_relocation_entries() for the type ARM_THUMB_32BIT_BRANCH. The
relocation entry put out is a local relocation entry that is pcrel, length
of long and an symbolnum of R_ABS (0). Also in fix_to_relocation_entries()
it sets fr_type to ARM_THUMB_32BIT_BRANCH so that
BFD_RELOC_THUMB_PCREL_BRANCH20 versions with no symbol gets set correctly.
Changes for the 5.26 release (the cctools-737 release):
- Changed the logic in fixup_section() in layout.c to correctly not add the
value of the symbol into the item to be relocated if it is an external
coalesced symbol or weak symbol. Radar bug #6505278.
- Changed frag_grow() in frags.c to not reduce the frags.chunk_size if 2 times
the number of bytes needed is less than that. Radar bug #6490903.
- Changed the arm assembler to pad sections with pure instructions with nop's
other than zero bytes. Radar bug #6492867.
- Changed arm.c and made the variable thumb_mode global and not static.
- Changed s_align() and added #ifdef ARM code to use specific nop values.
Changes for the 5.26 release (the cctools-736 release):
- Fixed misspelling of "can not" to "cannot" in 4 error messages in i386.c .
Radar bug #6447059.
- Fixed bugs in dwarf support:
- Changed layout_addresses() in layout.c to make two loops for calling
relax_section(). The first for non-debug sections so the sizes of their
sections are set before the second loop of the debug sections which might
have line number sections. Radar bug #6460301.
- To get the old .file and dwarf .file to both work the trick is that
dwarf2_directive_file() parse the old .file directive stuff and pass
it to s_app_file() to make this work.
- This requires that dwarf2_directive_file is in md_pseudo_table[].
- Changed s_file() in read.[ch] to s_app_file().
- Changed the entry in the pseudo_table[] table from "file" to "appfile"
in read.c
- Changed process_entries() to return a non-zero value if it output something.
- Changed out_debug_line() to check the return value of process_entries() and
if nothing was outputted then output the a DW_LNE_set_address sequence to
set the address to zero and the DW_LNE_end_sequence.
- Changed dwarf2_finish() to also test files_in_use == 0 as part of its check
that there is no line information and no non-empty .debug_info section. Then
also added a check of all_segs == NULL before the check of
all_segs->next == NULL, and added this same set of checks before calling
out_debug_*() routines at the end. This is need so the file table in the
line section is put out for something that has only data and no line
numbers.
- Changed fixup_section() in layout.c for x86_64 in debug sections to actually
set the value of the expression.
- Changed section_new() in sections.c to not set now_seg for section types
of S_ZEROFILL.
- Added the routine add_last_frags_to_sections() to layout.[ch] so that the
hacks in get_frag_fix() will work. As get_frag_fix() will return 0 as the
size of the last frag in a section.
- Changed is_assembly_time_constant_subtraction_expression() in layout.c
to treat expression with symbols from a debug section with FAKE_LABEL_NAMEs
to be constant.
- Changed fixup_section() in layout.c when it has a section difference
relocatable subtraction where a symbol is at the end of a section and
the section it is relocating is a debug section as an assembly time constant
section if both symbols have FAKE_LABEL_NAMEs.
- Changed section_set() in sections.c to simply call section_new() to fix
the problems that it was not correctly setting frag_now and dropping the
the fragments from the sections frag chain.
- Fixed using basename() back to using lbasename() and added an lbasename()
routine just before get_filenum() which returns a pointer into the path
as expected by get_filenum().
- Fixed section_symbol() in symbols.c that was incorrectly testing
frcP->section_symbol != NULL when it should have been == NULL.
- Fixed the call to frag_align() in out_debug_aranges() in to pass the
fill prameters correctly so it does not crash.
- In read.c changed these routines:
sizeof_sleb128()
sizeof_uleb128()
to be two routines:
sizeof_sleb128_32() and sizeof_sleb128_64()
sizeof_uleb128_32() and sizeof_uleb128_64()
which has parameters of proper 32 and 64 bit signed types so when values of
different types get passed they are correctly converted. Without this it
will cause hangs as a shift of a negative value will be an arithmetic
shift and will never become zero.
- And added new sizeof_leb128() and sizeof_uleb128() routines that are
#ifdef'ed ARCH64 to call the correct sized and typed new routines.
- Changed relax_inc_line_addr() to not use make_expr_symbol() which is
ifdef'ed to abort, and changed it to use symbol_temp_new() and create a
SECT_DIFF expression in that symbol.
- Changed dwarf2dbg_estimate_size_before_relax() to match what
relax_inc_line_addr() creates and evaluate the symbol's expression before
the symbol values are set using their frag values.
- Changed dwarf2dbg_convert_frag() to match what relax_inc_line_addr() creates
and evaluate the symbol's expression after the symbol values are set.
- Hooked up the dwarf support for .loc and .file in the i386, ppc and arm
assemblers. Radar bug #4298593.
- Removed the #ifndef NeXT_MOD in the md_pseudo_table[] in i386.c around the
.file and .loc directives that use dwarf2_directive_file() and
dwarf2_directive_loc() respectively.
- Removed the #ifdef OBJ_ELF in the md_pseudo_table[] in arm.c around the
.file, .loc and .loc_mark_labels directives that use dwarf2_directive_file()
dwarf2_directive_loc() and dwarf2_directive_loc_mark_labels() respectfully.
- Removed the #ifndef NeXT_MOD in output_insn() in i386.c around the call to
dwarf2_emit_insn() at the beginning of the routine.
- Removed the #ifdef OBJ_ELF in output_inst() in arm.c around the call to
dwarf2_emit_insn() at the end of the routine.
- Removed the #ifdef OBJ_ELF in output_relax_insn() in arm.c around the call
to dwarf2_emit_insn() at the beginning of the routine.
- Removed the #ifdef OBJ_ELF in arm_frob_label() in arm.c at the end of the
routine.
- Added a call to dwarf2_emit_insn() in md_assemble() in ppc.c after the
call to md_number_to_chars().
- Added a call to dwarf2_finish() in main() in as.c after the call to
perform_an_assembly_pass().
- Added a case label for rs_dwarf2dbg in layout_addresses() in layout.c
with a line of code:
dwarf2dbg_convert_frag(fragP);
- Added a case label for rs_dwarf2dbg in the first loop in relax_section()
in layout.c with a line of code:
address += dwarf2dbg_estimate_size_before_relax(fragP);
- Added a case label for rs_dwarf2dbg in the second loop in relax_section()
in layout.c with a line of code:
growth = dwarf2dbg_relax_frag(fragP);
Changes for the 5.26 release (the cctools-733 release):
- Changed s_weak_definition() in read.c, colon() in symbols.c and
layout_symbols() in write_object.c to not check for the symbol defined as
not being in a coalesced section. Then changed fix_to_relocation_entries()
in write_object.c to produce an external relocation entry if is is an external
coalesced symbol or weak symbol. And changed fixup_section() in layout.c to
not add the value of the symbol into the item to be relocated if it is an
external coalesced symbol or weak symbol. Radar bug #6422713.
Changes for the 5.26 release (the cctools-731 release):
- Fixed s_thumb_set() in arm.c to set the N_ARM_THUMB_DEF bit in the sy_desc
field of the symbol. Radar bug #6366869
Changes for the 5.26 release (the cctools-729 release):
- Put back in the change below and removed the #ifdef CHANGE_FOR_6233320 in
driver.c . Radar bug #6233320.
Changes for the 5.26 release (the cctools-728 release):
- Changed arm_force_relocation() in arm.c so the the first if() conditional
includes the check for the relocation type BFD_RELOC_THUMB_PCREL_BRANCH25 to
cause it to generate a relocation entry for thumb2 b.w version of the branch.
Radar bug #6297258.
- Changed md_apply_fix() in arm.c in the BFD_RELOC_ARM_SWI case if conditonal
from this:
if (fixP->tc_fix_data != 0)
to:
if (*((int*)fixP->tc_fix_data) != 0)
Radar bug #6337904.
- Backed out the change below with an #ifdef CHANGE_FOR_6233320 in driver.c
for now. Radar bug #6233320.
Changes for the 5.26 release (the cctools-727 release):
- Put back in the change below and removed the #ifdef CHANGE_FOR_6233320 in
driver.c . Radar bug #6233320.
Changes for the 5.26 release (the cctools-724 release):
- Backed out the change below with an #ifdef CHANGE_FOR_6233320 in driver.c
for now. Radar bug #6233320.
Changes for the 5.26 release (the cctools-723 release):
- Changed the assember driver, /usr/bin/as , when running on an Intel x86_64
host to default to -arch x86_64 instead of -arch i386. The change was in
driver.c in main() after it finds no arch_name set and calls
get_arch_from_host(). A test for cputype == CPU_TYPE_I386 is made and if
true the arch_flag gets set to the family flag for CPU_TYPE_X86_64. Radar
bug #6233320.
- radr://4545225 assembler should implement all dwarf-related GAS pseudo-ops
This is NOT yet hooked up as of cctools-723, but the code changes are put in
to top of tree to since they build and do not effect the operation of the
assembler's existing functionality.
- Moved now_seg definition from layout.c to sections.c and added the
definition of now_subseg to sections.c. Added their declrations to
sections.h .
- Changed section_new() sections.c to always set now_seg as the value from
the frch_nsect field in the struct frchain for the current section.
now_subseg is always left with a value of zero.
- This is done so now_seg can be used dwarf2dbg.c for the seg field
of the struct line_seg. And letting now_subseg remain zero so there are
just one line_subseg per line_seg.
- Added the new routines get_section_by_nsect(), get_section_by_name() and
section_set() to sections.[ch] .
- Added dwarf2dbg.c from GAS and ported it:
- The routine dwarf2_finish() got a bit of a rewrite to deal with Mach-O
objects. And now uses section_new() and get_section_by_name() to see if
it exists to replace the bfd_get_section_by_name() test.
- Changed the three uses of seg_info() to get_section_by_nsect() and
modified the code to use the returned struct frchain pointer.
- The routines to emit the dwarf sections: out_debug_line(),
out_debug_ranges(), out_debug_aranges(), out_debug_abbrev(),
out_debug_info() were changes to instead take segT parameters to take
struct frchain * parameters and there names changed to XXX_section from
XXX_seg.
- Changed subseg_set() to a new routine section_set() that was added to
change sections.
- Changed the call to resolve_symbol_value() in both
dwarf2dbg_estimate_size_before_relax() and dwarf2dbg_convert_frag()
to use the frag's values to calculate this.
- The use of O_subtract with expressionS and use of the X_op_symbol field for
the symbol value to subtract were change to use SEG_DIFFSECT and
X_subtract_symbol instead.
- Added rs_dwarf2dbg to relax_stateT enum in relax.h .
- Added a definition and declration of debug_type and enum debug_info_type
to as.[ch] and changed main() to set set debug_type to DEBUG_DWARF2 when
the argument --gdwarf2 is specified.
- Added an emit_expr() routine to dwarf2dbg.c based on the functionality of
the one in GAS's read.c but just for the ported uses in dwarf2dbg.c .
- Added to symbols.[ch] the routines: symbol_temp_new(),symbol_temp_new_now(),
symbol_temp_make() and symbol_set_value_now() needed in dwarf2dbg.c .
- Added a section_symbol field to the frchain struct in sections.h .
- Added to sections.[ch] section_symbol() to create and set a symbol at the
start of the section.
- Changed all the routines used in the poc_handler field of a pseudo_typeS to
change its parameter from int to uintptr_t to get rid of warnings. See
Radar bug #6219587 below.
Changes for the 5.26 release (the cctools-722 release):
- Changed s_fill() in read.c to allow .fill with a size of 8. Radar bug
#5783217.
Changes for the 5.26 release (the cctools-718 release):
- Fixed a bug in read.h which had an incorrect prameter type for the poc_handler
function pointer field of the pseudo_typeS typedef. The poc_val parameter was
an int type but is used to hold pointers so it needs to be a uintptr_t type.
Radar bug #6219587.
Changes for the 5.26 release (the cctools-716 release):
- Fixed a bug in s_zerofill() in read.c caused by the change in Radar bug
#6159238 where it needed to change to handle the case of an extra space after
the segment name before a comma in the parsing. Radar bug #6195283.
- Fixed a bug in parse_neon_el_struct_list() in arm.c where it was loosing
the information about the index used for things like "d0[]" and making it
look like "d0". In the loop to parse the argument list, in the is case
"{d0[]-d1[0]}" it called parse_typed_reg_or_scalar() which was passed
pointer to a neon_typed_alias struct, atype, to fill in. Then when it gets to
parsing the "-" it does not look at the atype.defined like the code after
the parsing of "-". So that same bit of code which looks like:
if ((atype.defined & NTA_HASINDEX) != 0)
{
if (lane == -1)
lane = atype.index;
else if (lane != atype.index)
{
first_error (_(type_error));
return FAIL;
}
}
else if (lane == -1)
...
to match what is done after the the parsing of "-" so it does not lose the
information about the index of "d0[]". What was happening for example was
this instruction:
vld1.f32 {d0[]-d1[]}, [r0]
was ending up as this instruction:
vld1.32 {d0-d1}, [r0]
Radar bug #6185952.
Changes for the 5.26 release (the cctools-715 release):
- Fixed the .thumb_set directive fails which was failing to parse its second
operand. The fix was in s_thumb_set() to include the call to pseudo_set()
outside the #ifdef. And to change pseudo_set() in read.[ch] to not be
static. Radar bug #6182961.
- Fixed the handling of -force_cpusubtype_ALL for the arm assembler. The code
near the end of md_begin() that tested for force_cpusubtype_ALL needed to
be changed to set the cpu_variant variable to allow everything and to set
archflag_cpusubtype to CPU_SUBTYPE_ARM_ALL as that is what will get set into
the resulting object file (not md_cpusubtype). Radar bug #6150777.
- Fixed a problem with the change to Radar bug #4141844 where it was trying to
allow a prefix on the same line as the instruction without a separating shash,
'/', character. The original change added a new "substate" variable to app.c.
But there were cases that it did not get set back it its initial value of zero
at the start of each line (state 0). So to make this a bit more robust all
places where the state gets set back to zero the substate is also set to zero.
Radar bug #6159238.
Changes for the 5.26 release (the cctools-714 release):
- Added the arm instructions vmrs and vmsr to be the same as fmrx and fmxr in
arm.c . Radar bug #6152739.
- Fixed a bug where the field sy_has_been_resolved in the symbol struct was
being used uninitialized. The fix is in symbol_new() in symbols.c .
Radar bug #6163385.
Changes for the 5.26 release (the cctools-713 release):
- Changed an error message in md_apply_fix() for the BFD_RELOC_ARM_OFFSET_IMM8
case to use as_bad_where() not as_bad() so a file and line number gets
printed. Radar bug #5890008.
Changes for the 5.26 release (the cctools-711 release):
- Fixed a problem with getting the correct value of a difference of two symbols
that it was divided by 2 for 32-bit targets. As the resulting value was
being calculated into a 64-bit signed type, signed_expr_t, with a += of an
expression of 2 32-bit unsigned types in fixup_section() in layout.c .
Radar bug #6096697.
Changes for the 5.26 release (the cctools-709 release):
- Fixed a problem with .quad expression values getting assembled with constants
larger than 32-bits being added to symbols. Two places were using 32-bit
types, the offset field in the fix struct in fixes.h and the local variable
value in fixup_section() in layout.c . Radar bug #5955044.
- Removed the call to abort() in reloc() in i386.c that should not be there
after a call to as_bad() to generate an error message. Radar bug #6029941.
Changes for the 5.26 release (the cctools-706 release):
- Fixed a problem with the ported to run on 64-bit hosts, where changes to
atof-generic.c had 3 places change to uint32_t which should have been int32_t.
Changes for the 5.26 release (the cctools-705 release):
- Ported to run on 64-bit hosts. Radar bug #6022298
Changes for the 5.26 release (the cctools-703 release):
- Added the .inlineasmstart and .inlineasmend directives to check that only
numeric lables are defined between them. And if not generate an error message
from the arguments of the .inlineasmstart directive. Radar bug #5896375.
- In read.c added the new routine s_inlineasm() and added it to the
pseudo_table[]. For the .inlineasmstart directive, s_inlineasm() is passed
a value of 1 and parses out the optional file_name, line_number and
column_number arguments. When the .inlineasmstart directive is see it sets
the new inlineasm_checks variable and fills in the new variables
inlineasm_file_name, inlineasm_line_number and inlineasm_column_number with
the arguments. And when the .inlineasmend directive is seen it clears
the new variables. The new variables are defined in read.[ch].
- The routine colon(), in symbols.[ch], was changed to take a new parameter
local_colon that is set when called by local_colon. And a test was added
if inlineasm_checks is set and local_colon() is not to cause an error.
If inlineasm_file_name is set then the new routine, in messages.[ch],
as_warn_where_with_column() is called else as_bad() is called to generate
the error message.
- The two plaves in read.c that call colon() had a 0 second argument added and
the call to colon() from local_colon() in symbols.c had a 1 second argument
added.
- Improved the x86_64 assembly time by caching the value of the linked list
search of x86_64_resolve_local_symbol() in i386.c. Added two fields,
sy_has_been_resolved and sy_prev_resolved, to the symbol structure in
struc-symbol.h. Radar bug #5902686.
Changes for the 5.26 release (the cctools-701 release):
- Fixed a problem in expr() in expr.c when it is combining a symbol that is a
set to a substract expression, SEG_DIFFSECT, with a constant. In that case
it gets incorrectly combined to a SEG_ABSOLUTE instead of retaining the
SEG_DIFFSECT expression type. Also Changed the x86_64 assembler to allow a
32-bit displacement operand as a symbol set to the subtraction of two
instructions. Code was added in output_disp() in i386.c to test for this
specific case and create the needed fix for it. Radar bug #5683920.
- Fixed a problem in pseudo_set() in read.c when it is setting a symbol to
another symbol that was already a section difference expression. In that case
expression() returns an expression of just one symbol and a type of
SEG_DIFFSECT. Code at the case for SEG_DIFFSECT needed to be added to handle
this case and copy the expression of the symbol in to the new symbol being
set. Radar bug #5684077.
- Changed the x86_64 assembler to allow a 32-bit displacement operand as a
subtraction of two instructions. Code was added in output_disp() in i386.c
to test for this specific case and create the needed relocation entries for
it. Radar bug #5622258.
Changes for the 5.25 release (the cctools-697 release):
Changes for the 5.25 release (the cctools-689 release):
- Fixed a place that needed force_cpusubtype_ALL tested in md_assemble() in
arm.c . Radar bug #5817558.
- Added support for the special assembly time constant expression of the diff-
erence of two symbols defined in the same section then divided by exactly 2.
Radar bug #5819903.
- Added the field X_sectdiff_divide_by_two to the expressionS typedef in
expr.h .
- In expr() in expr.c added code to look for an expression with a SEG_DIFFSECT
and a left side with an SEG_ABSOLUTE X_add_number == 2 and a O_divide right
side and if so set the new X_sectdiff_divide_by_two field.
- Added the field fx_sectdiff_divide_by_two to the struct fix.
- Changed cons() in read.c to save the fix for a SEG_DIFFSECT and to copy the
the new X_sectdiff_divide_by_two field from the expression to the new
fx_sectdiff_divide_by_two field in the fix.
- In fixup_section() in layout.c added code to test the news
fx_sectdiff_divide_by_two field and if so divide the value by 2. And check
to see that the expression is an assembly time constant.
Changes for the 5.25 release (the cctools-686 release):
- Changed arm_relax_frag() in arm.c to take a section number for the section it
is operating on. That is in turn now passed to relax_branch() in arm.c and
in there if the symbol that is being branched to is not for that section it
returns 4 for the size of the instruction.
- In md_apply_fix() for BFD_RELOC_THUMB_PCREL_BRANCH25 the fx_r_type type is
set to BFD_RELOC_THUMB_PCREL_BRANCH23 so it can be output as a known
relocation type (ARM_THUMB_RELOC_BR22).
Changes for the 5.25 release (the cctools-685 release):
- Changed relaxed_symbol_addr() in arm.c to also add in
fragp->fr_symbol->sy_nlist.n_value when calculating the addr. Without this
it can some times get "branch out of range" errors since it appears this
to have the address of the symbol from the start of the fragment.
- Changed the check in md_assemble() in arm.c for "width suffixes are invalid
in ARM mode" to allow them with unified syntax.
- Changed the arm APPLE LOCAL trap instruction opcode entry in arm.c from this:
CE(trap, 7ffdefe, 0, (), noargs),
to this:
TUE(trap, 7ffdefe, defe, 0, (), noargs, noargs),
which allows it to work with thumb mode.
- Changed relax_section() in layout.c in two plaves #ifdef'ed ARM. In the first
look it now sets frag's relax_marker field to zero. In the "Do relax" loop
it flips the relax_marker field with an "^= 1". This is to fix a problem in
relaxed_symbol_addr() so it knows to add the stretch amount to the addr it
is returning.
- Added a second call to as_file_and_line() in frag_var() in frags.c which was
#ifdef ARM at the top of the file to set the fr_file and fr_line fields of
the current frag before creating a new frag. This is needed because as the
line number of the last instruction which is the "relaxable" instruction is
the line number that is needed (not the first one). So when relax_section()
calls arm_relax_frag() which calls md_convert_frag() it will call
fix_new_exp() and fix_new_internal() so a new fix may be created after all
code is read for the last instruction in the frag. So md_convert_frag()
copies the fr_file and fr_line fields of the frag into the fx_file and fx_line
fields of the fix so line numbers can be reported.
Changes for the 5.25 release (the cctools-684 release):
- Changed relaxed_symbol_addr() in arm.c to not use the hack ported
S_GET_VALUE() as at that point the symbol does not have the value set. The
code was #ifdef NOTYET to use fragp->fr_symbol->sy_frag->fr_address to get
the address from the symbol's frag. The hack ported S_GET_VALUE() routine
was removed from symbols.[ch].
- Changed relax_addsub() where it had a memcpy() hack in place of the
#ifdef NOTYET out call to bfd_get_16() to a call to a call to
md_chars_to_number() to retrive the old thumb opcode into op that
output_relax_insn() put into the frag.
- Added a call to as_file_and_line() in frag_var() in frags.c which was #ifdef
ARM to set the fr_file and fr_line fields of the frag. This is needed because
when relax_section() calls arm_relax_frag() which calls md_convert_frag()
it will call fix_new_exp() and fix_new_internal() so a new fix will be created
after all code is read. So md_convert_frag() copies the fr_file and fr_line
fields of the frag into the fx_file and fx_line fields of the fix so line
numbers can be reported.
Changes for the 5.25 release (the cctools-683 release):
- Fixed a bug in fix_new_arm() in arm.c after the call to fix_new_exp() before
the default: case label as missing a break; statement.
- Fixed a bug in relax_section() in layout.c for the "Do relax" loop where
the call to arm_relax_frag() should have set the variable growth not
incremented. The "+=" was change to just an assignment, "=".
Changes for the 5.25 release (the cctools-682 release):
- Changed relax_section() in layout.c for the "Do relax" loop and #ifdef'ed
ARM the rs_machine_dependent case to call arm_relax_frag().
- Changed md_convert_frag() where it had a memcpy() hack in place of the
#ifdef NOTYET out call to bfd_get_16() to a call to a call to
md_chars_to_number() to retrive the old thumb opcode into old_op that
output_relax_insn() put into the frag.
Changes for the 5.25 release (the cctools-681 release):
- Removed all __OPEN_SOURCE__ references in Makefile, notes, as.c, layout.c,
messages.c, symbols.c, write_object.c, frags.h and symbols.h .
- Fixed a bug in the table entries for the pextrq and pinsrq instructions as
they should have had Size64 to force the REX.W byte.
- Fixed a bug in the table entries for the i386 SSE4.1 instruction insertps
where it had the second parameter as a general register not a xmm register.
The change was in i386-opcode.h to change Reg32 and Reg64 to RegXMM.
Radar bug #5650346.
- Changed output_relax_insn() in arm.c to just used the undefined symbol instead
of calling make_expr_symbol() if the expression is a just an undefined symbol
and no subtract symbol.
Changes for the 5.25 release (the cctools-680 release):