The global symbol offset is a loop invariant; no need to compute it
over and over. The compiler probably will not be able to do this for
us due to global variables and function calls.
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
Clean up the arithmetic for global symbols; in particular, make it
clear where the dependency on the debug format comes from (DWARF needs
three symbol table entries all by itself.)
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
Bump GLOBAL_TEMP_BASE so far that it *cannot* conflict for ELF32
(which has 24-bit symbol table indicies) and is *unlikely* to conflict
for ELF64 (which has 32-bit symbol table indicies.)
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
Eliminiate hard-coded section numbers, at least to the best of our
ability. There is still a very odd piece of computation in
elf_build_reltab() which I can't really figure out...
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
The dependency machinery relies on properly rooted includes, so give
it to them... the path syntax munging machinery in the dependency
script handles it from a Makefile syntax perspective, and then we can
hope that C compilers are smart enough to deal with forward-slash
paths even when that is not the native syntax.
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
Add something approaching real ELF header files.
Begin merging the common ELF code, beginning with the section name
detection.
Drop automatic generation of .comment section, and in particular the
treatment of .common as a special section (if we decide generating
.comment is still a good idea, we should just do it as a macro.)
Augment the list of known sections, and make it table-driven.
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
Some vendors get nervous about parameterized printf patterns;
furthermore, it's completely unnecessary in this case.
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
Fix the SAFESEH directive for the specific case of a symbol internal
to the program. With the optimizer enabled, it would otherwise fail
unless the symbol is external.
Fix crash caused by uninitialised memory that lead to dangling pointer
in the rbtree. This can be seen by compiling zsnes 1.50, with a file
that define many symbols, such as fxemu2c.asm.
We have a number of all-zero buffers in the code. Put a single
all-zero buffer in nasmlib.c. Additionally, add fwritezero()
which can be used to write an arbitrary number of all-zero bytes;
this prevents the situation where the all-zero buffer is simply
too small.
When a section other than .text has a start < origin, we would
segfault; fix that.
Furthermore, at bin_cleanup() we don't have usable file/line
information, so pass ERR_NOFILE to the error() function. Perhaps less
than ideal, but better than printing a null pointer.
Previously, the ELF backends silently ignored incorrect or unknown
attributes on section declarations, and therefore used default values
in cases where the user had make an error in attempting to specify
custom values.
I needed entry point support with the as86 format, and after looking through
the archives found a similar desire from someone in 2002. For some reason
such a patch never made it into the code, even though the required flag
value is present, so I offer the a patch of my own.
I compared against what is done in the .obj format and the approaches are
quite similar which I hope will aid in its acceptability. While I have
tested it extensively it does do the job asked, and I'm honestly not sure
what extensive testing of the change would look like.
Several projects have taken to using .text to store read-only data
when building on Mac OS X due to crashes in SSE code from the .rodata
section being mis-aligned. It seems there was a misunderstanding about
how ld/ld64 handles section alignment in outmacho.c so I wrote a patch
to fix it. I tested it against x264 git, modified it to use ".rodata
align=16" for the data section and use movdqa instructions (guaranteed
to crash when built with unpatched nasm) and it passed all tests in
its checkasm tool.
If you want more data I can provide, but it's late and I've had a
couple glasses of mulled wine :)
-DrD-
Linear searches are evil, so use an llrbtree to search for symbols by
offset. This doesn't change the preexisting behaviour that we only
look for global symbols.
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
nasm.c should respect the default debug format of the output format,
instead of replacing it with the first format in the list.
This is cleaner and allows the list to be sorted normally.
This commit rewrites commit 116994111b which was very fragile.
Move all the version strings to a single compilation unit, ver.c; this
does not include the version macros, which are fed into macros.c.
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
Set default attributes for .tdata and .tbss sections
Implement new attribute 'tls' for arbitrary section names
Flag variables in sections with tls attribute with STT_TLS
Add a common file, outlib.c, for output formats. Add the function
realsize() instead of open-coded variants in almost every backend.
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
Set STT_TLS in symbol table for symbols declared
in thread local storage sections. Note that,
for now at least, such symbols must also be
declared as GLOBAL.
The "bin" format was misinterpreting the overloading of the "size"
argument to out(), which caused another source of 64-bit relative
offset errors.
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
It is unclear if we will ever see any "naked" (absolute bytes)
OUT_REL*ADR coming from the assembler, but if we do, we should
generate them correctly.
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
Now when the assembler is properly generating the address that we push
down to the backend, enable requesting an exact value for these
relocations (these are pointing to a specific GOT or PLT slot; the
addend is used to adjust the computed value in the instruction, not
for offset for the symbol.)
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
The "size" argument to the OUT_REL*ADR output types is actually
intra-instruction offset, not the actual size. Thus, emit the size
properly.
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
Fix the arithmetic for relative GOT/PLT references.
We still can't enable exactitude, because of the assumption that
"size" is always the proper adjustment for the offset of the
displacement inside the instruction, which is wrong in the case of
displacements that are followed by an immediate. This also affects
the list file, so it really should be fixed.
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
GOTOFF64 is used for local variables (as a 64-bit offset from the GOT;
only needed in the Medium PIC or Large PIC models.) It therefore
should *not* be a elf_add_gsym_reloc() invocation.
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
I am having a bit of a hard time understanding the proper operation of
the "exact" flag to elf_add_gsym_reloc(). We apparently won't
generate proper GOTOFF64 relocations with this flag set; it is
possible that there are *no* proper uses of this flag. This clearly
needs to be figured out.
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
GOT and PLT references need a symbol; after all, they reference a GOT
or PLT slot. Thus, they need elf_add_gsym_reloc(). Mungify the
interface so that they can communicate the need for the PC-shifted
offset into the relocation.
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
When generating an address that is *not* tied to a symbol, we just
want to emit the bytes. I believe the assembler is already supposed
to do that for us, but just in case, do it right here too.
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
The x86-64 ABI wants the symbol addend to reside in the addend field
of the RELA relocation, not in the code stream. Apparently it's
something one can get away with, but the linker would still botch it
for some cases. Change it so we pass the proper output and emit zero
into the code stream.
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
Try to make the various GOT relocations do the right thing in ELF64,
including erring out when appropriate.
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
Expressions like
mov r15,[rel integer wrt ..got]
lea rax,[rel integer wrt ..gotoff]
now assemble correctly.
In addition, a fix has been made to the corresponding
abs relocations.
Both of these areas still need additional testing.
Somehow the win32 and win64 aliases got listed on Mach-O, not on
COFF. This doesn't have any effect on the current code, but might in
the future. Correct.
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
Don't use explicit L's for things which are really size_t; not only is
it unnecessarily ugly, but it's wrong in a lot of ways. Do some other
minor stylistic cleanups.
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
We would leave the output symbol type uninitialized. Explicitly
initialize it to zero (T_NULL, meaning no symbol type information),
since that's what was effectively done.
Signed-off-by: H. Peter Anvin <hpa@zytor.com>
The testcase illustrates the problem. After "nasm -f obj
alonesym.nasm"
let's look to dump:
======
PUBDEF386(91) recnum:5, offset:0000005bh, len:03f9h, chksum:bbh(bb)
Group: 0, Seg: 1
00020000h - 'sym0000' Type:0
00020004h - 'sym0001' Type:0
....
00020134h - 'sym0077' Type:0
PUBDEF(90) recnum:6, offset:00000457h, len:000ah, chksum:b6h(b6)
Group: 0, Seg: 1
00000138h - 's' Type:2
0000b600h - '' Type:0
======
The problem is while 's' offset is 20138h it is marked as type 90h not
91h. The root cause is located in obj_x():
static ObjRecord *obj_x(ObjRecord * orp, uint32_t val)
{
if (orp->type & 1)
orp->x_size = 32;
if (val > 0xFFFF)
orp = obj_force(orp, 32);
if (orp->x_size == 32)
return (obj_dword(orp, val));
orp->x_size = 16;
return (obj_word(orp, val));
}
It sets up x_size and than writes data. In the testcase data are the
offset and this offset overflows a record. In this case the record is
emitted and its x_size is cleared. Because this is last PUBDEF the new
record with only 's' symbol is emitted also but its x_size is not 32
(it's still zero) so obj_fwrite doesn't switch to 91h type.
The problem seems to be very generic and expected to be occurred on
many other record types as well.
----
And the fix is simple:
if (orp->x_size == 32)
{
ObjRecord * nxt = obj_dword(orp, val);
nxt->x_size = 32; /* x_size is cleared when a record overflows */
return nxt;
}
ctype functions take an *int*, which the user is expected to have
taken the input character from getc() and friends, or taken a
character and cast it to (unsigned char).
We don't care about EOF (-1), so use macros that cast to (unsigned
char) for us.
Move the handling of "extra" macros (i.e. output format macros) into
the macros.pl mechanism. This allows us to change the format of the
internal macro store in the future - e.g. to a single byte store
without redundant pointers.
Also, stop using indicies into a long array when there is no good
reason to not just use different arrays.
