This makes available the __builtin_va_list type and __builtin variants
of va_start, va_arg, va_copy and va_end. We do this via a header file
that's prepended to all compilations always (except if merely
preprocessing): tcc_predefs.h. That header could also be used
for predefining other builtins in the future.
We don't need the define hacks for musl anymore with this.
Also fix x86_64 gfunc_prologue to reserve enoug space for the
full va_list structure, not just 16 bytes.
this partly reverts 1803762e3 to fix stdarg on x86-64 again. I've tried
to retain the apple specific changes from that commit.
Also include stdarg.h first in tcc.h, maybe that helps as well.
When compiling on macOS (at least in version 10.12) the TCC compiler failed
to compile libtcc1.a. Three problems were solved:
- The predefined macro "__APPLE__" is now available, as it is tested in the
libc darwin header files
- the libtcc1 Makefile defined _ANSI_SOURCE, although it used signals
- stdargs.h defined va_list differently from the darwin libc.
If the darwin standard library was included BEFORE stdargs this caused
problems.
- the darwin libc generated a warning if GCC < 4 was used
- additional defines are predefined now to make darwin libc headers compile.
as there's overlap between handling types for binary and ternay
operations. Factor this into a single routine (combine_types).
This uses the structure that gen_op was following, and expr_cond
was using as well in the past, which I find easier to reconvene
with the standard language. But it reuses the new functions for
diagnostics to improve (a little) on what GCC or clang produce :)
the support for the macro GCC_MAJOR is gone since 2017, and it's
fairly doubtful that anyone serious is using gcc 2.95.
Also adds a test for the ternary ops typing rules: 'x?bool:bool' has
to promote to int.
as per testcase. We must not reset token.sym_label twice with
kept symbols. This is no problem for non-label symbols because those
aren't generated on demand when mentioning them.
On windows. there is no long double really IOW it is the
same as double. However setting the VT_LONG flag in
combination with VT_DOUBLE allows to keep track of the
original type for the purpose of '_Generic() or more
accurate type warnings.
... in order to avoid fp stack overflow (see test below).
#include <math.h>
#include <stdio.h>
int main()
{
printf("%f %f %f %f\n", trunc(1.2), rint(1.2), trunc(1.2), rint(1.2));
printf("%f %f %f %f\n", trunc(1.2), rint(1.2), trunc(1.2), rint(1.2));
printf("%f %f %f %f\n", trunc(1.2), rint(1.2), trunc(1.2), rintl(1.2));
}
Also in rintl:
- 'long double' is not a ten-byte float on windows.
- Updated msvcrt.def with symbols from 64bit version of dll - it contains the float math functions missing in the 32bit dll.
- Made sure this patch only apply to to WIN32 and WIN64. For WIN32 float functions calls the double variants, on 64bit they are called natively.
by calling the double functions and removing unsupported __asm implementations using "=t".
This patch also enables the double version of logb(double).
Note: None of the 'long double' variants works though - could easily be fixed
by aliasing the double equvalents for completeness/compability.
Fix static assert to support literal string instead of just printing
the sring of the current token as it use to be
so we can now use _Static_assert(0, "0" "1") which will print
__FILE__ __LINE__ error: 01
commit ef42295fe8 intrduced 4 .o files to
win32/lib. But they (bcheck.o, bt-dll.o, bt-exe.o and bt-log.o) are not
removed by "win32/build-tcc.bat -clean". This commit should fix it.
- revert const-folding in gvtst() and put it back into
expr_landor(). Although it did make sense, one reason
not to do it is __builtin_constant_p() which may return
true when it shouldn't because of nocode_wanted, see test.
- tccgen_init() can do init_prec(), also for tcc -E.
- for nostalgic reasons, keep the original expression parser
functions in the source.
- Makefile: remove stale stuff
so that it also is called from the precedence parser. This
is complicated by the fact that something needs to be done before
the second operand is parsed in a single pass compiler, so it
doesn't quite fit into expr_infix itself. It turns out the smallest
code changes result when expr_landor remains separate. But it can
be tidied a bit.
This is smaller and uses less stack depth per expression (eight function
calls from expr_or to get down to a unary). It's a tiny bit faster
depending on how good the branch predictor is, on my machine a wash.
ELF files that refer to shared libs containing sym-versions, but
don't refer to any dynamic symbols with symbol versions (should happen
only with very simple shared libs) would generate an empty .gnu.version_r
section. Some dynamic linker contain bugs in that they don't check
the section size or DT_VERNEEDNUM (which are both zero for such files
we generate) before accessing the first entry, and then bail out with
a message like
./a.exe: error while loading shared libraries: ./a1.so: unsupported
version 25960 of Verneed record
(where the "version number" actually comes from neighboring bytes
from different sections).
So, there's not much choice, we simply must not generate such section.
- tests2/113_btdll.c: test handling multiple stabs infos
Also:
- libtcc.c: remove _ISOC99_SOURCE pre-defines. It is causing
strange warnings such as 'strdup not declared'
- i386/x86_64-gen.c cleanup bounds_pro/epilog. This discards
the extra code for main's argv. If needed, __argv might be
processed instead.
- tccgen.c:block(): reduce stackspace usage. For example with
code like "if (..) ... else if (..) ... else if (..)... "
considerable numbers of nested block() calls may occur.
Before that most stack space used when compiling itself was
for libtcc.c:tcc_set_linker().
Now it's rather this construct at tccpp.c:2765: in next_nomacro1():
if (!((isidnum_table[c - CH_EOF] & (IS_ID|IS_NUM))
|| c == '.'
|| ((c == '+' || c == '-')
...
the strcat checker first checks dest for overlap, then src.
If the padding byte between b[] and the pad[] arrays happens to be zero
the dest check would have succeeded and the src test failed. If that
padding byte would be zero the dest check would trigger first.
As we can't influence the padding byte (only the b[] and pad[] arrays)
it was random if the dest or src checks triggered.
This makes it reliably trigger the dest check first.
This makes it possible to get backtraces with executables
(including DLLs/SOs) like we had it already with -g -run.
Option -b includes -bt, and -bt includes -g.
- new file lib/bt-exe.c: used to link rt_printline and the
exception handler from tccrun.c into executables/DLLs.
- new file lib/bt-log.c: provides a function that may be
called from user code to print out a backtrace with a
message (currently for i386/x86_64 only):
int (*tcc_backtrace)(const char *fmt, ...);
As an extra hack, if 'fmt' is prefixed like "^file.c^..."
then the backtrace will skip calls from within 'file.c'.
- new file lib/bt-dll.c: used on win32 to link the backtrace
and bcheck functions with the main module at runtime
- bcheck.c: now uses the tcc_backtrace function from above
- tccgen.c: minor cleanups
- tccelf.c: stab sections get SHF_ALLOC for easy access.
Also in relocate_section(): 64bit relocations for stabs
in DLLs cannot work. To find DLL addresses, the DLL base
is added manually in tccrun.c via rc.prog_base instead.
- tccpe.c: there are some changes to allow merging sections,
used to merge .finit_array into .data in the first place.
- tccpp.c: tcc -run now #defines __TCC_RUN__
also: refactor a line in tal_realloc that was incompatible
with bcheck
- tcctest.c: fixed a problem with r12 which tcc cannot preserve
as well as gcc does.
- tests2/112_backtrace.c: test the feature and the bcheck test18
that previously was in boundtest.c
remove quadratic loops by not using side tables; address-taken
can simply be a flag per local sym, and the lbounds section can
be filled after symbols go out of scope at which point we know
if the address was taken, so that there's no need to compress it
again after the funcion is done.
we were emitting error messages for something like
'static int i = 2 || 1/0', even though the exception would be in
the unevaluated part. This doesn't destroy const-ness, so we must
accept it. This requires splitting the nocode_wanted values a bit more,
so that nocode_wanted due to const_wanted can be differentiated from
nocode_wanted due to non-evaluation.
Add __attribute__((constructor)) to __bounds_init.
- remove tcc_add_bcheck from i386-link.c and x86_64-link.c
- add simplified tcc_add_bcheck to tccelf.c
- Update tccrun.c to call constructor/destructor.
Set dynsym sh_info to number of local symbols in tccelf.c
Reduce stack size when bounds checking is enabled.
Added variable TCC_LIBBCHECK for windows support.
Add signal stack to detect stack overflow.
Add all & parameters in lbound_section and remove them if not used.
Close fd in tcc_relocate in tccrun.c
Fix section type constructor/destructor in tccelf.c
Add check code in tests/boundtest.c for mem/str functions.
Remove -ba from documentation.
Add bounds check signal info in documentation.
bcheck.c:
- Fix initial_pool alignment.
. Fix printf statements.
. Add prototypes for all external interface functions.
- Add TCC_BOUNDS_WARN_POINTER_ADD environment variable.
. Add ctype and errno data.
- Fix alloca when multithreading is used.
- Add lock for __bound_checking and __bound_never_fatal.
- Catch pthread_create and use locks when called.
- Detect in loaded in shared lib and use locks when found
- Use spin locks instead of semaphore locks.
- Make spin locked code as small as possible.
- Fix mem/str functions checking.
- Fix overlap checking mem/str functions.
this is a bit complicated: for i386 and x86-64 we really need to
extend return values ourself, as the common code now does. For arm64
this at least preserves old behaviour. For riscv64 we don't have to
extend ourself but can expect things to be extended up to int (this
matters for var-args tests, when the sign-extension to int64 needs to
happen explicitely). As the extensions are useless, don't do them.
And for arm32 we actually can't express GCC behaviour: the callee side
expects the return value to be correctly extended to int32, but
remembers the original type. In case the ultimate target type for the
call result is only int, no further extension is done. But in case
the target type is e.g. int64 an extension happens, but not from int32
but from the original type. We don't know the ultimate target type,
so we have to choose a type to put into vtop:
* original type (plus VT_MUSTCAST) - this looses when the ultimate
target is int (GCC: no cast, TCC: a cast)
* int (without MUSTCAST) - this looses when the ultimate target is
int64 (GCC: cast from original type, TCC: cast from int)
This difference can only be seen with undefined sources, like the
testcases, so it doesn't seem worthwhile to try an make it work, just
disable the test on arm and choose the second variant as that generates
less code.
This allows adding files or libraries from
#pragma comment(option, ...)
Also, {f}/file.c will be expanded with the directory of
the current source, that is the file that has the #pragma
