error: invalid conversion between vector type '__m128i' (vector of 2
'long long' values) and integer type 'uintptr_t' (aka 'unsigned long')
of different size

in 8 xmm chunks (8*128).
todo: check cpuid globally, cast aligned sp to __m128i vector.
we have 8 xmm0-7 regs, so spread it over 8 parallel copies.

some compilers have such a bad optimizer, that they cannot get the
object_size of static vars, like pgcc 17.4 and pgcc 19.4 on linux and
darwin. (nvidia)

Adjust the tests to detect that limitations.  In many cases we have now 3
different compiler-dependent results on overflows: msvcrt (different native secapi),
pgcc (no bos detection) and normal compilers with proper object_size
compile-time detection.

but dont exit on failing tests. null slack is different with RMAX+1.
see branch pgi-slack

The preprocessor define changed, the crashing bug still exists

for gcc-8 and linux.  No changes, just the linux glibc is now faster than
the darwin libc.

RSIZE_MAX_STR + 1 leads to a compile-time abort

PGC-S-0037-Syntax error: Recovery attempted by deleting <wstring>.
a preprocessor problem

but only temporary, because we define/undef it case by case.
without it set the probe is useless

so that the perf binaries are still fast.
Closes #71

with older compilers

Closes #60

esp. EXPECT_BOS_TODO on fail. make it a TODO then.

compile-time check for NULL and zero, not just
overflow

and rename the wcsnlen_s args for consistency

oops. use word size

pgroup.com pgcc does not support -pedantic -Werror,
probe for it on non-GNUC compatible compilers.
Add @PEDANTIC@ and @WARN_CFLAGS_TESTS@

Tested ok with nvidia PGI pgcc 17.4 and 19.4 on darwin.

when we pass same pointers on purpose. This should warn.

refactor the EXPSTR ind local.

there are still inconsistencies vs a clean install.

ignore the msys2 SECAPI. it does not honor MINGW_HAS_SECURE_API anymore.
link to our qsort_s and asctime_s implementations.

newer gcc (9):
test_private.h:329:5: warning: macro expands to multiple statements [-Wmultistatement-macros]
     ind = strcmp((s1), (s2));                                                  \
     ^~~
test_gets_s.c:81:10: note: in expansion of macro 'EXPSTR'
     else EXPSTR(sub, "1234");
          ^~~~~~
test_gets_s.c:81:5: note: some parts of macro expansion are not guarded by this 'else' clause
     else EXPSTR(sub, "1234");

looking for conflicting packages...
:: mingw-w64-x86_64-pdcurses and mingw-w64-x86_64-ncurses are in conflict. Remove mingw-w64-x86_64-ncurses? [Y/n]
:: mingw-w64-x86_64-pdcurses and mingw-w64-x86_64-termcap are in conflict. Remove mingw-w64-x86_64-termcap? [Y/n]
error: failed to prepare transaction (could not satisfy dependencies)

which fails with ENOENT when TZ is set.

by clang-format.
skip those few tests files which do need EXPECT_BOS on the same line
clang-format-all.sh takes no argument now

e.g. my windows is stuck at an earlier unicode casetable.
check for a new unicode 12 casefold.

The -uni16 tables for windows has now overflows (-1) with indirection,
add code for that. Previously we only needed overflow _exc handling
for the compat tables. cperl had that code already, because utf8
had much longer entries.

This should fix the windows crashes.

both branches were empty in most cases

we may always declare _ctime_s_chk, just skip the ctime_s define when its already defined

use the gcc-4 builtin __sync_synchronize, not
ASM_VOLATILE("memory_barrier" ::: "memory"), which is documented to work,
but does not on the arm-linux-gnueabihf-gcc 8.3 cross compiler

bump so version in smoke.sh