Make 9d0ab74e9e more responsible by limiting permission changes to
pages that are intended to be both writable and executable for code
generation. That way, the signal handler doesn't just reopen holes in
loaded foreign libraries that W^X would close.
A new resource section was aligned based on the old section's
virtual address, instead of the PE's specified section
alignment. That could make alignment round up too far, leaving
a disallowed gap in the sections' virtual addresses.
Conform to W^X by using a signal handler that switches between W and X
mode on any fault. That's not the spirit of W^X, certainly, but it
should make Racket work without special configuration.
Beware that this change can turn some crashes into infinite loops.
It may be possible to detect those loops, but I didn't find a
good and portable way, so far.
Creating an executable with embedded DLLs means that the executable
can be truly stand-alone, instead of needing to be kept with its
DLLs in a relative subdirectory.
DLL embedding works by bypassing the OS's LoadLibrary and
GetProcAddress functions, and instead maps the DLL into memory
and performs relocations explicitly. Joachim Bauch's MemoryModule
(Mozilla license) implements those steps.
Perpetuate a failure to make Windows paths behave reasonably with
path-manipulation functions.
In one case, the new implementation seemed better than the old one, so
I've changed the old implementation (by deleting code) and test cases.
The old code would split "x /y" to "\\?\REL\x " and "y", and the new
one splits to "x /" and "y"; the trailing separator is now enough to
preserve the space character, and it also preserves the directoryness
of the path. Of course, "x /" splits to 'relative and "\\?\REL\x " as
it strips away the trailing "/".
A remaining problem in both implementations: some Windows API
functions implicitly erase a trailing "." in a directory name, making
"x./y" equivalent to "x/y". The Racket path-manipulation functions
don't do that, so splitting and recombining "x./y" does not access the
same path as the original. This apparently hasn't been a problem in
practice, and there are so many terrible hacks already, so I left it
alone.
The new implementation perpetuates also the implementation mistake of
representing paths internally as byte strings. If, in some terrible
universe, I'm forced to do this again, the right choice is probably to
keep the path in a parsed form with enough information to reconstruct
the original, but with the information sorted nicely to make various
normalizations and combinations easy.
Since only the expander is compiled in unsafe mode right now, the
checks are skipped only when the implementations of `reverse`, `map`,
etc., are part of the flattened expander.
Add a cache on binding lookup that is like the old expander --- a
small cache that is consulted before the more general cache that is
already in place.
The new cache layer primarily helps when a single identifier is
compared to a sequence of other identifiers.
