What changed in EDuke32, build by build

Rotation-fixing happens for a couple of hard-coded statnums that presumably never move (DEFAULT, STANDABLE, FX, FALLER, LIGHT), but for actors it wouldn't make sense since the common case is that they do move. For this reason, bit 4 was introduced in r1934. The position of such useractors will not diverge due to error roundoff accumulation in rotating sectors (SE0, train).

The former is really only a workaround. Walls/vertical sprites/pskies with ysize 512 (and presumably greater, but this was not tested) are rendered with one shade higher at the borders (1 pixel vlines) because of a certain assumption in the ASM (see comments there). With very dark shades, the palookup[] buffer is accessed oob. We simply allocate 256 bytes more at the end for each. The latter is only for CLASSIC_NONPOW2_YSIZE_WALLS builds, which is not enabled yet. It seems to matter only for the uncommon case where the such pskies repeat in the height. A 1680x1050 window fully covered with such a sky is then rendered at about 60/85 the FPS for me (mostly due to not using the 4 pixel vline routines), so it may be leaning a bit too much on the side of correctness. A compilation switch DEBUG_TILESIZY_512 is introduced in engine.c for demonstration purposes.

Reverts

705cc72

It was caused by the definition of c_dfDIJoystick using literal 24 and 16 values in place of sizeof(DIDATAFORMAT) and sizeof(DIOBJECTDATAFORMAT), which include pointers. On 64-bit, the values end up being 32 and 24, causing a discrepancy in which DIERR_INVALIDPARAM was thrown. https://gitorious.org/0xdroid/external_libsdl-12/commit/1b1e05db06edcbe64599c6a5ad703795e0e02acb/diffs

- Updated to use StackWalk64 function call, requiring some addition of headers from MinGW-w64 to compile with MinGW. - Code added to support 64-bit executables: functionality added and one warning fixed. - New DLL binaries compiled using i686-MinGW-w64 and x86_64-MinGW-w64. (Oddly, the DLL built with MinGW is 5 MB while MinGW-w64's is 1 MB.)

This global option will set bit 1024 and clear bits 256 and 512 for all rotatesprite calls, fixing complex HUD drawing code relying on precise alignment of individual elements (widescreen rotatesprite is entirely unsuitable for this purpose).

This includes a complete Windows header and library refresh, including the addition of 64-bit compiled libs: *libogg 1.3.0 *libvorbis 1.3.3 *zlib 1.2.7 *libpng 1.5.13 *libvpx 9a3de881c0e681ba1a79a166a86308bbc84b4acd *SDL_mixer 1.2.12 (for RENDERTYPE=SDL) *DirectX import libraries: dsound and dxguid (now included) To build in 64-bit, you essentially need MinGW's MSYS (but not MinGW itself) and MinGW-w64 at the top of your PATH. The target is automatically detected using `$(CC) -dumpmachine`. The EDukeWiki will get detailed instrucitons. All compiler and linker warnings when building in 64-bit mode have been fixed. Remaining 64-bit to-do: - The ebacktrace dll does not build under 64-bit. It uses code specific to the format of 32-bit executables and will have to be ported to work with 64-bit executables. A future 64-bit version will be named ebacktrace1-64.dll. - RENDERTYPE=SDL crashes in SDL_mixer's Mix_Linked_Version(). - DirectInput gives an error and does not function. This only affects joysticks, and the error never happens without any plugged in. - Port the classic renderer ASM to 64-bit. (Just kidding, this is way out of my league.) This commit includes a fair bit of Makefile development spanning all platforms, including simplifying the SDLCONFIG code, fixing build on Mac OS X (thanks rhoenie!), globally factoring Apple brew/port inclusion, enforcing that all -L come before all -l, and ensuring that $(shell ) is always :='d. In addition, I have resurrected the old GCC_MAJOR and GCC_MINOR detection using `$(CC) -dumpversion`, but I have made it failsafe in case the command fails or the version is manually specified. I have applied this new fine-grained detection where applicable, including allowing LTO, and restraining -W's to versions that support them.

Also, fix deficient logic in Gv_Free and Gv_Clear (both M32 and CON) so that gamevar and gamearray erasure results are (closer to) correct, and so that the game does not crash when system arrays are accessed from CON because they all have been nulled.

This fixes 2nd and following rotatesprite calls drawing differently from the first in a context with custom set-up aspect, as reported by Lezing here: http://forums.duke4.net/topic/775-eduke32-20-and-polymer/page__view__findpost__p__143090 Also, back up and restore the aspect state across EVENT_DISPLAYREST.

That is, fix some oversights introduced in the preceding runs. listglobals.sh is a helper to find global accesses. Translator: add "number-conversion" warning option, max. error limit.

- more codegen - make more members const, some char unsigned - fix some "geom" metamethods - '^' operator

1) see if we actually received anything (if http://eduke32.sourceforge.net/VERSION is unavailable, this returns now instead of continuing with nonsensical data) 2) check if we're getting a token from strtok()

Commit breaks MSVC builds by failing to provide NOWARN macros to MSVC.

The redundant ones are in code like this: s->cstat = (int16_t)32768; Because the value is eventually converted to the type of "s->cstat", any casts to integral types having at least as many bits are no-ops, signedness being irrelevant due to (probably any two's complement arch targeting compiler's) bit-pattern preserving semantics of these conversions. The now incorrect one is: if (lotag == (int16_t) 65535), "int32_t lotag" being read from a wall or sprite struct directly earlier. Now, with these members being unsigned, and (int16_t)65535 equalling -1, the check always fails. The correction fixes normal switches having such a lotag ending the level immediately. In short: integer casts before assignments are unnecessary, those in reads highly relevant!

Rewriting them in the obvious way, i.e. by casting the expression to int16_t first. (That is, this commit is the reverse of r3174, but with casts applied.) This fixes at least one regression: a FIREEXT with a hitag of 0 should not be linked with same- (that is, zero-) tagged SEENINEs or OOZFILTERs. Mind the corner cases!