qfcc now does local common subexpression elimination. It seems to work, but
is optional (default off): use -O to enable. Also, uninitialized variable
detection is finally back :)
The progs engine now has very basic valgrind-like functionality for
checking pointer accesses. Enable with pr_boundscheck 2
Getting everything right with an enum proved to be too difficult if not
impossible. Also use better tests for equivalence and intersection.
Many more tests have been added. All pass :)
Also move the ALLOC/FREE macros from qfcc.h to QF/alloc.h (needed to for
set.c).
Both modules are more generally useful than just for qfcc (eg, set
builtins for ruamoko).
Aliasing the jump table to an integer broke statement_get_targetlist with
the new alias def handling, and was really wrong anyway. I probably did
that due to being fed up with things and wanting to get qfcc working again
rather than spending time getting jumpb right.
The depth limits in the gl and glsl renderers and in the trace code really
bothered me, but then the fix hit me: at load-time, recurse the trees
normally and record the depth in the appropriate place. The node stacks can
then be allocated as necessary (I chose to add a paranoia buffer of 2, but
I expect the maximum depth will rarely be used).
While accessing short foo[2][4]; as foo[0][0..7] should work in theory, who
knows what gcc does with it when optimizing. I don't know if this will fix
johnnyonflame's bsp loading problem, but no point in having rhinodemonic
code hanging around.
This necessitated disabling the id2 padding, but it's only commented out
incase there's more growth. Now the (compiler) error in -addObjectNoRetain
is caught ealier.
Using "=" was rather confusing, so changing it to "<CONV>" seems to be a
good idea. As the string is used only for selecting opcodes at compile
time, only qfcc is affected.
Using "=" was rather confusing, so changing it to "<CONV>" seems to be a
good idea. As the string is used only for selecting opcodes at compile
time, only qfcc is affected.
Going by "standard" Objective-C, retainCount really doesn't belong in
Object itself. The way GNUStep does it is to stash retainCount in memory
just below the object by allocating extra bytes for the count and returning
a pointer just beyond those extra bytes. Now Ruamoko does the same. This
fixes the inconsistencies in structure layouts for Protocol and class
structs between qfcc generated (internal) structs and user visible structs.
The attached patch (against quakeforge git) changes the [con]width,
[con]height, and most importantly the rowbytes members of viddef_t
from unsigned to signed int, like in q2. This allows for a properly
negative vid.rowbytes which may be needed in, e.g. a DIB sections
windows driver if needed. Along with it, I changed a few places
where unsigned int is used along with comparisons against the relevant
vid.* members.
One thing I am not 100% sure is the signedness requirements of
d_zrowbytes and d_zwidth: q2 has them as unsigned but I am not sure
whether that is because they are needed as unsigned or it was just an
oversight of the id developers. They do look like they should be OK
as signed int to me, though: comments?
==
Note from Bill Currie: I had to do some extra changes as many
signed/unsigned comparisons were somehow missed.
This fixes the horribly different results between optimized and unoptimized
qfbsp (there is still a difference of 1 brushface). Unfortunately, it also
severely limits the maximum size of a map.
It turns out the expected orientation of the sky cube is exactly that of
Blender's default cube looked at from the front view (num-1) and the front
face being the nearest face. This put's Marcher's sun nicely in the view
when exiting the cave.
Rearrange the sky_suffix and sky_coords arrays and remove the sky_target
array such that the faces can be loaded using
GL_TEXTURE_CUBE_MAP_POSITIVE_X + i (apparently certain drivers break if
the faces aren't loaded in the correct order).
Also, the nomalization of the direction vector in the fragment isn't
necessary.
Need to subtract the size of the bsp_t/bsp29_t struct. Now old and new
qfbsp produce identical bsps (so long as they're both unoptimized, or
(probably) both optimized).
All of the nastiness is hidden in bspfile.c (including the old bsp29
specific data types). However, the conversions between bsp29 and bsp2 are
implemented but not yet hooked up properly. This commit just gets the data
structures in place and the obvious changes necessary to the rest of the
engine to get it to compile, plus a few obvious "make it work" changes.
Certain versions of qcc (fteqcc comes to mind :P) strip the names from
builtin functions. This breaks saved games that happen to have a builtin
function in a saved function variable. The earlier builtin name
reconstruction patch happened to fix the writing of save games for such
progs, and this one fixes the reading.
The setup had been lost at some stage, thus shadows were always directly
under the entity. Unlike the original quake shadow code, the vector is
correctly transformed into the entity's space.
I finally found the cause of Despair's gl shadows non-rendering+segfault...
the shadow code expected triangle fans and strips but was getting simple
triangles. Oops.
Nothing in the main program currently uses Key_Progs_Init, so the object
file wasn't getting pulled into the link. However, it's quite necessary for
the client console plugin :/
That is, the descriptors loaded from the progs file. Some compilers (eg,
fteqcc :P) strip builtin names from the progs, which makes debugging
difficult.
LordHavoc had made lighting positive for sw32, but I had done something in
the plugin code that broke that (probably something to do with the
colormap loading). Going back to id's original code fixes the issue.
This reverts commit e170f4ee75.
It turns out I messed up something in the patch. I noticed the problem
while playing digs04.bsp: many sub-model surfaces, particularly those with
animated textures, were not being transformed correctly. As this patch did
not make a large performance difference, it's probably better to just
revert it. I might revisit it later.
Since the backtile is loaded into a scrap and used as a subtexture, we
can't use GL's texture wrapping, thus do the wrapping ourselves. There are
some minor issues with the wrong part of the scrap being drawn: need to
investigate where the bug is (vrect, make_quad, etc).
Rather than checking the raw edict count in the entities file against the
progs' max_edicts, check the allocated entity's number. This allows loading
of sophisticated maps (eg, digs04) that prune many of their entities.
In the end, it turns out this is the correct fix for the gl seg on
overkill, because build_skin will correctly use the fully setup player skin
if the glskin doesn't have a texture associated with it.
It turns out glsl, sw and sw32 weren't getting any benefit from R_CullBox
because the frustum wasn't setup :P. Get another 8% out of bigass1
(174->184fps). bigass1 now runs 2x as fast as it did before I started this
optimisation run :)
This severely reduces the calles to BindTexture, and more importantly,
glUseProgram, EnableVertexAttribArray etc. The biggest changes are:
o icons and text are all in the one giant texture
o icons and text are mixed in the one queue
This gave ~9% speedup for bigass1 (159->174fps).
It seems recent(?) 64-bit strcpy implementations of strcpy don't work
properly for overlapping regions even when moving down. Quite the
surprise, as I thought that would always work. *shrug*
I didn't like the way client/server code was poking around at the
implementation. Instead, provide a couple of accessor functions for the
same information.
There will normally be only one unnamed field (if any), and it's always the
null field. This will put an eventual end to the "'' is not a field"
messages.
Double converting texcoords results in 0,0 for all affected texcoords. Mr
Fixit was looking rather ill. Now he looks weird (something wrong in the
renderer?).
gl, sw and sw32 use blend palettes, so share the code. This also abandons
the optimization for transforming verts in sw (had all sorts of problems
anyway). sw still doesn't work, though.
There are still many issues to sort out, but the basics are working.
Problems:
rendered fullbright (no lighting done)
normals are ignored
extra textures (glow etc) not used/loaded
4 models on the screen don't seem to be a problem.
Though the bsp loader doesn't yet support colored lighting, the ambient
light will be colored when it does. With this, I guess iqm model support is
done for glsl until I figure out how I want to do dual quaternion support.
There's still a problem with his finger tips and feet, but the rest of his
limbs seem to be working well. Much thanks to Spike for encouraging me to
do a dump of the matices that are actually sent to the card.
It turns out that animated joints remain relative right up to the last
moment.
Since iqm vertex arrays are variable, and I don't want to calculate the
stride every time I render a model, cached the value used when building the
arrays.
And the tests really exercised VectorShear (first attempt had things
messed up when more than one shear value was non-zero). Also,
Mat4Decompose wasn't orthogonalizing the z axis row. Oops. Anyway,
Mat4Decompose is now known to work well, and the usage of its output is
understood :)
It seems (some versions of) windows vsnprintf don't count the terminating 0
when limiting the number of chars written to the buffer. Nor do they
guarantee the output string will be terminated.
I got the idea from blender when I discovered by accident that quat * vect
produces the same result as quat * qvect * quat* and looked up the code to
check what was going on. While matrix/vector multiplication still beats the
pants off quaternion/vector multiplication, QuatMultVec is a slight
optimization over quat * qvect * quat* (17+,24* vs 24+,32*, plus no need to
to generate quat*).
This avoids sending invalid pose data to the renderer. The symptom was a
vertex array offset higher than the vertex array size. Discovered by calim
of nouveau while he was debugging a driver problem found by QF. Many
thanks.
One's an actual bug, the other a bit of error checking (not sure how
necessary it is, but it's in code that we don't /want/ to run, so it can't
hurt :)
While this particular tigger of the real bug was caused by 659d95221e
(hopefully fix both the "get stuck waiting for 3d" bug and the null
worldmode bug.), the real bug was lurking in the code since the dawn of
time (from sw32's perspective). This fix is as per LordHavoc's suggestion
(heh, despite the years, he knows his code), but I spent the time hunting
down the trigger to understand just what was going on.
It turns out that (0,0,0) is too close to a wall (probably on, but the
slight default offset is too close) and the above commit changed the first
rendered frame to be before the player origin was set rather than after.
This fix feels correct to me because noclipping around with the sw32
renderer would probably hit the same bug with a bit of bad luck. Thus
ensure the index resulting from zi never exceeds 65535.
While checking the shaders to see if there might be anything obvious to
work around the current nouveau shader issues, I found a 1 that should have
been a 1.0. I'm surprised it ever compiled.
It doesn't seem to have any useful effect in QF (even before the plugin
project) other than setting the number of frames to update. I'm not sure if
it's a useless variable or one where the user is supposed to match it to
the system configuration. Anyway, with this, the glsl plugin now works.
This allows the vid module to load the render module and access render
specific functions before the renderer initializes, which happens to need
an initialized vid module...
The renderer now gets initialized and things sort of work (qw-client will
idle, though nothing is displayed). However, as the viddef stuff is broken,
it segs on trying to run the overkill demo.