It now addressing_mode cleaning up store instructions to use ptr+offset
instead of lea;store ptr...
Entity.field addressing has been impelmented as well.
Move instructions still generate sub-optimal code in that they use an
add instruction instead of lea.
This allows the code handling simple pointer dereferences to recurse
along an alias chain that resulted from casting between different
pointer types (such chains could probably be eliminated by replacing the
type in the original pointer expression, but it wasn't worth it at this
stage).
Aliasing an alias expression to the same type as the original aliased
expression is a no-op, so drop the alias entirely in order to simplify
code generation.
Simply dereferencing a pointer does not need to go through array_expr
and thus collect a 0 offset that will only be constant-folded out again.
Really just a minor optimization in qfcc, but at one stage in today's
modification, it resulted in some unwanted aliasing chains.
While this does make the generated code a little worse, load is behaving
nicely), the two are at least consistent with each other and when I fix
one, I'll fix both. I missed this change the other day when I did the
address_expr cleanup. Yay near-duplicate code :P
This is what using new_ret_expr would result in, but new_ret_expr is no
longer used for referencing .return (except in pascal, but I haven't
gotten around to sorting that out) due to the recent changes for Ruamoko
progs. Fixes an ICE when compiling (with optimization) something like
the following (dir is a vector):
dir /= sqrt (dir * dir);
return dir * speed;
It turns out the sorting wasn't working properly and I've decided that
anything that actually needs the defs to be sorted by address (such as a
debugger searching for defs by address) can do the sorting itself. Fixes
a weird swapping of def names.
This is necessary to get statement disassembly working, and likely
debugging in general. locals is the total size of the stack frame and
thus reaches above the function-entry stack pointer, and params_start is
the local space relative start of the parameters. Thus, knowing the
function-entry stack pointer, the bottom of the locals space can be
found by subtracting params_start, and the top of the locals space by
adding (locals - params_start).
This gets all the sections of the progs file nicely aligned and the code
easier to read with the offset and size calculations not being spread
through the function. ivar-struct-return now works when compiled for
Ruamoko.
This cleans up dprograms_t, making it easier to read and see what chunks
are in it (I was surprised to see only 6, the explicit pairs made it
seem to have more).
While I think the reason the dags code moved an instruction before
adjstk and with was they shared a constant with that instruction (which
is a different bug), this ensures other instructions cannot get
reordered in front of adjstk and with, as doing so would cause any such
instructions to access incorrect data.
The goal was to get lea being used for locals in ruamoko progs because
lea takes the base registers into account while the constant pointer
defs used by v6p cannot. Pointer defs are still used for gobals as they
may be out of reach of 16-bit addressing.
address_expr() has been simplified in that it no longer takes an offset:
the vast majority of the callers never passed one, and the few that did
have been reworked to use other mechanisms. In particular,
offset_pointer_expr does the manipulations needed to add an offset
(unscaled by type size) to a pointer. High-level pointer offsets still
apply a scale, though.
Alias expressions now do a better job of hanling aliasing of aliases by
simply replacing the target type when possible.
It's possible I lost the child printing when creating the return
expressions, but dot diagrams are much more useful when they don't have
nodes with just pointer values.
The parameter defs are allocated from the parameter space using a
minimum alignment of 4, and varargs functions get a va_list struct in
place of the ...
An "args" expression is unconditionally injected into the call arguments
list at the place where ... is in the list, with arguments passed
through ... coming after the ...
Arguments get through to functions now, but there's problems with taking
the address of local variables: currently done using constant pointer
defs, which can't work for the base register addressing used in Ruamoko
progs.
With the update to test-bi's printf (and a hack to qfcc for lea),
triangle.r actually works, printing the expected results (but -1 instead
of 1 for equality, though that too is actually expected). qfcc will take
a bit longer because it seems there are some design issues in address
expressions (ambiguity, and a few other things) that have pretty much
always been there.
The aux use ops need to be counted and given nodes explicitly as they
may refer to defs that are not accessed by other statements other than
by aliases, and those aliases need to be marked live as well as the used
def.
This is part of the work for #26 (Record resource pointer with builtin
function data). Currently, the data pointer gets as far as the
per-instance VM function table (I don't feel like tackling the job of
converting all the builtin functions tonight). All the builtin modules
that register a resources data block pass that block on to
PR_RegisterBuiltins.
This will make it possible for the engine to set up their parameter
pointers when running Ruamoko progs. At this stage, it doesn't matter
*too* much, except for varargs functions, because no builtin yet takes
anything larger than a float quaternion, but it will be critical when
double or long vec3 and vec4 values are passed.
Storing a variable into a dereference pointer (*p = x) is not marking
the variable as used (due to a mistake while converting to Ruamoko
statement format) resulting in assignments to that variable being
dropped due to it being a dead assignment as the assignment to the
variable and the storing need to be in separate basic blocks (thus the
call in the test, though an if would have worked, I think) for the bug
to trigger.
The problem was a missed change when switching the internal statement
format to Ruamoko: I "used" the statement's operands directly rather
than the rotated ones when emitting v6p progs. Fixes a compile segfault
when NOT optimizing.
There was an out-by-one where attempting to run a program with only one
argument would result in the argument not being passed to the program
(two worked). This is actually the source of the error fixed in
9347e4f901 because test-harness.c was the
basis for qwaq's main.c
While all base registers can be used for any purpose at any time (this
is why the with instruction has hard-absolute modes: you can never get
permanently lost), qfcc currently uses the convention of register 0 for
globals and register 1 for stack locals (params, locals, function args).
The register used to access a def is stored in the def and that is used
to set the register bits in the instruction opcode.
The def code actually doesn't know anything about any conventions: it
assumes all defs are global for non-temp defs (the function code updates
the defs before emitting code) and the current function provides the
register to use for any temp defs allocated while emitting code.
Seems to work well, but debug is utterly messed up (not surprised, that
will be tricky).
Still need to get the base register index into the instructions, but I
think this is it for basic code generation. I should be able to start
testing Ruamoko properly fairly soon :)
Thanks to the use/def/kill lists attached to statements for pseudo-ops,
it turned out to be a lot easier to implement flow analysis (and thus
dags processing) than I expected. I suspect I should go back and make
the old call code use them too, and probably several other places, as
that will greatly simplify the edge setting.
The means that the actual call expression is not in the statement lint
of the enclosing block expression, but just its result, whether the call
is void or not. This actually simplifies several things, but most
importantly will make Ruamoko calls easier to implement.
The test is because I had some trouble with double-calls, and is how I
found the return-postop issue :P
Commit 76b3bedb72 broke more than just the
swap test, but at least I know I need to get an edge in the dag.
Currently, the following code is generated: return and add are reversed.
../tools/qfcc/test/return-postop.r:8: return counter++;
0001 store.i counter, .tmp0
0002 return .tmp0
0003 add.i .tmp0, (1), counter
However, I don't want to deal with it right now, so it's marked XFAIL.
Since Ruamoko now uses the stack for parameters and locals, parameters
need to come after locals in the address space (instead of before, as in
v6 progs). Thus use separate spaces for parameters and locals regardless
of the target, then stitch them together appropriately for the target.
The third space is used for allocating stack space for arguments to
called functions. It us not used for v6 progs, and comes before locals
in Ruamoko progs.
Other than the return value, and optimization (ice, not implemented)
calls in Ruamoko look like they'll work.
Thanks to me having done something right 20 years ago, that was pretty
easy :). The two boolean types aren't supported yet because I haven't
decided on just how to represent their types in qfcc.
This seems to be the most reasonable approach to allocating space for
function call parameters without using push and pop (or adding to the
stack pointer), though it's probably good even when using push and pop
to help keep things aligned.
My little test program now builds with the Ruamoko ISA :)
void cp (int *dst, int *src, int count)
{
while (count--) {
*dst++ = *src++;
}
}
Calls are broken (unimplemented), and non-void returns are not likely to
work either (only partially implemented).
Operand width is encoded in the instruction opcode, so the width needs
to be accounted for in order to select the correct instruction. With
this, my little test generates correct code for the ruamoko ISA (except
for return, still fails).
For the most part, it wasn't too bad as it's just a rotation of the
operands for some instructions (store, assign, branch), but dealing with
all the direct accesses to specific operands was a small pain. I am very
glad I made all those automated tests :)
This makes the v6p instruction table consistent with the ruamoko
instruction table, and clears up some of the ugliness with the load,
store, and assign instructions (. .= and = are now spelled out). I think
I'd still prefer an enum code (faster) but at least this is more
readable.
Missed this case in duplicate_type. Allows "short foo" and
"sizeof(short)" (even though qfcc and the engine have two ideas of the
size: I expect trouble later).
long is ignored for double, and v6p progs are stuck with 32 bits for
longs (don't feel like extending v6p any further), but the basics are
there for Ruamoko.
short is ignored for ints because the minimum size is 32, and signed is
just noise for ints anyway (and no chars, so...).
unsigned, however, is finally implemented properly (or at least seems to
be working correctly: tests pass after getting things compiling again,
and lt.u is used where it should be :)
Attempting to add ev_ushort caused ptraliasenc to break, but that was
because it was already broken: I had implemented the scan of the xdef
table incorrectly, thus adding only 1 ev type resulted in the walked
pointer being out of phase with its data due to it first passing over
the type encodings (which is why adding long and ulong didn't cause any
obvious trouble).
And other related fields so integer is now int (and uinteger is uint). I
really don't know why I went with integer in the first place, but this
will make using macros easier for dealing with types.
They are both gone, and pr_pointer_t is now pr_ptr_t (pointer may be a
little clearer than ptr, but ptr is consistent with things like intptr,
and keeps the type name short).
I don't know why they were ever signed (oversight at id and just
propagated?). Anyway, this resulted in "unsigned" spreading a bit, but
all to reasonable places.
This includes calls and unconditional jumps, relative and through a
table. The parameters are all lumped into the one object, with some
being unused by the different types (eg, args and ret_type used only by
call expressions). Just having nice names for the parameters (instead of
e1 and e2) makes it nice, even with all the sub-types lumped together.
No mysterious type aliasing bugs this time ;)
The move operator names are definitely obsolete (due to dropping the
expressions a year or two ago) and the precedence checks seem to be
handled elsewhere. Memset and state expressions went away a while back
too.
While this was a pain to get working, that pain only went to prove the
value of using proper "types" (even if only an enum) for different
expression types: just finding all the places to edit was a chore, and
easy to make mistakes (forgetting bits here and there).
Strangely enough, this exposed a pile of *type* aliasing bugs (next
commit).
v6 vs v6p are more or less as before, with ruamoko added in. qfcc will
now try (and fail, due to the opcode table opnames being wrong) to
create ruamoko progs when given the ruamoko target option.
At this stage, I doubt emit.c will need to know the details of the
target (v6, v6p, ruamoko) since the instruction formats are identical,
just different meanings for the opcode itself.
This allows v6, v6p (older QF VM) or ruamoko (new QF VM) to be targeted.
Currently defaults to v6p to allow QF to continue building without too
much hassle.
While qfcc dealing sensibly with mixed target VMs in the object files
has always been an outstanding issue, with the new instruction set it
has become a priority. Most importantly, this should allow QF to
continue building while I work on qfcc targeting the new IS.
It does little good for documentation to refer to fields that don't
exist (because a certain someone forgot to change the docs when changing
the field names, I wonder who :P).
And partial implementations in qfcc (most places will generate an
internal error (not implemented) or segfault, but some low-hanging fruit
has already been implemented).
This allows the VM to select the right execution loop and qfcc currently
still produces only the old IS (it doesn't know how to deal with the new
IS yet)
build_struct was unconditionally setting the type's alignment. This was
not a problem before because no types were requesting alignments larger
than those requested by their members (for structs). However, with the
upcoming new instruction set, quaternions need to be 4-word aligned.
For int, long, float and double. I've been meaning to add them for a
while, and they're part of the new Ruamoko instructions set (which is
progressing nicely).
The opcode table is a nightmare to maintain, but this does clean it up
and speed up opcode lookups since they can now be indexed. Of course, it
turns out I had missed adding several instructions, so had to fix that,
and qfcc needed a bit of a re-jigger to get the opcode out of the table.
The assignment to the node's variable must come after any uses of that
node, which the node's parent set indicates. In the swap test, this was
not a problem as the node had no parents, and in the link order test, it
just happened(?) to work.
While using just the label node's reachable set was sufficient for a
simple swap (t = a; a = b; b = t;), it is not sufficient for
read-before-write dependencies such as found in linked-list building:
{ o = array[ind]; o.next = obj; obj = o; }
The assignment to o.next uses obj, but that use is hidden because obj's
reachable nodes does not include o thus assigning o to obj causes the
array dereference to be assigned directly to obj and thus o.next winds
up pointing to o instead of whatever obj was. The parent nodes of obj's
node are its users, so any new assigned to obj must come after those
parents as well as any node reachable by obj's node.
Fixes a runaway loop error when adding a frikbot to the server.
qfo_to_progs was modifying the space data pointers in the input qfo,
making it impossible to reuse the qfo. However, qfo_relocate_refs needs
the updated pointers, thus do a shallow copy of the qfo and its spaces
(but not any of the data)
build_builtin_function does the right thing, and it was only legacy
syntax functions that were affected anyway. Certainly, external
variables should not be initialized, but klik uses @extern { } wrapped
around several builtin functions and I had added the feature to allow
just this as it is rather convenient.
I decided that the check for whether control reaches the end of the
function without performing some necessary action (eg, invoking
[super dealoc] in a derived -dealoc) is conceptually the return
statement using a pseudo operand and the necessary action defining that
pseudo operand and thus is the same as checking for uninitialised
variables. Thus, add a pseudo operand type and use one to represent the
invocation of [super alloc], with a special function to call when the
"used" pseudo operand is "uninitialised".
While I currently don't know what else pseudo operands could be used
for, the system should be flexible enough to add any check.
Fixes#24
I want to use the function's pseudo address that was used for managing
aliased temporary variables for other pseudo operands as well. The new
name seems to better reflect the variable's purpose even without the
other pseudo operands as temporary variables are, effectively, pseudo
operands until they are properly allocated.
Forgetting to invoke [super dealloc] in a derived class's -dealloc
method has caused me to waste far too much time chasing down the
resulting memory leaks and crashes. This is actually the main focus of
issue #24, but I want to take care of multiple paths before I consider
the issue to be done.
However, as a bonus, four cases were found :)
While get_selector does the job of getting a selector from a selector
reference expression, I have long considered lumping various expression
types under ex_expr to be a mistake. Not only is this a step towards
sorting that out, it will make working on #24 easier.
I have gotten tired of chasing memory leaks caused by me forgetting to
add [super dealloc] to my dealloc methods, so getting qfcc to chew me
out when I do seems to be a good idea (having such a warning would have
saved me many hours, just as missing return warnings have).
Well... it could be done better, but this works for now assuming it's in
/usr/include (and it's correct for mxe builts). Does need proper
autoconfiscation, though.
The portal flow stack nodes contain a simd vector, which requires
16-byte alignment. However, on 32-bit Windows, malloc returns 8-byte
aligned memory, leading to eventual segfaults. Since pstack_t is 48
bytes on 32-bit systems, it fits nicely into a 64-byte aligned cache
line (or two on 64-bit systems due to being 80 bytes).
For most (if not all) maps. The heapsort is needed only if the clustered
leafs are not contiguous, but most bsp compilers output contiguous leaf
clusters, so is just a bit of protection. The difference isn't really
noticeable on a fast machine, but no point in doing more work than
necessary.
Now that only 3852 clusters need to be checked for each cluster, fat-pvs
construction for ad_tears completes in about 0.7s, most of which seems
to be loading, conversion, compression and writing. O(N^3) cuts both
ways (hurts like crazy when N increases, does wonders when N decreases,
especially by a factor of 25). And then throw in improved cache
performance...
I suspect having an off-line compiler is still useful, but even if
qfvis's implementation never actually gets used, if cluster
reconstruction is put in the engine, large maps will be feasible even
for quakeworld. Just the reduced memory requirements alone will be a
huge benefit (~3GB down to 1.8MB).
This is only the first half (vertical) in that the vis bits are still
for the leafs rather than the clusters, but ad_tears goes from 500s to
7s for calculating the fat pvs (3852 clusters).
While this doesn't give as much of a boost as does basic sphere culling
(since it's just culling sphere tests), it took ad_tears' base vis from
1000s to 720s on my machine.
This removes the last of the arbitrary limits from qfvis. The goal is
not so much supporting crazy maps, but more about better data usage
(cluster_t is now 24 (or 16) bytes instead of 1048 (or 528). And
passages isn't used (yet?)...
It turns out cmem is not so good for many large allocations (probably a
bug in handling the blocks), but was really meant for lots of little
churning allocations anyway. After an analysis of winding lifetimes, it
became clear that the hunk allocator would work very well. The base
windings are allocated from a global hunk (currently 1GB, plenty for
even ad_tears), and ephemeral windings are allocated from a per-thread
hunk of 1MB (seems to be way more than enough: gmsp3v2 uses a maximum of
only 56064 bytes, and ad_tears got through 30% before I gave up on it).
Any speed difference (for gmsp3v2) seems to be lost in the noise: still
completing in 38.4s on my machine.
The output fat-pvs data is the *difference* between the base pvs and fat
pvs. This currently makes for about 64kB savings for marcher.bsp, and
about 233MB savings for ad_tears.bsp (or about 50% (470.7MB->237.1MB)).
I expect using utf-8 encoding for the run lengths to make for even
bigger savings (the second output fat-pvs leaf of marcher.bsp is all 0s,
or 6 bytes in the file, which would reduce to 3 bytes using utf-8).
After seeing set_size and thinking it redundant (thought it returned the
capacity of the set until I checked), I realized set_count would be a
much better name (set_count (node->successors) in qfcc does make much
more sense).
Extremely large maps take a very long time to process their PVS sets for
PHS or shadows, so having an off-line compiler seems like a good idea.
The data isn't written out yet, and the fat pvs code may not be optimal
for cache access, but it gets through ad_tears in about 500s (12
threads, compared to 2100s single-threaded in the qw server).
This reduces the overhead needed to manage the memory blocks as the
blocks are guaranteed to be page-aligned. Also, the superblock is now
alllocated from within one of the memory blocks it manages. While this
does slightly reduce the available cachelines within the first block (by
one or two depending on 32 vs 64 bit pointers), it removes the need for
an extra memory allocation (probably via malloc) for the superblock.
When moving an identifier label from one node to another, the first node
must be evaluated before the second node, which the edge guarantees.
However, code for swapping two variables
t = a; a = b; b = t;
creates a dependency cycle. The solution is to create a new leaf node
for the source operand of the assignment. This fixes the swap.r test
without pessimizing postop code.
This takes care of the core problem in #3, but there is still room for
improvement in that the load/store can be combined into a move.
This reverts commit 2fcda44ab0.
Killing the node is not the correcgt answer as it blocks many
optimization opportunities. The correct answer is adding edges to
describe the temporal dependencies. Of course, this breaks the swap.r
test.
In order to correctly handle swap-style code
{ t = a; a = b; b = t; }
edges need to be created for each of the assignments moving an
identifier lable, but the dag must remain acyclic (the above example
wants to create a cycle). Having the reachable nodes recorded makes
checking for potential loops a quick operation.
Identifiers can be constants. I don't remember quite what it fixed other
than some bogus kill relations in the dags (which might have caused
issues later).
If the src type is not a class, there is no inheritance chain to walk.
Fixes a segfault when returning self after a syntax error in the
following:
+(EditStatus *)withRect:(Rect)rect
{
return [[[self alloc] initWithRect:rect]:
}
-setCursorMode:(CursorMode)mode
{
cursorMode = mode;
return self;
}
GCC does a fairly nice job of producing code for vector types when the
hardware doesn't support SIMD, but it seems to break certain math
optimization rules due to excess precision (?). Still, it works well
enough for the core engine, but may not be well suited to the tools.
However, so far, only qfvis uses vector types (and it's not tested yet),
and tools should probably be used on suitable machines anyway (not
forces, of course).
This fixes the mightsee updates never occurring, but it doesn't make a
huge difference (though I suppose it might have back in the 90s, or with
a different map).
The stats were being updated before UpdateMightsee was getting called,
and it was incrementing the wrong value (so it would not have been
thread-safe).
While whether it's any faster is debatable (it's slightly slower, but
many more portals are being tested due to different rounding in the base
vis stage), it's certainly easier to read.
While the main bulk of the improvement (36s down from 42s for
gmsp3v2.bsp on my i7-6850K) comes from using a high-tide allocator for
the windings (which necessitated using a fixed size), it is ever so
slightly faster than using malloc as the back-end.
This is for the conversion /to/ paletted textures. The conversion is
necessary for csqc support. In the process, the conversion has been sped up
by implementing a color cache for the conversion process. I haven't
measured the difference yet, but Mr Fixit does seem to load much faster for
the sw renderer than it did before the change (many months old memory).
The server edict arrays are now stored outside of progs memory, only the
entity data itself (ie data accessible to progs via ent.fld) is stored in
progs memory. Many of the changes were due to code accessing edicts and
entity fields directly rather than through the provided macros.
Double benefit, actually: faster when building a fat PVS (don't need to
copy as much) and can be used in multiple threads. Also, default visiblity
can be set, and the buffer size has its own macro.
