Commit graph

1613 commits

Author SHA1 Message Date
David Given bb17aea73a You can now mark a register as corrupting a certain register class; calls work,
or at least look like they work. The bad news is that the register allocator
has a rare talent for putting things in the wrong register.
2016-10-15 01:15:08 +02:00
David Given 98fe70a7de Output register equality constraints work. 2016-10-14 22:17:02 +02:00
David Given f06b51c981 Keep track of register types as well as attributes --- the type being how we
find new registers when evicting values. Input constraints work (they were
being ignored before). Various bug fixing so they actually work.
2016-10-12 22:58:46 +02:00
David Given 23c3575f0f The register allocator now makes a spirited attempt to honour register
attributes when allocating. Unfortunately, backward edges don't work (because
the limited def-use chain stuff doesn't work across basic blocks). Needs more
thought.
2016-10-09 15:09:34 +02:00
David Given cfe5312fcc Predicates can now take numeric arguments. The PowerPC predicates have been
turned into generic ones (as they'll be useful everywhere). Node arguments for
predicates require the '%' prefix for consistency. Hex numbers are permitted.
2016-10-09 12:32:36 +02:00
David Given 2198db69b1 Instruction predicates work now. 2016-10-08 11:35:33 +02:00
David Given 7f901aa4d0 We're not using 'allocates' any more; clean up. 2016-10-06 21:33:43 +02:00
David Given 88fb231d6e Better constraint syntax; mcgg now passes register usage information up to mcg;
mcg can track individual hop inputs and outputs (needed for live range
analysis!); the register allocator now puts the basic blocks into the right
order in preparation for live range analysis.
2016-10-05 22:56:25 +02:00
David Given 92502901a7 Better management of register data. Add struct hreg. 2016-10-05 21:00:28 +02:00
David Given 8d4186130d Oops --- hadn't updated the nts array for the new child order. 2016-10-04 21:32:28 +02:00
David Given bd28bddb92 Massive rewrite of how emitters and the instruction selector works, after I
realised that the existing approach wasn't working. Now, hopefully, tracks the
instruction trees generated during selection properly.
2016-10-04 00:16:06 +02:00
David Given 68f98cbad7 Instruction selection now happens on a shadow tree, rather than on the IR tree
itself. Currently it's semantically the same but the implementation is cleaner.
2016-10-03 20:52:36 +02:00
David Given 288ee56203 Get quite a long way towards basic output-register equality constraints (needed
to make special nodes like NOP work properly). Realise that the way I'm dealing
with the instruction selector is all wrong; I need to physically copy chunks of
tree to give to burg (so I can terminate them correctly).
2016-10-02 23:25:54 +02:00
David Given 3aa30e50d1 Come up with a syntax for register constraints. 2016-10-02 21:51:25 +02:00
David Given c079e97492 Perform SSA conversion of locals. Much, *much* better code now, at least
inasmuch as it looks better before register allocation. Basic blocks now know
their own successors and predecessors (after a certain point in the IR
processing).
2016-10-02 17:50:34 +02:00
David Given 21898f784a We're going to need some type inference after all, I think. Let's do a little
for now and see how it goes.
2016-10-01 19:10:22 +02:00
David Given 91e277e046 Predicates work; we now have prefers and requires clauses. Predicates must be
functions. Not convinced that semantic types are actually working --- there are
still problems with earlier statements leaving things in the wrong registers.
2016-10-01 13:56:52 +02:00
David Given 4a3a9a98dc It doesn't really make a lot of sense to have BURG nonterminal names different
to register classes, so combine them. Refactor the map code.
2016-10-01 12:17:14 +02:00
David Given 3a973a19f3 Move fatal(), warning() and aprintf() into the new data module (because they're
really useful).
2016-09-30 19:10:30 +02:00
David Given b32883b013 More properly keep track of register classes. 2016-09-29 22:32:43 +02:00
David Given b27758b7de Error check fragment rules which don't emit anything. 2016-09-29 22:14:11 +02:00
David Given 0d246c0d73 Much better handling of fragments (no run-time code needed to distinguish them
from registers) and better handling of individual hops within a paragraph ---
no more ghastly hacks to try and distinguish the input from the output.
2016-09-29 22:06:04 +02:00
David Given a0131fdb47 You know what, the type inference stuff is a complete red herring. What this
actually needs is a more intelligent register allocator. So, remove the type
inference.
2016-09-29 19:58:02 +02:00
David Given cc176e5183 Keep more data around about ir instructions. Implement a half-baked type
inference routine to propagate information about floats up the tree, so we know
whether to put floats into special registers as early as possible.
2016-09-26 22:12:46 +02:00
David Given 416b13fd76 Start factoring out the hardware op code. 2016-09-25 23:29:59 +02:00
David Given 39aa672422 Sort of keep track of registers and register classes. Start walking the
generated instruction tree --- holy cow, they look like instructions!
2016-09-25 22:17:14 +02:00
David Given bde5792b1a Collapse several rule arrays into one; actually generate the array properly. 2016-09-25 17:14:54 +02:00
David Given bcc74ba18d Stupid stringlist is stupid. Use a proper data structure, properly abstracted
out (so other things can use it).
2016-09-25 12:18:39 +02:00
David Given 9f78e0b36b Rethink the way patterns are mapped to rules; generate emitters (probably
badly).
2016-09-25 11:49:51 +02:00
David Given 7c028bdd45 We now record the code fragments to be emitted by each rule. 2016-09-25 00:21:46 +02:00
David Given 629e0ddfc6 Some instruction selection is now happening. 2016-09-24 22:46:08 +02:00
David Given c8fcbe282a More grammar changes. 2016-09-24 19:03:55 +02:00
David Given 2acc4ed29d IR codes are now owned by mcgg; ir terminals are inserted into the table during
compilation (so you can refer to them).
2016-09-24 18:31:35 +02:00
David Given 1516657907 Crudely bolt on mcgg to mcg itself. 2016-09-24 17:20:40 +02:00
David Given 13132128a1 Parameters are parsed with getopt. Simplify, constify. 2016-09-24 16:59:49 +02:00
David Given 434eafd35d Change the predicate stuff to use costs instead; now you can use when clauses
on leaves. Remove an iburg premature optimisation (required for above).
2016-09-24 13:33:59 +02:00
David Given d96ceea08a Lots of exploratory new grammar for instruction definitions and string and
fragment emission (none of which is hooked up to anything yet).
2016-09-24 13:08:17 +02:00
David Given 960259f0b0 Add support for labelled tree nodes. 2016-09-24 12:11:30 +02:00
David Given 4546dd5f22 Massive grammar overhaul and refactor. Hacked in support for predicates, where
instructions can be turned on and off based on their parameters. New lexer
using a lexer. Now quite a lot of the way towards being a real instruction
selector.
2016-09-21 00:43:10 +02:00
David Given 2183c6c622 Run through clang-format. 2016-09-20 21:00:16 +02:00
David Given 03b7202e54 Strip out surplus files. Rewrite README. 2016-09-20 20:46:45 +02:00
David Given 5cb3fbb3d3 Import iburg. 2016-09-20 20:44:51 +02:00
David Given 13c117d15d Import iburg. 2016-09-20 20:37:16 +02:00
George Koehler b1d1b5e1f8 Fix bugs with memory allocation in ego.
cf/cf_loop.c and share/put.c tried to read the next pointer in an
element of a linked list after freeing the element.  ud/ud_copy.c
tried to read beyond the end of the _defs_ array: it only has
_nrexpldefs_ elements, not _nrdefs_ elements.

These bugs caused core dumps on OpenBSD.  Its malloc() put _defs_ near
the end of a page, so reading beyond the end crossed into an unmapped
page.  Its free() wrote junk bytes and changed the next pointer to
0xdfdfdfdfdfdfdfdf.
2016-09-09 23:37:43 -04:00
David Given f67c98e239 Distributions are a pain --- let's not bother any more. Instead, we just tag
the repository and download a complete snapshot, old and ancient stuff and all.
2016-09-02 23:00:38 +02:00
David Given 612e38f1c6 Remove the old make-based build system, plus some big chunks of horribly
obsolete protomake build system.
2016-09-02 22:17:51 +02:00
David Given 2b6d251dec Fix a fun bug where, every now again, ego would get its temporary files mangled
and generate invalid calls to the optimisers.

Previously ego would generate a temporary file template that looked like
/tmp/ego.A.BB.XXXXXX, call mktemp() on it to randomise the XXXXXX, and then
replace A and BB with data.

However, it used strrchr to find the A and B. Which would fine, except when
mktemp produced an A or a B in the randomised part...

This code was written on 4 March 1991. I was 16.
2016-08-22 23:53:01 +02:00
David Given 2a95b1c5e3 Forgot to check a file in. 2016-08-22 22:45:32 +02:00
David Given 5bae29a00c ego now builds and is used.
This needed lots of refactoring to ego --- not all platforms have ego descr
files, and ego will just crash if you invoke it without one. I think originally
it was never intended that these platforms would be used at -O2 or above.

Plats now only specify the ego descr file if they have one.
2016-08-21 22:01:19 +02:00
David Given 84ee7c9cc4 Fix linking bug where em_decode would generate binary data instead of text. 2016-08-21 20:27:52 +02:00