Fixed 5.22 problems during my compiler port
I uncovered and fixed many 5.22 problems with cperl already, but in the last months I was busy to port the 3 compilers B::C, B::CC and B::Bytecode to 5.22.
As I said in my interview it's my belief that if all current p5p core committers would stop committing their bad code it would be actually be the best for the perl5 project. They weren't able to implemented any of the already properly designed features from perl6 in the last 12 years, and every feature they did implement is just so horrifibly bad, making our already bad code base, which led to reimplementation efforts of perl6/parrot with a better core, even worse. With cperl I can only undo a little, but when they start breaking the API and planned features in an incompatible way they should just stop.
Nevertheless, 5.22 added a significant improvement from outside, syber's monomorphic inline caching for method calls besides the internal improvement of multideref by Dave Mitchell.
Now to the problems I had to fix in the last months with that 5.22.0 release:
1. Father broke ByteLoader
This is something I cannot fix in the compiler. I updated my perl patcher
App::perlall with new --patches=Compiler patches to fix this, and cperl of course
also has this fix.
I had to write a complicated
for ByteLoader to check if the used perl5.22 version is already
patched or not. Probing a to-be-built XS submodule is not that easy. A
typical chicken and egg problem. I could use my already existing
B::C::Flags helper config, which allows custom compiler settings.
There I initialize the variable
undef, and when the XS modules are all built I call a helper script to
probe for a working ByteLoader, and patch
$B::C::Flags::have_byteloader to 0 or 1. I can use this then in the
tests to skip or run the bytecode tests. And I had to put this helper
script into the hints directory to skip it from being
installed. Messing with EUMM libscan() was too dirty for me.
The internal compiler op.c creates a new main or eval environment with
newPROG(), setting the entry points
from the intermediate parsed
PL_compcv. In the case of en empty source
the parser always adds a final
; semicolon, which leads to an empty optree
But with commit
for [perl #77452] the compiler now always adds a LINESEQ in front of
the STUB, but the logic in newPROG for source filters which already
PL_main_root wasnot changed, which led to
a broken ByteLoader.
This is an interesting commit as it added a lot of wrong comments about the inner working of this, but didn't update the logic.
I can only guess that p5p blocked me again, because they didn't like me to call them incompetent. Blocking bug reports and fixes is worse than just incompetence, but I got used to that recently. They blocked my simple fix for the horrific double-readonly system, and they proudly announced last week some new optimization regarding faster arithmetic, but didn't have a look into my fast arithmetic optimizations which I wrote half a year ago, and which makes them look very bad in the end. Everybody applauded poor Dave for this "fantastic breakthrough". The guys are really that simple. Looking through my improvements would have wasted less time and would have improved it upstream by 30% not just 10%.
2. Dave couldn't implemented multideref access for the compiler
Multideref merges sequential hash or array access into one compressed op. This is a pretty good compiler optimization, if the B design would not be so bad.
The upstream design of the new 5.22
deviates significantly from proper B design. aux_list requires the
curcv to be provided, which is not trivial to do for a B module, and
it needs this to resolve shared SVs beforehand. Requiring the curcv
to resolve the padoffset is unneeded and does not help B and any of
its clients. Clients need the padoffset and resolving it e.g. in
B::Deparse is to be done in B as with all other threaded and shared SV
Thanksfully I can patch most of B bugs by myself, and don't have to fork it publicly into a worse name. B is already a good enough name, and I don't want to deviate from that, even if p5p consistently refused to maintain B properly in the last years. There was some short time a few years ago where I could work without a patched B, but this period only lasted very shortly, and none of my fixes were applied, while other new horrific mistakes made it in.
3. Missing HV::ENAMES api
Stashes can be aliased to seperate namespaces, and the ENAMES API to access this names never made it into B, and thus never into a compiler. Namespaces aliases are rather seldom, so it caused not too much trouble, but now I added ENAMES and could hereby fix most of the remaining compiler limitations, even for 5.14.
4. Missing PADNAME B api
I explained that technically in my interview. Currently we limit the max name length of lexical variables to 60, because we statically allocate the buffers for them. It is not a practical problem, and I'll optimize that sooner or later to smaller static structs.
5. Fixed HEK assertions
The cperl code to support static HEKs is already there, but I still need to add compiler code and probes to support that.
6. Broken B::RV->FLAGS for GVOP_gv -> CVREF
5.22 has a wrong RV->FLAGS for a GVOP_gv pointing to a CVREF gv(cv ref:). It returns the flags for a GV (0x808009) where it should be just 0x801, a ROK RV. This is suddenly broken in 5.22 because it's a new optimization they did, and of course wrong.
I haven't fixed that yet in cperl, I just a workaround in the compiler with this patch
Overall we are very happy with the new 5.22 compiler, though we are not yet using the much more advanced cperl optimizations. The B::C optimizations alone lead to ~20% less memory, with cperl and its compiled readonly hashes for Config and warnings and its upcoming support for static GV/AV/CV/PAD/HEK layout it's much more dramatic. This will be a real COW (copy-on-write) mechanism then, being able to statically allocate readonly buffers, and copy it to the heap, when it's being changed. For the compiler we only need to ensure that static buffers are not freed, which is trivial with the added flag.
-m support for perlcc, compiling to modules, not single binaries is also improving. This can split various optimizations per module/.pm file, so we can use B::CC compiled modules or even rperl compiled modules, compile-times should go down from 20min to ~5min, with much faster smoker feedbacks, and pushing updates live is much faster, because they will be much smaller. The old compile times were 2 hours.
But since fixing B::C for 5.22 needed so much more time than expected I couldn't add most of the planned cperl optimizations for the upcoming cperl-5.22.2 release and B-C-1.53 release.