Replying to
@joe@f.duriansoftware.com
Oh yes Pascal code too.
Pascal code look 100x better than rust code for some reason. Maybe it is the let syntax which just makes it so much harder to read rust code.
Remote
Andrea (Drea) Tamar Pinski
@pinskia@hachyderm.io
mastodon 4.7.3279 Followers
493 Following
50 Posts
Joined November 16, 2023
Well, trying to improve std::vector::push_back for gcc and llvm. (Disclaimer with gcc's libstc++, didn't look into libc++ code though).
Llvm change: https://github.com/llvm/llvm-project/pull/222727 (from my team).
Gcc change i will be posting tonight or tomorrow depending on me getting it done.
We have always been at war^w alience with
Today has been a day where I get no work done. I hate it.
RE: @Elizafox@social.treehouse.systems
My branch predictor is my local llm.
GCC patches posted:
https://gcc.gnu.org/pipermail/gcc-patches/2026-September/731050.html
https://gcc.gnu.org/pipermail/gcc-patches/2026-September/731051.html
The first one is exactly what the LLVM patch does. But the second one is needed for GCC code generation; I have not looked into why LLVM does not need a similar patch (a reduced testcase for LLVM fails in the same way as GCC did).
I looked into libc++ code generation with push_back and I see the slow path is NOT inlined which might be causing performance issues too. libstdc++'s push_back inlined for both GCC and LLVM. GCC was also to shrink wrap the code but the LLVM bug there is https://github.com/llvm/llvm-project/issues/123120 .
at first I thought it was a different patch which was causing my bug. But it turns out it was something I Just totally forgot about. Anyways I know understand what is going wrong. A RaW dependancy that I missed could happen.
Today reminds me that I am not a super human and that I have only so much time in the day.
I try to have a good balance between cleanups and improvements for gcc.
In fact most of my cleanups are directly related to the improvements I am doing
I find the code that uses visit style programming bad.
Sorry LLVM; I think it is badly written code and is a bad style of code in general.
Yes some folks are adding that style to GCC for the rust front-end.
But I think there are better ways of writing the code. Especially when there are things like statements that would share the same code.
Yes I would say C++ code popularized it and I think it is poorly thought out code.
Replying to
@brouhaha@mastodon.social @revk@toot.me.uk
>I am slightly curious as to whether the GCC implementation uses a display vector, or a static link in the stack frame.
It is a stack chain but it is "unnested" so the stack chain is really just a struct these days. For pre-GCC 4.0, it was a full stack link that was passed to the functions.
The "unnesting" is done in tree-nested.cc.
For nested function which are called indirectly information is below:
For some ABIs (e.g. powerpc64 ABIv1 [maybe ABIv2 too]), since there is already a fat pointer for function pointers (with an context pointer), it will use that.
For others, it uses the stack and for others still it will use statically allocated function pointers (aarch64 darwin) [this is swappable to compile time even].
Look prototyping something should be fun and rewarding. Having a computer do it for you is boring and not rewarding.
Yes experience can make the prototyping faster but that is just like doing an outline of a book or an outline for drawing. Both are types of prototyping. You lay down the infrastructure of a drawing or book first via these outlines.
You could claim that your english instructions to the computer is an outline (yes it is but for a different reason than you think). And that the computer takes that outline and makes something. But it didn't make anything but rather it just took the words/tokens and did a search of previous work and put that all together. It created a derivative work of many things all mashed together. It didn't create anything new. Unlike when you create something .
[RFC PATCH 0/3] wasm: New backend (for GCC)
https://inbox.sourceware.org/gcc-patches/20260505223045.347444-2-feedabl3@gmail.com/
Replying to
@ireneista@adhd.irenes.space @lcamtuf@infosec.exchange
"Netscape went bankrupt trying to re-write their software from scratch"
It is also why Microsoft Edge went from something written from scratch to be a fork of chromium. The story is the same and even more it is about the similar product. Plus it is a recent example of the whole starting from scratch issues.
Sent out an RFC for GCC to turn off `-ftrapping-math` by default for C and C++.
This will make GCC inline for clang/LLVM.
A few years ago I would have recommened doing the opposite that is an RFC for Clang to change its default but looking into GCC more; GCC really does not implement -ftrapping-math correctly and seemly never has.
Canonical ceo Mark is huge tech bro who peaked on high school as obvious by his questions about experience during high school.
So his company choosing a half finished utils is just peak white cis male who peaked in high school.
"Woman are supposed to submissive."
That is the thinking of why tech bros call their chatbot by women pronouns.
Plain and simple.
It is the same reason why ships and cars are woman names too.
Nothing new here.
It is also why transwomen are not women to them because transwomen correct men.
It is always me I am the top of the food chain be submissive to me.
The whole chucked, calling men girls, etc. Happens too.
It is all interconnected.
Replying to
@mgorny
On a slightly related subject I think many free/open source projects do have a LTS issue.
Even gcc which only does 1 major release a year. Gcc has a support cycle of 4 years. Is that long enough or does it really need to be longer.
Some folks complain gcc does not release often enough. While others complain gcc should support much older releases.
There are some bugs which are hard (not impossible) to fix in older releases of the cycle due to code infrastructure changes. I can provide a few examples which I fixed in the last year, backporting the fix would be hard. But I can also say those cases didnt show up in a bug report by an user but rather myself looking for a specific case at the time (not even found by a fuzzier). The others were due to a fuzzier (with some options which most folks dont use or some odd code which is possible in real code but I would doubt show up).
To do the backports would require to spend a few days to do them and if it was an obvious issue from code in the wild, I would spend that time.
Now if you look at llvm, there is what a few minor releases after the major one and then off to the major one.
Debugging a vrp miss optimization regression.
Lucky it is just a regression in gcc 16.
And only shows up in a small cases that I highly doubt it will affect real code.
Replying to
@CursedSilicon@social.restless.systems @thephd@pony.social
No but we (the GCC project) are testing a Forgejo based forge which can be accessed at:
https://forge.sourceware.org/gcc/gcc-TEST
Anyone can sign up for an account and then get access to have patches reviewed that way (patches are still sent to the email list though).
See https://gcc.gnu.org/wiki/UsingForge for more information.
Replying to
@notbobbytables@infosec.exchange
glibc side of things is discussed here:
https://inbox.sourceware.org/libc-alpha/PAWPR08MB8982DA2C405BDC8CE4DC66AE83F42@PAWPR08MB8982.eurprd08.prod.outlook.com/T/
gdb side of things is discussed here:
https://inbox.sourceware.org/gdb-patches/20260629153203.10142-1-muhammad.kamran@arm.com/
Replying to
@notbobbytables@infosec.exchange
So what happens is gdb breaks. Because it overloads malloc. It is a bug in gdb and a newer gdb will not be released for a while. So it was decided to revert the glibc patch for now. And wait till the next release cycle for the glibc patch.
The gdb fix went in as it does not break anything else too.
I would provide a link but sourceware.org is under a huge ddos attack due to Ai bots.
Replying to
@JamesWidman@mastodon.social
Oh p4 is another language which does not take into account non ascii identifiers into account. It was never even brought up as a problem.
The main reason why many of these languages do that is because they depend on flex rather than hand written lexers.
Replying to
@JamesWidman@mastodon.social
The weird thing i saw about p4 is the developers were like we like our bison/flex parser/lexer. As someone wrote their parser and lexer for p4 I was super confused why they like it. The lexing nor the parsing part was the big issue with p4 but rather the semantics were not always fully defined. Especially their type matching for function calls. They left it up for implemented defined there. Especially with generics being included. It was fun to implement and get similar semantics and even prove the semantics and implementations dont line up.
If architecture of a building an art then is so architecture of a program an art.
User-interface is art as is the design of a building is art.
Internals of say how floors are put together and pipes are also considered art. So code is art too.
Replying to
@SRAZKVT@tech.lgbt yes oop vistor pattern.
What I find in the code that does those kind of is missing part of the point of oop.
Most of the time the vistor code should belong with the class itself rather than having a seperate vistor class.
A big example of over use of the oop vistor code is the p4c compiler.
When I wrote my own P4 front-end.
I used oop/classes but things like cloning and "generics" replacement and folding was handled by each AST itself rather than a vistor class for each of those things.
Replying to
@thephd@pony.social @CursedSilicon@social.restless.systems
As I mentioned anyone can sign up and get an account. We are trying to encourage folks to use it especially if they rather use it than emailing patches.
We have a more open policy giving access to folks have their own repo hosted here than on the main repo.
The only thing we dont want is private repositories and code that is not related to projects hosted on sourceware (there might be exceptions but those are handled on case by case). But we want folks to have their own branches and such. And not worry about the "qualify for an account" idea either.
Replying to
@thesamesamhttps://sourceware.org/pipermail/binutils/2015-February/087626.html
So it was closer to 10 years rather than 15.
So gcc 6 or 7.
Replying to
Backporting just via git cherry-pick is very very bad.
Even yesterday in GCC there needed a fixup for backport (it applied cleanly via git cherry-pick too but produced wrong code) just on newly branched GCC 16. https://gcc.gnu.org/pipermail/gcc-patches/2026-May/715488.html :).
Do you want some toast?
How about an English muffin?
Replying to
@mirabilos@toot.mirbsd.org @joe@f.duriansoftware.com
The important check is:
Libcalls.getLibcallImpl(RTLIB::BZERO) == RTLIB::Unsupported
Replying to
@mirabilos@toot.mirbsd.org @joe@f.duriansoftware.comhttps://developer.arm.com/documentation/109697/2025_12/Feature-descriptions/The-Armv8-8-architecture-extension
"FEAT_MOPS, Standardization of memory operations"
Replying to
So I use both GUI editors and vi. For quick and simple (and most of the time small) stuff, it is faster to use vi (for me).
For an example most of my GCC testcases are almost all written via vi. Creating a new file is easier using vi rather than a GUI; especially copying the file into the correct location afterwards and having to edit it after testing it (make check-gcc RUNTESTFLAGS=...).
Larger changes and changes which I have to think about I use an editor with a GUI (notepad++).
I also use vi to read code in many cases more than a GUI editor but that is because edit vs command view is more natural for that. But that is also if I was searching for something quickly. Rather than something which I need to keep in view for a lot longer and then I use notepad++.
For me I use both for different purposes and for different lengths. I won't leave open a vi session for hours on end while I will leave open a few notepad++ sessions for days on end. (yes I use screen too).
Replying to
@giflian@techhub.social long covid?
Replying to
@slava@mathstodon.xyz @andnull@social.nouveau.community
Well metrowerks might had been acquired by apple rather than Freescale. Or did I get the time lines wrong.
Replying to
@slava@mathstodon.xyz @andnull@social.nouveau.community
No i get it right.
Be was not using gcc until post 1997/98.
Replying to
Of all things that get miscompiled while making changes, it is graphite. Seriously everything else in the gcc testsuite works.
Oh it works at stage1 so I know my patch is the cause.
Will debug this tomorrow.
I am 99% sure i am miscompiling ISL. But where and why only isl.
In this one section of push_back code, I have now found 4/5 different missed optimizations. For both gcc and llvm. The last one is a saturated add/multiply which is slightly more complex to handle though but should be doable.
Basically this code is trying to do a sat_mult(elements, 2*sizeof(element)) with a minimum value of sizeof(element).
The sat_mult part i figured out tonight.
I wonder how much this could improve the code.
Replying to
@malwareminigun@infosec.exchange
"Everyone has a different opinion " Again see the eposide of Silicon Valley.
Anyways there is an option for GCC to change the tab stop. But most don't know about it. -ftabstop=.
Replying to
@glyph@mastodon.social
It has to do with his blog entry on Bambu Lab. I suspect.
I thought this addition to genmatch was going to be much harder. But it turns out not to be so bad. Still need to add some error checking; into the generated code too.
Replying to
@malwareminigun@infosec.exchange @chandlerc@hachyderm.io
See the eposide of Silicon Valley. It was a mistake that this fight even happens.
Anyways tabs with spaces rules the world. A tab is always every 8th column; nothing else makes sense. And windows and modern editors are broken.
Replying to
@joe@f.duriansoftware.com @mirabilos@toot.mirbsd.org
I assume for aarch64 with the mops instruction it always inlines memset anyways so what does it matter there :).
/me goes off and hides since he only cares about targets with MOPs now.
Finally understood how to implement this piece of code and the old code too.
Plus it should give an infrastructure to implement another piece of code. That a coworker was implementing.
So this VRP regression is one of these; this is odd and you keep on thinking why is this code causing issues later on.
And then you remember there is a cache in place. And then you think can reset the cache after this code is done; not exactly so things get complex.
Basically time to hand this over to the person who designed this whole thing; maybe they can think of a decent way of fixing this.
Replying to
@SRAZKVT@tech.lgbt did swift use it? I have not looked into their generics yet. I know rust does.
I have not looked into zig either.