Shrinking While Linking(tweag.io)
tweag.io
Shrinking While Linking
https://www.tweag.io/blog/2025-11-27-shrinking-static-libs/
32 comments
I'd be very surprised if the C/C++ code being written today was comparable in volume to the collective pile of JavaScript, Java, C#, Go, Python, and PHP.
Contrast to the heyday of C and C++ when not much else got a look in (Pascal perhaps? Perl probably...)
I think it's fair.
Contrast to the heyday of C and C++ when not much else got a look in (Pascal perhaps? Perl probably...)
I think it's fair.
Yes, surely much more code is being written in other languages than in C, but probably more C is being written now than ever before, too.
Sure, but I think that "back in the day your compiler would ..." is a reasonable characterisation.
Back in the day, yes, your compiler almost certainly would do that. Now, your compiler might possibly do that. For most working stiffs it won't though.
Ok, my side of the shed's painted now :D
Back in the day, yes, your compiler almost certainly would do that. Now, your compiler might possibly do that. For most working stiffs it won't though.
Ok, my side of the shed's painted now :D
I agree.
hmm how can I reuse this useful Go library in python... Oh I can't.. hmm and how can I reuse this useful java library in php ? Oh I can't. Oh and which of the programming languages you mentioned can and do use C libraries? All of them.
Reminds me of that coworker who thought that OpenCV was basically written in python.
Reminds me of that coworker who thought that OpenCV was basically written in python.
I'm not claiming that there is no C or C++ out there. But it's such a nit pick when for most developers, no, their day-to-day work absolutely does not involve the creation of object files.
Sure, akshuwally, there are still C and C++ devs out there. Meanwhile a friend has just embarked upon a career as a pro COBOL developer. What of it?
Edit: Also, in the spirit of akshewally, I have just googled up this monster! My word, PHP and Java AND XML... it's like the unholy trinity of HackerNewsbane... https://php-java-bridge.sourceforge.net/pjb/
Sure, akshuwally, there are still C and C++ devs out there. Meanwhile a friend has just embarked upon a career as a pro COBOL developer. What of it?
Edit: Also, in the spirit of akshewally, I have just googled up this monster! My word, PHP and Java AND XML... it's like the unholy trinity of HackerNewsbane... https://php-java-bridge.sourceforge.net/pjb/
Looks genuinely useful. Bookmarking.
NativeAOT compiled C# library can create C compatible exports which can be used with any language that supports C libraries.
C++ is still wildly popular and will remain so for the foreseeable future. Your JavaScript interpreter/JIT compiler is written in it, so is your Java JVM, and I don't need to mention CPython in more detail.
Those all get compiled into object files and then linked.
Those all get compiled into object files and then linked.
But not by most devs. Back in the day it was true for most devs.
"If you compiled your code with support for link-time optimization (LTO), it contains another copy (in the form of LLVM bitcode — more on that later) of all our code and the code of all our dependencies."
The author then goes and removes the lto-related code via objcopy. Which is fine, and great, and useful if you care about size more than having link-time optimization enabled for your end users.
But if you are just going to turn around and remove it, you probably shouldn't include it in the first place, which makes the original analysis of 132MB vs 15MB a little misleading. It should be more like 50MB vs 15.
The author then goes and removes the lto-related code via objcopy. Which is fine, and great, and useful if you care about size more than having link-time optimization enabled for your end users.
But if you are just going to turn around and remove it, you probably shouldn't include it in the first place, which makes the original analysis of 132MB vs 15MB a little misleading. It should be more like 50MB vs 15.
These are extremely practical tips on using binutils to shrink your libraries. It might also be worthwhile to compile the library without per-function sections in the first place, which hopefully can be done without patching rustc. On many platforms this produces significantly smaller code in GCC.
I had a quick look and couldn't find a way to turn off per-function sections in rustc. But I think it's a pretty good default for rust, because the unit of compilation is pretty large and so it's common to have a lot of unused functions. It's really only a problem for distributing static libraries, since binaries and shared libraries already lose their per-function sections.
I would rather say it's really only potentially beneficial for distributing static libraries, since binaries and shared libraries already lose their per-function sections. So they pay the cost of the fluffier code without getting much benefit, if any.
> A static library is nothing but a bundle of object files, wrapped in an ancient and never-quite-standardized archive format.
To this day I'm astonished that it's not just tar. Or pax. Or even cpio! Or literally any file format that has any other use.
To this day I'm astonished that it's not just tar. Or pax. Or even cpio! Or literally any file format that has any other use.
Wasn't there first 'ar' as a general purpose archiver and later 'tar', "tape archiver"?
Why do you find it surprising that the archive format from that time was used to archive a bunch of files?
I wasn't alive but I'm pretty sure ar wasn't only used for this purpose in unix.
Why do you find it surprising that the archive format from that time was used to archive a bunch of files?
I wasn't alive but I'm pretty sure ar wasn't only used for this purpose in unix.
> Why do you find it surprising that the archive format from that time was used to archive a bunch of files?
It's surprising because we still use it today, not because it was used at the time.
It's surprising because we still use it today, not because it was used at the time.
Why would we break backwards compatibility, when it works just fine?
This response answers too much--why did we invent and adopt tar for all other use cases, then? Why do we not still use ar for its original general archiving purpose in any other situation but this one? Obviously there is some reason! I'm curious what that reason is. You can make regular archives with ar today but we don't--"it still works" wasn't good enough reason for any other archive use case. What are the general archiving requirements that don't exist for object archives that meant ar is still fine for this one use case?
Because we care about metadata and robustness for data files, but don't for some automatically generated artifacts that are consumed by the next program anyways? What do you need your static archives to have, that ar doesn't provide?
Because e.g. a.out didn't survive and was replaced? Several times, even?
That is a good point. Although I can only guess, it intuitively makes sense to me why the executable format would more quickly evolve than the archive format which "just works" still.
tar/pax are kind of terrible formats. They are hard to implement correctly. I'm glad they are not used more often.
cpio is pretty reasonable though.
zip is actually pretty great and I've been growing increasingly fond of it over the years.
cpio is pretty reasonable though.
zip is actually pretty great and I've been growing increasingly fond of it over the years.
The thing is, there is always tar(1) even in the most basic of distributions. And everyone uses tar.gz's or .bz2's or whatever for distributing all kinds of things, so tar is pretty ubiquitous. But the moment you want to do some C development, or any binutils-related, nope, install and use ar(1) which is used for literally one single purpose and nothing else. Because reasons.
Im not sure how ar does it, but tar has no centralised directory. The only way to get file 100 is to walk trough the 99 files before. This kills random access speed.
Ar puts a file called "/" as the first file of the archive. Inside, there is a number N, then a list of N file offsets, and then a list of N null-terminated strings. It's a symbol table of sorts: each null-terminated string is a symbol name, and the corresponding file offset points at the archive header for the object file that contains the symbol. The filenames themselves are not recorded centrally since it's not really needed.
> tar/pax are kind of terrible formats. They are hard to implement correctly. I'm glad they are not used more often.
I'll grant you "kind of terrible", but what's hard to correctly implement about tar? It's just a bunch of files concatenated together with a tiny chunk of metadata stuck on the front of each.
I'll grant you "kind of terrible", but what's hard to correctly implement about tar? It's just a bunch of files concatenated together with a tiny chunk of metadata stuck on the front of each.
Having never done it myself, I don't know, but I do know that the "microtar" library I picked up off GitHub is buggy when expanding GNU Tar archives but perfect when expanding its own archives. Correctly creating one valid archive is a lot easier than reliably extracting all valid archives. The code appeared competent, I assume tar just has a bunch of historical baggage that you can get wrong or fail to implement.
The real problem are generics. They cause a blowup in size because they are instantiated at every type combination that is actually called. This can sometimes be a performance boost (typically not because it avoids an indirect call, but because it enables optimizations such as inlining). But it can also makes code slower (not by much), or have little effect on performance
Rust lets you choose between generics and trait objects, but this is a viral change that sometimes means that large sections of code must be rewritten. There is also an optimization that turns generics into virtual dispatch if deemed beneficial, but I'm not sure how well it works
Rust lets you choose between generics and trait objects, but this is a viral change that sometimes means that large sections of code must be rewritten. There is also an optimization that turns generics into virtual dispatch if deemed beneficial, but I'm not sure how well it works
Lots and lots of code is still C and C++. That's not really "back in the day".