Larger binaries produced by zig-linux-x86_64-0.14.0-dev.2414+ba37a4369

[fwsGonzo]

Zig Version

0.14.0-dev.2414+ba37a4369

Steps to Reproduce and Observed Behavior

It looks like a change has made binaries much bigger from one nightly to the next.

zig-linux-x86_64-0.14.0-dev.2384+cbc05e0b1

$ ls -lah bin
total 4,5M
drwxrwxr-x 2 gonzo gonzo 4,0K des.   8 18:50 .
drwxrwxr-x 7 gonzo gonzo 4,0K des.   8 18:50 ..
-rwxrwxr-x 1 gonzo gonzo 589K des.   8 18:49 asm
-rwxrwxr-x 1 gonzo gonzo 615K des.   8 18:50 cjit
-rwxrwxr-x 1 gonzo gonzo 231K des.   8 18:49 hello_world
-rwxrwxr-x 1 gonzo gonzo 762K des.   8 18:49 luajit
-rwxrwxr-x 1 gonzo gonzo 875K des.   8 18:49 mirjit
-rwxrwxr-x 1 gonzo gonzo 1,4M des.   8 18:50 robust_weight_transfer

zig-linux-x86_64-0.14.0-dev.2414+ba37a4369

$ ls -lah bin
total 13M
drwxrwxr-x 2 gonzo gonzo 4,0K des.   9 07:36 .
drwxrwxr-x 7 gonzo gonzo 4,0K des.   9 07:36 ..
-rwxrwxr-x 1 gonzo gonzo 739K des.   9 07:36 asm
-rwxrwxr-x 1 gonzo gonzo 1,5M des.   9 07:36 cjit
-rwxrwxr-x 1 gonzo gonzo 398K des.   9 07:36 hello_world
-rwxrwxr-x 1 gonzo gonzo 929K des.   9 07:36 luajit
-rwxrwxr-x 1 gonzo gonzo 1,0M des.   9 07:36 mirjit
-rwxrwxr-x 1 gonzo gonzo 8,5M des.   9 07:36 robust_weight_transfer

I haven't changed anything in my build systems, other than updating Zig. Could this be related to #22167 somehow?

The build is riscv64-linux-musl with no -march or -mabi. Plain Release CMake build.

Expected Behavior

Same incredibly small build sizes as before.

[Rexicon226]

There are no steps to reproduce your bug report, just the observed behaviour. What's the easiest way you've found to reproduce this size increase?

[fwsGonzo]

After some testing I can indeed reduce it down to something very concrete. Just a stroke of luck, really:

The empty atomic library in the CMake build was all that was needed. I have the small binaries again. A bit confused why this works.

add_library(atomic STATIC "${CMAKE_SOURCE_DIR}/ext/atomic.c")
...
$ file ../ext/atomic.c 
../ext/atomic.c: empty

There are still a lot of atomic operations in the binaries:

$ riscv64-linux-gnu-objdump -drl bin/robust_weight_transfer | grep amo | wc -l
406

The sizes are small again with this change and the new Zig nightly, just like shown in the initial post.
Not sure what to make of this. 🤷‍♂️

[Rexicon226]

How interesting 🤔. This seems to be because of the PR you mentioned.
cc @alexrp

[fwsGonzo]

Right, but I don't think that PR broke anything anymore. I just think I discovered a way to reduce the sizes. The programs all seem to be working fine. Applying -flto I could reduce them even further. It's really quite something.

But I don't know what the way forward is with this. It's strange that pulling in empty atomic reduced a single program from 8.5MB down to 1.4MB. I can reproduce this without a toolchain file both locally and in CI. All I need is an empty atomic.c.

If you want to inspect the binaries they are here now: https://github.com/libriscv/godot-sandbox-programs/releases/tag/v0.5

[alexrp]

Can you show the compile and link command lines that CMake generates, both before and after the issue started happening?