Many tests fail with musl libc about mold

[rui314]

@Logarithmus They failed because a stack frame was not properly aligned to a 16-byte boundary. The x86-64 psABI requires that when you make a function call (i.e. when you execute call instruction), %rsp must be aligned to a 16-byte boundary.

call instruction pushes a 8-byte return address to the stack. So, at the beginning of a function, %rsp is not aligned to a 16-byte boundary, so you need to adjust %rsp (e.g. pushing %rsp by 8 bytes) before making another call. I didn't do that.

If you do not do that, some instructions sensitive to alignment requirement may fail. glibc's printf is not sensitive to the alignment requirement unless you try to print out floating-point values (if you try to do that, it segfaults). It looks like musl's printf is always sensitive. So the tests failed.

[rui314]

No, it's odd. Is there any way to download the object files from the CI or run the same thing on my local environment?

[Logarithmus]

@rui314 https://hub.docker.com/r/voidlinux/voidlinux-musl

[rui314]

Thanks. I'll give it a try.

[rui314]

A noob question, but the environment built by docker pull voidlinux/voidlinux-musl doesn't contain any developer tools. How can I install a compiler, make and so on into it? I'm not familiar with Void Linux's package manager.

[Logarithmus]

@rui314 #67 (comment)

[rui314]

This test assumed that the default output is position-dependent, which actually depends on how your compiler was built. I made it explicit by adding -no-pie.

[Logarithmus]

@rui314 what about a new tag?

[rui314]

Do you mean a new version of the linker, like 0.9.3?

[Logarithmus]

@rui314 yeah, you know I need a new release to use this fix in a package

[rui314]

OK, let me create 0.9.3-pre1 so that you can test it.

[rui314]

Please test with this tag. https://github.com/rui314/mold/tree/v0.9.3-rc1

[rui314]

I created a new release candidate https://github.com/rui314/mold/tree/v0.9.3-rc2
It includes a fix of --sysroot flag handling.

[Logarithmus]

@rui314 see #89 (comment)

[rui314]

I somehow thought that test/canonical-plt.sh is the only failing test. I'll take a look at other failures.

[Logarithmus]

@rui314 I think some of them can't run successfully due to musl lacking some features of glibc e. g ifunc. So it would be nice to detect system libc somewhere in test/Makefile and conditionally skip those tests. Though some tests can be fixed I guess.

[Logarithmus]

@rui314 this thread might be helpful: https://news.ycombinator.com/item?id=22680577

[rui314]

@Logarithmus Is there any way to install i386 dev tools to musl Void Linux? Or is it 64-bit only?

[rui314]

Except i386 tests, I believe I fixed all musl-related failures in the main branch. Since cherrypicking these changes will be more riskier than making a full release, 0.9.3 will be a full release instead of bug-fix-only release.

[rui314]

I think mold at the head of the git main branch is now cleanly built and tested on Void Linux with musl.

[rui314]

@Logarithmus I tagged v0.9.3-rc3 as a new release candidate.

[Logarithmus]

@Logarithmus Is there any way to install i386 dev tools to musl Void Linux? Or is it 64-bit only?

Unfortunately, musl doesn't support multilib (running 32-bit programs with 64-bit libc). Also Void Linux doesn't have i686-musl variant, only glibc one. I don't know why.

[Logarithmus]

@Logarithmus I tagged v0.9.3-rc3 as a new release candidate.

OK thanks, I'll check it out!

[Logarithmus]

@rui314 do you know why those pieces of assembly used to fail with musl?280892f#diff-5725ceb876566e601501d9b70812c960b4fd5efaf2e8919e73a41c436fc9c308
Maybe it has to do with @PLT stuff? What does call printf@PLT mean by the way?

[rui314]

I'm closing this bug as the bug has been fixed.