[ndossche]

Although the issue mentioned FreeBSD, this is a broader problem:
the current ARM64 code to load the TLS offset assumes a setup with
the non-default TLS model. This problem can also apply on some
configurations on other platforms.

[arnaud-lb]

I think the code to fetch the TCB offset is good, but unfortunately I don't think we can cache the TCB offset unless we are using the local-exec or initial-exec models, because this offset may be different in each thread (as the dynamic TLS block of the module is allocated separately in each thread).

We do it on some platforms, but I believe that are relying on surplus Static TLS Space in these platforms.

When loading apache mod_php on freebsd aarch64, our TLS block is allocated dynamically. I found about this because the assertion tsrm_ls_cache_tcb_offset <= LDR_STR_PIMM64 fails.

We should probably return 0 in the TSRM_TLS_MODEL_DEFAULT case, and update the JIT to cache only the TLS descriptor as we do in x86. (This code is broken currently on FreeBSD, but this should fix it: #13928. Additionally, on ARM the descriptor is 3 words.)

[ndossche]

Interesting, thanks for the review!
I guess I'll wait for your merge to adapt this. I'm currently busy with something else anyway...

[arnaud-lb]

I've merged #13928

[ndossche]

Thanks, I will be able to allocate time for this tomorrow to rebase and finish this PR.

[ndossche]

It took a little while longer because the software emulation slowed me down, but hopefully I did it right. At least it seems to work locally, but best double check.

[arnaud-lb]

LGTM!

[arnaud-lb]

We may mention the ABI here, as it specifies that where points to two GOT entries, and the second one holds the information we need. This information is platform-specific though:

The first entry holds a pointer to the variable's TLS descriptor resolver function and the second entry holds a platform-specific offset or pointer.

(https://github.com/ARM-software/abi-aa/blob/2a70c42d62e9c3eb5887fa50b71257f20daca6f9/aaelf64/aaelf64.rst)

[ndossche]

Thanks for the link & review, updated the comment.
Looks like for master we'll need changes in the IR repo. So I'm proposing the following:

• Merge this into 8.2 & 8.3, empty merge into master
• Get a PR to IR to support the codegen part
• Once IR is merged back into PHP master, create a PR that essentially recreates this PR on top of JIT-IR.

Sounds good?

[arnaud-lb]

Yes this sound like a good plan

[dstogov]

I can't check this. Feel free to merge this. (the IR part already in master).