fix: enable linux shared-library-safe v8 tls mode by default

[mplemay]

Summary

Fixes Linux shared-library (cdylib) linking failures caused by V8 TLS relocations when rusty_v8 static archives are linked into downstream shared objects.

Closes #1706.

Root cause

On Linux, V8 uses fast TLS access unless V8_TLS_USED_IN_LIBRARY is defined.
Without that define, V8 code can emit relocations such as R_X86_64_TPOFF32, which are invalid when linking into -shared and fail downstream cdylib builds (for example PyO3 extension modules).

What changed

• build.rs:
  • Linux builds now inject:
    • extra_cflags=["-DV8_TLS_USED_IN_LIBRARY"]
  • Injection is skipped if user-provided GN_ARGS already contains V8_TLS_USED_IN_LIBRARY.
• README.md:
  • Documented the Linux default and the exact injected GN arg.
  • Documented that Linux prebuilt release archives are built with shared-library-compatible TLS mode.

Why this approach

This fixes the issue without requiring changes to vendored V8 GN logic and avoids enabling broader build-mode switches (for example v8_monolithic=true) just to get TLS library mode behavior.

Compatibility and risk

• Linux-only behavior change.
• No Rust API changes.
• No C++ binding surface changes.
• Expected tradeoff: slight TLS access overhead in exchange for reliable downstream shared-library linking.

Override behavior

If needed, users can still fully control TLS-related flags via GN_ARGS; build.rs only injects the define when it is not already present.

[CLAassistant]

All committers have signed the CLA.

[mplemay]

@fraidev or @bartlomieju - when you get a chance, could you take a look at this?

[kajukitli]

looks like the right default to me

linux cdylib consumers hitting TLS relocation failures is a real footgun, and setting V8_TLS_USED_IN_LIBRARY by default for linux source builds is the least ugly fix.

one thing i'd still call out: the GN_ARGS parsing here is pretty flimsy.

for arg in args.split_whitespace() {
  if arg.contains("V8_TLS_USED_IN_LIBRARY") {
    has_tls_library_mode_define = true;
  }
  gn_args.push(arg.to_string());
}

GN_ARGS is not actually whitespace-safe in the general case, so this can misparse quoted args and arrays. rusty_v8 already has this problem elsewhere in build.rs, so it's not a regression unique to this PR, but it's worth keeping in mind.

for this specific change though, i think the behavior is good and the override logic is reasonable.

[bartlomieju]

I'm gonna defer to @devsnek for final approval

[mplemay]

looks like the right default to me

linux cdylib consumers hitting TLS relocation failures is a real footgun, and setting V8_TLS_USED_IN_LIBRARY by default for linux source builds is the least ugly fix.

one thing i'd still call out: the GN_ARGS parsing here is pretty flimsy.

for arg in args.split_whitespace() {
  if arg.contains("V8_TLS_USED_IN_LIBRARY") {
    has_tls_library_mode_define = true;
  }
  gn_args.push(arg.to_string());
}

GN_ARGS is not actually whitespace-safe in the general case, so this can misparse quoted args and arrays. rusty_v8 already has this problem elsewhere in build.rs, so it's not a regression unique to this PR, but it's worth keeping in mind.

for this specific change though, i think the behavior is good and the override logic is reasonable.

Hey @kajukitli - thanks for the quick review. I went ahead and bolstered the logics so it does a better job with the edge cases. Let me know if there is anything else I should do to improve it.

@devsnek - let me know what you think. Happy to make any additional edits to get this shipped.