[BOLT] Err when linking objects of different architectures

[mtvec]

This could happen, for example, when instrumenting an AArch64 binary on an x86 host because the instrumentation library is always built for the host.

Note that this check will probably need to be refined in the future as merely having the same architecture does not guarantee objects can be linked. For example, on RISC-V, the float ABI of all objects should match.

[aaupov]

Good catch, thanks for the added check. We should err but that doesn't address the rootcause. In principle, we should support something akin to LLVM_RUNTIME_TARGETS.

Can you please add a test with x86 host (REQUIRES: target=x86_64{{.*}}) instrumenting aarch64 binary?

[mtvec]

Good catch, thanks for the added check. We should err but that doesn't address the rootcause. In principle, we should support something akin to LLVM_RUNTIME_TARGETS.

Indeed, this would be nice. I have been working around this by manually building the runtime library for RISC-V in order to use it on an x86 host. It would be better to integrate this in the build system so I might try this out soon.

Can you please add a test with x86 host (REQUIRES: target=x86_64{{.*}}) instrumenting aarch64 binary?

Done! It was a bit tricky though to produce an AArch64 binary suitable for instrumentation on x86. I had to:

• Force a dummy relocation;
• Add a dummy _fini function to force DT_FINI to be emitted;
• Link with -pie (to avoid a static binary being created).

Please let me know if there's an easier way to accomplish this.

[yota9]

Yes, these are the minimum requirements. Usually I prefer to use -Wl,-fini flag though, but it doesn't really matter

[mtvec]

Yes, these are the minimum requirements. Usually I prefer to use -Wl,-fini flag though, but it doesn't really matter

Thanks!

As a side note: is AArch64 using DT_FINI_ARRAY instead of DT_FINI? I have a patch ready to let BOLT use the former if the latter is not available; I'll try to submit it today.

[yota9]

The BOLT uses DT_FINI currently only, the DT_FINI_ARRAY tag is not even read right now. If you mean at runtime - I think as for any other platforms all FINI_ARRAY and FINI is getting called in a sequence

[mtvec]

The BOLT uses DT_FINI currently only, the DT_FINI_ARRAY tag is not even read right now.

I was wondering if AArch64 doesn't use DT_FINI by default and you have to work around this in BOLT by forcing the linker to emit it through -fini or something similar. That's the case for RISC-V (it only use DT_FINI_ARRAY by default) so that's why I created the patch I talked about before. But maybe we should continue this discussion once I post that patch :)

[yota9]

AFAICT the symbol is came from libc and it does nothing but return. But since BOLT tests doesn't use libc the _fini symbol is absent and that is why we have to create it manually in our tests