[Orc][examples] Fix lljit-with-remote-debugging test failure

[RoboTux]

Test lljit-with-remote-debugging fails on AArch64 because the
argc_sub1_elf.ll input file have a x86_64 target triple. Since there is
no way to set the target triple when invoking LLJITWithRemoteDebugging
let's require x86_64 target to be the default target.

[lhames]

@RoboTux @weliveindetail I don't think that argc_sub1_elf.ll contains anything x86-64 specific -- what if we just removed the triple instead?

[weliveindetail]

what if we just removed the triple instead?

Yes, we could try that. Otherwise, running the test only on x86_64 would be fine too.

[RoboTux]

@RoboTux @weliveindetail I don't think that argc_sub1_elf.ll contains anything x86-64 specific -- what if we just removed the triple instead?

I tried but then it complains that there is no target "". Perhaps it's worth fixing orc to support an empty target.

[RoboTux]

Dropping since #74831 has been merged.

[lhames]

@RoboTux @weliveindetail I don't think that argc_sub1_elf.ll contains anything x86-64 specific -- what if we just removed the triple instead?

I tried but then it complains that there is no target "". Perhaps it's worth fixing orc to support an empty target.

I missed this earlier -- that's a good idea. LLJIT seems like the right place to handle this: addIRModule can check for an empty triple and replace it with the triple reported through the ExecutionSession.

[RoboTux]

I missed this earlier -- that's a good idea. LLJIT seems like the right place to handle this: addIRModule can check for an empty triple and replace it with the triple reported through the ExecutionSession.

That's been done by Stefan in d86a937.

[lhames]

I missed this earlier -- that's a good idea. LLJIT seems like the right place to handle this: addIRModule can check for an empty triple and replace it with the triple reported through the ExecutionSession.

That's been done by Stefan in d86a937.

Stefan's change was to the testcase. I was proposing to change LLJIT itself, though now that I've taken a look I can see that it wouldn't have made a difference here: The testcase triple was being used to construct the target (via JITTargetMachineBuilder) and that's why the empty triple was causing a "no target" error. Normally the target is constructed using the ExecutionSession's target triple, and LLVM optimizations and CodeGen will happily process a module with an empty triple.

[RoboTux]

Stefan's change was to the testcase. I was proposing to change LLJIT itself, though now that I've taken a look I can see that it wouldn't have made a difference here: The testcase triple was being used to construct the target (via JITTargetMachineBuilder) and that's why the empty triple was causing a "no target" error. Normally the target is constructed using the ExecutionSession's target triple, and LLVM optimizations and CodeGen will happily process a module with an empty triple.

Oh right, I didn't realize LLJITWithRemoteDebugging is a test only binary. My bad.

[weliveindetail]

I started drafting a PR with some maintenance on implementation and tests for advanced LLJIT examples here: #76236