Don't override target passed on command-line with target from sysimg for processor_fallback

[vchuravy]

Until #26252 is fixed mark the --cpu-target tests as broken.

Don't override target passed on command-line with target from sysimg for processor_fallback

[yuyichao]

This check is done by llvm not us. Is jl_get_llvm_target not setting the JL_TARGET_UNKNOWN_NAME flag? Or is MSTI->isCPUStringValid not working?

[vchuravy]

jl_get_llvm_target uses the zeroth entry in jit_targets (

julia/src/processor_fallback.cpp

Line 110 in b376982

flags = jit_targets[0].en.flags;

), which is initialized by sysimg_init_cb

julia/src/processor_fallback.cpp

Lines 50 to 51 in b376982

	target = sysimg[best_idx];
	jit_targets.push_back(std::move(target));

The cmdline target doesn't match -C invalid doesn't match any of the sysimg targets and thus we use the zeroth sysimg target,
which is why the code in jl_init_llvm then sees as a valid target.

[yuyichao]

Ah, ok. Yeah, the unconditional target = sysimg[best_idx]; there is a bit wrong. Right now the cmdline flag is mostly ignored.

The system image check should only be used to find the index for system image initialization but not to change the compilation target unless there is ABI issue, e.g. register size. Since we know nothing about the ABI we can just assume there's no such thing. The fix should be simply removing that line.

[yuyichao]

As a bonus, if we found the CPU name in the system image we can probably remove the unknown_name flag. We can assume whoever compiled the system image did so successfully and know what they are doing .....