Some sanitizers should be target modifiers

[Darksonn]

Some sanitizers need to be target modifiers, and some do not. For now, we should mark all sanitizers as target modifiers except for these: AddressSanitizer, LeakSanitizer

For kCFI, the helper flag -Zsanitizer-cfi-normalize-integers should also be a target modifier. Same applies to #138311 when it lands.

The shadow-call-stack sanitizer is currently being considered as a target modifier for simplicity. However, there are cases where it does not need to be a target modifier.

Please see the target modifiers tracking issue for more details. See also the Tracking Issue for stabilizing the sanitizers.

cc @rcvalle @azhogin @maurer
@rustbot label F-target_modifiers A-sanitizers

[tmiasko]

MemorySanitizer and ThreadSanitizer should be considered target modifiers as well.

[Darksonn]

Ok, thanks. Updated.

[azhogin]

Also, from zulip thread:

they should be target modifiers iff they lead to unsoundness when different values of the flag are used for different parts of what eventually becomes a single binary

18:20
I think -Csoft-float should just be removed, though if someone cares about the feature it could be kept as a target modifier (but the implementation would need to be changed for that, currently rustc assumes it can just look at the target spec to determine whether the current arm target uses hardfloat or softfloat ABI)

18:21
for -Ctarget-feature=-neon, we'll have to decide what to do with target features currently marked as "forbidden": do we still allow overwriting them with a target modifier, and if so how?

18:22
I'm open to the idea of allowing that, though it would need some careful design around what to do for nonsense combination where there just isn't a well-defined ABI. maybe that should be a separate -C flag though.

18:25
-neon on an aarch64 hardfloat target is a "nonsense ABI" case so it's unclear what we should even do when allowing the flag, I think it should just be a hard error

18:25
the more interesting question is atomics-32 on arm and forced-atomics on riscv; those probably should become target modifiers

18:27
long-term we should probably treat all target features that we do not know as target modifiers

[azhogin]

Draft PR: #138736. Has problems with tests using sanitizers inconsistent with std deps.

[tmiasko]

Using -C unsafe-allow-abi-mismatch in existing tests should be fine. Ideally tests would rebuild the standard library, but it is not clear to me whether it is worth it given the complexity and compile time cost.

[azhogin]

-C unsafe-allow-abi-mismatch

Yes, fixed.