LOH conflicting with MachineOutliner

Code Generation AArch64
12

I’ve identified a problem with running LOH analysis after the MachineOutliner. Since outlined function do not follow calling conversions, we can’t even be sure temporary registers don’t live across function calls.

In this example, the MachineOutliner creates the the thunk _OUTLINED_FUNCTION_0 and immediately moves x16 to the first argument of bar or goo.

_OUTLINED_FUNCTION_0:
        adrp    x19, _A@PAGE
        add     x19, x19, _A@PAGEOFF
        ldr     w20, [x19]
        b       _foo
...
        bl      _OUTLINED_FUNCTION_0
        mov     x0, x19
        bl      _bar

If we run LOH now, it will try to add an .loh AdrpAddLdr directive for the adrp, add, ldr triplet, which could turn into this in the linker.

        nop
        nop
        ldr     w20, [x19, <_A offset>]

This leaves x16 undefined.

I don’t think it’s reasonable to expect LOH to analyze register liveness across functions, so it’s probably best to keep LOH before the MachineOutliner. Instead, I think we should teach the MachineOutliner how to handle these LOH directives, instead of allowing the to block outlining. Either by removing LOH directives for outlined functions, or maybe analyzing whether they are valid and fixing them up if possible.