ccall: Add probecall for user-defined probe points

[jpsamaroo]

Adds a new probecall calling convention to ccall, which, when compiled, generates a new USDT-compatible probe slot (and associated semaphore), which, when executed via ccall, calls the loaded probe with the provided arguments.

Todo:

• Register added probes in runtime (@pchintalapudi @vchuravy)
• Add runtime functions for querying and controlling probes
• Emit USDT notes
• Fix semaphore semantics
• Save probe names during precompilation, restore probe data during init
• Validate probes work with bpftrace/BPFnative.jl
• Switch from ccall special form to builtin
• Re-introduce probes lost in port partr multiq to julia #44653 (will fix USDT probes are missing on master #44746)

[vchuravy]

Suggested change

ALIAS(probecall, probecall) \

[jpsamaroo]

I think we still need this to define Base.probecall?

[vchuravy]

but we don't need to define it? It's ccall("name", probecall). Base.llvmcall !== ccall("name", llvmcall)

[vtjnash]

Making this a new special function form hijacking onto ccall seems a terrible idea 😬

[vchuravy]

But turning it into a call ABI should be fine right? The reason is that for BpF we need to make sure the args are in the right registers

[vtjnash]

Making it a builtin function would likely be better. Regardless, need to figure out what this does in the interpreter also

[jpsamaroo]

Makes sense to me; I'll get it all properly working with ccall first, and then switch to a builtin.

For the interpreter, we can always dynamically lookup the probe and semaphore values from the runtime, so that should be pretty straightforward.

[vchuravy]

On the profiling call today @gbaraldi and I were discussing that on could potentially use LLVM patchpoints

declare void @llvm.experimental.patchpoint.void(i64, i32, ptr, i32, ...)

define i32 @square(i32) local_unnamed_addr #0 {
    %2 = mul nsw i32 %0, %0
    call anyregcc void (i64, i32, i8*, i32, ...) @llvm.experimental.patchpoint.void(i64 78, i32 16, ptr null, i32 0)
    ret i32 %2
}

This will generate an elf section of _LLVM.Stackmaps that we would need to translate into staps format. Or we add an intrinsic in that style to LLVM.