AArch64: support closure cfunctions (missing LLVM support for trampolines)

[vchuravy]

The good news is that the AArch64 builds are back, the bad new is that the test are failing.
From https://build.julialang.org/#/builders/86/builds/206/steps/2/logs/stdio

      From worker 7:	LLVM ERROR: Cannot select: 0x3b0167f8: ch = init_trampoline 0x30660570, 0x34ac35e0, 0x3b2fd438, 0x3af9f838, SrcValue:ch<0x3b333810>, SrcValue:ch<0x3b194f08>
      From worker 7:	  0x34ac35e0: i64,ch = CopyFromReg 0x30660570, Register:i64 %0
      From worker 7:	    0x3b2fd980: i64 = Register %0
      From worker 7:	  0x3b2fd438: i64 = AArch64ISD::WrapperLarge TargetGlobalAddress:i64<%jl_value_t addrspace(10)* (%jl_value_t addrspace(10)**, i8*)* @jlcapi_cfunction_38022> 0 [TF=3], TargetGlobalAddress:i64<%jl_value_t addrspace(10)* (%jl_value_t addrspace(10)**, i8*)* @jlcapi_cfunction_38022> 0 [TF=36], TargetGlobalAddress:i64<%jl_value_t addrspace(10)* (%jl_value_t addrspace(10)**, i8*)* @jlcapi_cfunction_38022> 0 [TF=37], TargetGlobalAddress:i64<%jl_value_t addrspace(10)* (%jl_value_t addrspace(10)**, i8*)* @jlcapi_cfunction_38022> 0 [TF=38]
      From worker 7:	    0x3b0187b0: i64 = TargetGlobalAddress<%jl_value_t addrspace(10)* (%jl_value_t addrspace(10)**, i8*)* @jlcapi_cfunction_38022> 0 [TF=3]
      From worker 7:	    0x34b21630: i64 = TargetGlobalAddress<%jl_value_t addrspace(10)* (%jl_value_t addrspace(10)**, i8*)* @jlcapi_cfunction_38022> 0 [TF=36]
      From worker 7:	    0x3b2fd7e0: i64 = TargetGlobalAddress<%jl_value_t addrspace(10)* (%jl_value_t addrspace(10)**, i8*)* @jlcapi_cfunction_38022> 0 [TF=37]
      From worker 7:	    0x3b016e78: i64 = TargetGlobalAddress<%jl_value_t addrspace(10)* (%jl_value_t addrspace(10)**, i8*)* @jlcapi_cfunction_38022> 0 [TF=38]
      From worker 7:	  0x3af9f838: i64,ch = CopyFromReg 0x30660570, Register:i64 %1
      From worker 7:	    0x3b018678: i64 = Register %1
      From worker 7:	In function: jlcapi_cfunction_38022make

@vtjnash is this the issue with trampoline suppport on ARM that you mentioned a while back?

[vtjnash]

Yeah, we need to edit our lowering to emit an explicit “not yet implemented on ARM”

[oschulz]

Just stumbled on this via FunctionWrappers.jl - is this still an issue? Just wondering since it may affect the design of a project of ours.

[vchuravy]

Yes still an issue on PowerPC and AArch64 afaik

[oschulz]

Will this have any practical implications for Apple-Silicon systems?

[musm]

We have reports of users running into these problem at JuliaIO/HDF5.jl#811 (comment)
Does LLVM 11 improve the situation here at all?

[vchuravy]

I just checked. It seems PPC has gained support for trampoline's, but we end up calling abort from within the __init_trampoline call (https://android.googlesource.com/toolchain/gcc/+/master/gcc-4.9/libgcc/config/rs6000/tramp.S#181).

Grepping for INIT_TRAMPOLINE shows that only the X86 and PPC backend have lowerings.

[giordano]

If I understand it correctly, this might be addressed upstream by llvm/llvm-project#70267