Issues related to the purity modeling PR

[gbaraldi]

This is more of a reminding issue than something to hold up development.

The PR causes an LLVM error in the CodeInstances ccall in boot.jl when building in the M1, this might be related to the fact that the codeinstances struct uses a nested anonymous union. (This has a hacky fix for now, see #44279 and #44107 for details)

The same union also causes warnings when building with clang and apparently violates TBAA. (see #44255)

Edit:

This is what tells LLVM that it can't do a tailcall for the CodeInstances ccall https://github.com/JuliaLang/llvm-project/blob/bc5644ee74f4cb42042257ac129d2be1c252e3f2/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp#L5935

[gbaraldi]

To debug the M1 issue I used an llvm build with some debug prints added, see Make.user:

override LLVM_SHA1=8480406c80c0bf5f7288b07d0df4a030319c5a14
USE_BINARYBUILDER_LLVM=0
LLVM_GIT_URL=https://github.com/gbaraldi/llvm-project.git
LLVM_BRANCH=add-debug
LLVM_ASSERTIONS=1
LLVM_DEBUG=Release

[JeffBezanson]

Could you paste the error output? I'm also not sure what you mean about CodeInstance:

julia> fieldtypes(Core.CodeInstance)
(Core.MethodInstance, Core.CodeInstance, UInt64, UInt64, Any, Any, Any, UInt8, UInt8, Any, Bool, Bool, Ptr{Nothing}, Ptr{Nothing}, UInt8)

What is the union you're referring to?

[gbaraldi]

The union is in C

julia/src/julia.h

Lines 377 to 426 in 1b47efb

	typedef struct _jl_code_instance_t {
	JL_DATA_TYPE
	jl_method_instance_t *def; // method this is specialized from
	_Atomic(struct _jl_code_instance_t*) next; // pointer to the next cache entry
	// world range for which this object is valid to use
	size_t min_world;
	size_t max_world;
	// inference state cache
	jl_value_t *rettype; // return type for fptr
	jl_value_t *rettype_const; // inferred constant return value, or null
	jl_value_t *inferred; // inferred jl_code_info_t, or jl_nothing, or null
	//TODO: jl_array_t *edges; // stored information about edges from this object
	//TODO: uint8_t absolute_max; // whether true max world is unknown
	// purity results
	union {
	uint8_t ipo_purity_bits;
	struct {
	uint8_t ipo_consistent:2;
	uint8_t ipo_effect_free:2;
	uint8_t ipo_nothrow:2;
	uint8_t ipo_terminates:2;
	} ipo_purity_flags;
	};
	union {
	uint8_t purity_bits;
	struct {
	uint8_t consistent:2;
	uint8_t effect_free:2;
	uint8_t nothrow:2;
	uint8_t terminates:2;
	} purity_flags;
	};
	jl_value_t *argescapes; // escape information of call arguments
	// compilation state cache
	uint8_t isspecsig; // if specptr is a specialized function signature for specTypes->rettype
	_Atomic(uint8_t) precompile; // if set, this will be added to the output system image
	_Atomic(jl_callptr_t) invoke; // jlcall entry point
	union _jl_generic_specptr_t {
	_Atomic(void*) fptr;
	_Atomic(jl_fptr_args_t) fptr1;
	// 2 constant
	_Atomic(jl_fptr_sparam_t) fptr3;
	// 4 interpreter
	} specptr; // private data for `jlcall entry point
	uint8_t relocatability; // nonzero if all roots are built into sysimg or tagged by module key
	} jl_code_instance_t;

and it used to fail when this was called

julia/base/boot.jl

Lines 421 to 431 in 1b47efb

	function CodeInstance(
	mi::MethodInstance, @nospecialize(rettype), @nospecialize(inferred_const),
	@nospecialize(inferred), const_flags::Int32, min_world::UInt, max_world::UInt,
	ipo_effects::UInt8, effects::UInt8, @nospecialize(argescapes#=::Union{Nothing,Vector{ArgEscapeInfo}}=#),
	relocatability::UInt8)
	return ccall(:jl_new_codeinst, Ref{CodeInstance},
	(Any, Any, Any, Any, Int32, UInt, UInt, UInt8, UInt8, Any, UInt8),
	mi, rettype, inferred_const, inferred, const_flags, min_world, max_world,
	ipo_effects, effects, argescapes,
	relocatability)
	end

The error output came from deep in LLVM but it's

LLVM ERROR: failed to perform tail call elimination on a call site marked musttail

[JeffBezanson]

Thanks.

[giordano]

@gbaraldi can you check whether #44807 helps?

[gbaraldi]

It does not unfortunately