Use at-consistent_overlay for 1.11 compatibility.

[maleadt]

Address another part of #2241

@aviatesk I just realized though that we violate the requirements of @consistent_overlay without @device_function macro, which is used to implement intrinsics that are not CPU compatible (they make LLVM abort) so we make the origin method error()while the overlay performs the expected operation:

CUDA.jl/src/device/utils.jl

Lines 16 to 36 in 123977e

	macro device_override(ex)
	ex = macroexpand(__module__, ex)
	esc(quote
	Base.Experimental.@overlay(CUDA.method_table, $ex)
	end)
	end
	macro device_function(ex)
	ex = macroexpand(__module__, ex)
	def = splitdef(ex)
	# generate a function that errors
	def[:body] = quote
	error("This function is not intended for use on the CPU")
	end
	esc(quote
	$(combinedef(def))
	@device_override $ex
	end)
	end

IIUC this is not valid with @consistent_overlay. Do you have another suggestion to implement this in a way that's compatible with the effects system?

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 73.26%. Comparing base (991e23a) to head (30fe600).
Report is 1 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #2492      +/-   ##
==========================================
+ Coverage   73.06%   73.26%   +0.20%     
==========================================
  Files         157      157              
  Lines       14930    14930              
==========================================
+ Hits        10908    10938      +30     
+ Misses       4022     3992      -30     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[aviatesk]

Exactly, I think using @consistent_overlay for @device_function would very likely lead to mis-compilation. Would it be possible to change @device_override to a 2-arg macro and simply have @device_function use @consistent instead of @consistent_overlay?

[maleadt]

Yeah. Hopefully that doesn't lead to inference issues a la JuliaLang/julia#52938.