More special, eigen, and bitarray tests

[kshyatt]

Got error throwing to work for eigmin and eigmax. Added tests for a bunch of untested branches in log and airy and erfinv.

[tkelman]

should this still have a message?

[kshyatt]

It throws an error with the message.

[tkelman]

but a generic domain error, as opposed to the specific "complex eigenvalues cannot be ordered" information right?

[kshyatt]

I'm confused. If you put the string between the parens it will freak out.

[tkelman]

Oh, bummer, DomainError doesn't have an extra message field like I thought it would. (So

julia/base/linalg/cholesky.jl

Line 89 in 71f08f3

throw(DomainError("x must be positive semidefinite"))

is probably untested.) Throwing a more specific exception type is good, but so is providing additional info to the user. Looks like DomainErrors get specially handled here

julia/base/replutil.jl

Lines 96 to 113 in 71f08f3

	function showerror(io::IO, ex::DomainError, bt; backtrace=true)
	print(io, "DomainError:")
	for b in bt
	code = ccall(:jl_lookup_code_address, Any, (Ptr{Void}, Cint), b-1, true)
	if length(code) == 5 && !code[4] # code[4] == fromC
	if code[1] in (:log, :log2, :log10, :sqrt) # TODO add :besselj, :besseli, :bessely, :besselk
	println(io,"\n$(code[1]) will only return a complex result if called with a complex argument.")
	print(io, "try $(code[1]) (complex(x))")
	elseif (code[1] == :^ && code[2] == symbol("intfuncs.jl")) || code[1] == :power_by_squaring
	println(io, "\nCannot raise an integer x to a negative power -n. Make x a float by adding")
	print(io, "a zero decimal (e.g. 2.0^-n instead of 2^-n), or write 1/x^n, float(x)^-n, or (x//1)^-n")
	end
	break
	end
	end
	backtrace && show_backtrace(io, bt)
	nothing
	end

[tkelman]

The untested cholesky method sure (probably better to use PosDefException for that anyway), though the loss of information in this error (and the nonstandard indentation on this line) are unfortunate, and would be easy to forget to revisit. Maybe a TODO comment?

[andreasnoack]

I think it can make an exception much more costly if contains a string (or any other pointer to a Julia object). These costs are probably okay for expensive linear algebra functions but not in scalar math functions such as sqrt where DomainError is used.

[simonster]

We could have DomainError{T<:Union(None,ByteString)} where DomainError() would be a DomainError{None}. I don't think that would be any more expensive when there's no string to be thrown. Would need some changes on the codegen side, though...

[yuyichao]

I think the right approach is just to fix the performance issue. Ref #11508

[simonster]

I'm not sure removing the GC frame would be enough. It would eliminate the additional cost for a DomainError with a string parameter in code that doesn't throw a DomainError, but you'd still have additional overhead if you actually throw a DomainError, which might matter.

[carlobaldassi]

B and fB seem to be unused.

[kshyatt]

Yes. I'll remove and rebase. Thanks for catching this.

[kshyatt]

Anyone know what the deal is with that OSX failure? Does this look ok?

[tkelman]

#12127. Still need to decide what to do about https://github.com/JuliaLang/julia/pull/12121/files#r34417008 - at a bare minimum, fix the indentation and put in a TODO comment to put the informational error message back

[kshyatt]

The TODO is already there. I'll fix my indenting.

@tkelman, is everything ok with you?

[tkelman]

Still a bit disappointed about losing the complex eigenvalues cannot be ordered information in the error, but I guess it's alright for now.

[kshyatt]

@tkelman my impression was that we (someone?) need to make a decision about DomainError that's kind out of out-of-scope for this PR. Would you agree? I can open an issue for that and mention this PR so putting the info back in doesn't fall off the radar.

[tkelman]

ok