All entries for Wednesday 20 March 2024

It's sort of pass by reference, ish

When teaching Fortran, it's often tempting to describe it as a "pass by reference" language. This is almost true. Unfortunately as the phrase has it, "almost only counts in horseshoes and handgrenades" and "almost true" in programming terms isn't usually good enough. All too often, the sort of complicated technical thing that should only matter to an expert trips up plenty of beginners who, like the Inexperienced Swordsman "doesn't do the thing he ought to do"

So what's the truth? The truth is that most of the time things work just like pass-by-reference because they are expected to look just as if they were. But actually, they are allowed to do a "copy-in copy-back" process, as long as the final result has the expected changes. Copies can be expensive, so compilers don't tend to actually do this without a good reason.

The place this usually comes to our notice is when we use slices of arrays. For example, we can take every second element of an array in Fortran, very easily. Or we can take half of a 2-dimensional array (in the row direction). Both of these are valid "array slices" but both consist of data that is not contiguous in memory. Items do not follow each other one after the other in memory.

But in Fortran, we can pass these to some function that expects just "an array". Now imagine what would have to happen to do a true "pass by reference" here. The compiler would have to pass the array, and then enough information to "slice out" only the values we wanted. This would be pretty weird, and if we passed that value on to another function also expecting an array, it could easily get out of control. So instead, the compiler will tend to generate the code to copy the values in the slice into a temporary array, use them there, and copy them back when we are done. To us, everything will work seamlessly.

That is, as long as we do what we're supposed to, everything works. But if we start breaking rules, we can get some very odd behaviour from this. Compilers are like puppies - they trust us to keep our promises. In fact, they rely on this so much that they simply assume that we will! When we don't funny things happen.

The following code is modified from this old post https://groups.google.com/g/comp.lang.fortran/c/z11RW0ezojE?pli=1 to use modern Fortran.

MODULE mymod


CONTAINS


SUBROUTINE set_to_two(B, C)
REAL, DIMENSION(10) :: B
REAL, DIMENSION(5) :: C
INTEGER :: i
DO I = 1, 10
B(I) = 2.0
ENDDO
END


END MODULE


PROGRAM main
USE mymod
REAL, DIMENSION(10) :: A


A = 1.0
CALL set_to_two(A, A(1:10:2))
PRINT *, A


END


Pretty simple code - it takes an array and sets every value to 2.0. For some reason it also takes a second unused irrelevant parameter, but we can ignore that one, surely?

Run that code, as written with current gfortran, and this is the result:

1.00000000  2.00000000  1.00000000  2.00000000  1.00000000  2.00000000  1.00000000  2.00000000  1.00000000  2.00000000  

The code is bad - passing A as two different parameters to a function violates a promise we make as programmers not to alias variables (call them by two different names). But even knowing this we're pretty surprised to get that result! In particular, on my machine, if I call the function with CALL set_to_two(A, A(2:6)) which violates the aliasing rules just as badly, nothing odd happens at all, and all I get is A = 2.0. In this case, the compiler is able to avoid a potentially costly copy as the data is contiguous even though it's a slice.

It's pretty obvious what is actually happening once we know about the copy-back idea. Because the compiler trusted us not to have two names for the same piece of data (the names here being B and C, and the data being A) it happily copies data for the C argument, copying it back at the end. This copy is never affected by the update to B so its content remains 1.0. That gets copied back into A, overwriting the changes we'd made.

This can happen even though we never use C in the function, so nothing actually changes - that second irrelevant argument is not so irrelevant after all.

Take Home Point

The real take-home point here is not to upset your compiler - don't do things wrong and these sorts of details wont matter. But when things do go wrong, it can be pretty helpful to understand what is actually going on.

Honestly a lot of the bugs we write as programmers are a case of miscommunication - what we thought we were writing is not what we actually have said. This is expecially true with modern optimising compilers, which very liberally agree to produce a program that works "as if" it was the one we wrote, and will happily omit things that do not affect the results, or, as in this case, will assume that we are writing correct code and act accordingly.

Fortran is usually a lot nicer to us than C/C++ in terms of undefined behaviour, but as this example shows, it will still do strange things if we break the rules.

Heather Ratcliffe : 20 Mar 2024 19:32 |  Tags: Bugs Fortran Programming Undefined-Behaviour |  Comments (0) |  Close comments |  Report a problem