Let lifting breaks things in the presence of set!

[eholk]

This came up while working on a library version of reduce:

(define (vector_sum xs)
    (while (> (length xs) 1)
      ;;(print "vector_sum length = ");
      ;;(print (length xs))
      ;;(print "\n")
      (let ((len (/ (length xs) 2)))
        (set! xs (kernel ((i (iota len)))
                   (+ (vector-ref xs i) (vector-ref xs (+ i len)))))))
    (vector-ref xs 0))

With the print statements uncommented, the code works correctly. It seems the (let ((len (/ (length xs) 2))) ...) section gets lifted above the while loop, since the lifter assumes xs never changes.

This is probably a good argument for removing set! entirely in favor of named let.

[calvis]

This can be solved without removing set! entirely; i.e. we could just do an assigned-variable analysis. I'm not saying that is necessarily the right way to go, but doing that analysis would probably help out some other passes as well.

[eholk]

Yeah, that makes sense. I was worried that the assigned variable analysis might be kind of intricate, but I guess not. As you say, the analysis will probably be helpful anyway. For example, named let will probably expand into set!s and gotos.

[eholk]

I suspect this bug explains why our Mandelbrot programs only generate white images now.