Different results for EXPT between compiled and interpreted code

[rtoy]

Consider

(defun sqr (x)
  (declare (type (complex double-float) x))
  (expt x 2))

Compare the results:

* (sqr #c(0d0 1d0))
#C(-1d0 0d0)

* (expt #c(0d0 1d0) 2)
#C(-1.0d0 1.2246467991473532d-16)

The difference is caused by a deftransform for expt that converts (expt x 2) to (* x x).

Perhaps the expt should be modified to do the same transformation when the power is one of the special values in the deftransform? (The deftransform handles 2, 3, 4, and 1/2 and their negatives.)

[toy.raymond@…]

commit 93656b6aa0ef4e939a84dfb62a6f088f58d3ff62 Author: Raymond Toy <toy.raymond@…> Date: Wed Jan 23 21:22:24 2013 -0800

Fix ticket:65. Implement the deftransform in the expt function.

[rtoy]

Some cases were missed in the previous patch. Here is a test that should cover all the cases.

(defun test-expt-xfrm ()
  (let (failures)
    (dolist (base '(2 2f0 2d0 2w0 #c(0 1) #c(0f0 1) #c(0d0 1) #c(0w0 1)))
      (dolist (power '(2 3 1/2 -2 -3 -1/2 5))
	(dolist (power-type '(rational single-float double-float double-double-float
			      (complex single-float) (complex double-float)
			      (complex double-double-float)))
	  (let* ((pp (coerce power power-type))
		 (interp (expt base pp))
		 (compiled (funcall (compile nil `(lambda (b)
						    (declare (type ,(type-of base) b))
						    (expt b ,pp)))
				    base)))
	    (unless (= interp compiled)
	      (push (list base pp interp compiled) failures)
	      (format t "~S^~S =~%  ~S~%  ~S~%" base pp interp compiled))))))
    failures))

This should return NIL if everything is working correctly.

[toy.raymond@…]

commit 719e87b7d103d3201c031412de576653a42daff7 Author: Raymond Toy <toy.raymond@…> Date: Thu Jan 24 20:55:45 2013 -0800

Fix ticket:65 some more.

Apply the expt transform in more places. The test script in the ticket now passes.