From: Taylor R Campbell <campbell@mumble.net>
Date: Sat, 27 Oct 2018 00:23:12 +0000 (+0000)
Subject: Simple-minded criterion to compute x^-y by 1/x^y was too simple.
X-Git-Tag: mit-scheme-pucked-10.1.2~16^2~150
X-Git-Url: https://birchwood-abbey.net/git?a=commitdiff_plain;h=ecf3e6cb79574d90c0c4a89fd398e513bed9e943;p=mit-scheme.git

Simple-minded criterion to compute x^-y by 1/x^y was too simple.

For any |x| > 1, we can't do this without comparing the size of
log(x) and y well enough to discern what might yield subnormal, so
might as well just compute the general case then.
---

diff --git a/src/runtime/arith.scm b/src/runtime/arith.scm
index 134e99908..71c66b399 100644
--- a/src/runtime/arith.scm
+++ b/src/runtime/arith.scm
@@ -1265,12 +1265,15 @@ USA.
 				      (loop x y answer)))))))))
 		 (cond ((int:positive? y) (exact-method y))
 		       ((int:negative? y)
-			(if (int:< y microcode-id/floating-exponent-min)
-			    ;; The exact method cannot generate
-			    ;; subnormals because the negated exponent
-			    ;; overflows, so use the general case.
-			    (general-case x (int:->flonum y))
-			    (flo:/ flo:1 (exact-method (int:negate y)))))
+			(if (flo:< (flo:abs x) flo:1)
+			    (flo:/ flo:1 (exact-method (int:negate y)))
+			    ;; For any base above 1 and sufficiently large
+			    ;; exponents, or for any negative exponent and
+			    ;; sufficiently large base, the division above
+			    ;; would overflow into infinite and then yield
+			    ;; 0 when it should yield a subnormal.  So use
+			    ;; the general case.
+			    (general-case x (int:->flonum y))))
 		       (else flo:1))))
 	      (else
 	       (general-case x (rat:->inexact y))))
diff --git a/tests/runtime/test-arith.scm b/tests/runtime/test-arith.scm
index 7017c7e0e..9384402e9 100644
--- a/tests/runtime/test-arith.scm
+++ b/tests/runtime/test-arith.scm
@@ -245,4 +245,8 @@ USA.
         (let ((x (vector-ref v 0))
               (y (vector-ref v 1))
               (x^y (string->number (vector-ref v 2))))
-          (assert-eqv (expt x y) x^y))))))
\ No newline at end of file
+          (assert-eqv (expt x y) x^y)
+          ;; For all the inputs, reciprocal is exact.
+          (assert-eqv (expt (/ 1 x) (- y)) x^y)
+          (assert-eqv (expt (* 2 x) (/ y 2)) x^y)
+          (assert-eqv (expt (/ 1 (* 2 x)) (- (/ y 2))) x^y))))))
\ No newline at end of file