fix continuation-attachment optimization and non-single-value primitives

original commit: 6180395ad38887c6278e96a6cea9e0a4aad2453a

mats/4.ms

@@ -3630,6 +3630,24 @@
                  (lambda ()
                    (k (lambda () #t)))))])
     (proc))
+
+  (equal?
+   '(2 2)
+   (let ([f (parameterize ([run-cp0
+                            ;; Disable cp0 inlining of `for-each`:
+                            (lambda (cp0 e) e)])
+              (eval '(lambda (g l)
+                       (let-values ([(x y)
+                                     (call-setting-continuation-attachment
+                                      'hi
+                                      (lambda ()
+                                        ;; Could be by `cpnanopoass`, but make
+                                        ;; sure it's constrained by the fact that
+                                        ;; `for-each` can tail-call `g`:
+                                        (#3%for-each g l)))])
+                         (list x y)))))])
+     (f (lambda (x) (values x x)) (list 1 2))))
+
  )
  
 ;;; section 4-7:

s/cpnanopass.ss

@@ -2880,9 +2880,9 @@
            (cond
              [(and
                (or (not (info-call-shift-attachment? info))
-                   ;; FIXME: need a less fragile way to avoid multiple results
-                   ;; Exclude inlined primitives that return more than one value:
-                   (not (memq (primref-name pr) '(values call/cc call-with-current-continuation call/1cc))))
+                   ;; Note: single-valued also implies that the primitive doesn't
+                   ;; tail-call an arbitary function (which might inspect attachments):
+                   (all-set? (prim-mask single-valued) (primref-flags pr)))
                (handle-prim (info-call-src info) (info-call-sexpr info) (primref-level pr) (primref-name pr) e*))
               => (lambda (e)
                    (let ([e (Expr e)])