1.0.14.36: faster PROPAGATE-FROM-SETS

 20-25% improvement for the test-case in bug 188.

 * New slot in LAMBDA-VAR: LAST-INITIAL-TYPE, which holds the last
   initial-type for that variable seen by PROPAGATE-FROM-SETS.

 * Be lazy, and don't PROPAGATE-TO-REFS unless something of
   interest has happened, to wit:

    -- One of the CSET nodes has a new, more specific type.

    -- INITIAL-TYPE has become more specific.

   This also allows us elide TYPE-UNION computation in the
   uninteresting cases.

 * Requires having NODE-REOPTIMIZE set when IR1-OPTIMIZE-SET
   is called.

BUGS

@@ -479,6 +479,11 @@ WORKAROUND:
  (print (incf start 22))
  (print (incf start 26))))))
 
+ [ Update: 1.0.14.36 improved this quite a bit (20-25%) by
+ eliminating useless work from PROPAGATE-FROM-SETS -- but as alluded
+ below, maybe we should be smarter about when to decide a derived
+ type is "good enough". ]
+
  This example could be solved with clever enough constraint
  propagation or with SSA, but consider
 

NEWS

@@ -11,6 +11,10 @@ changes in sbcl-1.0.15 relative to sbcl-1.0.14:
  when SB-DEBUG:*SHOW-ENTRY-POINT-DETAILS* is NIL.
  * unreadably printed representation of hash-tables now includes
  weakness if any.
+ * bug fix: partially fixed #188: type propagation from assignments
+ is now more efficient.
+ * bug fix: fixed #407: (COERCE X 'SINGLE-FLOAT) and (COERCE X
+ 'DOUBLE-FLOAT) are not flushable.
  * bug fix: on x86 and x86-64 pointer based EQ-hashing now uses the
  full address of the object, and none of the tag bits.
  * bug fix: readably printing hash-tables now respects other printer

src/compiler/ir1opt.lisp

@@ -306,6 +306,9 @@
  (when value
  (derive-node-type node (lvar-derived-type value)))))
  (cset
+ ;; PROPAGATE-FROM-SETS can do a better job if NODE-REOPTIMIZE
+ ;; is accurate till the node actually has been reoptimized.
+ (setf (node-reoptimize node) t)
  (ir1-optimize-set node))
  (cast
  (ir1-optimize-cast node)))))
@@ -1342,17 +1345,22 @@
 ;;; the union of the INITIAL-TYPE and the types of all the set
 ;;; values and to a PROPAGATE-TO-REFS with this type.
 (defun propagate-from-sets (var initial-type)
- (collect ((res initial-type type-union))
- (dolist (set (basic-var-sets var))
+ (let ((changes (not (csubtypep (lambda-var-last-initial-type var) initial-type)))
+ (types nil))
+ (dolist (set (lambda-var-sets var))
  (let ((type (lvar-type (set-value set))))
- (res type)
+ (push type types)
  (when (node-reoptimize set)
- (derive-node-type set (make-single-value-type type))
+ (let ((old-type (node-derived-type set)))
+ (unless (values-subtypep old-type type)
+ (derive-node-type set (make-single-value-type type))
+ (setf changes t)))
  (setf (node-reoptimize set) nil))))
- (let ((res (res)))
- (awhen (maybe-infer-iteration-var-type var initial-type)
- (setq res it))
- (propagate-to-refs var res)))
+ (when changes
+ (setf (lambda-var-last-initial-type var) initial-type)
+ (let ((res-type (or (maybe-infer-iteration-var-type var initial-type)
+ (apply #'type-union initial-type types))))
+ (propagate-to-refs var res-type))))
  (values))
 
 ;;; If a LET variable, find the initial value's type and do
@@ -1368,9 +1376,9 @@
  (initial-type (lvar-type initial-value)))
  (setf (lvar-reoptimize initial-value) nil)
  (propagate-from-sets var initial-type))))))
-
  (derive-node-type node (make-single-value-type
  (lvar-type (set-value node))))
+ (setf (node-reoptimize node) nil)
  (values))
 
 ;;; Return true if the value of REF will always be the same (and is

src/compiler/node.lisp

@@ -1114,6 +1114,8 @@
  ;; determine that this is a set closure variable, and is thus not a
  ;; good subject for flow analysis.
  (constraints nil :type (or sset null))
+ ;; Initial type of a LET variable as last seen by PROPAGATE-FROM-SETS.
+ (last-initial-type *universal-type* :type ctype)
  ;; The FOP handle of the lexical variable represented by LAMBDA-VAR
  ;; in the fopcompiler.
  (fop-value nil))

version.lisp-expr

@@ -17,4 +17,4 @@
 ;;; checkins which aren't released. (And occasionally for internal
 ;;; versions, especially for internal versions off the main CVS
 ;;; branch, it gets hairier, e.g. "0.pre7.14.flaky4.13".)
-"1.0.14.35"
+"1.0.14.36"