Optimize raw-instance-slots-equalp for #-complex-float-vops.
[sbcl.git] / src / compiler / float-tran.lisp
index 8d2eaed..2ae1e33 100644 (file)
   (frob %random-single-float single-float)
   (frob %random-double-float double-float))
 
-;;; Mersenne Twister RNG
-;;;
-;;; FIXME: It's unpleasant to have RANDOM functionality scattered
-;;; through the code this way. It would be nice to move this into the
-;;; same file as the other RANDOM definitions.
+;;; Return an expression to generate an integer of N-BITS many random
+;;; bits, using the minimal number of random chunks possible.
+(defun generate-random-expr-for-power-of-2 (n-bits state)
+  (declare (type (integer 1 #.sb!vm:n-word-bits) n-bits))
+  (multiple-value-bind (n-chunk-bits chunk-expr)
+      (cond ((<= n-bits n-random-chunk-bits)
+             (values n-random-chunk-bits `(random-chunk ,state)))
+            ((<= n-bits (* 2 n-random-chunk-bits))
+             (values (* 2 n-random-chunk-bits) `(big-random-chunk ,state)))
+            (t
+             (error "Unexpectedly small N-RANDOM-CHUNK-BITS")))
+    (if (< n-bits n-chunk-bits)
+        `(logand ,(1- (ash 1 n-bits)) ,chunk-expr)
+        chunk-expr)))
+
+;;; This transform for compile-time constant word-sized integers
+;;; generates an accept-reject loop to achieve equidistribution of the
+;;; returned values. Several optimizations are done: If NUM is a power
+;;; of two no loop is needed. If the random chunk size is half the word
+;;; size only one chunk is used where sufficient. For values of NUM
+;;; where it is possible and results in faster code, the rejection
+;;; probability is reduced by accepting all values below the largest
+;;; multiple of the limit that fits into one or two chunks and and doing
+;;; a division to get the random value into the desired range.
 (deftransform random ((num &optional state)
-                      ((integer 1 #.(expt 2 sb!vm::n-word-bits)) &optional *))
-  ;; FIXME: I almost conditionalized this as #!+sb-doc. Find some way
-  ;; of automatically finding #!+sb-doc in proximity to DEFTRANSFORM
-  ;; to let me scan for places that I made this mistake and didn't
-  ;; catch myself.
-  "use inline (UNSIGNED-BYTE 32) operations"
-  (let ((type (lvar-type num))
-        (limit (expt 2 sb!vm::n-word-bits))
-        (random-chunk (ecase sb!vm::n-word-bits
-                        (32 'random-chunk)
-                        (64 'sb!kernel::big-random-chunk))))
-    (if (numeric-type-p type)
-        (let ((num-high (numeric-type-high (lvar-type num))))
-          (aver num-high)
-          (cond ((constant-lvar-p num)
-                 ;; Check the worst case sum absolute error for the
-                 ;; random number expectations.
-                 (let ((rem (rem limit num-high)))
-                   (unless (< (/ (* 2 rem (- num-high rem))
-                                 num-high limit)
-                              (expt 2 (- sb!kernel::random-integer-extra-bits)))
-                     (give-up-ir1-transform
-                      "The random number expectations are inaccurate."))
-                   (if (= num-high limit)
-                       `(,random-chunk (or state *random-state*))
-                       #!-(or x86 x86-64)
-                       `(rem (,random-chunk (or state *random-state*)) num)
-                       #!+(or x86 x86-64)
-                       ;; Use multiplication, which is faster.
-                       `(values (sb!bignum::%multiply
-                                 (,random-chunk (or state *random-state*))
-                                 num)))))
-                ((> num-high random-fixnum-max)
-                 (give-up-ir1-transform
-                  "The range is too large to ensure an accurate result."))
-                #!+(or x86 x86-64)
-                ((< num-high limit)
-                 `(values (sb!bignum::%multiply
-                           (,random-chunk (or state *random-state*))
-                           num)))
-                (t
-                 `(rem (,random-chunk (or state *random-state*)) num))))
-        ;; KLUDGE: a relatively conservative treatment, but better
-        ;; than a bug (reported by PFD sbcl-devel towards the end of
-        ;; 2004-11.
-        (give-up-ir1-transform
-         "Argument type is too complex to optimize for."))))
+                      ((constant-arg (integer 1 #.(expt 2 sb!vm:n-word-bits)))
+                       &optional *)
+                      *
+                      :policy (and (> speed compilation-speed)
+                                   (> speed space)))
+  "optimize to inlined RANDOM-CHUNK operations"
+  (let ((num (lvar-value num)))
+    (if (= num 1)
+        0
+        (flet ((chunk-n-bits-and-expr (n-bits)
+                 (cond ((<= n-bits n-random-chunk-bits)
+                        (values n-random-chunk-bits
+                                '(random-chunk (or state *random-state*))))
+                       ((<= n-bits (* 2 n-random-chunk-bits))
+                        (values (* 2 n-random-chunk-bits)
+                                '(big-random-chunk (or state *random-state*))))
+                       (t
+                        (error "Unexpectedly small N-RANDOM-CHUNK-BITS")))))
+          (if (zerop (logand num (1- num)))
+              ;; NUM is a power of 2.
+              (let ((n-bits (integer-length (1- num))))
+                (multiple-value-bind (n-chunk-bits chunk-expr)
+                    (chunk-n-bits-and-expr n-bits)
+                  (if (< n-bits n-chunk-bits)
+                      `(logand ,(1- (ash 1 n-bits)) ,chunk-expr)
+                      chunk-expr)))
+              ;; Generate an accept-reject loop.
+              (let ((n-bits (integer-length num)))
+                (multiple-value-bind (n-chunk-bits chunk-expr)
+                    (chunk-n-bits-and-expr n-bits)
+                  (if (or (> (* num 3) (expt 2 n-chunk-bits))
+                          (logbitp (- n-bits 2) num))
+                      ;; Division can't help as the quotient is below 3,
+                      ;; or is too costly as the rejection probability
+                      ;; without it is already small (namely at most 1/4
+                      ;; with the given test, which is experimentally a
+                      ;; reasonable threshold and cheap to test for).
+                      `(loop
+                         (let ((bits ,(generate-random-expr-for-power-of-2
+                                       n-bits '(or state *random-state*))))
+                           (when (< bits num)
+                             (return bits))))
+                      (let ((d (truncate (expt 2 n-chunk-bits) num)))
+                        `(loop
+                           (let ((bits ,chunk-expr))
+                             (when (< bits ,(* num d))
+                               (return (values (truncate bits ,d)))))))))))))))
+
 \f
 ;;;; float accessors
 
                                        (if (< x ,most-negative)
                                            ,most-negative
                                            (coerce x ',type)))
-                                     (numeric-type-low num)))
+                                     (numeric-type-low num)
+                                     nil))
                      (hi (bound-func (lambda (x)
                                        (if (< ,most-positive x )
                                            ,most-positive
                                            (coerce x ',type)))
-                                     (numeric-type-high num))))
+                                     (numeric-type-high num)
+                                     nil)))
                 (specifier-type `(,',type ,(or lo '*) ,(or hi '*)))))
 
             (defoptimizer (,fun derive-type) ((num))
                  ;; Process the intersection.
                  (let* ((low (interval-low intersection))
                         (high (interval-high intersection))
-                        (res-lo (or (bound-func fun (if increasingp low high))
+                        (res-lo (or (bound-func fun (if increasingp low high) nil)
                                     default-low))
-                        (res-hi (or (bound-func fun (if increasingp high low))
+                        (res-hi (or (bound-func fun (if increasingp high low) nil)
                                     default-high))
                         (format (case (numeric-type-class arg)
                                   ((integer rational) 'single-float)
                    (int-hi (if hi
                                (ceiling (type-bound-number hi))
                                '*))
-                   (f-lo (or (bound-func #'float lo)
+                   (f-lo (or (bound-func #'float lo nil)
                              '*))
-                   (f-hi (or (bound-func #'float hi)
+                   (f-hi (or (bound-func #'float hi nil)
                              '*)))
               (specifier-type `(or (rational ,int-lo ,int-hi)
                                 (single-float ,f-lo, f-hi)))))
            (float
             ;; A positive integer to a float power is a float.
-            (modified-numeric-type y-type
-                                   :low (interval-low bnd)
-                                   :high (interval-high bnd)))
+            (let ((format (numeric-type-format y-type)))
+              (aver format)
+              (modified-numeric-type
+               y-type
+               :low (coerce-numeric-bound (interval-low bnd) format)
+               :high (coerce-numeric-bound (interval-high bnd) format))))
            (t
             ;; A positive integer to a number is a number (for now).
             (specifier-type 'number))))
                    (int-hi (if hi
                                (ceiling (type-bound-number hi))
                                '*))
-                   (f-lo (or (bound-func #'float lo)
+                   (f-lo (or (bound-func #'float lo nil)
                              '*))
-                   (f-hi (or (bound-func #'float hi)
+                   (f-hi (or (bound-func #'float hi nil)
                              '*)))
               (specifier-type `(or (rational ,int-lo ,int-hi)
                                 (single-float ,f-lo, f-hi)))))
            (float
             ;; A positive rational to a float power is a float.
-            (modified-numeric-type y-type
-                                   :low (interval-low bnd)
-                                   :high (interval-high bnd)))
+            (let ((format (numeric-type-format y-type)))
+              (aver format)
+              (modified-numeric-type
+               y-type
+               :low (coerce-numeric-bound (interval-low bnd) format)
+               :high (coerce-numeric-bound (interval-high bnd) format))))
            (t
             ;; A positive rational to a number is a number (for now).
             (specifier-type 'number))))
            ((or integer rational)
             ;; A positive float to an integer or rational power is
             ;; always a float.
-            (make-numeric-type
-             :class 'float
-             :format (numeric-type-format x-type)
-             :low (interval-low bnd)
-             :high (interval-high bnd)))
+            (let ((format (numeric-type-format x-type)))
+              (aver format)
+              (make-numeric-type
+               :class 'float
+               :format format
+               :low (coerce-numeric-bound (interval-low bnd) format)
+               :high (coerce-numeric-bound (interval-high bnd) format))))
            (float
             ;; A positive float to a float power is a float of the
             ;; higher type.
-            (make-numeric-type
-             :class 'float
-             :format (float-format-max (numeric-type-format x-type)
-                                       (numeric-type-format y-type))
-             :low (interval-low bnd)
-             :high (interval-high bnd)))
+            (let ((format (float-format-max (numeric-type-format x-type)
+                                            (numeric-type-format y-type))))
+              (aver format)
+              (make-numeric-type
+               :class 'float
+               :format format
+               :low (coerce-numeric-bound (interval-low bnd) format)
+               :high (coerce-numeric-bound (interval-high bnd) format))))
            (t
             ;; A positive float to a number is a number (for now)
             (specifier-type 'number))))
                               :complexp :real
                               :low (numeric-type-low type)
                               :high (numeric-type-high type))))))
-#-sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
+
 (defoptimizer (realpart derive-type) ((num))
   (one-arg-derive-type num #'realpart-derive-type-aux #'realpart))
+
 (defun imagpart-derive-type-aux (type)
   (let ((class (numeric-type-class type))
         (format (numeric-type-format type)))
                               :complexp :real
                               :low (numeric-type-low type)
                               :high (numeric-type-high type))))))
-#-sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
+
 (defoptimizer (imagpart derive-type) ((num))
   (one-arg-derive-type num #'imagpart-derive-type-aux #'imagpart))
 
               ;; exactly the same way as the functions themselves do
               ;; it.
               (if (csubtypep arg domain)
-                  (let ((res-lo (bound-func fun (numeric-type-low arg)))
-                        (res-hi (bound-func fun (numeric-type-high arg))))
+                  (let ((res-lo (bound-func fun (numeric-type-low arg) nil))
+                        (res-hi (bound-func fun (numeric-type-high arg) nil)))
                     (unless increasingp
                       (rotatef res-lo res-hi))
                     (make-numeric-type