From a6f03ee38f0e735b1a7004a826427e8fe7d7b9d5 Mon Sep 17 00:00:00 2001
From: Ryan Culpepper <ryanc@racket-lang.org>
Date: Tue, 18 Dec 2012 09:44:22 -0500
Subject: [PATCH] use popcount table

---
 collects/data/bit-vector.rkt                   | 16 ++++++++++++++--
 collects/data/private/count-bits-in-fixnum.rkt |  7 +++----
 2 files changed, 17 insertions(+), 6 deletions(-)

diff --git a/collects/data/bit-vector.rkt b/collects/data/bit-vector.rkt
index 59595b26a4..56443b494f 100644
--- a/collects/data/bit-vector.rkt
+++ b/collects/data/bit-vector.rkt
@@ -1,8 +1,9 @@
 #lang racket/base
 (require (for-syntax racket/base)
+         (for-syntax "private/count-bits-in-fixnum.rkt")
          racket/private/vector-wraps
          racket/match
-         racket/dict         
+         racket/dict
          racket/contract/base
          racket/fixnum
          racket/unsafe/ops
@@ -90,9 +91,20 @@
            (bit-vector-copy bv start end)])])
     bit-vector-copy))
 
+(define popcount-table
+  (let ()
+    (define-syntax (make-table stx)
+      (with-syntax ([(elt ...)
+                     (for/list ([i (in-range 256)])
+                       (fxpopcount i))])
+        ;; Literal immutable vector allocated once (?)
+        #'(quote #(elt ...))))
+    (make-table)))
+
 (define (bit-vector-popcount bv)
   (for/sum ([b (in-bytes (bit-vector-words bv))])
-    (unsafe-fxpopcount* b 8)))
+    #| (unsafe-fxpopcount* b 8) |#
+    (unsafe-vector-ref popcount-table b)))
 
 (define (bit-vector->list bv)
   (define len (bit-vector-size bv))
diff --git a/collects/data/private/count-bits-in-fixnum.rkt b/collects/data/private/count-bits-in-fixnum.rkt
index 3e85bf65cd..f9b7deb38e 100644
--- a/collects/data/private/count-bits-in-fixnum.rkt
+++ b/collects/data/private/count-bits-in-fixnum.rkt
@@ -1,6 +1,6 @@
 #lang racket/base
 (require racket/unsafe/ops
-         (for-syntax racket/base racket/fixnum racket/vector))
+         (for-syntax racket/base racket/fixnum))
 (provide fxpopcount
          unsafe-fxpopcount*)
 ;; Count set bits for 30 bit number in 5 steps.
@@ -43,13 +43,12 @@
              (cond [(<= bits 8) lut8]
                    [(<= bits 30) lut30]
                    [(<= bits 62) lut62]
-                   [else (raise-syntax-error "bit width too big" stx #'bits0)])]
-            [flut (vector-map bitwise-not lut)])
+                   [else (raise-syntax-error "bit width too big" stx #'bits0)])])
        ;; Unroll the addition loop
        #`(let ([n expr])
            (let* #,(for/list ([m (in-vector lut)]
-                              [f (in-vector flut)]
                               [b (in-naturals)])
+                     (define f (bitwise-not m))
                      #`[n (unsafe-fx+ (unsafe-fxrshift (unsafe-fxand n #,m)
                                                        #,(fxlshift 1 b))
                                       (unsafe-fxand n #,f))])