;;;; -*- coding: utf-8; -*-
+changes in sbcl-0.9.16 relative to sbcl-0.9.15:
+ * optimization: faster LOGCOUNT implementation on x86 and x86-64
+ (thanks to Lutz Euler)
+
changes in sbcl-0.9.15 relative to sbcl-0.9.14:
* added support for the ucs-2 external format. (contributed by Ivan
Boldyrev)
(defun bignum-logcount (bignum)
(declare (type bignum-type bignum))
- (let* ((length (%bignum-length bignum))
- (plusp (%bignum-0-or-plusp bignum length))
- (result 0))
+ (let ((length (%bignum-length bignum))
+ (result 0))
(declare (type bignum-index length)
(fixnum result))
(do ((index 0 (1+ index)))
- ((= index length) result)
+ ((= index length)
+ (if (%bignum-0-or-plusp bignum length)
+ result
+ (- (* length digit-size) result)))
(let ((digit (%bignum-ref bignum index)))
(declare (type bignum-element-type digit))
- (incf result (logcount (if plusp digit (%lognot digit))))))))
+ (incf result (logcount digit))))))
\f
;;;; logical operations
(inst xor res res)
DONE))
-
(define-vop (unsigned-byte-64-count)
(:translate logcount)
(:note "inline (unsigned-byte 64) logcount")
(:policy :fast-safe)
- (:args (arg :scs (unsigned-reg)))
+ (:args (arg :scs (unsigned-reg) :target result))
(:arg-types unsigned-num)
(:results (result :scs (unsigned-reg)))
(:result-types positive-fixnum)
- (:temporary (:sc unsigned-reg :from (:argument 0)) temp)
- (:temporary (:sc unsigned-reg :from (:argument 0)) t1)
- (:generator 60
+ (:temporary (:sc unsigned-reg) temp)
+ (:temporary (:sc unsigned-reg) mask)
+ (:generator 14
+ ;; See the comments below for how the algorithm works. The tricks
+ ;; used can be found for example in AMD's software optimization
+ ;; guide or at "http://www.hackersdelight.org/HDcode/pop.cc" in the
+ ;; function "pop1", for 32-bit words. The extension to 64 bits is
+ ;; straightforward.
+ ;; Calculate 2-bit sums. Note that the value of a two-digit binary
+ ;; number is the sum of the right digit and twice the left digit.
+ ;; Thus we can calculate the sum of the two digits by shifting the
+ ;; left digit to the right position and doing a two-bit subtraction.
+ ;; This subtraction will never create a borrow and thus can be made
+ ;; on all 32 2-digit numbers at once.
(move result arg)
- (move t1 arg)
-
- (inst mov temp result)
- (inst shr temp 1)
- (inst and result #x55555555) ; note these masks will restrict the
- (inst and temp #x55555555) ; count to the lower half of arg
- (inst add result temp)
-
- (inst mov temp result)
+ (move temp arg)
+ (inst shr result 1)
+ (inst mov mask #x5555555555555555)
+ (inst and result mask)
+ (inst sub temp result)
+ ;; Calculate 4-bit sums by straightforward shift, mask and add.
+ ;; Note that we shift the source operand of the MOV and not its
+ ;; destination so that the SHR and the MOV can execute in the same
+ ;; clock cycle.
+ (inst mov result temp)
(inst shr temp 2)
- (inst and result #x33333333)
- (inst and temp #x33333333)
- (inst add result temp)
-
- (inst mov temp result)
- (inst shr temp 4)
- (inst and result #x0f0f0f0f)
- (inst and temp #x0f0f0f0f)
+ (inst mov mask #x3333333333333333)
+ (inst and result mask)
+ (inst and temp mask)
(inst add result temp)
-
+ ;; Calculate 8-bit sums. Since each sum is at most 8, which fits
+ ;; into 4 bits, we can apply the mask after the addition, saving one
+ ;; instruction.
(inst mov temp result)
- (inst shr temp 8)
- (inst and result #x00ff00ff)
- (inst and temp #x00ff00ff)
+ (inst shr result 4)
(inst add result temp)
-
- (inst mov temp result)
- (inst shr temp 16)
- (inst and result #x0000ffff)
- (inst and temp #x0000ffff)
- (inst add result temp)
-
- ;;; now do the upper half
- (inst shr t1 32)
-
- (inst mov temp t1)
- (inst shr temp 1)
- (inst and t1 #x55555555)
- (inst and temp #x55555555)
- (inst add t1 temp)
-
- (inst mov temp t1)
- (inst shr temp 2)
- (inst and t1 #x33333333)
- (inst and temp #x33333333)
- (inst add t1 temp)
-
- (inst mov temp t1)
- (inst shr temp 4)
- (inst and t1 #x0f0f0f0f)
- (inst and temp #x0f0f0f0f)
- (inst add t1 temp)
-
- (inst mov temp t1)
- (inst shr temp 8)
- (inst and t1 #x00ff00ff)
- (inst and temp #x00ff00ff)
- (inst add t1 temp)
-
- (inst mov temp t1)
- (inst shr temp 16)
- (inst and t1 #x0000ffff)
- (inst and temp #x0000ffff)
- (inst add t1 temp)
- (inst add result t1)))
-
-
+ (inst mov mask #x0f0f0f0f0f0f0f0f)
+ (inst and result mask)
+ ;; Add all 8 bytes at once by multiplying with #256r11111111.
+ ;; We need to calculate only the lower 8 bytes of the product.
+ ;; Of these the most significant byte contains the final result.
+ ;; Note that there can be no overflow from one byte to the next
+ ;; as the sum is at most 64 which needs only 7 bits.
+ (inst mov mask #x0101010101010101)
+ (inst imul result mask)
+ (inst shr result 56)))
\f
;;;; binary conditional VOPs
(:translate logcount)
(:note "inline (unsigned-byte 32) logcount")
(:policy :fast-safe)
- (:args (arg :scs (unsigned-reg)))
+ (:args (arg :scs (unsigned-reg) :target result))
(:arg-types unsigned-num)
(:results (result :scs (unsigned-reg)))
(:result-types positive-fixnum)
- (:temporary (:sc unsigned-reg :from (:argument 0)) temp)
- (:generator 30
+ (:temporary (:sc unsigned-reg) temp)
+ (:generator 14
+ ;; See the comments below for how the algorithm works. The tricks
+ ;; used can be found for example in AMD's software optimization
+ ;; guide or at "http://www.hackersdelight.org/HDcode/pop.cc" in the
+ ;; function "pop1".
+ ;; Calculate 2-bit sums. Note that the value of a two-digit binary
+ ;; number is the sum of the right digit and twice the left digit.
+ ;; Thus we can calculate the sum of the two digits by shifting the
+ ;; left digit to the right position and doing a two-bit subtraction.
+ ;; This subtraction will never create a borrow and thus can be made
+ ;; on all 16 2-digit numbers at once.
(move result arg)
-
- (inst mov temp result)
- (inst shr temp 1)
+ (move temp arg)
+ (inst shr result 1)
(inst and result #x55555555)
- (inst and temp #x55555555)
- (inst add result temp)
-
- (inst mov temp result)
+ (inst sub temp result)
+ ;; Calculate 4-bit sums by straightforward shift, mask and add.
+ ;; Note that we shift the source operand of the MOV and not its
+ ;; destination so that the SHR and the MOV can execute in the same
+ ;; clock cycle.
+ (inst mov result temp)
(inst shr temp 2)
(inst and result #x33333333)
(inst and temp #x33333333)
(inst add result temp)
-
+ ;; Calculate 8-bit sums. Since each sum is at most 8, which fits
+ ;; into 4 bits, we can apply the mask after the addition, saving one
+ ;; instruction.
(inst mov temp result)
- (inst shr temp 4)
- (inst and result #x0f0f0f0f)
- (inst and temp #x0f0f0f0f)
+ (inst shr result 4)
(inst add result temp)
-
+ (inst and result #x0f0f0f0f)
+ ;; Calculate the two 16-bit sums and the 32-bit sum. No masking is
+ ;; necessary inbetween since the final sum is at most 32 which fits
+ ;; into 6 bits.
(inst mov temp result)
- (inst shr temp 8)
- (inst and result #x00ff00ff)
- (inst and temp #x00ff00ff)
+ (inst shr result 8)
(inst add result temp)
-
(inst mov temp result)
- (inst shr temp 16)
- (inst and result #x0000ffff)
- (inst and temp #x0000ffff)
- (inst add result temp)))
+ (inst shr result 16)
+ (inst add result temp)
+ (inst and result #xff)))
\f
;;;; binary conditional VOPs
(funcall (lambda ()
(declare (notinline logxor))
(min (logxor 0 0 0 286142502))))))
+
+;; Small bugs in LOGCOUNT can still allow SBCL to be built and thus go
+;; unnoticed, so check more thoroughly here.
+(with-test (:name :logcount)
+ (flet ((test (x n)
+ (unless (= (logcount x) n)
+ (error "logcount failure for ~a" x))))
+ ;; Test with some patterns with well known number of ones/zeroes ...
+ (dotimes (i 128)
+ (let ((x (ash 1 i)))
+ (test x 1)
+ (test (- x) i)
+ (test (1- x) i)))
+ ;; ... and with some random integers of varying length.
+ (flet ((test-logcount (x)
+ (declare (type integer x))
+ (do ((result 0 (1+ result))
+ (x (if (minusp x)
+ (lognot x)
+ x)
+ (logand x (1- x))))
+ ((zerop x) result))))
+ (dotimes (i 200)
+ (let ((x (random (ash 1 i))))
+ (test x (test-logcount x))
+ (test (- x) (test-logcount (- x))))))))
;;; 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".)
-"0.9.15"
+"0.9.15.1"
projects / sbcl.git / commitdiff