-;;;; the VM definition arithmetic VOPs for the x86
+;;;; the VM definition of arithmetic VOPs for the x86
;;;; This software is part of the SBCL system. See the README file for
;;;; more information.
(:translate ,translate)
(:generator ,untagged-penalty
(move r x)
- (inst ,op r y))))))
-
- ;;(define-binop + 4 add)
+ ,(if (eq translate 'logand)
+ ;; for the -C/UNSIGNED=>UNSIGNED VOP, this case
+ ;; is optimized away as an identity somewhere
+ ;; along the lines. However, this VOP is used in
+ ;; -C/SIGNED=>UNSIGNED, below, when the
+ ;; higher-level lisp code can't optimize away the
+ ;; non-trivial identity.
+ `(unless (= y #.(1- (ash 1 n-word-bits)))
+ (inst ,op r y))
+ `(inst ,op r y)))))))
(define-binop - 4 sub)
(define-binop logand 2 and)
(define-binop logior 2 or)
(define-binop logxor 2 xor))
-
;;; Special handling of add on the x86; can use lea to avoid a
;;; register load, otherwise it uses add.
(define-vop (fast-+/fixnum=>fixnum fast-safe-arith-op)
(move r x)
(inst add r y)))))
-
;;;; Special logand cases: (logand signed unsigned) => unsigned
(define-vop (fast-logand/signed-unsigned=>unsigned
;; The result-type ensures us that this shift will not overflow.
(inst shl result :cl)))
-(define-vop (fast-ash-c)
+(define-vop (fast-ash-c/signed=>signed)
(:translate ash)
(:policy :fast-safe)
- (:args (number :scs (signed-reg unsigned-reg) :target result
- :load-if (not (and (sc-is number signed-stack unsigned-stack)
- (sc-is result signed-stack unsigned-stack)
+ (:args (number :scs (signed-reg) :target result
+ :load-if (not (and (sc-is number signed-stack)
+ (sc-is result signed-stack)
(location= number result)))))
(:info amount)
- (:arg-types (:or signed-num unsigned-num) (:constant integer))
- (:results (result :scs (signed-reg unsigned-reg)
- :load-if (not
- (and (sc-is number signed-stack unsigned-stack)
- (sc-is result signed-stack unsigned-stack)
- (location= number result)))))
- (:result-types (:or signed-num unsigned-num))
+ (:arg-types signed-num (:constant integer))
+ (:results (result :scs (signed-reg)
+ :load-if (not (and (sc-is number signed-stack)
+ (sc-is result signed-stack)
+ (location= number result)))))
+ (:result-types signed-num)
(:note "inline ASH")
(:generator 3
(cond ((and (= amount 1) (not (location= number result)))
(inst lea result (make-ea :dword :index number :scale 8)))
(t
(move result number)
- (cond ((plusp amount)
- ;; We don't have to worry about overflow because of the
- ;; result type restriction.
- (inst shl result amount))
- ((sc-is number signed-reg signed-stack)
- ;; If the amount is greater than 31, only shift by 31. We
- ;; have to do this because the shift instructions only look
- ;; at the low five bits of the result.
- (inst sar result (min 31 (- amount))))
- (t
- (inst shr result (min 31 (- amount)))))))))
+ (cond ((plusp amount) (inst shl result amount))
+ (t (inst sar result (min 31 (- amount)))))))))
+
+(define-vop (fast-ash-c/unsigned=>unsigned)
+ (:translate ash)
+ (:policy :fast-safe)
+ (:args (number :scs (unsigned-reg) :target result
+ :load-if (not (and (sc-is number unsigned-stack)
+ (sc-is result unsigned-stack)
+ (location= number result)))))
+ (:info amount)
+ (:arg-types unsigned-num (:constant integer))
+ (:results (result :scs (unsigned-reg)
+ :load-if (not (and (sc-is number unsigned-stack)
+ (sc-is result unsigned-stack)
+ (location= number result)))))
+ (:result-types unsigned-num)
+ (:note "inline ASH")
+ (:generator 3
+ (cond ((and (= amount 1) (not (location= number result)))
+ (inst lea result (make-ea :dword :index number :scale 2)))
+ ((and (= amount 2) (not (location= number result)))
+ (inst lea result (make-ea :dword :index number :scale 4)))
+ ((and (= amount 3) (not (location= number result)))
+ (inst lea result (make-ea :dword :index number :scale 8)))
+ (t
+ (move result number)
+ (cond ((< -32 amount 32)
+ ;; this code is used both in ASH and ASH-MOD32, so
+ ;; be careful
+ (if (plusp amount)
+ (inst shl result amount)
+ (inst shr result (- amount))))
+ (t (if (sc-is result unsigned-reg)
+ (inst xor result result)
+ (inst mov result 0))))))))
+
+(define-vop (fast-ash-left/signed=>signed)
+ (:translate ash)
+ (:args (number :scs (signed-reg) :target result
+ :load-if (not (and (sc-is number signed-stack)
+ (sc-is result signed-stack)
+ (location= number result))))
+ (amount :scs (unsigned-reg) :target ecx))
+ (:arg-types signed-num positive-fixnum)
+ (:temporary (:sc unsigned-reg :offset ecx-offset :from (:argument 1)) ecx)
+ (:results (result :scs (signed-reg) :from (:argument 0)
+ :load-if (not (and (sc-is number signed-stack)
+ (sc-is result signed-stack)
+ (location= number result)))))
+ (:result-types signed-num)
+ (:policy :fast-safe)
+ (:note "inline ASH")
+ (:generator 4
+ (move result number)
+ (move ecx amount)
+ (inst shl result :cl)))
-(define-vop (fast-ash-left)
+(define-vop (fast-ash-left/unsigned=>unsigned)
(:translate ash)
- (:args (number :scs (signed-reg unsigned-reg) :target result
- :load-if (not (and (sc-is number signed-stack unsigned-stack)
- (sc-is result signed-stack unsigned-stack)
+ (:args (number :scs (unsigned-reg) :target result
+ :load-if (not (and (sc-is number unsigned-stack)
+ (sc-is result unsigned-stack)
(location= number result))))
(amount :scs (unsigned-reg) :target ecx))
- (:arg-types (:or signed-num unsigned-num) positive-fixnum)
+ (:arg-types unsigned-num positive-fixnum)
(:temporary (:sc unsigned-reg :offset ecx-offset :from (:argument 1)) ecx)
- (:results (result :scs (signed-reg unsigned-reg) :from (:argument 0)
- :load-if (not
- (and (sc-is number signed-stack unsigned-stack)
- (sc-is result signed-stack unsigned-stack)
- (location= number result)))))
- (:result-types (:or signed-num unsigned-num))
+ (:results (result :scs (unsigned-reg) :from (:argument 0)
+ :load-if (not (and (sc-is number unsigned-stack)
+ (sc-is result unsigned-stack)
+ (location= number result)))))
+ (:result-types unsigned-num)
(:policy :fast-safe)
(:note "inline ASH")
(:generator 4
(move result number)
(move ecx amount)
- ;; The result-type ensures us that this shift will not overflow.
(inst shl result :cl)))
-(define-vop (fast-ash)
+(define-vop (fast-ash/signed=>signed)
(:translate ash)
(:policy :fast-safe)
- (:args (number :scs (signed-reg unsigned-reg) :target result)
+ (:args (number :scs (signed-reg) :target result)
(amount :scs (signed-reg) :target ecx))
- (:arg-types (:or signed-num unsigned-num) signed-num)
- (:results (result :scs (signed-reg unsigned-reg) :from (:argument 0)))
- (:result-types (:or signed-num unsigned-num))
+ (:arg-types signed-num signed-num)
+ (:results (result :scs (signed-reg) :from (:argument 0)))
+ (:result-types signed-num)
(:temporary (:sc signed-reg :offset ecx-offset :from (:argument 1)) ecx)
(:note "inline ASH")
(:generator 5
(move result number)
- (move ecx amount)
+ (move ecx amount)
(inst or ecx ecx)
(inst jmp :ns positive)
(inst neg ecx)
(inst jmp :be okay)
(inst mov ecx 31)
OKAY
- (sc-case number
- (signed-reg (inst sar result :cl))
- (unsigned-reg (inst shr result :cl)))
+ (inst sar result :cl)
(inst jmp done)
POSITIVE
(inst shl result :cl)
DONE))
+
+(define-vop (fast-ash/unsigned=>unsigned)
+ (:translate ash)
+ (:policy :fast-safe)
+ (:args (number :scs (unsigned-reg) :target result)
+ (amount :scs (signed-reg) :target ecx))
+ (:arg-types unsigned-num signed-num)
+ (:results (result :scs (unsigned-reg) :from (:argument 0)))
+ (:result-types unsigned-num)
+ (:temporary (:sc signed-reg :offset ecx-offset :from (:argument 1)) ecx)
+ (:note "inline ASH")
+ (:generator 5
+ (move result number)
+ (move ecx amount)
+ (inst or ecx ecx)
+ (inst jmp :ns positive)
+ (inst neg ecx)
+ (inst cmp ecx 31)
+ (inst jmp :be okay)
+ (inst xor result result)
+ (inst jmp done)
+ OKAY
+ (inst shr result :cl)
+ (inst jmp done)
+
+ POSITIVE
+ ;; The result-type ensures us that this shift will not overflow.
+ (inst shl result :cl)
+
+ DONE))
+
+(in-package "SB!C")
+
+(defknown %lea (integer integer (member 1 2 4 8) (signed-byte 32))
+ integer
+ (foldable flushable movable))
+
+(defoptimizer (%lea derive-type) ((base index scale disp))
+ (when (and (constant-lvar-p scale)
+ (constant-lvar-p disp))
+ (let ((scale (lvar-value scale))
+ (disp (lvar-value disp))
+ (base-type (lvar-type base))
+ (index-type (lvar-type index)))
+ (when (and (numeric-type-p base-type)
+ (numeric-type-p index-type))
+ (let ((base-lo (numeric-type-low base-type))
+ (base-hi (numeric-type-high base-type))
+ (index-lo (numeric-type-low index-type))
+ (index-hi (numeric-type-high index-type)))
+ (make-numeric-type :class 'integer
+ :complexp :real
+ :low (when (and base-lo index-lo)
+ (+ base-lo (* index-lo scale) disp))
+ :high (when (and base-hi index-hi)
+ (+ base-hi (* index-hi scale) disp))))))))
+
+(defun %lea (base index scale disp)
+ (+ base (* index scale) disp))
+
+(in-package "SB!VM")
+
+(define-vop (%lea/unsigned=>unsigned)
+ (:translate %lea)
+ (:policy :fast-safe)
+ (:args (base :scs (unsigned-reg))
+ (index :scs (unsigned-reg)))
+ (:info scale disp)
+ (:arg-types unsigned-num unsigned-num
+ (:constant (member 1 2 4 8))
+ (:constant (signed-byte 32)))
+ (:results (r :scs (unsigned-reg)))
+ (:result-types unsigned-num)
+ (:generator 5
+ (inst lea r (make-ea :dword :base base :index index
+ :scale scale :disp disp))))
+
+(define-vop (%lea/signed=>signed)
+ (:translate %lea)
+ (:policy :fast-safe)
+ (:args (base :scs (signed-reg))
+ (index :scs (signed-reg)))
+ (:info scale disp)
+ (:arg-types signed-num signed-num
+ (:constant (member 1 2 4 8))
+ (:constant (signed-byte 32)))
+ (:results (r :scs (signed-reg)))
+ (:result-types signed-num)
+ (:generator 4
+ (inst lea r (make-ea :dword :base base :index index
+ :scale scale :disp disp))))
+
+(define-vop (%lea/fixnum=>fixnum)
+ (:translate %lea)
+ (:policy :fast-safe)
+ (:args (base :scs (any-reg))
+ (index :scs (any-reg)))
+ (:info scale disp)
+ (:arg-types tagged-num tagged-num
+ (:constant (member 1 2 4 8))
+ (:constant (signed-byte 32)))
+ (:results (r :scs (any-reg)))
+ (:result-types tagged-num)
+ (:generator 3
+ (inst lea r (make-ea :dword :base base :index index
+ :scale scale :disp disp))))
+
+;;; FIXME: before making knowledge of this too public, it needs to be
+;;; fixed so that it's actually _faster_ than the non-CMOV version; at
+;;; least on my Celeron-XXX laptop, this version is marginally slower
+;;; than the above version with branches. -- CSR, 2003-09-04
+(define-vop (fast-cmov-ash/unsigned=>unsigned)
+ (:translate ash)
+ (:policy :fast-safe)
+ (:args (number :scs (unsigned-reg) :target result)
+ (amount :scs (signed-reg) :target ecx))
+ (:arg-types unsigned-num signed-num)
+ (:results (result :scs (unsigned-reg) :from (:argument 0)))
+ (:result-types unsigned-num)
+ (:temporary (:sc signed-reg :offset ecx-offset :from (:argument 1)) ecx)
+ (:temporary (:sc any-reg :from (:eval 0) :to (:eval 1)) zero)
+ (:note "inline ASH")
+ (:guard (member :cmov *backend-subfeatures*))
+ (:generator 4
+ (move result number)
+ (move ecx amount)
+ (inst or ecx ecx)
+ (inst jmp :ns positive)
+ (inst neg ecx)
+ (inst xor zero zero)
+ (inst shr result :cl)
+ (inst cmp ecx 31)
+ (inst cmov :nbe result zero)
+ (inst jmp done)
+
+ POSITIVE
+ ;; The result-type ensures us that this shift will not overflow.
+ (inst shl result :cl)
+
+ DONE))
\f
;;; Note: documentation for this function is wrong - rtfm
(define-vop (signed-byte-32-len)
(move result prev)
(inst shrd result next :cl)))
-(define-source-transform 32bit-logical-not (x)
+(define-source-transform word-logical-not (x)
`(logand (lognot (the (unsigned-byte 32) ,x)) #.(1- (ash 1 32))))
-(deftransform 32bit-logical-and ((x y))
+(deftransform word-logical-and ((x y))
'(logand x y))
-(define-source-transform 32bit-logical-nand (x y)
- `(32bit-logical-not (32bit-logical-and ,x ,y)))
+(define-source-transform word-logical-nand (x y)
+ `(word-logical-not (word-logical-and ,x ,y)))
-(deftransform 32bit-logical-or ((x y))
+(deftransform word-logical-or ((x y))
'(logior x y))
-(define-source-transform 32bit-logical-nor (x y)
- `(32bit-logical-not (32bit-logical-or ,x ,y)))
+(define-source-transform word-logical-nor (x y)
+ `(word-logical-not (word-logical-or ,x ,y)))
-(deftransform 32bit-logical-xor ((x y))
+(deftransform word-logical-xor ((x y))
'(logxor x y))
-(define-source-transform 32bit-logical-eqv (x y)
- `(32bit-logical-not (32bit-logical-xor ,x ,y)))
+(define-source-transform word-logical-eqv (x y)
+ `(word-logical-not (word-logical-xor ,x ,y)))
-(define-source-transform 32bit-logical-orc1 (x y)
- `(32bit-logical-or (32bit-logical-not ,x) ,y))
+(define-source-transform word-logical-orc1 (x y)
+ `(word-logical-or (word-logical-not ,x) ,y))
-(define-source-transform 32bit-logical-orc2 (x y)
- `(32bit-logical-or ,x (32bit-logical-not ,y)))
+(define-source-transform word-logical-orc2 (x y)
+ `(word-logical-or ,x (word-logical-not ,y)))
-(define-source-transform 32bit-logical-andc1 (x y)
- `(32bit-logical-and (32bit-logical-not ,x) ,y))
+(define-source-transform word-logical-andc1 (x y)
+ `(word-logical-and (word-logical-not ,x) ,y))
-(define-source-transform 32bit-logical-andc2 (x y)
- `(32bit-logical-and ,x (32bit-logical-not ,y)))
+(define-source-transform word-logical-andc2 (x y)
+ `(word-logical-and ,x (word-logical-not ,y)))
;;; Only the lower 5 bits of the shift amount are significant.
(define-vop (shift-towards-someplace)
(:translate +-mod32))
(define-vop (fast-+-mod32-c/unsigned=>unsigned fast-+-c/unsigned=>unsigned)
(:translate +-mod32))
+(define-modular-fun --mod32 (x y) - 32)
+(define-vop (fast---mod32/unsigned=>unsigned fast--/unsigned=>unsigned)
+ (:translate --mod32))
+(define-vop (fast---mod32-c/unsigned=>unsigned fast---c/unsigned=>unsigned)
+ (:translate --mod32))
+
+(define-modular-fun *-mod32 (x y) * 32)
+(define-vop (fast-*-mod32/unsigned=>unsigned fast-*/unsigned=>unsigned)
+ (:translate *-mod32))
+;;; (no -C variant as x86 MUL instruction doesn't take an immediate)
+
+(define-vop (fast-ash-left-mod32-c/unsigned=>unsigned
+ fast-ash-c/unsigned=>unsigned)
+ (:translate ash-left-mod32))
+
+(in-package "SB!C")
+
+(defknown sb!vm::%lea-mod32 (integer integer (member 1 2 4 8) (signed-byte 32))
+ (unsigned-byte 32)
+ (foldable flushable movable))
+
+(define-modular-fun-optimizer %lea ((base index scale disp) :width width)
+ (when (and (<= width 32)
+ (constant-lvar-p scale)
+ (constant-lvar-p disp))
+ (cut-to-width base width)
+ (cut-to-width index width)
+ 'sb!vm::%lea-mod32))
+
+#+sb-xc-host
+(defun sb!vm::%lea-mod32 (base index scale disp)
+ (ldb (byte 32 0) (%lea base index scale disp)))
+#-sb-xc-host
+(defun sb!vm::%lea-mod32 (base index scale disp)
+ (let ((base (logand base #xffffffff))
+ (index (logand index #xffffffff)))
+ ;; can't use modular version of %LEA, as we only have VOPs for
+ ;; constant SCALE and DISP.
+ (ldb (byte 32 0) (+ base (* index scale) disp))))
+
+(in-package "SB!VM")
+
+(define-vop (%lea-mod32/unsigned=>unsigned
+ %lea/unsigned=>unsigned)
+ (:translate %lea-mod32))
;;; logical operations
(define-modular-fun lognot-mod32 (x) lognot 32)
(define-vop (fast-logxor-mod32-c/unsigned=>unsigned
fast-logxor-c/unsigned=>unsigned)
(:translate logxor-mod32))
+
+(define-source-transform logeqv (&rest args)
+ (if (oddp (length args))
+ `(logxor ,@args)
+ `(lognot (logxor ,@args))))
+(define-source-transform logandc1 (x y)
+ `(logand (lognot ,x) ,y))
+(define-source-transform logandc2 (x y)
+ `(logand ,x (lognot ,y)))
+(define-source-transform logorc1 (x y)
+ `(logior (lognot ,x) ,y))
+(define-source-transform logorc2 (x y)
+ `(logior ,x (lognot ,y)))
+(define-source-transform lognor (x y)
+ `(lognot (logior ,x ,y)))
+(define-source-transform lognand (x y)
+ `(lognot (logand ,x ,y)))
\f
;;;; bignum stuff
(define-vop (bignum-length get-header-data)
- (:translate sb!bignum::%bignum-length)
+ (:translate sb!bignum:%bignum-length)
(:policy :fast-safe))
(define-vop (bignum-set-length set-header-data)
- (:translate sb!bignum::%bignum-set-length)
+ (:translate sb!bignum:%bignum-set-length)
(:policy :fast-safe))
(define-full-reffer bignum-ref * bignum-digits-offset other-pointer-lowtag
- (unsigned-reg) unsigned-num sb!bignum::%bignum-ref)
+ (unsigned-reg) unsigned-num sb!bignum:%bignum-ref)
(define-full-setter bignum-set * bignum-digits-offset other-pointer-lowtag
- (unsigned-reg) unsigned-num sb!bignum::%bignum-set)
+ (unsigned-reg) unsigned-num sb!bignum:%bignum-set)
(define-vop (digit-0-or-plus)
- (:translate sb!bignum::%digit-0-or-plusp)
+ (:translate sb!bignum:%digit-0-or-plusp)
(:policy :fast-safe)
(:args (digit :scs (unsigned-reg)))
(:arg-types unsigned-num)
;;; 4. This is easy to deal with and may save a fixnum-word
;;; conversion.
(define-vop (add-w/carry)
- (:translate sb!bignum::%add-with-carry)
+ (:translate sb!bignum:%add-with-carry)
(:policy :fast-safe)
(:args (a :scs (unsigned-reg) :target result)
(b :scs (unsigned-reg unsigned-stack) :to :eval)
;;; Note: the borrow is the oppostite of the x86 convention - 1 for no
;;; borrow and 0 for a borrow.
(define-vop (sub-w/borrow)
- (:translate sb!bignum::%subtract-with-borrow)
+ (:translate sb!bignum:%subtract-with-borrow)
(:policy :fast-safe)
(:args (a :scs (unsigned-reg) :to :eval :target result)
(b :scs (unsigned-reg unsigned-stack) :to :result)
(define-vop (bignum-mult-and-add-3-arg)
- (:translate sb!bignum::%multiply-and-add)
+ (:translate sb!bignum:%multiply-and-add)
(:policy :fast-safe)
(:args (x :scs (unsigned-reg) :target eax)
(y :scs (unsigned-reg unsigned-stack))
(move lo eax)))
(define-vop (bignum-mult-and-add-4-arg)
- (:translate sb!bignum::%multiply-and-add)
+ (:translate sb!bignum:%multiply-and-add)
(:policy :fast-safe)
(:args (x :scs (unsigned-reg) :target eax)
(y :scs (unsigned-reg unsigned-stack))
(define-vop (bignum-mult)
- (:translate sb!bignum::%multiply)
+ (:translate sb!bignum:%multiply)
(:policy :fast-safe)
(:args (x :scs (unsigned-reg) :target eax)
(y :scs (unsigned-reg unsigned-stack)))
(move lo eax)))
(define-vop (bignum-lognot lognot-mod32/unsigned=>unsigned)
- (:translate sb!bignum::%lognot))
+ (:translate sb!bignum:%lognot))
(define-vop (fixnum-to-digit)
- (:translate sb!bignum::%fixnum-to-digit)
+ (:translate sb!bignum:%fixnum-to-digit)
(:policy :fast-safe)
(:args (fixnum :scs (any-reg control-stack) :target digit))
(:arg-types tagged-num)
(inst sar digit 2)))
(define-vop (bignum-floor)
- (:translate sb!bignum::%floor)
+ (:translate sb!bignum:%floor)
(:policy :fast-safe)
(:args (div-high :scs (unsigned-reg) :target edx)
(div-low :scs (unsigned-reg) :target eax)
(move rem edx)))
(define-vop (signify-digit)
- (:translate sb!bignum::%fixnum-digit-with-correct-sign)
+ (:translate sb!bignum:%fixnum-digit-with-correct-sign)
(:policy :fast-safe)
(:args (digit :scs (unsigned-reg unsigned-stack) :target res))
(:arg-types unsigned-num)
(inst shl res 2))))
(define-vop (digit-ashr)
- (:translate sb!bignum::%ashr)
+ (:translate sb!bignum:%ashr)
(:policy :fast-safe)
(:args (digit :scs (unsigned-reg unsigned-stack) :target result)
(count :scs (unsigned-reg) :target ecx))
(inst sar result :cl)))
(define-vop (digit-lshr digit-ashr)
- (:translate sb!bignum::%digit-logical-shift-right)
+ (:translate sb!bignum:%digit-logical-shift-right)
(:generator 1
(move result digit)
(move ecx count)
(inst shr result :cl)))
(define-vop (digit-ashl digit-ashr)
- (:translate sb!bignum::%ashl)
+ (:translate sb!bignum:%ashl)
(:generator 1
(move result digit)
(move ecx count)
(inst mov tmp y)
(inst shr tmp 18)
(inst xor y tmp)))
+
+(in-package "SB!C")
+
+;;; This is essentially a straight implementation of the algorithm in
+;;; "Strength Reduction of Multiplications by Integer Constants",
+;;; Youfeng Wu, ACM SIGPLAN Notices, Vol. 30, No.2, February 1995.
+(defun basic-decompose-multiplication (arg num n-bits condensed)
+ (case (aref condensed 0)
+ (0
+ (let ((tmp (min 3 (aref condensed 1))))
+ (decf (aref condensed 1) tmp)
+ `(logand #xffffffff
+ (%lea ,arg
+ ,(decompose-multiplication
+ arg (ash (1- num) (- tmp)) (1- n-bits) (subseq condensed 1))
+ ,(ash 1 tmp) 0))))
+ ((1 2 3)
+ (let ((r0 (aref condensed 0)))
+ (incf (aref condensed 1) r0)
+ `(logand #xffffffff
+ (%lea ,(decompose-multiplication
+ arg (- num (ash 1 r0)) (1- n-bits) (subseq condensed 1))
+ ,arg
+ ,(ash 1 r0) 0))))
+ (t (let ((r0 (aref condensed 0)))
+ (setf (aref condensed 0) 0)
+ `(logand #xffffffff
+ (ash ,(decompose-multiplication
+ arg (ash num (- r0)) n-bits condensed)
+ ,r0))))))
+
+(defun decompose-multiplication (arg num n-bits condensed)
+ (cond
+ ((= n-bits 0) 0)
+ ((= num 1) arg)
+ ((= n-bits 1)
+ `(logand #xffffffff (ash ,arg ,(1- (integer-length num)))))
+ ((let ((max 0) (end 0))
+ (loop for i from 2 to (length condensed)
+ for j = (reduce #'+ (subseq condensed 0 i))
+ when (and (> (- (* 2 i) 3 j) max)
+ (< (+ (ash 1 (1+ j))
+ (ash (ldb (byte (- 32 (1+ j)) (1+ j)) num)
+ (1+ j)))
+ (ash 1 32)))
+ do (setq max (- (* 2 i) 3 j)
+ end i))
+ (when (> max 0)
+ (let ((j (reduce #'+ (subseq condensed 0 end))))
+ (let ((n2 (+ (ash 1 (1+ j))
+ (ash (ldb (byte (- 32 (1+ j)) (1+ j)) num) (1+ j))))
+ (n1 (1+ (ldb (byte (1+ j) 0) (lognot num)))))
+ `(logand #xffffffff
+ (- ,(optimize-multiply arg n2) ,(optimize-multiply arg n1))))))))
+ ((dolist (i '(9 5 3))
+ (when (integerp (/ num i))
+ (when (< (logcount (/ num i)) (logcount num))
+ (let ((x (gensym)))
+ (return `(let ((,x ,(optimize-multiply arg (/ num i))))
+ (logand #xffffffff
+ (%lea ,x ,x (1- ,i) 0)))))))))
+ (t (basic-decompose-multiplication arg num n-bits condensed))))
+
+(defun optimize-multiply (arg x)
+ (let* ((n-bits (logcount x))
+ (condensed (make-array n-bits)))
+ (let ((count 0) (bit 0))
+ (dotimes (i 32)
+ (cond ((logbitp i x)
+ (setf (aref condensed bit) count)
+ (setf count 1)
+ (incf bit))
+ (t (incf count)))))
+ (decompose-multiplication arg x n-bits condensed)))
+
+(defun *-transformer (y)
+ (cond
+ ((= y (ash 1 (integer-length y)))
+ ;; there's a generic transform for y = 2^k
+ (give-up-ir1-transform))
+ ((member y '(3 5 9))
+ ;; we can do these multiplications directly using LEA
+ `(%lea x x ,(1- y) 0))
+ ((member :pentium4 *backend-subfeatures*)
+ ;; the pentium4's multiply unit is reportedly very good
+ (give-up-ir1-transform))
+ ;; FIXME: should make this more fine-grained. If nothing else,
+ ;; there should probably be a cutoff of about 9 instructions on
+ ;; pentium-class machines.
+ (t (optimize-multiply 'x y))))
+
+(deftransform * ((x y)
+ ((unsigned-byte 32) (constant-arg (unsigned-byte 32)))
+ (unsigned-byte 32))
+ "recode as leas, shifts and adds"
+ (let ((y (lvar-value y)))
+ (*-transformer y)))
+
+(deftransform sb!vm::*-mod32
+ ((x y) ((unsigned-byte 32) (constant-arg (unsigned-byte 32)))
+ (unsigned-byte 32))
+ "recode as leas, shifts and adds"
+ (let ((y (lvar-value y)))
+ (*-transformer y)))
+
+;;; FIXME: we should also be able to write an optimizer or two to
+;;; convert (+ (* x 2) 17), (- (* x 9) 5) to a %LEA.
projects / sbcl.git / blobdiff