-;;;; the VM definition of arithmetic VOPs for the x86
+;;;; the VM definition of arithmetic VOPs for the x86-64
;;;; This software is part of the SBCL system. See the README file for
;;;; more information.
(define-vop (fast-lognot/fixnum fixnum-unop)
(:translate lognot)
- (:generator 2
+ (:generator 1
(move res x)
(inst xor res (fixnumize -1))))
(define-vop (fast-lognot/signed signed-unop)
(:translate lognot)
- (:generator 1
+ (:generator 2
(move res x)
(inst not res)))
\f
(:args (x :scs (unsigned-reg) :target eax)
(y :scs (unsigned-reg unsigned-stack)))
(:arg-types unsigned-num unsigned-num)
- (:temporary (:sc unsigned-reg :offset eax-offset :target result
+ (:temporary (:sc unsigned-reg :offset eax-offset :target r
:from (:argument 0) :to :result) eax)
(:temporary (:sc unsigned-reg :offset edx-offset
:from :eval :to :result) edx)
(:ignore edx)
- (:results (result :scs (unsigned-reg)))
+ (:results (r :scs (unsigned-reg)))
(:result-types unsigned-num)
(:note "inline (unsigned-byte 64) arithmetic")
(:vop-var vop)
(:generator 6
(move eax x)
(inst mul eax y)
- (move result eax)))
+ (move r eax)))
(define-vop (fast-truncate/fixnum=>fixnum fast-safe-arith-op)
(:vop-var vop)
(:save-p :compute-only)
(:generator 31
- (let ((zero (generate-error-code vop division-by-zero-error x y)))
+ (let ((zero (generate-error-code vop 'division-by-zero-error x y)))
(if (sc-is y any-reg)
(inst test y y) ; smaller instruction
(inst cmp y 0))
(:vop-var vop)
(:save-p :compute-only)
(:generator 33
- (let ((zero (generate-error-code vop division-by-zero-error x y)))
+ (let ((zero (generate-error-code vop 'division-by-zero-error x y)))
(if (sc-is y unsigned-reg)
(inst test y y) ; smaller instruction
(inst cmp y 0))
(:vop-var vop)
(:save-p :compute-only)
(:generator 33
- (let ((zero (generate-error-code vop division-by-zero-error x y)))
+ (let ((zero (generate-error-code vop 'division-by-zero-error x y)))
(if (sc-is y signed-reg)
(inst test y y) ; smaller instruction
(inst cmp y 0))
(:note "inline ASH")
(:generator 2
(cond ((and (= amount 1) (not (location= number result)))
- (inst lea result (make-ea :qword :index number :scale 2)))
+ (inst lea result (make-ea :qword :base number :index number)))
((and (= amount 2) (not (location= number result)))
(inst lea result (make-ea :qword :index number :scale 4)))
((and (= amount 3) (not (location= number result)))
(inst lea result (make-ea :qword :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))
- (t
- ;; Since the shift instructions take the shift amount
- ;; modulo 64 we must special case amounts of 64 and more.
- ;; Because fixnums have only 61 bits, the result is 0 or
- ;; -1 for all amounts of 60 or more, so use this as the
- ;; limit instead.
- (inst sar result (min (- n-word-bits n-fixnum-tag-bits 1)
- (- amount)))
- (inst and result (lognot fixnum-tag-mask))))))))
+ (cond ((< -64 amount 64)
+ ;; this code is used both in ASH and ASH-SMOD61, so
+ ;; be careful
+ (if (plusp amount)
+ (inst shl result amount)
+ (progn
+ (inst sar result (- amount))
+ (inst and result (lognot fixnum-tag-mask)))))
+ ((plusp amount)
+ (if (sc-is result any-reg)
+ (inst xor result result)
+ (inst mov result 0)))
+ (t (inst sar result 63)
+ (inst and result (lognot fixnum-tag-mask))))))))
(define-vop (fast-ash-left/fixnum=>fixnum)
(:translate ash)
(:note "inline ASH")
(:generator 3
(cond ((and (= amount 1) (not (location= number result)))
- (inst lea result (make-ea :qword :index number :scale 2)))
+ (inst lea result (make-ea :qword :base number :index number)))
((and (= amount 2) (not (location= number result)))
(inst lea result (make-ea :qword :index number :scale 4)))
((and (= amount 3) (not (location= number result)))
(:note "inline ASH")
(:generator 3
(cond ((and (= amount 1) (not (location= number result)))
- (inst lea result (make-ea :qword :index number :scale 2)))
+ (inst lea result (make-ea :qword :base number :index number)))
((and (= amount 2) (not (location= number result)))
(inst lea result (make-ea :qword :index number :scale 4)))
((and (= amount 3) (not (location= number result)))
(inst shl result amount)
(inst shr result (- amount))))
(t (if (sc-is result unsigned-reg)
- (inst xor result result)
+ (zeroize result)
(inst mov result 0))))))))
(define-vop (fast-ash-left/signed=>signed)
(inst neg ecx)
(inst cmp ecx 63)
(inst jmp :be OKAY)
- (inst xor result result)
+ (zeroize result)
(inst jmp DONE)
OKAY
(inst shr result :cl)
(inst or ecx ecx)
(inst jmp :ns POSITIVE)
(inst neg ecx)
- (inst xor zero zero)
+ (zeroize zero)
(inst shr result :cl)
(inst cmp ecx 63)
(inst cmov :nbe result zero)
(:result-types unsigned-num)
(:generator 28
(move res arg)
- (inst cmp res 0)
+ (if (sc-is res unsigned-reg)
+ (inst test res res)
+ (inst cmp res 0))
(inst jmp :ge POS)
(inst not res)
POS
(inst inc res)
(inst jmp DONE)
ZERO
- (inst xor res res)
+ (zeroize res)
DONE))
(define-vop (unsigned-byte-64-len)
(inst inc res)
(inst jmp DONE)
ZERO
- (inst xor res res)
+ (zeroize 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 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 4)
- (inst and result #x0f0f0f0f)
- (inst and temp #x0f0f0f0f)
+ (inst shr result 4)
(inst add result temp)
-
- (inst mov temp result)
- (inst shr temp 8)
- (inst and result #x00ff00ff)
- (inst and temp #x00ff00ff)
- (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
(define-vop (fast-conditional)
- (:conditional)
- (:info target not-p)
+ (:conditional :e)
+ (:info)
(:effects)
(:affected)
(:policy :fast-safe))
(define-vop (fast-conditional-c/fixnum fast-conditional/fixnum)
(:args (x :scs (any-reg control-stack)))
(:arg-types tagged-num (:constant (signed-byte 29)))
- (:info target not-p y))
+ (:info y))
(define-vop (fast-conditional/signed fast-conditional)
(:args (x :scs (signed-reg)
(define-vop (fast-conditional-c/signed fast-conditional/signed)
(:args (x :scs (signed-reg signed-stack)))
(:arg-types signed-num (:constant (signed-byte 31)))
- (:info target not-p y))
+ (:info y))
(define-vop (fast-conditional/unsigned fast-conditional)
(:args (x :scs (unsigned-reg)
(define-vop (fast-conditional-c/unsigned fast-conditional/unsigned)
(:args (x :scs (unsigned-reg unsigned-stack)))
(:arg-types unsigned-num (:constant (unsigned-byte 31)))
- (:info target not-p y))
+ (:info y))
(macrolet ((define-conditional-vop (tran cond unsigned not-cond not-unsigned)
`(progn
(format nil "~:@(FAST-CONDITIONAL~A~)"
suffix)))
(:translate ,tran)
+ (:conditional ,(if signed cond unsigned))
(:generator ,cost
(inst cmp x
,(if (eq suffix '-c/fixnum)
'(fixnumize y)
- 'y))
- (inst jmp (if not-p
- ,(if signed
- not-cond
- not-unsigned)
- ,(if signed
- cond
- unsigned))
- target))))
+ 'y)))))
'(/fixnum -c/fixnum /signed -c/signed /unsigned -c/unsigned)
; '(/fixnum /signed /unsigned)
'(4 3 6 5 6 5)
(define-vop (fast-if-eql/signed fast-conditional/signed)
(:translate eql)
(:generator 6
- (inst cmp x y)
- (inst jmp (if not-p :ne :e) target)))
+ (inst cmp x y)))
(define-vop (fast-if-eql-c/signed fast-conditional-c/signed)
(:translate eql)
(cond ((and (sc-is x signed-reg) (zerop y))
(inst test x x)) ; smaller instruction
(t
- (inst cmp x y)))
- (inst jmp (if not-p :ne :e) target)))
+ (inst cmp x y)))))
(define-vop (fast-if-eql/unsigned fast-conditional/unsigned)
(:translate eql)
(:generator 6
- (inst cmp x y)
- (inst jmp (if not-p :ne :e) target)))
+ (inst cmp x y)))
(define-vop (fast-if-eql-c/unsigned fast-conditional-c/unsigned)
(:translate eql)
(cond ((and (sc-is x unsigned-reg) (zerop y))
(inst test x x)) ; smaller instruction
(t
- (inst cmp x y)))
- (inst jmp (if not-p :ne :e) target)))
+ (inst cmp x y)))))
;;; EQL/FIXNUM is funny because the first arg can be of any type, not just a
;;; known fixnum.
(:note "inline fixnum comparison")
(:translate eql)
(:generator 4
- (inst cmp x y)
- (inst jmp (if not-p :ne :e) target)))
+ (inst cmp x y)))
+
(define-vop (generic-eql/fixnum fast-eql/fixnum)
(:args (x :scs (any-reg descriptor-reg)
:load-if (not (and (sc-is x control-stack)
(:arg-types * tagged-num)
(:variant-cost 7))
-
(define-vop (fast-eql-c/fixnum fast-conditional/fixnum)
(:args (x :scs (any-reg control-stack)))
(:arg-types tagged-num (:constant (signed-byte 29)))
- (:info target not-p y)
+ (:info y)
(:translate eql)
(:generator 2
(cond ((and (sc-is x any-reg) (zerop y))
(inst test x x)) ; smaller instruction
(t
- (inst cmp x (fixnumize y))))
- (inst jmp (if not-p :ne :e) target)))
+ (inst cmp x (fixnumize y))))))
(define-vop (generic-eql-c/fixnum fast-eql-c/fixnum)
(:args (x :scs (any-reg descriptor-reg control-stack)))
\f
;;;; 32-bit logical operations
-(define-vop (merge-bits)
- (:translate merge-bits)
- (:args (shift :scs (signed-reg unsigned-reg) :target ecx)
- (prev :scs (unsigned-reg) :target result)
- (next :scs (unsigned-reg)))
- (:arg-types tagged-num unsigned-num unsigned-num)
- (:temporary (:sc signed-reg :offset ecx-offset :from (:argument 0)) ecx)
- (:results (result :scs (unsigned-reg) :from (:argument 1)))
- (:result-types unsigned-num)
- (:policy :fast-safe)
- (:generator 4
- (move ecx shift)
- (move result prev)
- (inst shrd result next :cl)))
-
;;; Only the lower 6 bits of the shift amount are significant.
(define-vop (shift-towards-someplace)
(:policy :fast-safe)
\f
;;;; Modular functions
+(defmacro define-mod-binop ((name prototype) function)
+ `(define-vop (,name ,prototype)
+ (:args (x :target r :scs (unsigned-reg signed-reg)
+ :load-if (not (and (or (sc-is x unsigned-stack)
+ (sc-is x signed-stack))
+ (or (sc-is y unsigned-reg)
+ (sc-is y signed-reg))
+ (or (sc-is r unsigned-stack)
+ (sc-is r signed-stack))
+ (location= x r))))
+ (y :scs (unsigned-reg signed-reg unsigned-stack signed-stack)))
+ (:arg-types untagged-num untagged-num)
+ (:results (r :scs (unsigned-reg signed-reg) :from (:argument 0)
+ :load-if (not (and (or (sc-is x unsigned-stack)
+ (sc-is x signed-stack))
+ (or (sc-is y unsigned-reg)
+ (sc-is y unsigned-reg))
+ (or (sc-is r unsigned-stack)
+ (sc-is r unsigned-stack))
+ (location= x r)))))
+ (:result-types unsigned-num)
+ (:translate ,function)))
+(defmacro define-mod-binop-c ((name prototype) function)
+ `(define-vop (,name ,prototype)
+ (:args (x :target r :scs (unsigned-reg signed-reg)
+ :load-if (not (and (or (sc-is x unsigned-stack)
+ (sc-is x signed-stack))
+ (or (sc-is r unsigned-stack)
+ (sc-is r signed-stack))
+ (location= x r)))))
+ (:info y)
+ (:arg-types untagged-num (:constant (or (unsigned-byte 31) (signed-byte 32))))
+ (:results (r :scs (unsigned-reg signed-reg) :from (:argument 0)
+ :load-if (not (and (or (sc-is x unsigned-stack)
+ (sc-is x signed-stack))
+ (or (sc-is r unsigned-stack)
+ (sc-is r unsigned-stack))
+ (location= x r)))))
+ (:result-types unsigned-num)
+ (:translate ,function)))
+
(macrolet ((def (name -c-p)
(let ((fun64 (intern (format nil "~S-MOD64" name)))
(vopu (intern (format nil "FAST-~S/UNSIGNED=>UNSIGNED" name)))
(vopcu (intern (format nil "FAST-~S-C/UNSIGNED=>UNSIGNED" name)))
(vopf (intern (format nil "FAST-~S/FIXNUM=>FIXNUM" name)))
(vopcf (intern (format nil "FAST-~S-C/FIXNUM=>FIXNUM" name)))
- (vop64u (intern (format nil "FAST-~S-MOD64/UNSIGNED=>UNSIGNED" name)))
+ (vop64u (intern (format nil "FAST-~S-MOD64/WORD=>UNSIGNED" name)))
(vop64f (intern (format nil "FAST-~S-MOD64/FIXNUM=>FIXNUM" name)))
- (vop64cu (intern (format nil "FAST-~S-MOD64-C/UNSIGNED=>UNSIGNED" name)))
+ (vop64cu (intern (format nil "FAST-~S-MOD64-C/WORD=>UNSIGNED" name)))
(vop64cf (intern (format nil "FAST-~S-MOD64-C/FIXNUM=>FIXNUM" name)))
(sfun61 (intern (format nil "~S-SMOD61" name)))
(svop61f (intern (format nil "FAST-~S-SMOD61/FIXNUM=>FIXNUM" name)))
(svop61cf (intern (format nil "FAST-~S-SMOD61-C/FIXNUM=>FIXNUM" name))))
`(progn
- (define-modular-fun ,fun64 (x y) ,name :unsigned 64)
- (define-modular-fun ,sfun61 (x y) ,name :signed 61)
- (define-vop (,vop64u ,vopu) (:translate ,fun64))
+ (define-modular-fun ,fun64 (x y) ,name :untagged nil 64)
+ (define-modular-fun ,sfun61 (x y) ,name :tagged t 61)
+ (define-mod-binop (,vop64u ,vopu) ,fun64)
(define-vop (,vop64f ,vopf) (:translate ,fun64))
(define-vop (,svop61f ,vopf) (:translate ,sfun61))
,@(when -c-p
- `((define-vop (,vop64cu ,vopcu) (:translate ,fun64))
+ `((define-mod-binop-c (,vop64cu ,vopcu) ,fun64)
(define-vop (,svop61cf ,vopcf) (:translate ,sfun61))))))))
(def + t)
(def - t)
(signed-byte 61)
(foldable flushable movable))
-(define-modular-fun-optimizer %lea ((base index scale disp) :unsigned :width width)
+(define-modular-fun-optimizer %lea ((base index scale disp) :untagged nil :width width)
(when (and (<= width 64)
(constant-lvar-p scale)
(constant-lvar-p disp))
- (cut-to-width base :unsigned width)
- (cut-to-width index :unsigned width)
+ (cut-to-width base :untagged width nil)
+ (cut-to-width index :untagged width nil)
'sb!vm::%lea-mod64))
-(define-modular-fun-optimizer %lea ((base index scale disp) :signed :width width)
+(define-modular-fun-optimizer %lea ((base index scale disp) :tagged t :width width)
(when (and (<= width 61)
(constant-lvar-p scale)
(constant-lvar-p disp))
- (cut-to-width base :signed width)
- (cut-to-width index :signed width)
+ (cut-to-width base :tagged width t)
+ (cut-to-width index :tagged width t)
'sb!vm::%lea-smod61))
#+sb-xc-host
(:translate %lea-smod61))
;;; logical operations
-(define-modular-fun lognot-mod64 (x) lognot :unsigned 64)
+(define-modular-fun lognot-mod64 (x) lognot :untagged nil 64)
(define-vop (lognot-mod64/unsigned=>unsigned)
(:translate lognot-mod64)
(:args (x :scs (unsigned-reg unsigned-stack) :target r
(move r x)
(inst not r)))
-(define-modular-fun logxor-mod64 (x y) logxor :unsigned 64)
-(define-vop (fast-logxor-mod64/unsigned=>unsigned
- fast-logxor/unsigned=>unsigned)
- (:translate logxor-mod64))
-(define-vop (fast-logxor-mod64-c/unsigned=>unsigned
- fast-logxor-c/unsigned=>unsigned)
- (:translate logxor-mod64))
-(define-vop (fast-logxor-mod64/fixnum=>fixnum
- fast-logxor/fixnum=>fixnum)
- (:translate logxor-mod64))
-(define-vop (fast-logxor-mod64-c/fixnum=>fixnum
- fast-logxor-c/fixnum=>fixnum)
- (:translate logxor-mod64))
-
(define-source-transform logeqv (&rest args)
(if (oddp (length args))
`(logxor ,@args)
(define-full-reffer bignum-ref * bignum-digits-offset other-pointer-lowtag
(unsigned-reg) unsigned-num sb!bignum:%bignum-ref)
-
+(define-full-reffer+offset bignum--ref-with-offset * bignum-digits-offset
+ other-pointer-lowtag (unsigned-reg) unsigned-num
+ sb!bignum:%bignum-ref-with-offset)
(define-full-setter bignum-set * bignum-digits-offset other-pointer-lowtag
(unsigned-reg) unsigned-num sb!bignum:%bignum-set)
(:policy :fast-safe)
(:args (digit :scs (unsigned-reg)))
(:arg-types unsigned-num)
- (:conditional)
- (:info target not-p)
+ (:conditional :ns)
(:generator 3
- (inst or digit digit)
- (inst jmp (if not-p :s :ns) target)))
+ (inst or digit digit)))
;;; For add and sub with carry the sc of carry argument is any-reg so
:load-if (not (and (sc-is result unsigned-stack)
(location= digit result)))))
(:result-types unsigned-num)
- (:generator 1
+ (:generator 2
(move result digit)
(move ecx count)
(inst sar result :cl)))
projects / sbcl.git / blobdiff