X-Git-Url: http://repo.macrolet.net/gitweb/?a=blobdiff_plain;f=src%2Fcompiler%2Fgeneric%2Fvm-tran.lisp;h=5be78b7d913eec00a1f1e3b4ca71062c6f512161;hb=75b52379bdc2269961af6a1308eca63610f38ac3;hp=1d93ebafc282f4f0c6400c147afeb489a4301110;hpb=1e9966d5f24709d227e20911b4e1ddd27c87a00e;p=sbcl.git diff --git a/src/compiler/generic/vm-tran.lisp b/src/compiler/generic/vm-tran.lisp index 1d93eba..5be78b7 100644 --- a/src/compiler/generic/vm-tran.lisp +++ b/src/compiler/generic/vm-tran.lisp @@ -50,7 +50,7 @@ ((simple-array character (*)) (data-vector-ref string index)) ((simple-array nil (*)) (data-vector-ref string index)))))) -(deftransform hairy-data-vector-ref ((array index) (array t) * :important t) +(deftransform hairy-data-vector-ref ((array index) (array t) *) "avoid runtime dispatch on array element type" (let ((element-ctype (extract-upgraded-element-type array)) (declared-element-ctype (extract-declared-element-type array))) @@ -103,8 +103,7 @@ (deftransform hairy-data-vector-set ((array index new-value) (array t t) - * - :important t) + *) "avoid runtime dispatch on array element type" (let ((element-ctype (extract-upgraded-element-type array)) (declared-element-ctype (extract-declared-element-type array))) @@ -153,8 +152,7 @@ (deftransform %data-vector-and-index ((%array %index) (simple-array t) - * - :important t) + *) ;; KLUDGE: why the percent signs? Well, ARRAY and INDEX are ;; respectively exported from the CL and SB!INT packages, which ;; means that they're visible to all sorts of things. If the @@ -260,16 +258,16 @@ (setf (%raw-bits result-bit-array index) (,',wordfun (%raw-bits bit-array-1 index) (%raw-bits bit-array-2 index)))))))))) - (def bit-and 32bit-logical-and) - (def bit-ior 32bit-logical-or) - (def bit-xor 32bit-logical-xor) - (def bit-eqv 32bit-logical-eqv) - (def bit-nand 32bit-logical-nand) - (def bit-nor 32bit-logical-nor) - (def bit-andc1 32bit-logical-andc1) - (def bit-andc2 32bit-logical-andc2) - (def bit-orc1 32bit-logical-orc1) - (def bit-orc2 32bit-logical-orc2)) + (def bit-and word-logical-and) + (def bit-ior word-logical-or) + (def bit-xor word-logical-xor) + (def bit-eqv word-logical-eqv) + (def bit-nand word-logical-nand) + (def bit-nor word-logical-nor) + (def bit-andc1 word-logical-andc1) + (def bit-andc2 word-logical-andc2) + (def bit-orc1 word-logical-orc1) + (def bit-orc2 word-logical-orc2)) (deftransform bit-not ((bit-array result-bit-array) @@ -284,26 +282,26 @@ bit-array result-bit-array)))) (let ((length (length result-bit-array))) (if (= length 0) - ;; We avoid doing anything to 0-length bit-vectors, or - ;; rather, the memory that follows them. Other - ;; divisible-by-32 cases are handled by the (1- length), - ;; below. CSR, 2002-04-24 + ;; We avoid doing anything to 0-length bit-vectors, or rather, + ;; the memory that follows them. Other divisible-by + ;; n-word-bits cases are handled by the (1- length), below. + ;; CSR, 2002-04-24 result-bit-array (do ((index sb!vm:vector-data-offset (1+ index)) (end-1 (+ sb!vm:vector-data-offset - ;; bit-vectors of length 1-32 need precisely - ;; one (SETF %RAW-BITS), done here in the - ;; epilogue. - CSR, 2002-04-24 + ;; bit-vectors of length 1 to n-word-bits need + ;; precisely one (SETF %RAW-BITS), done here in + ;; the epilogue. - CSR, 2002-04-24 (truncate (truly-the index (1- length)) sb!vm:n-word-bits)))) ((= index end-1) (setf (%raw-bits result-bit-array index) - (32bit-logical-not (%raw-bits bit-array index))) + (word-logical-not (%raw-bits bit-array index))) result-bit-array) (declare (optimize (speed 3) (safety 0)) (type index index end-1)) (setf (%raw-bits result-bit-array index) - (32bit-logical-not (%raw-bits bit-array index)))))))) + (word-logical-not (%raw-bits bit-array index)))))))) (deftransform bit-vector-= ((x y) (simple-bit-vector simple-bit-vector)) `(and (= (length x) (length y)) @@ -331,32 +329,73 @@ (:big-endian '(- sb!vm:n-word-bits extra)))) (%raw-bits y i)))) - (declare (type (integer 0 31) extra) - (type (unsigned-byte 32) mask numx numy)) + (declare (type (mod #.sb!vm:n-word-bits) + extra) + (type sb!vm:word mask numx numy)) (= numx numy))) (declare (type index i end-1)) (let ((numx (%raw-bits x i)) (numy (%raw-bits y i))) - (declare (type (unsigned-byte 32) numx numy)) + (declare (type sb!vm:word numx numy)) (unless (= numx numy) (return nil)))))))) +(deftransform count ((item sequence) (bit simple-bit-vector) * + :policy (>= speed space)) + `(let ((length (length sequence))) + (if (zerop length) + 0 + (do ((index sb!vm:vector-data-offset (1+ index)) + (count 0) + (end-1 (+ sb!vm:vector-data-offset + (truncate (truly-the index (1- length)) + sb!vm:n-word-bits)))) + ((= index end-1) + (let* ((extra (mod length sb!vm:n-word-bits)) + (mask (1- (ash 1 extra))) + (bits (logand (ash mask + ,(ecase sb!c:*backend-byte-order* + (:little-endian 0) + (:big-endian + '(- sb!vm:n-word-bits extra)))) + (%raw-bits sequence index)))) + (declare (type (mod #.sb!vm:n-word-bits) extra)) + (declare (type sb!vm:word mask bits)) + ;; could consider LOGNOT for the zero case instead of + ;; doing the subtraction... + (incf count ,(if (constant-lvar-p item) + (if (zerop (lvar-value item)) + '(- extra (logcount bits)) + '(logcount bits)) + '(if (zerop item) + (- extra (logcount bits)) + (logcount bits)))))) + (declare (type index index count end-1) + (optimize (speed 3) (safety 0))) + (incf count ,(if (constant-lvar-p item) + (if (zerop (lvar-value item)) + '(- sb!vm:n-word-bits (logcount (%raw-bits sequence index))) + '(logcount (%raw-bits sequence index))) + '(if (zerop item) + (- sb!vm:n-word-bits (logcount (%raw-bits sequence index))) + (logcount (%raw-bits sequence index))))))))) + (deftransform fill ((sequence item) (simple-bit-vector bit) * :policy (>= speed space)) (let ((value (if (constant-lvar-p item) (if (= (lvar-value item) 0) 0 - #.(1- (ash 1 32))) - `(if (= item 0) 0 #.(1- (ash 1 32)))))) + #.(1- (ash 1 sb!vm:n-word-bits))) + `(if (= item 0) 0 #.(1- (ash 1 sb!vm:n-word-bits)))))) `(let ((length (length sequence)) (value ,value)) (if (= length 0) sequence (do ((index sb!vm:vector-data-offset (1+ index)) (end-1 (+ sb!vm:vector-data-offset - ;; bit-vectors of length 1-32 need precisely - ;; one (SETF %RAW-BITS), done here in the - ;; epilogue. - CSR, 2002-04-24 + ;; bit-vectors of length 1 to n-word-bits need + ;; precisely one (SETF %RAW-BITS), done here + ;; in the epilogue. - CSR, 2002-04-24 (truncate (truly-the index (1- length)) sb!vm:n-word-bits)))) ((= index end-1) @@ -370,18 +409,21 @@ :policy (>= speed space)) (let ((value (if (constant-lvar-p item) (let* ((char (lvar-value item)) - (code (sb!xc:char-code char))) - (logior code (ash code 8) (ash code 16) (ash code 24))) + (code (sb!xc:char-code char)) + (accum 0)) + (dotimes (i sb!vm:n-word-bytes accum) + (setf accum (logior accum (ash code (* 8 i)))))) `(let ((code (sb!xc:char-code item))) - (logior code (ash code 8) (ash code 16) (ash code 24)))))) + (logior ,@(loop for i from 0 below sb!vm:n-word-bytes + collect `(ash code ,(* 8 i)))))))) `(let ((length (length sequence)) (value ,value)) (multiple-value-bind (times rem) - (truncate length 4) + (truncate length sb!vm:n-word-bytes) (do ((index sb!vm:vector-data-offset (1+ index)) (end (+ times sb!vm:vector-data-offset))) ((= index end) - (let ((place (* times 4))) + (let ((place (* times sb!vm:n-word-bytes))) (declare (fixnum place)) (dotimes (j rem sequence) (declare (index j)) @@ -444,26 +486,67 @@ (define-good-modular-fun logior) ;;; FIXME: XOR? ANDC1, ANDC2? -- CSR, 2003-09-16 -#!-alpha -(progn - (defknown sb!vm::ash-left-mod32 (integer (integer 0)) (unsigned-byte 32) - (foldable flushable movable)) - (define-modular-fun-optimizer ash ((integer count) :width width) - (when (and (<= width 32) - (constant-lvar-p count) ; ? - (plusp (lvar-value count))) - (cut-to-width integer width) - 'sb!vm::ash-left-mod32))) -#!+alpha -(progn - (defknown sb!vm::ash-left-mod64 (integer (integer 0)) (unsigned-byte 64) - (foldable flushable movable)) - (define-modular-fun-optimizer ash ((integer count) :width width) - (when (and (<= width 64) - (constant-lvar-p count) ; ? - (plusp (lvar-value count))) - (cut-to-width integer width) - 'sb!vm::ash-left-mod64))) +(macrolet + ((def (name width) + `(progn + (defknown ,name (integer (integer 0)) (unsigned-byte ,width) + (foldable flushable movable)) + (define-modular-fun-optimizer ash ((integer count) :width width) + (when (and (<= width ,width) + (or (and (constant-lvar-p count) + (plusp (lvar-value count))) + (csubtypep (lvar-type count) + (specifier-type '(and unsigned-byte + fixnum))))) + (cut-to-width integer width) + ',name)) + (setf (gethash ',name *modular-versions*) `(ash ,',width))))) + ;; This should really be dependent on SB!VM:N-WORD-BITS, but since we + ;; don't have a true Alpha64 port yet, we'll have to stick to + ;; SB!VM:N-MACHINE-WORD-BITS for the time being. --njf, 2004-08-14 + #!+#.(cl:if (cl:= 32 sb!vm:n-machine-word-bits) '(and) '(or)) + (def sb!vm::ash-left-mod32 32) + #!+#.(cl:if (cl:= 64 sb!vm:n-machine-word-bits) '(and) '(or)) + (def sb!vm::ash-left-mod64 64)) + + +;;;; word-wise logical operations + +;;; These transforms assume the presence of modular arithmetic to +;;; generate efficient code. + +(define-source-transform word-logical-not (x) + `(logand (lognot (the sb!vm:word ,x)) #.(1- (ash 1 sb!vm:n-word-bits)))) + +(deftransform word-logical-and ((x y)) + '(logand x y)) + +(deftransform word-logical-nand ((x y)) + '(logand (lognand x y) #.(1- (ash 1 sb!vm:n-word-bits)))) + +(deftransform word-logical-or ((x y)) + '(logior x y)) + +(deftransform word-logical-nor ((x y)) + '(logand (lognor x y) #.(1- (ash 1 sb!vm:n-word-bits)))) + +(deftransform word-logical-xor ((x y)) + '(logxor x y)) + +(deftransform word-logical-eqv ((x y)) + '(logand (logeqv x y) #.(1- (ash 1 sb!vm:n-word-bits)))) + +(deftransform word-logical-orc1 ((x y)) + '(logand (logorc1 x y) #.(1- (ash 1 sb!vm:n-word-bits)))) + +(deftransform word-logical-orc2 ((x y)) + '(logand (logorc2 x y) #.(1- (ash 1 sb!vm:n-word-bits)))) + +(deftransform word-logical-andc1 ((x y)) + '(logand (logandc1 x y) #.(1- (ash 1 sb!vm:n-word-bits)))) + +(deftransform word-logical-andc2 ((x y)) + '(logand (logandc2 x y) #.(1- (ash 1 sb!vm:n-word-bits)))) ;;; There are two different ways the multiplier can be recoded. The