X-Git-Url: http://repo.macrolet.net/gitweb/?a=blobdiff_plain;f=src%2Fcompiler%2Fgeneric%2Fvm-tran.lisp;h=29c5c060cb3261372c50e6844ca7bb9b403a4491;hb=ace140856e6b3f92bb06597092a59753f1e59142;hp=ad74c20e18706db86d9eef41d925aa92c973e2f2;hpb=5eb97830eca716fef626c6e12429c99c9b97e3c8;p=sbcl.git diff --git a/src/compiler/generic/vm-tran.lisp b/src/compiler/generic/vm-tran.lisp index ad74c20..29c5c06 100644 --- a/src/compiler/generic/vm-tran.lisp +++ b/src/compiler/generic/vm-tran.lisp @@ -11,46 +11,49 @@ (in-package "SB!C") -;;; FIXME: It would be good to implement SB!XC:DEFCONSTANT, and use -;;; use that here, so that the compiler is born knowing this value. -;;; FIXME: Add a comment telling whether this holds for all vectors -;;; or only for vectors based on simple arrays (non-adjustable, etc.). -(defconstant vector-data-bit-offset - (* sb!vm:vector-data-offset sb!vm:word-bits)) - -;;; We need to define these predicates, since the TYPEP source transform picks -;;; whichever predicate was defined last when there are multiple predicates for -;;; equivalent types. -(def-source-transform short-float-p (x) `(single-float-p ,x)) +;;; We need to define these predicates, since the TYPEP source +;;; transform picks whichever predicate was defined last when there +;;; are multiple predicates for equivalent types. +(define-source-transform short-float-p (x) `(single-float-p ,x)) #!-long-float -(def-source-transform long-float-p (x) `(double-float-p ,x)) +(define-source-transform long-float-p (x) `(double-float-p ,x)) -(def-source-transform compiled-function-p (x) +(define-source-transform compiled-function-p (x) `(functionp ,x)) -(def-source-transform char-int (x) +(define-source-transform char-int (x) `(char-code ,x)) (deftransform abs ((x) (rational)) '(if (< x 0) (- x) x)) ;;; The layout is stored in slot 0. -(def-source-transform %instance-layout (x) +(define-source-transform %instance-layout (x) `(truly-the layout (%instance-ref ,x 0))) -(def-source-transform %set-instance-layout (x val) +(define-source-transform %set-instance-layout (x val) `(%instance-set ,x 0 (the layout ,val))) ;;;; character support ;;; In our implementation there are really only BASE-CHARs. -(def-source-transform characterp (obj) +(define-source-transform characterp (obj) `(base-char-p ,obj)) ;;;; simplifying HAIRY-DATA-VECTOR-REF and HAIRY-DATA-VECTOR-SET -(deftransform hairy-data-vector-ref ((array index) (array t) * :important t) +(deftransform hairy-data-vector-ref ((string index) (simple-string t)) + (let ((ctype (lvar-type string))) + (if (array-type-p ctype) + ;; the other transform will kick in, so that's OK + (give-up-ir1-transform) + `(etypecase string + ((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) *) "avoid runtime dispatch on array element type" - (let ((element-ctype (extract-upgraded-element-type array))) + (let ((element-ctype (extract-upgraded-element-type array)) + (declared-element-ctype (extract-declared-element-type array))) (declare (type ctype element-ctype)) (when (eq *wild-type* element-ctype) (give-up-ir1-transform @@ -61,95 +64,114 @@ ;; to hand-expand it ourselves.) (let ((element-type-specifier (type-specifier element-ctype))) `(multiple-value-bind (array index) - ;; FIXME: All this noise should move into a - ;; %DATA-VECTOR-AND-INDEX function, and there should be - ;; DEFTRANSFORMs for %DATA-VECTOR-AND-INDEX to optimize the - ;; function call away when the array is known to be simple, - ;; and to specialize to - ;; %DATA-VECTOR-AND-INDEX-IN-VECTOR-CASE when the array is - ;; known to have only one dimension. - (if (array-header-p array) - (%with-array-data array index nil) - (let ((array array)) - (declare (type (simple-array ,element-type-specifier 1) - array)) - (%check-bound array 0 index) - (values array index))) + (%data-vector-and-index array index) (declare (type (simple-array ,element-type-specifier 1) array)) - (data-vector-ref array index))))) + ,(let ((bare-form '(data-vector-ref array index))) + (if (type= element-ctype declared-element-ctype) + bare-form + `(the ,(type-specifier declared-element-ctype) + ,bare-form))))))) -;;; MNA: open-coded-simple-array patch (deftransform data-vector-ref ((array index) (simple-array t)) - (let ((array-type (continuation-type array))) + (let ((array-type (lvar-type array))) (unless (array-type-p array-type) (give-up-ir1-transform)) (let ((dims (array-type-dimensions array-type))) - (when (and (consp dims) (= (length dims) 1)) + (when (or (atom dims) (= (length dims) 1)) (give-up-ir1-transform)) - (let* ((el-type (array-type-element-type array-type)) - (total-size (if (or (atom dims) (member '* dims)) + (let ((el-type (array-type-specialized-element-type array-type)) + (total-size (if (member '* dims) '* - (reduce #'* dims))) - (type-sp `(simple-array ,(type-specifier el-type) - (,total-size)))) - (if (atom dims) - `(let ((a (truly-the ,type-sp (%array-simp array)))) - (data-vector-ref a index)) - `(let ((a (truly-the ,type-sp (%array-data-vector array)))) - (data-vector-ref a index))))))) + (reduce #'* dims)))) + `(data-vector-ref (truly-the (simple-array ,(type-specifier el-type) + (,total-size)) + (%array-data-vector array)) + index))))) + +(deftransform hairy-data-vector-set ((string index new-value) + (simple-string t t)) + (let ((ctype (lvar-type string))) + (if (array-type-p ctype) + ;; the other transform will kick in, so that's OK + (give-up-ir1-transform) + `(etypecase string + ((simple-array character (*)) + (data-vector-set string index new-value)) + ((simple-array nil (*)) + (data-vector-set string index new-value)))))) (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))) + (let ((element-ctype (extract-upgraded-element-type array)) + (declared-element-ctype (extract-declared-element-type array))) (declare (type ctype element-ctype)) (when (eq *wild-type* element-ctype) (give-up-ir1-transform "Upgraded element type of array is not known at compile time.")) (let ((element-type-specifier (type-specifier element-ctype))) `(multiple-value-bind (array index) - ;; FIXME: All this noise should move into a - ;; %DATA-VECTOR-AND-INDEX function, and there should be - ;; DEFTRANSFORMs for %DATA-VECTOR-AND-INDEX to optimize the - ;; function call away when the array is known to be simple, - ;; and to specialize to - ;; %DATA-VECTOR-AND-INDEX-IN-VECTOR-CASE when the array is - ;; known to have only one dimension. - (if (array-header-p array) - (%with-array-data array index nil) - (let ((array array)) - (declare (type (simple-array ,element-type-specifier 1) - array)) - (%check-bound array 0 index) - (values array index))) - (data-vector-set (truly-the (simple-array ,element-type-specifier 1) - array) - index - new-value))))) + (%data-vector-and-index array index) + (declare (type (simple-array ,element-type-specifier 1) array) + (type ,element-type-specifier new-value)) + ,(if (type= element-ctype declared-element-ctype) + '(data-vector-set array index new-value) + `(truly-the ,(type-specifier declared-element-ctype) + (data-vector-set array index + (the ,(type-specifier declared-element-ctype) + new-value)))))))) -;;; MNA: open-coded-simple-array patch (deftransform data-vector-set ((array index new-value) - (simple-array t t)) - (let ((array-type (continuation-type array))) + (simple-array t t)) + (let ((array-type (lvar-type array))) (unless (array-type-p array-type) (give-up-ir1-transform)) (let ((dims (array-type-dimensions array-type))) - (when (and (consp dims) (= (length dims) 1)) - (give-up-ir1-transform)) - (let* ((el-type (array-type-element-type array-type)) - (total-size (if (or (atom dims) (member '* dims)) - '* - (reduce #'* dims))) - (type-sp `(simple-array ,(type-specifier el-type) - (,total-size)))) - (if (atom dims) - `(let ((a (truly-the ,type-sp (%array-simp array)))) - (data-vector-set a index new-value)) - `(let ((a (truly-the ,type-sp (%array-data-vector array)))) - (data-vector-set a index new-value))))))) + (when (or (atom dims) (= (length dims) 1)) + (give-up-ir1-transform)) + (let ((el-type (array-type-specialized-element-type array-type)) + (total-size (if (member '* dims) + '* + (reduce #'* dims)))) + `(data-vector-set (truly-the (simple-array ,(type-specifier el-type) + (,total-size)) + (%array-data-vector array)) + index + new-value))))) + +(defoptimizer (%data-vector-and-index derive-type) ((array index)) + (let ((atype (lvar-type array))) + (when (array-type-p atype) + (values-specifier-type + `(values (simple-array ,(type-specifier + (array-type-specialized-element-type atype)) + (*)) + index))))) + +(deftransform %data-vector-and-index ((%array %index) + (simple-array 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 + ;; compiler can prove that the call to ARRAY-HEADER-P, below, either + ;; returns T or NIL, it will delete the irrelevant branch. However, + ;; user code might have got here with a variable named CL:ARRAY, and + ;; quite often compiler code with a variable named SB!INT:INDEX, so + ;; this can generate code deletion notes for innocuous user code: + ;; (DEFUN F (ARRAY I) (DECLARE (SIMPLE-VECTOR ARRAY)) (AREF ARRAY I)) + ;; -- CSR, 2003-04-01 + + ;; We do this solely for the -OR-GIVE-UP side effect, since we want + ;; to know that the type can be figured out in the end before we + ;; proceed, but we don't care yet what the type will turn out to be. + (upgraded-element-type-specifier-or-give-up %array) + + '(if (array-header-p %array) + (values (%array-data-vector %array) %index) + (values %array %index))) ;;; transforms for getting at simple arrays of (UNSIGNED-BYTE N) when (< N 8) ;;; @@ -160,7 +182,7 @@ #+nil (macrolet ((frob (type bits) - (let ((elements-per-word (truncate sb!vm:word-bits bits))) + (let ((elements-per-word (truncate sb!vm:n-word-bits bits))) `(progn (deftransform data-vector-ref ((vector index) (,type *)) @@ -168,7 +190,7 @@ (floor index ,',elements-per-word) (ldb ,(ecase sb!vm:target-byte-order (:little-endian '(byte ,bits (* bit ,bits))) - (:big-endian '(byte ,bits (- sb!vm:word-bits + (:big-endian '(byte ,bits (- sb!vm:n-word-bits (* (1+ bit) ,bits))))) (%raw-bits vector (+ word sb!vm:vector-data-offset))))) (deftransform data-vector-set ((vector index new-value) @@ -178,7 +200,7 @@ (setf (ldb ,(ecase sb!vm:target-byte-order (:little-endian '(byte ,bits (* bit ,bits))) (:big-endian - '(byte ,bits (- sb!vm:word-bits + '(byte ,bits (- sb!vm:n-word-bits (* (1+ bit) ,bits))))) (%raw-bits vector (+ word sb!vm:vector-data-offset))) new-value))))))) @@ -186,47 +208,66 @@ (frob (simple-array (unsigned-byte 2) (*)) 2) (frob (simple-array (unsigned-byte 4) (*)) 4)) -;;;; bit vector hackery +;;;; BIT-VECTOR hackery -;;; SIMPLE-BIT-VECTOR bit-array operations are transformed to a word loop that -;;; does 32 bits at a time. +;;; SIMPLE-BIT-VECTOR bit-array operations are transformed to a word +;;; loop that does 32 bits at a time. ;;; -;;; FIXME: This is a lot of repeatedly macroexpanded code. It should be a -;;; function call instead. And do it with DEF-FROB instead of DOLIST. -(dolist (x '((bit-and 32bit-logical-and) - (bit-ior 32bit-logical-or) - (bit-xor 32bit-logical-xor) - (bit-eqv 32bit-logical-eqv) - (bit-nand 32bit-logical-nand) - (bit-nor 32bit-logical-nor) - (bit-andc1 32bit-logical-andc1) - (bit-andc2 32bit-logical-andc2) - (bit-orc1 32bit-logical-orc1) - (bit-orc2 32bit-logical-orc2))) - (destructuring-bind (bitfun wordfun) x - (deftransform bitfun - ((bit-array-1 bit-array-2 result-bit-array) - '(simple-bit-vector simple-bit-vector simple-bit-vector) '* - :eval-name t :node node :policy (>= speed space)) - `(progn - ,@(unless (policy node (zerop safety)) - '((unless (= (length bit-array-1) (length bit-array-2) - (length result-bit-array)) - (error "Argument and/or result bit arrays are not the same length:~ +;;; FIXME: This is a lot of repeatedly macroexpanded code. It should +;;; be a function call instead. +(macrolet ((def (bitfun wordfun) + `(deftransform ,bitfun ((bit-array-1 bit-array-2 result-bit-array) + (simple-bit-vector + simple-bit-vector + simple-bit-vector) + * + :node node :policy (>= speed space)) + `(progn + ,@(unless (policy node (zerop safety)) + '((unless (= (length bit-array-1) + (length bit-array-2) + (length result-bit-array)) + (error "Argument and/or result bit arrays are not the same length:~ ~% ~S~% ~S ~% ~S" - bit-array-1 bit-array-2 result-bit-array)))) - (do ((index sb!vm:vector-data-offset (1+ index)) - (end (+ sb!vm:vector-data-offset - (truncate (the index - (+ (length bit-array-1) - sb!vm:word-bits -1)) - sb!vm:word-bits)))) - ((= index end) result-bit-array) - (declare (optimize (speed 3) (safety 0)) - (type index index end)) - (setf (%raw-bits result-bit-array index) - (,wordfun (%raw-bits bit-array-1 index) - (%raw-bits bit-array-2 index)))))))) + bit-array-1 + bit-array-2 + 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 + 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 + (truncate (truly-the index (1- length)) + sb!vm:n-word-bits)))) + ((= index end-1) + (setf (%raw-bits result-bit-array index) + (,',wordfun (%raw-bits bit-array-1 index) + (%raw-bits bit-array-2 index))) + result-bit-array) + (declare (optimize (speed 3) (safety 0)) + (type index index end-1)) + (setf (%raw-bits result-bit-array index) + (,',wordfun (%raw-bits bit-array-1 index) + (%raw-bits bit-array-2 index)))))))))) + (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) @@ -239,46 +280,191 @@ (error "Argument and result bit arrays are not the same length:~ ~% ~S~% ~S" bit-array result-bit-array)))) - (do ((index sb!vm:vector-data-offset (1+ index)) - (end (+ sb!vm:vector-data-offset - (truncate (the index - (+ (length bit-array) - (1- sb!vm:word-bits))) - sb!vm:word-bits)))) - ((= index end) result-bit-array) - (declare (optimize (speed 3) (safety 0)) - (type index index end)) - (setf (%raw-bits result-bit-array index) - (32bit-logical-not (%raw-bits bit-array index)))))) + (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 + 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 + (truncate (truly-the index (1- length)) + sb!vm:n-word-bits)))) + ((= index end-1) + (setf (%raw-bits result-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) + (word-logical-not (%raw-bits bit-array index)))))))) + +(deftransform bit-vector-= ((x y) (simple-bit-vector simple-bit-vector)) + `(and (= (length x) (length y)) + (let ((length (length x))) + (or (= length 0) + (do* ((i sb!vm:vector-data-offset (+ i 1)) + (end-1 (+ sb!vm:vector-data-offset + (floor (1- length) sb!vm:n-word-bits)))) + ((= i end-1) + (let* ((extra (mod length sb!vm:n-word-bits)) + (mask (1- (ash 1 extra))) + (numx + (logand + (ash mask + ,(ecase sb!c:*backend-byte-order* + (:little-endian 0) + (:big-endian + '(- sb!vm:n-word-bits extra)))) + (%raw-bits x i))) + (numy + (logand + (ash mask + ,(ecase sb!c:*backend-byte-order* + (:little-endian 0) + (:big-endian + '(- sb!vm:n-word-bits extra)))) + (%raw-bits y i)))) + (declare (type (integer 0 31) 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 sb!vm:word numx numy)) + (unless (= numx numy) + (return nil)))))))) + +(deftransform count ((sequence item) (simple-bit-vector bit) * + :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 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)))))) + `(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 + (truncate (truly-the index (1- length)) + sb!vm:n-word-bits)))) + ((= index end-1) + (setf (%raw-bits sequence index) value) + sequence) + (declare (optimize (speed 3) (safety 0)) + (type index index end-1)) + (setf (%raw-bits sequence index) value)))))) + +(deftransform fill ((sequence item) (simple-base-string base-char) * + :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))) + `(let ((code (sb!xc:char-code item))) + (logior code (ash code 8) (ash code 16) (ash code 24)))))) + `(let ((length (length sequence)) + (value ,value)) + (multiple-value-bind (times rem) + (truncate length 4) + (do ((index sb!vm:vector-data-offset (1+ index)) + (end (+ times sb!vm:vector-data-offset))) + ((= index end) + (let ((place (* times 4))) + (declare (fixnum place)) + (dotimes (j rem sequence) + (declare (index j)) + (setf (schar sequence (the index (+ place j))) item)))) + (declare (optimize (speed 3) (safety 0)) + (type index index)) + (setf (%raw-bits sequence index) value)))))) -;;;; primitive translator for BYTE-BLT +;;;; %BYTE-BLT -(def-primitive-translator byte-blt (src src-start dst dst-start dst-end) - `(let ((src ,src) - (src-start (* ,src-start sb!vm:byte-bits)) - (dst ,dst) - (dst-start (* ,dst-start sb!vm:byte-bits)) - (dst-end (* ,dst-end sb!vm:byte-bits))) - (let ((length (- dst-end dst-start))) - (etypecase src - (system-area-pointer - (etypecase dst - (system-area-pointer - (system-area-copy src src-start dst dst-start length)) - ((simple-unboxed-array (*)) - (copy-from-system-area src src-start - dst (+ dst-start ,vector-data-bit-offset) - length)))) - ((simple-unboxed-array (*)) - (etypecase dst - (system-area-pointer - (copy-to-system-area src (+ src-start ,vector-data-bit-offset) - dst dst-start - length)) - ((simple-unboxed-array (*)) - (bit-bash-copy src (+ src-start ,vector-data-bit-offset) - dst (+ dst-start ,vector-data-bit-offset) - length)))))))) +;;; FIXME: The old CMU CL code used various COPY-TO/FROM-SYSTEM-AREA +;;; stuff (with all the associated bit-index cruft and overflow +;;; issues) even for byte moves. In SBCL, we're converting to byte +;;; moves as problems are discovered with the old code, and this is +;;; currently (ca. sbcl-0.6.12.30) the main interface for code in +;;; SB!KERNEL and SB!SYS (e.g. i/o code). It's not clear that it's the +;;; ideal interface, though, and it probably deserves some thought. +(deftransform %byte-blt ((src src-start dst dst-start dst-end) + ((or (simple-unboxed-array (*)) system-area-pointer) + index + (or (simple-unboxed-array (*)) system-area-pointer) + index + index)) + ;; FIXME: CMU CL had a hairier implementation of this (back when it + ;; was still called (%PRIMITIVE BYTE-BLT). It had the small problem + ;; that it didn't work for large (>16M) values of SRC-START or + ;; DST-START. However, it might have been more efficient. In + ;; particular, I don't really know how much the foreign function + ;; call costs us here. My guess is that if the overhead is + ;; acceptable for SQRT and COS, it's acceptable here, but this + ;; should probably be checked. -- WHN + '(flet ((sapify (thing) + (etypecase thing + (system-area-pointer thing) + ;; FIXME: The code here rather relies on the simple + ;; unboxed array here having byte-sized entries. That + ;; should be asserted explicitly, I just haven't found + ;; a concise way of doing it. (It would be nice to + ;; declare it in the DEFKNOWN too.) + ((simple-unboxed-array (*)) (vector-sap thing))))) + (declare (inline sapify)) + (without-gcing + (memmove (sap+ (sapify dst) dst-start) + (sap+ (sapify src) src-start) + (- dst-end dst-start))) + (values))) ;;;; transforms for EQL of floating point values @@ -289,3 +475,72 @@ '(and (= (double-float-low-bits x) (double-float-low-bits y)) (= (double-float-high-bits x) (double-float-high-bits y)))) + +;;;; modular functions +(define-good-modular-fun logand) +(define-good-modular-fun logior) +;;; FIXME: XOR? ANDC1, ANDC2? -- CSR, 2003-09-16 + +(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 32) + (constant-lvar-p count) ;? + (plusp (lvar-value count))) + (cut-to-width integer width) + ',name)) + (setf (gethash ',name *modular-versions*) `(ash ,',width))))) + #!-alpha (def sb!vm::ash-left-mod32 32) + #!+alpha (def sb!vm::ash-left-mod64 64)) + +;;; There are two different ways the multiplier can be recoded. The +;;; more obvious is to shift X by the correct amount for each bit set +;;; in Y and to sum the results. But if there is a string of bits that +;;; are all set, you can add X shifted by one more then the bit +;;; position of the first set bit and subtract X shifted by the bit +;;; position of the last set bit. We can't use this second method when +;;; the high order bit is bit 31 because shifting by 32 doesn't work +;;; too well. +(defun ub32-strength-reduce-constant-multiply (arg num) + (declare (type (unsigned-byte 32) num)) + (let ((adds 0) (shifts 0) + (result nil) first-one) + (labels ((add (next-factor) + (setf result + (if result + (progn (incf adds) `(+ ,result ,next-factor)) + next-factor)))) + (declare (inline add)) + (dotimes (bitpos 32) + (if first-one + (when (not (logbitp bitpos num)) + (add (if (= (1+ first-one) bitpos) + ;; There is only a single bit in the string. + (progn (incf shifts) `(ash ,arg ,first-one)) + ;; There are at least two. + (progn + (incf adds) + (incf shifts 2) + `(- (ash ,arg ,bitpos) + (ash ,arg ,first-one))))) + (setf first-one nil)) + (when (logbitp bitpos num) + (setf first-one bitpos)))) + (when first-one + (cond ((= first-one 31)) + ((= first-one 30) (incf shifts) (add `(ash ,arg 30))) + (t + (incf shifts 2) + (incf adds) + (add `(- (ash ,arg 31) + (ash ,arg ,first-one))))) + (incf shifts) + (add `(ash ,arg 31)))) + (values (if (plusp adds) + `(logand ,result #.(1- (ash 1 32))) ; using modular arithmetic + result) + adds + shifts)))