X-Git-Url: http://repo.macrolet.net/gitweb/?a=blobdiff_plain;f=src%2Fcompiler%2Fgeneric%2Fvm-tran.lisp;h=7031de38919acc1fcc89d5b09132181525f0ec7c;hb=11b8fcf55c80cb2686fb49663fa4d96f9b152ce4;hp=6d90e85f93adcb6309310c8bbf3874ca6fb78698;hpb=cea4896b2482b7b2b429c1631d774b4cfbc0efba;p=sbcl.git diff --git a/src/compiler/generic/vm-tran.lisp b/src/compiler/generic/vm-tran.lisp index 6d90e85..7031de3 100644 --- a/src/compiler/generic/vm-tran.lisp +++ b/src/compiler/generic/vm-tran.lisp @@ -11,39 +11,32 @@ (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 @@ -61,20 +54,7 @@ ;; 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))))) @@ -86,7 +66,7 @@ (let ((dims (array-type-dimensions array-type))) (when (or (atom dims) (= (length dims) 1)) (give-up-ir1-transform)) - (let ((el-type (array-type-element-type array-type)) + (let ((el-type (array-type-specialized-element-type array-type)) (total-size (if (member '* dims) '* (reduce #'* dims)))) @@ -107,42 +87,53 @@ "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) + (%data-vector-and-index array index) + (declare (type (simple-array ,element-type-specifier 1) array) + (type ,element-type-specifier new-value)) + (data-vector-set array index new-value))))) (deftransform data-vector-set ((array index new-value) - (simple-array t t)) + (simple-array t t)) (let ((array-type (continuation-type array))) (unless (array-type-p array-type) (give-up-ir1-transform)) (let ((dims (array-type-dimensions array-type))) (when (or (atom dims) (= (length dims) 1)) - (give-up-ir1-transform)) - (let ((el-type (array-type-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))))) + (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 (continuation-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) + * + :important t) + + ;; 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) ;;; @@ -153,7 +144,7 @@ #+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 *)) @@ -161,7 +152,7 @@ (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) @@ -171,7 +162,7 @@ (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))))))) @@ -179,85 +170,66 @@ (frob (simple-array (unsigned-byte 2) (*)) 2) (frob (simple-array (unsigned-byte 4) (*)) 4)) -;;;; simple string transforms -;;;; -;;;; Note: CMU CL had more of these, including transforms for -;;;; functions which cons. In SBCL, we've gotten rid of the transforms -;;;; for functions which cons, since our GC overhead is sufficiently -;;;; large that it doesn't seem worth it to try to economize on -;;;; function call overhead or on the overhead of runtime type -;;;; dispatch in AREF. - -;;; FIXME: Shouldn't we be testing for legality of -;;; * START1, START2, END1, and END2 indices? -;;; * size of copied string relative to destination string? -;;; (Either there should be tests conditional on SAFETY>=SPEED, or -;;; the transform should be conditional on SPEED>SAFETY.) -;;; -;;; FIXME: Also, the transform should probably be dependent on -;;; SPEED>SPACE. -(deftransform replace ((string1 string2 &key (start1 0) (start2 0) - end1 end2) - (simple-string simple-string &rest t)) - `(locally - (declare (optimize (safety 0))) - (bit-bash-copy string2 - (the index - (+ (the index (* start2 sb!vm:byte-bits)) - ,vector-data-bit-offset)) - string1 - (the index - (+ (the index (* start1 sb!vm:byte-bits)) - ,vector-data-bit-offset)) - (the index - (* (min (the index (- (or end1 (length string1)) - start1)) - (the index (- (or end2 (length string2)) - start2))) - sb!vm:byte-bits))) - string1)) - -;;;; 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 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)) (deftransform bit-not ((bit-array result-bit-array) @@ -270,46 +242,67 @@ (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) + (32bit-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)))))))) -;;;; 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))) + nil)) ;;;; transforms for EQL of floating point values @@ -319,3 +312,4 @@ (deftransform eql ((x y) (double-float double-float)) '(and (= (double-float-low-bits x) (double-float-low-bits y)) (= (double-float-high-bits x) (double-float-high-bits y)))) +