X-Git-Url: http://repo.macrolet.net/gitweb/?a=blobdiff_plain;f=src%2Fcompiler%2Fknownfun.lisp;h=bce0879531babce20841b10a21a1321ff39380f2;hb=d306e2d23b38487488eb93881dad836e439e0c77;hp=220bd57231457f9f934a80f2b9a6c04658fe24d1;hpb=0b5610d8a220a4b20cbeac958953ca4d67c00038;p=sbcl.git diff --git a/src/compiler/knownfun.lisp b/src/compiler/knownfun.lisp index 220bd57..bce0879 100644 --- a/src/compiler/knownfun.lisp +++ b/src/compiler/knownfun.lisp @@ -1,6 +1,6 @@ ;;;; This file contains stuff for maintaining a database of special ;;;; information about functions known to the compiler. This includes -;;;; semantic information such as side-effects and type inference +;;;; semantic information such as side effects and type inference ;;;; functions as well as transforms and IR2 translators. ;;;; This software is part of the SBCL system. See the README file for @@ -14,179 +14,270 @@ (in-package "SB!C") +(/show0 "knownfun.lisp 17") + ;;; IR1 boolean function attributes ;;; -;;; There are a number of boolean attributes of known functions which we like -;;; to have in IR1. This information is mostly side effect information of a -;;; sort, but it is different from the kind of information we want in IR2. We -;;; aren't interested in a fine breakdown of side effects, since we do very -;;; little code motion on IR1. We are interested in some deeper semantic -;;; properties such as whether it is safe to pass stack closures to. -(def-boolean-attribute ir1 - ;; May call functions that are passed as arguments. In order to determine - ;; what other effects are present, we must find the effects of all arguments - ;; that may be functions. +;;; There are a number of boolean attributes of known functions which +;;; we like to have in IR1. This information is mostly side effect +;;; information of a sort, but it is different from the kind of +;;; information we want in IR2. We aren't interested in a fine +;;; breakdown of side effects, since we do very little code motion on +;;; IR1. We are interested in some deeper semantic properties such as +;;; whether it is safe to pass stack closures to. +;;; +;;; FIXME: This whole notion of "bad" explicit attributes is bad for +;;; maintenance. How confident are we that we have no defknowns for functions +;;; with functional arguments that are missing the CALL attribute? Much better +;;; to have NO-CALLS, as it is much less likely to break accidentally. +(!def-boolean-attribute ir1 + ;; may call functions that are passed as arguments. In order to + ;; determine what other effects are present, we must find the + ;; effects of all arguments that may be functions. call - ;; May incorporate function or number arguments into the result or somehow - ;; pass them upward. Note that this applies to any argument that *might* be - ;; a function or number, not just the arguments that always are. - unsafe - ;; May fail to return during correct execution. Errors are O.K. + ;; may fail to return during correct execution. Errors are O.K. + ;; UNUSED, BEWARE OF BITROT. unwind - ;; The (default) worst case. Includes all the other bad things, plus any - ;; other possible bad thing. If this is present, the above bad attributes - ;; will be explicitly present as well. + ;; the (default) worst case. Includes all the other bad things, plus + ;; any other possible bad thing. If this is present, the above bad + ;; attributes will be explicitly present as well. any - ;; May be constant-folded. The function has no side effects, but may be - ;; affected by side effects on the arguments. e.g. SVREF, MAPC. Functions - ;; that side-effect their arguments are not considered to be foldable. - ;; Although it would be "legal" to constant fold them (since it "is an error" - ;; to modify a constant), we choose not to mark these functions as foldable - ;; in this database. + ;; all arguments are safe for dynamic extent. + ;; (We used to have an UNSAFE attribute, which was basically the inverse + ;; of this, but it was unused and bitrotted, so when we started making + ;; use of the information we flipped the name and meaning the safe way + ;; around.) + dx-safe + ;; may be constant-folded. The function has no side effects, but may + ;; be affected by side effects on the arguments. e.g. SVREF, MAPC. + ;; Functions that side-effect their arguments are not considered to + ;; be foldable. Although it would be "legal" to constant fold them + ;; (since it "is an error" to modify a constant), we choose not to + ;; mark these functions as foldable in this database. foldable - ;; May be eliminated if value is unused. The function has no side effects - ;; except possibly CONS. If a function is defined to signal errors, then it - ;; is not flushable even if it is movable or foldable. + ;; may be eliminated if value is unused. The function has no side + ;; effects except possibly cons. If a function might signal errors, + ;; then it is not flushable even if it is movable, foldable or + ;; unsafely-flushable. Implies UNSAFELY-FLUSHABLE. (In safe code + ;; type checking of arguments is always performed by the caller, so + ;; a function which SHOULD signal an error if arguments are not of + ;; declared types may be FLUSHABLE.) flushable - ;; May be moved with impunity. Has no side effects except possibly CONS, and - ;; is affected only by its arguments. + ;; unsafe call may be eliminated if value is unused. The function + ;; has no side effects except possibly cons and signalling an error + ;; in the safe code. If a function MUST signal errors, then it is + ;; not unsafely-flushable even if it is movable or foldable. + unsafely-flushable + ;; return value is important, and ignoring it is probably a mistake. + ;; Unlike the other attributes, this is used only for style + ;; warnings and has no effect on optimization. + important-result + ;; may be moved with impunity. Has no side effects except possibly + ;; consing, and is affected only by its arguments. + ;; UNUSED, BEWARE OF BITROT. movable - ;; Function is a true predicate likely to be open-coded. Convert any - ;; non-conditional uses into (IF T NIL). + ;; The function is a true predicate likely to be open-coded. Convert + ;; any non-conditional uses into (IF T NIL). Not usually + ;; specified to DEFKNOWN, since this is implementation dependent, + ;; and is usually automatically set by the DEFINE-VOP :CONDITIONAL + ;; option. predicate - ;; Inhibit any warning for compiling a recursive definition. (Normally the - ;; compiler warns when compiling a recursive definition for a known function, - ;; since it might be a botched interpreter stub.) + ;; Inhibit any warning for compiling a recursive definition. + ;; (Normally the compiler warns when compiling a recursive + ;; definition for a known function, since it might be a botched + ;; interpreter stub.) recursive - ;; Function does explicit argument type checking, so the declared type should - ;; not be asserted when a definition is compiled. - explicit-check) - -(defstruct (function-info #-sb-xc-host (:pure t)) - ;; Boolean attributes of this function. - (attributes (required-argument) :type attributes) - ;; A list of Transform structures describing transforms for this function. + ;; The function does explicit argument type checking, so the + ;; declared type should not be asserted when a definition is + ;; compiled. + explicit-check + ;; The function should always be translated by a VOP (i.e. it should + ;; should never be converted into a full call). This is used strictly + ;; as a consistency checking mechanism inside the compiler during IR2 + ;; transformation. + always-translatable) + +(defstruct (fun-info #-sb-xc-host (:pure t)) + ;; boolean attributes of this function. + (attributes (missing-arg) :type attributes) + ;; TRANSFORM structures describing transforms for this function (transforms () :type list) - ;; A function which computes the derived type for a call to this function by - ;; examining the arguments. This is null when there is no special method for - ;; this function. + ;; a function which computes the derived type for a call to this + ;; function by examining the arguments. This is null when there is + ;; no special method for this function. (derive-type nil :type (or function null)) - ;; A function that does various unspecified code transformations by directly - ;; hacking the IR. Returns true if further optimizations of the call - ;; shouldn't be attempted. + ;; a function that does various unspecified code transformations by + ;; directly hacking the IR. Returns true if further optimizations of + ;; the call shouldn't be attempted. ;; - ;; KLUDGE: This return convention (non-NIL if you shouldn't do further - ;; optimiz'ns) is backwards from the return convention for transforms. - ;; -- WHN 19990917 + ;; KLUDGE: This return convention (non-NIL if you shouldn't do + ;; further optimiz'ns) is backwards from the return convention for + ;; transforms. -- WHN 19990917 (optimizer nil :type (or function null)) - ;; If true, a special-case LTN annotation method that is used in place of the - ;; standard type/policy template selection. It may use arbitrary code to - ;; choose a template, decide to do a full call, or conspire with the - ;; IR2-Convert method to do almost anything. The Combination node is passed - ;; as the argument. + ;; a function computing the constant or literal arguments which are + ;; destructively modified by the call. + (destroyed-constant-args nil :type (or function null)) + ;; If true, a special-case LTN annotation method that is used in + ;; place of the standard type/policy template selection. It may use + ;; arbitrary code to choose a template, decide to do a full call, or + ;; conspire with the IR2-CONVERT method to do almost anything. The + ;; COMBINATION node is passed as the argument. (ltn-annotate nil :type (or function null)) - ;; If true, the special-case IR2 conversion method for this function. This - ;; deals with funny functions, and anything else that can't be handled using - ;; the template mechanism. The Combination node and the IR2-Block are passed - ;; as arguments. + ;; If true, the special-case IR2 conversion method for this + ;; function. This deals with funny functions, and anything else that + ;; can't be handled using the template mechanism. The COMBINATION + ;; node and the IR2-BLOCK are passed as arguments. (ir2-convert nil :type (or function null)) - ;; A list of all the templates that could be used to translate this function + ;; If true, the function can stack-allocate the result. The + ;; COMBINATION node is passed as an argument. + (stack-allocate-result nil :type (or function null)) + ;; If true, the function can add flow-sensitive type information + ;; about the state of the world after its execution. The COMBINATION + ;; node is passed as an argument, along with the current set of + ;; active constraints for the block. The function returns a + ;; sequence of constraints; a constraint is a triplet of a + ;; constraint kind (a symbol, see (defstruct (constraint ...)) in + ;; constraint.lisp) and arguments, either LVARs, LAMBDA-VARs, or + ;; CTYPEs. If any of these arguments is NIL, the constraint is + ;; skipped. This simplifies integration with OK-LVAR-LAMBDA-VAR, + ;; which maps LVARs to LAMBDA-VARs. An optional fourth value in + ;; each constraint flips the meaning of the constraint if it is + ;; non-NIL. + (constraint-propagate nil :type (or function null)) + ;; If true, the function can add flow-sensitive type information + ;; depending on the truthiness of its return value. Returns two + ;; values, a LVAR and a CTYPE. The LVAR is of that CTYPE iff the + ;; function returns true. + ;; It may also return additional third and fourth values. Each is + ;; a sequence of constraints (see CONSTRAINT-PROPAGATE), for the + ;; consequent and alternative branches, respectively. + (constraint-propagate-if nil :type (or function null)) + ;; all the templates that could be used to translate this function ;; into IR2, sorted by increasing cost. (templates nil :type list) - ;; If non-null, then this function is a unary type predicate for this type. + ;; If non-null, then this function is a unary type predicate for + ;; this type. (predicate-type nil :type (or ctype null)) - ;; If non-null, use this function to annotate the known call for the byte - ;; compiler. If it returns NIL, then change the call to :full. - (byte-annotate nil :type (or function null)) - ;; If non-null, use this function to generate the byte code for this known - ;; call. This function can only give up if there is a byte-annotate function - ;; that arranged for the functional to be pushed onto the stack. - (byte-compile nil :type (or function null))) - -(defprinter (function-info) + ;; If non-null, the index of the argument which becomes the result + ;; of the function. + (result-arg nil :type (or index null))) + +(defprinter (fun-info) + (attributes :test (not (zerop attributes)) + :prin1 (decode-ir1-attributes attributes)) (transforms :test transforms) (derive-type :test derive-type) (optimizer :test optimizer) (ltn-annotate :test ltn-annotate) (ir2-convert :test ir2-convert) (templates :test templates) - (predicate-type :test predicate-type) - (byte-annotate :test byte-annotate) - (byte-compile :test byte-compile)) + (predicate-type :test predicate-type)) ;;;; interfaces to defining macros ;;; an IR1 transform (defstruct (transform (:copier nil)) - ;; the function-type which enables this transform - (type (required-argument) :type ctype) - ;; the transformation function. Takes the COMBINATION node and returns a - ;; lambda, or throws out. - (function (required-argument) :type function) - ;; string used in efficency notes - (note (required-argument) :type string) + ;; the function type which enables this transform. + ;; + ;; (Note that declaring this :TYPE FUN-TYPE probably wouldn't + ;; work because some function types, like (SPECIFIER-TYPE 'FUNCTION0 + ;; itself, are represented as BUILT-IN-TYPE, and at least as of + ;; sbcl-0.pre7.54 or so, that's inconsistent with being a + ;; FUN-TYPE.) + (type (missing-arg) :type ctype) + ;; the transformation function. Takes the COMBINATION node and + ;; returns a lambda expression, or throws out. + (function (missing-arg) :type function) + ;; string used in efficiency notes + (note (missing-arg) :type string) ;; T if we should emit a failure note even if SPEED=INHIBIT-WARNINGS. - (important nil :type (member t nil)) - ;; usable for byte code, native code, or both - (when :native :type (member :byte :native :both))) + (important nil :type (member t nil))) -(defprinter (transform) type note important when) +(defprinter (transform) type note important) -;;; Grab the FUNCTION-INFO and enter the function, replacing any old +;;; Grab the FUN-INFO and enter the function, replacing any old ;;; one with the same type and note. (declaim (ftype (function (t list function &optional (or string null) - (member t nil) (member :native :byte :both)) - *) - %deftransform)) -(defun %deftransform (name type fun &optional note important (when :native)) + (member t nil)) + *) + %deftransform)) +(defun %deftransform (name type fun &optional note important) (let* ((ctype (specifier-type type)) - (note (or note "optimize")) - (info (function-info-or-lose name)) - (old (find-if (lambda (x) - (and (type= (transform-type x) ctype) - (string-equal (transform-note x) note) - (eq (transform-important x) important) - (eq (transform-when x) when))) - (function-info-transforms info)))) - (if old - (setf (transform-function old) fun (transform-note old) note) - (push (make-transform :type ctype :function fun :note note - :important important :when when) - (function-info-transforms info))) + (note (or note "optimize")) + (info (fun-info-or-lose name)) + (old (find-if (lambda (x) + (and (type= (transform-type x) ctype) + (string-equal (transform-note x) note) + (eq (transform-important x) important))) + (fun-info-transforms info)))) + (cond (old + (style-warn 'sb!kernel:redefinition-with-deftransform + :transform old) + (setf (transform-function old) fun + (transform-note old) note)) + (t + (push (make-transform :type ctype :function fun :note note + :important important) + (fun-info-transforms info)))) name)) -;;; Make a FUNCTION-INFO structure with the specified type, attributes +;;; Make a FUN-INFO structure with the specified type, attributes ;;; and optimizers. (declaim (ftype (function (list list attributes &key - (:derive-type (or function null)) - (:optimizer (or function null))) - *) - %defknown)) -(defun %defknown (names type attributes &key derive-type optimizer) + (:derive-type (or function null)) + (:optimizer (or function null)) + (:destroyed-constant-args (or function null)) + (:result-arg (or index null)) + (:overwrite-fndb-silently boolean)) + *) + %defknown)) +(defun %defknown (names type attributes + &key derive-type optimizer destroyed-constant-args result-arg + overwrite-fndb-silently) (let ((ctype (specifier-type type)) - (info (make-function-info :attributes attributes - :derive-type derive-type - :optimizer optimizer)) - (target-env (or *backend-info-environment* *info-environment*))) + (info (make-fun-info :attributes attributes + :derive-type derive-type + :optimizer optimizer + :destroyed-constant-args destroyed-constant-args + :result-arg result-arg))) (dolist (name names) - (setf (info :function :type name target-env) ctype) - (setf (info :function :where-from name target-env) :declared) - (setf (info :function :kind name target-env) :function) - (setf (info :function :info name target-env) info))) + (unless overwrite-fndb-silently + (let ((old-fun-info (info :function :info name))) + (when old-fun-info + ;; This is handled as an error because it's generally a bad + ;; thing to blow away all the old optimization stuff. It's + ;; also a potential source of sneaky bugs: + ;; DEFKNOWN FOO + ;; DEFTRANSFORM FOO + ;; DEFKNOWN FOO ; possibly hidden inside some macroexpansion + ;; ; Now the DEFTRANSFORM doesn't exist in the target Lisp. + ;; However, it's continuable because it might be useful to do + ;; it when testing new optimization stuff interactively. + (cerror "Go ahead, overwrite it." + "~@" + old-fun-info name)))) + (setf (info :function :type name) ctype) + (setf (info :function :where-from name) :declared) + (setf (info :function :kind name) :function) + (setf (info :function :info name) info))) names) -;;; Return the FUNCTION-INFO for NAME or die trying. Since this is -;;; used by people who want to modify the info, and the info may be +;;; Return the FUN-INFO for NAME or die trying. Since this is +;;; used by callers who want to modify the info, and the info may be ;;; shared, we copy it. We don't have to copy the lists, since each ;;; function that has generators or transforms has already been ;;; through here. -(declaim (ftype (function (t) function-info) function-info-or-lose)) -(defun function-info-or-lose (name) - (let ((*info-environment* (or *backend-info-environment* - *info-environment*))) +(declaim (ftype (sfunction (t) fun-info) fun-info-or-lose)) +(defun fun-info-or-lose (name) + (let (;; FIXME: Do we need this rebinding here? It's a literal + ;; translation of the old CMU CL rebinding to + ;; (OR *BACKEND-INFO-ENVIRONMENT* *INFO-ENVIRONMENT*), + ;; and it's not obvious whether the rebinding to itself is + ;; needed that SBCL doesn't need *BACKEND-INFO-ENVIRONMENT*. + (*info-environment* *info-environment*)) (let ((old (info :function :info name))) (unless old (error "~S is not a known function." name)) - (setf (info :function :info name) (copy-function-info old))))) + (setf (info :function :info name) (copy-fun-info old))))) ;;;; generic type inference methods @@ -194,42 +285,124 @@ ;;; only be done when the result value is that argument. (defun result-type-first-arg (call) (declare (type combination call)) - (let ((cont (first (combination-args call)))) - (when cont (continuation-type cont)))) + (let ((lvar (first (combination-args call)))) + (when lvar (lvar-type lvar)))) (defun result-type-last-arg (call) (declare (type combination call)) - (let ((cont (car (last (combination-args call))))) - (when cont (continuation-type cont)))) + (let ((lvar (car (last (combination-args call))))) + (when lvar (lvar-type lvar)))) ;;; Derive the result type according to the float contagion rules, but -;;; always return a float. This is used for irrational functions that preserve -;;; realness of their arguments. +;;; always return a float. This is used for irrational functions that +;;; preserve realness of their arguments. (defun result-type-float-contagion (call) (declare (type combination call)) (reduce #'numeric-contagion (combination-args call) - :key #'continuation-type - :initial-value (specifier-type 'single-float))) + :key #'lvar-type + :initial-value (specifier-type 'single-float))) -;;; Return a closure usable as a derive-type method for accessing the N'th -;;; argument. If arg is a list, result is a list. If arg is a vector, result -;;; is a vector with the same element type. +;;; Return a closure usable as a derive-type method for accessing the +;;; N'th argument. If arg is a list, result is a list. If arg is a +;;; vector, result is a vector with the same element type. (defun sequence-result-nth-arg (n) - #'(lambda (call) - (declare (type combination call)) - (let ((cont (nth (1- n) (combination-args call)))) - (when cont - (let ((type (continuation-type cont))) - (if (array-type-p type) - (specifier-type - `(vector ,(type-specifier (array-type-element-type type)))) - (let ((ltype (specifier-type 'list))) - (when (csubtypep type ltype) - ltype)))))))) - -;;; Derive the type to be the type specifier which is the N'th arg. + (lambda (call) + (declare (type combination call)) + (let ((lvar (nth (1- n) (combination-args call)))) + (when lvar + (let ((type (lvar-type lvar))) + (if (array-type-p type) + (specifier-type + `(vector ,(type-specifier (array-type-element-type type)))) + (let ((ltype (specifier-type 'list))) + (when (csubtypep type ltype) + ltype)))))))) + +;;; Derive the type to be the type specifier which is the Nth arg. (defun result-type-specifier-nth-arg (n) (lambda (call) (declare (type combination call)) - (let ((cont (nth (1- n) (combination-args call)))) - (when (and cont (constant-continuation-p cont)) - (specifier-type (continuation-value cont)))))) + (let ((lvar (nth (1- n) (combination-args call)))) + (when (and lvar (constant-lvar-p lvar)) + (careful-specifier-type (lvar-value lvar)))))) + +;;; Derive the type to be the type specifier which is the Nth arg, +;;; with the additional restriptions noted in the CLHS for STRING and +;;; SIMPLE-STRING, defined to specialize on CHARACTER, and for VECTOR +;;; (under the page for MAKE-SEQUENCE). +(defun creation-result-type-specifier-nth-arg (n) + (lambda (call) + (declare (type combination call)) + (let ((lvar (nth (1- n) (combination-args call)))) + (when (and lvar (constant-lvar-p lvar)) + (let* ((specifier (lvar-value lvar)) + (lspecifier (if (atom specifier) (list specifier) specifier))) + (cond + ((eq (car lspecifier) 'string) + (destructuring-bind (string &rest size) + lspecifier + (declare (ignore string)) + (careful-specifier-type + `(vector character ,@(when size size))))) + ((eq (car lspecifier) 'simple-string) + (destructuring-bind (simple-string &rest size) + lspecifier + (declare (ignore simple-string)) + (careful-specifier-type + `(simple-array character ,@(if size (list size) '((*))))))) + (t + (let ((ctype (careful-specifier-type specifier))) + (if (and (array-type-p ctype) + (eq (array-type-specialized-element-type ctype) + *wild-type*)) + ;; I don't think I'm allowed to modify what I get + ;; back from SPECIFIER-TYPE; it is, after all, + ;; cached. Better copy it, then. + (let ((real-ctype (copy-structure ctype))) + (setf (array-type-element-type real-ctype) + *universal-type* + (array-type-specialized-element-type real-ctype) + *universal-type*) + real-ctype) + ctype))))))))) + +(defun remove-non-constants-and-nils (fun) + (lambda (list) + (remove-if-not #'lvar-value + (remove-if-not #'constant-lvar-p (funcall fun list))))) + +;;; FIXME: bad name (first because it uses 1-based indexing; second +;;; because it doesn't get the nth constant arguments) +(defun nth-constant-args (&rest indices) + (lambda (list) + (let (result) + (do ((i 1 (1+ i)) + (list list (cdr list)) + (indices indices)) + ((null indices) (nreverse result)) + (when (= i (car indices)) + (when (constant-lvar-p (car list)) + (push (car list) result)) + (setf indices (cdr indices))))))) + +;;; FIXME: a number of the sequence functions not only do not destroy +;;; their argument if it is empty, but also leave it alone if :start +;;; and :end bound a null sequence, or if :count is 0. This test is a +;;; bit complicated to implement, verging on the impossible, but for +;;; extra points (fill #\1 "abc" :start 0 :end 0) should not cause a +;;; warning. +(defun nth-constant-nonempty-sequence-args (&rest indices) + (lambda (list) + (let (result) + (do ((i 1 (1+ i)) + (list list (cdr list)) + (indices indices)) + ((null indices) (nreverse result)) + (when (= i (car indices)) + (when (constant-lvar-p (car list)) + (let ((value (lvar-value (car list)))) + (unless (or (typep value 'null) + (typep value '(vector * 0))) + (push (car list) result)))) + (setf indices (cdr indices))))))) + +(/show0 "knownfun.lisp end of file")