| blob | 863dbb395e5e0ece6f2ddbc569bc62a18de73277 |
1 ;;;; This file contains stuff for maintaining a database of special
2 ;;;; information about functions known to the compiler. This includes
3 ;;;; semantic information such as side effects and type inference
4 ;;;; functions as well as transforms and IR2 translators.
6 ;;;; This software is part of the SBCL system. See the README file for
7 ;;;; more information.
8 ;;;;
9 ;;;; This software is derived from the CMU CL system, which was
10 ;;;; written at Carnegie Mellon University and released into the
11 ;;;; public domain. The software is in the public domain and is
12 ;;;; provided with absolutely no warranty. See the COPYING and CREDITS
13 ;;;; files for more information.
19 ;;; IR1 boolean function attributes
20 ;;;
21 ;;; There are a number of boolean attributes of known functions which
22 ;;; we like to have in IR1. This information is mostly side effect
23 ;;; information of a sort, but it is different from the kind of
24 ;;; information we want in IR2. We aren't interested in a fine
25 ;;; breakdown of side effects, since we do very little code motion on
26 ;;; IR1. We are interested in some deeper semantic properties such as
27 ;;; whether it is safe to pass stack closures to.
28 (!def-boolean-attribute ir1
29 ;; may call functions that are passed as arguments. In order to
30 ;; determine what other effects are present, we must find the
31 ;; effects of all arguments that may be functions.
32 call
33 ;; may incorporate function or number arguments into the result or
34 ;; somehow pass them upward. Note that this applies to any argument
35 ;; that *might* be a function or number, not just the arguments that
36 ;; always are.
37 unsafe
38 ;; may fail to return during correct execution. Errors are O.K.
39 unwind
40 ;; the (default) worst case. Includes all the other bad things, plus
41 ;; any other possible bad thing. If this is present, the above bad
42 ;; attributes will be explicitly present as well.
43 any
44 ;; may be constant-folded. The function has no side effects, but may
45 ;; be affected by side effects on the arguments. e.g. SVREF, MAPC.
46 ;; Functions that side-effect their arguments are not considered to
47 ;; be foldable. Although it would be "legal" to constant fold them
48 ;; (since it "is an error" to modify a constant), we choose not to
49 ;; mark these functions as foldable in this database.
50 foldable
51 ;; may be eliminated if value is unused. The function has no side
52 ;; effects except possibly cons. If a function might signal errors,
53 ;; then it is not flushable even if it is movable, foldable or
54 ;; unsafely-flushable. Implies UNSAFELY-FLUSHABLE. (In safe code
55 ;; type checking of arguments is always performed by the caller, so
56 ;; a function which SHOULD signal an error if arguments are not of
57 ;; declared types may be FLUSHABLE.)
58 flushable
59 ;; unsafe call may be eliminated if value is unused. The function
60 ;; has no side effects except possibly cons and signalling an error
61 ;; in the safe code. If a function MUST signal errors, then it is
62 ;; not unsafely-flushable even if it is movable or foldable.
63 unsafely-flushable
64 ;; return value is important, and ignoring it is probably a mistake.
65 ;; Unlike the other attributes, this is used only for style
66 ;; warnings and has no effect on optimization.
67 important-result
68 ;; may be moved with impunity. Has no side effects except possibly
69 ;; consing, and is affected only by its arguments.
70 ;;
71 ;; Since it is not used now, its distribution in fndb.lisp is
72 ;; mere random; use with caution.
73 movable
74 ;; The function is a true predicate likely to be open-coded. Convert
75 ;; any non-conditional uses into (IF <pred> T NIL). Not usually
76 ;; specified to DEFKNOWN, since this is implementation dependent,
77 ;; and is usually automatically set by the DEFINE-VOP :CONDITIONAL
78 ;; option.
79 predicate
80 ;; Inhibit any warning for compiling a recursive definition.
81 ;; (Normally the compiler warns when compiling a recursive
82 ;; definition for a known function, since it might be a botched
83 ;; interpreter stub.)
84 recursive
85 ;; The function does explicit argument type checking, so the
86 ;; declared type should not be asserted when a definition is
87 ;; compiled.
88 explicit-check
89 ;; The function should always be translated by a VOP (i.e. it should
90 ;; should never be converted into a full call). This is used strictly
91 ;; as a consistency checking mechanism inside the compiler during IR2
92 ;; transformation.
93 always-translatable)
96 ;; boolean attributes of this function.
98 ;; TRANSFORM structures describing transforms for this function
100 ;; a function which computes the derived type for a call to this
101 ;; function by examining the arguments. This is null when there is
102 ;; no special method for this function.
104 ;; a function that does various unspecified code transformations by
105 ;; directly hacking the IR. Returns true if further optimizations of
106 ;; the call shouldn't be attempted.
107 ;;
108 ;; KLUDGE: This return convention (non-NIL if you shouldn't do
109 ;; further optimiz'ns) is backwards from the return convention for
110 ;; transforms. -- WHN 19990917
112 ;; a function computing the constant or literal arguments which are
113 ;; destructively modified by the call.
115 ;; If true, a special-case LTN annotation method that is used in
116 ;; place of the standard type/policy template selection. It may use
117 ;; arbitrary code to choose a template, decide to do a full call, or
118 ;; conspire with the IR2-CONVERT method to do almost anything. The
119 ;; COMBINATION node is passed as the argument.
121 ;; If true, the special-case IR2 conversion method for this
122 ;; function. This deals with funny functions, and anything else that
123 ;; can't be handled using the template mechanism. The COMBINATION
124 ;; node and the IR2-BLOCK are passed as arguments.
126 ;; If true, the function can stack-allocate the result. The
127 ;; COMBINATION node is passed as an argument.
129 ;; all the templates that could be used to translate this function
130 ;; into IR2, sorted by increasing cost.
132 ;; If non-null, then this function is a unary type predicate for
133 ;; this type.
146 \f
147 ;;;; interfaces to defining macros
149 ;;; an IR1 transform
151 ;; the function type which enables this transform.
152 ;;
153 ;; (Note that declaring this :TYPE FUN-TYPE probably wouldn't
154 ;; work because some function types, like (SPECIFIER-TYPE 'FUNCTION0
155 ;; itself, are represented as BUILT-IN-TYPE, and at least as of
156 ;; sbcl-0.pre7.54 or so, that's inconsistent with being a
157 ;; FUN-TYPE.)
159 ;; the transformation function. Takes the COMBINATION node and
160 ;; returns a lambda expression, or throws out.
162 ;; string used in efficiency notes
164 ;; T if we should emit a failure note even if SPEED=INHIBIT-WARNINGS.
169 ;;; Grab the FUN-INFO and enter the function, replacing any old
170 ;;; one with the same type and note.
173 *)
174 %deftransform))
193 name))
195 ;;; Make a FUN-INFO structure with the specified type, attributes
196 ;;; and optimizers.
201 *)
202 %defknown))
206 :derive-type derive-type
207 :optimizer optimizer
213 ;; This is handled as an error because it's generally a bad
214 ;; thing to blow away all the old optimization stuff. It's
215 ;; also a potential source of sneaky bugs:
216 ;; DEFKNOWN FOO
217 ;; DEFTRANSFORM FOO
218 ;; DEFKNOWN FOO ; possibly hidden inside some macroexpansion
219 ;; ; Now the DEFTRANSFORM doesn't exist in the target Lisp.
220 ;; However, it's continuable because it might be useful to do
221 ;; it when testing new optimization stuff interactively.
223 "~@<overwriting old FUN-INFO ~2I~_~S ~I~_for ~S~:>"
224 old-fun-info name)))
229 names)
231 ;;; Return the FUN-INFO for NAME or die trying. Since this is
232 ;;; used by callers who want to modify the info, and the info may be
233 ;;; shared, we copy it. We don't have to copy the lists, since each
234 ;;; function that has generators or transforms has already been
235 ;;; through here.
239 ;; translation of the old CMU CL rebinding to
240 ;; (OR *BACKEND-INFO-ENVIRONMENT* *INFO-ENVIRONMENT*),
241 ;; and it's not obvious whether the rebinding to itself is
242 ;; needed that SBCL doesn't need *BACKEND-INFO-ENVIRONMENT*.
247 \f
248 ;;;; generic type inference methods
250 ;;; Derive the type to be the type of the xxx'th arg. This can normally
251 ;;; only be done when the result value is that argument.
261 ;;; Derive the result type according to the float contagion rules, but
262 ;;; always return a float. This is used for irrational functions that
263 ;;; preserve realness of their arguments.
270 ;;; Return a closure usable as a derive-type method for accessing the
271 ;;; N'th argument. If arg is a list, result is a list. If arg is a
272 ;;; vector, result is a vector with the same element type.
284 ltype))))))))
286 ;;; Derive the type to be the type specifier which is the Nth arg.
294 ;;; Derive the type to be the type specifier which is the Nth arg,
295 ;;; with the additional restriptions noted in the CLHS for STRING and
296 ;;; SIMPLE-STRING, defined to specialize on CHARACTER, and for VECTOR
297 ;;; (under the page for MAKE-SEQUENCE).
308 lspecifier
314 lspecifier
323 ;; I don't think I'm allowed to modify what I get
324 ;; back from SPECIFIER-TYPE; it is, after all,
325 ;; cached. Better copy it, then.
331 real-ctype)
332 ctype)))))))))
339 ;;; FIXME: bad name (first because it uses 1-based indexing; second
340 ;;; because it doesn't get the nth constant arguments)
353 ;;; FIXME: a number of the sequence functions not only do not destroy
354 ;;; their argument if it is empty, but also leave it alone if :start
355 ;;; and :end bound a null sequence, or if :count is 0. This test is a
356 ;;; bit complicated to implement, verging on the impossible, but for
357 ;;; extra points (fill #\1 "abc" :start 0 :end 0) should not cause a
358 ;;; warning.