1.0.20.16: make LOCK and FS prefixes part of the affected instruction
[sbcl/pkhuong.git] / src / compiler / x86 / macros.lisp
blob9789ec21ba8028b98c2271d7498c14332cc71c13
1 ;;;; a bunch of handy macros for the x86
3 ;;;; This software is part of the SBCL system. See the README file for
4 ;;;; more information.
5 ;;;;
6 ;;;; This software is derived from the CMU CL system, which was
7 ;;;; written at Carnegie Mellon University and released into the
8 ;;;; public domain. The software is in the public domain and is
9 ;;;; provided with absolutely no warranty. See the COPYING and CREDITS
10 ;;;; files for more information.
12 (in-package "SB!VM")
14 ;;; We can load/store into fp registers through the top of stack
15 ;;; %st(0) (fr0 here). Loads imply a push to an empty register which
16 ;;; then changes all the reg numbers. These macros help manage that.
18 ;;; Use this when we don't have to load anything. It preserves old tos
19 ;;; value, but probably destroys tn with operation.
20 (defmacro with-tn@fp-top((tn) &body body)
21 `(progn
22 (unless (zerop (tn-offset ,tn))
23 (inst fxch ,tn))
24 ,@body
25 (unless (zerop (tn-offset ,tn))
26 (inst fxch ,tn))))
28 ;;; Use this to prepare for load of new value from memory. This
29 ;;; changes the register numbering so the next instruction had better
30 ;;; be a FP load from memory; a register load from another register
31 ;;; will probably be loading the wrong register!
32 (defmacro with-empty-tn@fp-top((tn) &body body)
33 `(progn
34 (inst fstp ,tn)
35 ,@body
36 (unless (zerop (tn-offset ,tn))
37 (inst fxch ,tn)))) ; save into new dest and restore st(0)
39 ;;;; instruction-like macros
41 (defmacro move (dst src)
42 #!+sb-doc
43 "Move SRC into DST unless they are location=."
44 (once-only ((n-dst dst)
45 (n-src src))
46 `(unless (location= ,n-dst ,n-src)
47 (inst mov ,n-dst ,n-src))))
49 (defmacro align-stack-pointer (tn)
50 #!-darwin (declare (ignore tn))
51 #!+darwin
52 ;; 16 byte alignment.
53 `(inst and ,tn #xfffffff0))
55 (defmacro make-ea-for-object-slot (ptr slot lowtag &optional (size :dword))
56 `(make-ea ,size :base ,ptr :disp (- (* ,slot n-word-bytes) ,lowtag)))
58 (defmacro loadw (value ptr &optional (slot 0) (lowtag 0))
59 `(inst mov ,value (make-ea-for-object-slot ,ptr ,slot ,lowtag)))
61 (defmacro storew (value ptr &optional (slot 0) (lowtag 0))
62 (once-only ((value value))
63 `(inst mov (make-ea-for-object-slot ,ptr ,slot ,lowtag) ,value)))
65 ;;; A handy macro for storing widetags.
66 (defmacro storeb (value ptr &optional (slot 0) (lowtag 0))
67 (once-only ((value value))
68 `(inst mov (make-ea-for-object-slot ,ptr ,slot ,lowtag :byte) ,value)))
70 (defmacro pushw (ptr &optional (slot 0) (lowtag 0))
71 `(inst push (make-ea-for-object-slot ,ptr ,slot ,lowtag)))
73 (defmacro popw (ptr &optional (slot 0) (lowtag 0))
74 `(inst pop (make-ea-for-object-slot ,ptr ,slot ,lowtag)))
76 (defmacro make-ea-for-vector-data (object &key (size :dword) (offset 0)
77 index (scale (ash (width-bits size) -3)))
78 `(make-ea ,size :base ,object :index ,index :scale ,scale
79 :disp (- (+ (* vector-data-offset n-word-bytes)
80 (* ,offset ,scale))
81 other-pointer-lowtag)))
83 ;;;; macros to generate useful values
85 (defmacro load-symbol (reg symbol)
86 `(inst mov ,reg (+ nil-value (static-symbol-offset ,symbol))))
88 (defmacro make-ea-for-symbol-value (symbol &optional (width :dword))
89 (declare (type symbol symbol))
90 `(make-ea ,width
91 :disp (+ nil-value
92 (static-symbol-offset ',symbol)
93 (ash symbol-value-slot word-shift)
94 (- other-pointer-lowtag))))
96 (defmacro load-symbol-value (reg symbol)
97 `(inst mov ,reg (make-ea-for-symbol-value ,symbol)))
99 (defmacro store-symbol-value (reg symbol)
100 `(inst mov (make-ea-for-symbol-value ,symbol) ,reg))
102 #!+sb-thread
103 (defmacro make-ea-for-symbol-tls-index (symbol)
104 (declare (type symbol symbol))
105 `(make-ea :dword
106 :disp (+ nil-value
107 (static-symbol-offset ',symbol)
108 (ash symbol-tls-index-slot word-shift)
109 (- other-pointer-lowtag))))
111 #!+sb-thread
112 (defmacro load-tl-symbol-value (reg symbol)
113 `(progn
114 (inst mov ,reg (make-ea-for-symbol-tls-index ,symbol))
115 (inst mov ,reg (make-ea :dword :base ,reg) :fs)))
116 #!-sb-thread
117 (defmacro load-tl-symbol-value (reg symbol) `(load-symbol-value ,reg ,symbol))
119 #!+sb-thread
120 (defmacro store-tl-symbol-value (reg symbol temp)
121 `(progn
122 (inst mov ,temp (make-ea-for-symbol-tls-index ,symbol))
123 (inst mov (make-ea :dword :base ,temp) ,reg :fs)))
124 #!-sb-thread
125 (defmacro store-tl-symbol-value (reg symbol temp)
126 (declare (ignore temp))
127 `(store-symbol-value ,reg ,symbol))
129 (defmacro load-binding-stack-pointer (reg)
130 #!+sb-thread
131 `(progn
132 (inst mov ,reg (make-ea :dword
133 :disp (* 4 thread-binding-stack-pointer-slot))
134 :fs))
135 #!-sb-thread
136 `(load-symbol-value ,reg *binding-stack-pointer*))
138 (defmacro store-binding-stack-pointer (reg)
139 #!+sb-thread
140 `(progn
141 (inst mov (make-ea :dword
142 :disp (* 4 thread-binding-stack-pointer-slot))
143 ,reg :fs))
144 #!-sb-thread
145 `(store-symbol-value ,reg *binding-stack-pointer*))
147 (defmacro load-type (target source &optional (offset 0))
148 #!+sb-doc
149 "Loads the type bits of a pointer into target independent of
150 byte-ordering issues."
151 (once-only ((n-target target)
152 (n-source source)
153 (n-offset offset))
154 (ecase *backend-byte-order*
155 (:little-endian
156 `(inst mov ,n-target
157 (make-ea :byte :base ,n-source :disp ,n-offset)))
158 (:big-endian
159 `(inst mov ,n-target
160 (make-ea :byte :base ,n-source
161 :disp (+ ,n-offset (1- n-word-bytes))))))))
163 ;;;; allocation helpers
165 ;;; Allocation within alloc_region (which is thread local) can be done
166 ;;; inline. If the alloc_region is overflown allocation is done by
167 ;;; calling the C alloc() function.
169 ;;; C calls for allocation don't /seem/ to make an awful lot of
170 ;;; difference to speed. On pure consing it's about a 25%
171 ;;; gain. Guessing from historical context, it looks like inline
172 ;;; allocation was introduced before pseudo-atomic, at which time all
173 ;;; calls to alloc() would have needed a syscall to mask signals for
174 ;;; the duration. Now we have pseudoatomic there's no need for that
175 ;;; overhead.
177 (defun allocation-dynamic-extent (alloc-tn size lowtag)
178 (inst sub esp-tn size)
179 ;; FIXME: SIZE _should_ be double-word aligned (suggested but
180 ;; unfortunately not enforced by PAD-DATA-BLOCK and
181 ;; WITH-FIXED-ALLOCATION), so that ESP is always divisible by 8 (for
182 ;; 32-bit lispobjs). In that case, this AND instruction is
183 ;; unneccessary and could be removed. If not, explain why. -- CSR,
184 ;; 2004-03-30
185 (inst and esp-tn (lognot lowtag-mask))
186 (aver (not (location= alloc-tn esp-tn)))
187 (inst lea alloc-tn (make-ea :byte :base esp-tn :disp lowtag))
188 (values))
190 (defun allocation-notinline (alloc-tn size)
191 (let* ((alloc-tn-offset (tn-offset alloc-tn))
192 ;; C call to allocate via dispatch routines. Each
193 ;; destination has a special entry point. The size may be a
194 ;; register or a constant.
195 (tn-text (ecase alloc-tn-offset
196 (#.eax-offset "eax")
197 (#.ecx-offset "ecx")
198 (#.edx-offset "edx")
199 (#.ebx-offset "ebx")
200 (#.esi-offset "esi")
201 (#.edi-offset "edi")))
202 (size-text (case size (8 "8_") (16 "16_") (t ""))))
203 (unless (or (eql size 8) (eql size 16))
204 (unless (and (tn-p size) (location= alloc-tn size))
205 (inst mov alloc-tn size)))
206 (inst call (make-fixup (concatenate 'string
207 "alloc_" size-text
208 "to_" tn-text)
209 :foreign))))
211 (defun allocation-inline (alloc-tn size)
212 (let ((ok (gen-label))
213 (done (gen-label))
214 (free-pointer
215 (make-ea :dword :disp
216 #!+sb-thread (* n-word-bytes thread-alloc-region-slot)
217 #!-sb-thread (make-fixup "boxed_region" :foreign)
218 :scale 1)) ; thread->alloc_region.free_pointer
219 (end-addr
220 (make-ea :dword :disp
221 #!+sb-thread (* n-word-bytes (1+ thread-alloc-region-slot))
222 #!-sb-thread (make-fixup "boxed_region" :foreign 4)
223 :scale 1))) ; thread->alloc_region.end_addr
224 (unless (and (tn-p size) (location= alloc-tn size))
225 (inst mov alloc-tn size))
226 (inst add alloc-tn free-pointer #!+sb-thread :fs)
227 (inst cmp alloc-tn end-addr #!+sb-thread :fs)
228 (inst jmp :be ok)
229 (let ((dst (ecase (tn-offset alloc-tn)
230 (#.eax-offset "alloc_overflow_eax")
231 (#.ecx-offset "alloc_overflow_ecx")
232 (#.edx-offset "alloc_overflow_edx")
233 (#.ebx-offset "alloc_overflow_ebx")
234 (#.esi-offset "alloc_overflow_esi")
235 (#.edi-offset "alloc_overflow_edi"))))
236 (inst call (make-fixup dst :foreign)))
237 (inst jmp-short done)
238 (emit-label ok)
239 ;; Swap ALLOC-TN and FREE-POINTER
240 (cond ((and (tn-p size) (location= alloc-tn size))
241 ;; XCHG is extremely slow, use the xor swap trick
242 (inst xor alloc-tn free-pointer #!+sb-thread :fs)
243 (inst xor free-pointer alloc-tn #!+sb-thread :fs)
244 (inst xor alloc-tn free-pointer #!+sb-thread :fs))
246 ;; It's easier if SIZE is still available.
247 (inst mov free-pointer alloc-tn #!+sb-thread :fs)
248 (inst sub alloc-tn size)))
249 (emit-label done))
250 (values))
253 ;;; Emit code to allocate an object with a size in bytes given by
254 ;;; SIZE. The size may be an integer or a TN. If Inline is a VOP
255 ;;; node-var then it is used to make an appropriate speed vs size
256 ;;; decision.
258 ;;; Allocation should only be used inside a pseudo-atomic section, which
259 ;;; should also cover subsequent initialization of the object.
261 ;;; (FIXME: so why aren't we asserting this?)
263 (defun allocation (alloc-tn size &optional inline dynamic-extent lowtag)
264 (cond
265 (dynamic-extent
266 (allocation-dynamic-extent alloc-tn size lowtag))
267 ((or (null inline) (policy inline (>= speed space)))
268 (allocation-inline alloc-tn size))
270 (allocation-notinline alloc-tn size)))
271 (when (and lowtag (not dynamic-extent))
272 (inst lea alloc-tn (make-ea :byte :base alloc-tn :disp lowtag)))
273 (values))
275 ;;; Allocate an other-pointer object of fixed SIZE with a single word
276 ;;; header having the specified WIDETAG value. The result is placed in
277 ;;; RESULT-TN.
278 (defmacro with-fixed-allocation ((result-tn widetag size &optional inline stack-allocate-p)
279 &body forms)
280 (unless forms
281 (bug "empty &body in WITH-FIXED-ALLOCATION"))
282 (once-only ((result-tn result-tn) (size size) (stack-allocate-p stack-allocate-p))
283 `(maybe-pseudo-atomic ,stack-allocate-p
284 (allocation ,result-tn (pad-data-block ,size) ,inline ,stack-allocate-p
285 other-pointer-lowtag)
286 (storew (logior (ash (1- ,size) n-widetag-bits) ,widetag)
287 ,result-tn 0 other-pointer-lowtag)
288 ,@forms)))
290 ;;;; error code
291 (defun emit-error-break (vop kind code values)
292 (assemble ()
293 #!-darwin
294 (inst int 3) ; i386 breakpoint instruction
295 ;; CLH 20060314
296 ;; On Darwin, we need to use #x0b0f instead of int3 in order
297 ;; to generate a SIGILL instead of a SIGTRAP as darwin/x86
298 ;; doesn't seem to be reliably firing SIGTRAP
299 ;; handlers. Hopefully this will be fixed by Apple at a
300 ;; later date.
301 #!+darwin
302 (inst word #x0b0f)
303 ;; The return PC points here; note the location for the debugger.
304 (when vop
305 (note-this-location vop :internal-error))
306 (inst byte kind) ; e.g. trap_xyyy
307 (with-adjustable-vector (vector) ; interr arguments
308 (write-var-integer code vector)
309 (dolist (tn values)
310 ;; classic CMU CL comment:
311 ;; zzzzz jrd here. tn-offset is zero for constant
312 ;; tns.
313 (write-var-integer (make-sc-offset (sc-number (tn-sc tn))
314 (or (tn-offset tn) 0))
315 vector))
316 (inst byte (length vector))
317 (dotimes (i (length vector))
318 (inst byte (aref vector i))))))
320 (defun error-call (vop error-code &rest values)
321 #!+sb-doc
322 "Cause an error. ERROR-CODE is the error to cause."
323 (emit-error-break vop error-trap (error-number-or-lose error-code) values))
325 (defun generate-error-code (vop error-code &rest values)
326 #!+sb-doc
327 "Generate-Error-Code Error-code Value*
328 Emit code for an error with the specified Error-Code and context Values."
329 (assemble (*elsewhere*)
330 (let ((start-lab (gen-label)))
331 (emit-label start-lab)
332 (emit-error-break vop error-trap (error-number-or-lose error-code) values)
333 start-lab)))
336 ;;;; PSEUDO-ATOMIC
338 ;;; This is used to wrap operations which leave untagged memory lying
339 ;;; around. It's an operation which the AOP weenies would describe as
340 ;;; having "cross-cutting concerns", meaning it appears all over the
341 ;;; place and there's no logical single place to attach documentation.
342 ;;; grep (mostly in src/runtime) is your friend
344 ;;; KLUDGE: since the stack on the x86 is treated conservatively, it
345 ;;; does not matter whether a signal occurs during construction of a
346 ;;; dynamic-extent object, as the half-finished construction of the
347 ;;; object will not cause any difficulty. We can therefore elide
348 (defmacro maybe-pseudo-atomic (not-really-p &body forms)
349 `(if ,not-really-p
350 (progn ,@forms)
351 (pseudo-atomic ,@forms)))
353 #!+sb-thread
354 (defmacro pseudo-atomic (&rest forms)
355 (with-unique-names (label)
356 `(let ((,label (gen-label)))
357 (inst or (make-ea :byte :disp (* 4 thread-pseudo-atomic-bits-slot))
358 (fixnumize 1) :fs)
359 ,@forms
360 (inst xor (make-ea :byte :disp (* 4 thread-pseudo-atomic-bits-slot))
361 (fixnumize 1) :fs)
362 (inst jmp :z ,label)
363 ;; if PAI was set, interrupts were disabled at the same
364 ;; time using the process signal mask.
365 (inst break pending-interrupt-trap)
366 (emit-label ,label))))
368 #!-sb-thread
369 (defmacro pseudo-atomic (&rest forms)
370 (with-unique-names (label)
371 `(let ((,label (gen-label)))
372 (inst or (make-ea-for-symbol-value *pseudo-atomic-bits* :byte)
373 (fixnumize 1))
374 ,@forms
375 (inst xor (make-ea-for-symbol-value *pseudo-atomic-bits* :byte)
376 (fixnumize 1))
377 (inst jmp :z ,label)
378 ;; if PAI was set, interrupts were disabled at the same
379 ;; time using the process signal mask.
380 (inst break pending-interrupt-trap)
381 (emit-label ,label))))
383 ;;;; indexed references
385 (defmacro define-full-compare-and-swap
386 (name type offset lowtag scs el-type &optional translate)
387 `(progn
388 (define-vop (,name)
389 ,@(when translate `((:translate ,translate)))
390 (:policy :fast-safe)
391 (:args (object :scs (descriptor-reg) :to :eval)
392 (index :scs (any-reg immediate unsigned-reg) :to :result)
393 (old-value :scs ,scs :target eax)
394 (new-value :scs ,scs))
395 (:arg-types ,type tagged-num ,el-type ,el-type)
396 (:temporary (:sc descriptor-reg :offset eax-offset
397 :from (:argument 2) :to :result :target value) eax)
398 (:results (value :scs ,scs))
399 (:result-types ,el-type)
400 (:generator 5
401 (move eax old-value)
402 (let ((ea (sc-case index
403 (immediate
404 (make-ea :dword :base object
405 :disp (- (* (+ ,offset (tn-value index))
406 n-word-bytes)
407 ,lowtag)))
408 (unsigned-reg
409 (make-ea :dword :base object :index index :scale 4
410 :disp (- (* ,offset n-word-bytes)
411 ,lowtag)))
413 (make-ea :dword :base object :index index
414 :disp (- (* ,offset n-word-bytes)
415 ,lowtag))))))
416 (inst cmpxchg ea new-value :lock))
417 (move value eax)))))
419 (defmacro define-full-reffer (name type offset lowtag scs el-type &optional translate)
420 `(progn
421 (define-vop (,name)
422 ,@(when translate
423 `((:translate ,translate)))
424 (:policy :fast-safe)
425 (:args (object :scs (descriptor-reg))
426 (index :scs (any-reg immediate unsigned-reg)))
427 (:arg-types ,type tagged-num)
428 (:results (value :scs ,scs))
429 (:result-types ,el-type)
430 (:generator 3 ; pw was 5
431 (sc-case index
432 (immediate
433 (inst mov value (make-ea :dword :base object
434 :disp (- (* (+ ,offset (tn-value index))
435 n-word-bytes)
436 ,lowtag))))
437 (unsigned-reg
438 (inst mov value (make-ea :dword :base object :index index :scale 4
439 :disp (- (* ,offset n-word-bytes)
440 ,lowtag))))
442 (inst mov value (make-ea :dword :base object :index index
443 :disp (- (* ,offset n-word-bytes)
444 ,lowtag)))))))))
446 (defmacro define-full-reffer+offset (name type offset lowtag scs el-type &optional translate)
447 `(progn
448 (define-vop (,name)
449 ,@(when translate
450 `((:translate ,translate)))
451 (:policy :fast-safe)
452 (:args (object :scs (descriptor-reg))
453 (index :scs (any-reg immediate unsigned-reg)))
454 (:arg-types ,type tagged-num
455 (:constant (constant-displacement ,lowtag sb!vm:n-word-bytes ,offset)))
456 (:info offset)
457 (:results (value :scs ,scs))
458 (:result-types ,el-type)
459 (:generator 3 ; pw was 5
460 (sc-case index
461 (immediate
462 (inst mov value (make-ea :dword :base object
463 :disp (- (* (+ ,offset
464 (tn-value index)
465 offset)
466 n-word-bytes)
467 ,lowtag))))
468 (unsigned-reg
469 (inst mov value (make-ea :dword :base object :index index :scale 4
470 :disp (- (* (+ ,offset offset)
471 n-word-bytes)
472 ,lowtag))))
474 (inst mov value (make-ea :dword :base object :index index
475 :disp (- (* (+ ,offset offset)
476 n-word-bytes)
477 ,lowtag)))))))))
479 (defmacro define-full-setter (name type offset lowtag scs el-type &optional translate)
480 `(progn
481 (define-vop (,name)
482 ,@(when translate
483 `((:translate ,translate)))
484 (:policy :fast-safe)
485 (:args (object :scs (descriptor-reg))
486 (index :scs (any-reg immediate))
487 (value :scs ,scs :target result))
488 (:arg-types ,type tagged-num ,el-type)
489 (:results (result :scs ,scs))
490 (:result-types ,el-type)
491 (:generator 4 ; was 5
492 (sc-case index
493 (immediate
494 (inst mov (make-ea :dword :base object
495 :disp (- (* (+ ,offset (tn-value index))
496 n-word-bytes)
497 ,lowtag))
498 value))
500 (inst mov (make-ea :dword :base object :index index
501 :disp (- (* ,offset n-word-bytes) ,lowtag))
502 value)))
503 (move result value)))))
505 (defmacro define-full-setter+offset (name type offset lowtag scs el-type &optional translate)
506 `(progn
507 (define-vop (,name)
508 ,@(when translate
509 `((:translate ,translate)))
510 (:policy :fast-safe)
511 (:args (object :scs (descriptor-reg))
512 (index :scs (any-reg immediate))
513 (value :scs ,scs :target result))
514 (:info offset)
515 (:arg-types ,type tagged-num
516 (:constant (constant-displacement ,lowtag sb!vm:n-word-bytes ,offset)) ,el-type)
517 (:results (result :scs ,scs))
518 (:result-types ,el-type)
519 (:generator 4 ; was 5
520 (sc-case index
521 (immediate
522 (inst mov (make-ea :dword :base object
523 :disp (- (* (+ ,offset (tn-value index) offset)
524 n-word-bytes)
525 ,lowtag))
526 value))
528 (inst mov (make-ea :dword :base object :index index
529 :disp (- (* (+ ,offset offset)
530 n-word-bytes) ,lowtag))
531 value)))
532 (move result value)))))
534 ;;; helper for alien stuff.
536 (def!macro with-pinned-objects ((&rest objects) &body body)
537 "Arrange with the garbage collector that the pages occupied by
538 OBJECTS will not be moved in memory for the duration of BODY.
539 Useful for e.g. foreign calls where another thread may trigger
540 collection."
541 (if objects
542 (let ((pins (make-gensym-list (length objects)))
543 (wpo (block-gensym "WPO")))
544 ;; BODY is stuffed in a function to preserve the lexical
545 ;; environment.
546 `(flet ((,wpo () (progn ,@body)))
547 ;; PINS are dx-allocated in case the compiler for some
548 ;; unfathomable reason decides to allocate value-cells
549 ;; for them -- since we have DX value-cells on x86oid
550 ;; platforms this still forces them on the stack.
551 (dx-let ,(mapcar #'list pins objects)
552 (multiple-value-prog1 (,wpo)
553 ;; TOUCH-OBJECT has a VOP with an empty body: compiler
554 ;; thinks we're using the argument and doesn't flush
555 ;; the variable, but we don't have to pay any extra
556 ;; beyond that -- and MULTIPLE-VALUE-PROG1 keeps them
557 ;; live till the body has finished. *whew*
558 ,@(mapcar (lambda (pin)
559 `(touch-object ,pin))
560 pins)))))
561 `(progn ,@body)))