1 ;;;; a bunch of handy macros for the x86
3 ;;;; This software is part of the SBCL system. See the README file for
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.
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)
22 (unless (zerop (tn-offset ,tn))
25 (unless (zerop (tn-offset ,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)
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)
43 "Move SRC into DST unless they are location=."
44 (once-only ((n-dst dst)
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))
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)
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))
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))
103 (defmacro make-ea-for-symbol-tls-index (symbol)
104 (declare (type symbol symbol))
107 (static-symbol-offset ',symbol)
108 (ash symbol-tls-index-slot word-shift)
109 (- other-pointer-lowtag))))
112 (defmacro load-tl-symbol-value (reg symbol)
114 (inst mov ,reg (make-ea-for-symbol-tls-index ,symbol))
115 (inst fs-segment-prefix)
116 (inst mov ,reg (make-ea :dword :base ,reg))))
118 (defmacro load-tl-symbol-value (reg symbol) `(load-symbol-value ,reg ,symbol))
121 (defmacro store-tl-symbol-value (reg symbol temp)
123 (inst mov ,temp (make-ea-for-symbol-tls-index ,symbol))
124 (inst fs-segment-prefix)
125 (inst mov (make-ea :dword :base ,temp) ,reg)))
127 (defmacro store-tl-symbol-value (reg symbol temp)
128 (declare (ignore temp))
129 `(store-symbol-value ,reg ,symbol))
131 (defmacro load-binding-stack-pointer (reg)
134 (inst fs-segment-prefix)
135 (inst mov ,reg (make-ea :dword
136 :disp (* 4 thread-binding-stack-pointer-slot))))
138 `(load-symbol-value ,reg *binding-stack-pointer*))
140 (defmacro store-binding-stack-pointer (reg)
143 (inst fs-segment-prefix)
144 (inst mov (make-ea :dword
145 :disp (* 4 thread-binding-stack-pointer-slot))
148 `(store-symbol-value ,reg *binding-stack-pointer*))
150 (defmacro load-type (target source &optional (offset 0))
152 "Loads the type bits of a pointer into target independent of
153 byte-ordering issues."
154 (once-only ((n-target target)
157 (ecase *backend-byte-order*
160 (make-ea :byte :base ,n-source :disp ,n-offset)))
163 (make-ea :byte :base ,n-source :disp (+ ,n-offset 3)))))))
165 ;;;; allocation helpers
167 ;;; Allocation within alloc_region (which is thread local) can be done
168 ;;; inline. If the alloc_region is overflown allocation is done by
169 ;;; calling the C alloc() function.
171 ;;; C calls for allocation don't /seem/ to make an awful lot of
172 ;;; difference to speed. On pure consing it's about a 25%
173 ;;; gain. Guessing from historical context, it looks like inline
174 ;;; allocation was introduced before pseudo-atomic, at which time all
175 ;;; calls to alloc() would have needed a syscall to mask signals for
176 ;;; the duration. Now we have pseudoatomic there's no need for that
179 (defun allocation-dynamic-extent (alloc-tn size)
180 (inst sub esp-tn size)
181 ;; FIXME: SIZE _should_ be double-word aligned (suggested but
182 ;; unfortunately not enforced by PAD-DATA-BLOCK and
183 ;; WITH-FIXED-ALLOCATION), so that ESP is always divisible by 8 (for
184 ;; 32-bit lispobjs). In that case, this AND instruction is
185 ;; unneccessary and could be removed. If not, explain why. -- CSR,
187 (inst and esp-tn (lognot lowtag-mask))
188 (aver (not (location= alloc-tn esp-tn)))
189 (inst mov alloc-tn esp-tn)
192 (defun allocation-notinline (alloc-tn size)
193 (let* ((alloc-tn-offset (tn-offset alloc-tn))
194 ;; C call to allocate via dispatch routines. Each
195 ;; destination has a special entry point. The size may be a
196 ;; register or a constant.
197 (tn-text (ecase alloc-tn-offset
203 (#.edi-offset "edi")))
204 (size-text (case size (8 "8_") (16 "16_") (t ""))))
205 (unless (or (eql size 8) (eql size 16))
206 (unless (and (tn-p size) (location= alloc-tn size))
207 (inst mov alloc-tn size)))
208 (inst call (make-fixup (concatenate 'string
213 (defun allocation-inline (alloc-tn size)
214 (let ((ok (gen-label))
217 (make-ea :dword :disp
218 #!+sb-thread (* n-word-bytes thread-alloc-region-slot)
219 #!-sb-thread (make-fixup "boxed_region" :foreign)
220 :scale 1)) ; thread->alloc_region.free_pointer
222 (make-ea :dword :disp
223 #!+sb-thread (* n-word-bytes (1+ thread-alloc-region-slot))
224 #!-sb-thread (make-fixup "boxed_region" :foreign 4)
225 :scale 1))) ; thread->alloc_region.end_addr
226 (unless (and (tn-p size) (location= alloc-tn size))
227 (inst mov alloc-tn size))
228 #!+sb-thread (inst fs-segment-prefix)
229 (inst add alloc-tn free-pointer)
230 #!+sb-thread (inst fs-segment-prefix)
231 (inst cmp alloc-tn end-addr)
233 (let ((dst (ecase (tn-offset alloc-tn)
234 (#.eax-offset "alloc_overflow_eax")
235 (#.ecx-offset "alloc_overflow_ecx")
236 (#.edx-offset "alloc_overflow_edx")
237 (#.ebx-offset "alloc_overflow_ebx")
238 (#.esi-offset "alloc_overflow_esi")
239 (#.edi-offset "alloc_overflow_edi"))))
240 (inst call (make-fixup dst :foreign)))
241 (inst jmp-short done)
243 ;; Swap ALLOC-TN and FREE-POINTER
244 (cond ((and (tn-p size) (location= alloc-tn size))
245 ;; XCHG is extremely slow, use the xor swap trick
246 #!+sb-thread (inst fs-segment-prefix)
247 (inst xor alloc-tn free-pointer)
248 #!+sb-thread (inst fs-segment-prefix)
249 (inst xor free-pointer alloc-tn)
250 #!+sb-thread (inst fs-segment-prefix)
251 (inst xor alloc-tn free-pointer))
253 ;; It's easier if SIZE is still available.
254 #!+sb-thread (inst fs-segment-prefix)
255 (inst mov free-pointer alloc-tn)
256 (inst sub alloc-tn size)))
261 ;;; Emit code to allocate an object with a size in bytes given by
262 ;;; SIZE. The size may be an integer or a TN. If Inline is a VOP
263 ;;; node-var then it is used to make an appropriate speed vs size
266 ;;; Allocation should only be used inside a pseudo-atomic section, which
267 ;;; should also cover subsequent initialization of the object.
269 ;;; (FIXME: so why aren't we asserting this?)
271 (defun allocation (alloc-tn size &optional inline dynamic-extent)
273 (dynamic-extent (allocation-dynamic-extent alloc-tn size))
274 ((or (null inline) (policy inline (>= speed space)))
275 (allocation-inline alloc-tn size))
276 (t (allocation-notinline alloc-tn size)))
279 ;;; Allocate an other-pointer object of fixed SIZE with a single word
280 ;;; header having the specified WIDETAG value. The result is placed in
282 (defmacro with-fixed-allocation ((result-tn widetag size &optional inline)
285 (bug "empty &body in WITH-FIXED-ALLOCATION"))
286 (once-only ((result-tn result-tn) (size size))
288 (allocation ,result-tn (pad-data-block ,size) ,inline)
289 (storew (logior (ash (1- ,size) n-widetag-bits) ,widetag)
292 (make-ea :byte :base ,result-tn :disp other-pointer-lowtag))
296 (eval-when (#-sb-xc :compile-toplevel :load-toplevel :execute)
297 (defun emit-error-break (vop kind code values)
298 (let ((vector (gensym)))
300 #-darwin (inst int 3) ; i386 breakpoint instruction
302 ;; On Darwin, we need to use #x0b0f instead of int3 in order
303 ;; to generate a SIGILL instead of a SIGTRAP as darwin/x86
304 ;; doesn't seem to be reliably firing SIGTRAP
305 ;; handlers. Hopefully this will be fixed by Apple at a
307 #+darwin (inst word #x0b0f))
308 ;; The return PC points here; note the location for the debugger.
311 (note-this-location vop :internal-error)))
312 (inst byte ,kind) ; eg trap_Xyyy
313 (with-adjustable-vector (,vector) ; interr arguments
314 (write-var-integer (error-number-or-lose ',code) ,vector)
315 ,@(mapcar (lambda (tn)
317 ;; classic CMU CL comment:
318 ;; zzzzz jrd here. tn-offset is zero for constant
320 (write-var-integer (make-sc-offset (sc-number
326 (inst byte (length ,vector))
327 (dotimes (i (length ,vector))
328 (inst byte (aref ,vector i))))))))
330 (defmacro error-call (vop error-code &rest values)
332 "Cause an error. ERROR-CODE is the error to cause."
334 (emit-error-break vop error-trap error-code values)))
336 (defmacro generate-error-code (vop error-code &rest values)
338 "Generate-Error-Code Error-code Value*
339 Emit code for an error with the specified Error-Code and context Values."
340 `(assemble (*elsewhere*)
341 (let ((start-lab (gen-label)))
342 (emit-label start-lab)
343 (error-call ,vop ,error-code ,@values)
349 ;;; This is used to wrap operations which leave untagged memory lying
350 ;;; around. It's an operation which the AOP weenies would describe as
351 ;;; having "cross-cutting concerns", meaning it appears all over the
352 ;;; place and there's no logical single place to attach documentation.
353 ;;; grep (mostly in src/runtime) is your friend
355 ;;; KLUDGE: since the stack on the x86 is treated conservatively, it
356 ;;; does not matter whether a signal occurs during construction of a
357 ;;; dynamic-extent object, as the half-finished construction of the
358 ;;; object will not cause any difficulty. We can therefore elide
359 (defmacro maybe-pseudo-atomic (really-p &body forms)
362 (pseudo-atomic ,@forms)))
365 (defmacro pseudo-atomic (&rest forms)
366 (with-unique-names (label)
367 `(let ((,label (gen-label)))
368 (inst fs-segment-prefix)
369 (inst or (make-ea :byte :disp (* 4 thread-pseudo-atomic-bits-slot))
372 (inst fs-segment-prefix)
373 (inst xor (make-ea :byte :disp (* 4 thread-pseudo-atomic-bits-slot))
376 ;; if PAI was set, interrupts were disabled at the same
377 ;; time using the process signal mask.
378 (inst break pending-interrupt-trap)
379 (emit-label ,label))))
382 (defmacro pseudo-atomic (&rest forms)
383 (with-unique-names (label)
384 `(let ((,label (gen-label)))
385 (inst or (make-ea-for-symbol-value *pseudo-atomic-bits* :byte)
388 (inst xor (make-ea-for-symbol-value *pseudo-atomic-bits* :byte)
391 ;; if PAI was set, interrupts were disabled at the same
392 ;; time using the process signal mask.
393 (inst break pending-interrupt-trap)
394 (emit-label ,label))))
396 ;;;; indexed references
398 (defmacro define-full-compare-and-swap
399 (name type offset lowtag scs el-type &optional translate)
402 ,@(when translate `((:translate ,translate)))
404 (:args (object :scs (descriptor-reg) :to :eval)
405 (index :scs (any-reg immediate unsigned-reg) :to :result)
406 (old-value :scs ,scs :target eax)
407 (new-value :scs ,scs))
408 (:arg-types ,type tagged-num ,el-type ,el-type)
409 (:temporary (:sc descriptor-reg :offset eax-offset
410 :from (:argument 2) :to :result :target value) eax)
411 (:results (value :scs ,scs))
412 (:result-types ,el-type)
417 (let ((ea (sc-case index
419 (make-ea :dword :base object
420 :disp (- (* (+ ,offset (tn-value index))
424 (make-ea :dword :base object :index index :scale 4
425 :disp (- (* ,offset n-word-bytes)
428 (make-ea :dword :base object :index index
429 :disp (- (* ,offset n-word-bytes)
431 (inst cmpxchg ea new-value))
434 (defmacro define-full-reffer (name type offset lowtag scs el-type &optional translate)
438 `((:translate ,translate)))
440 (:args (object :scs (descriptor-reg))
441 (index :scs (any-reg immediate unsigned-reg)))
442 (:arg-types ,type tagged-num)
443 (:results (value :scs ,scs))
444 (:result-types ,el-type)
445 (:generator 3 ; pw was 5
448 (inst mov value (make-ea :dword :base object
449 :disp (- (* (+ ,offset (tn-value index))
453 (inst mov value (make-ea :dword :base object :index index :scale 4
454 :disp (- (* ,offset n-word-bytes)
457 (inst mov value (make-ea :dword :base object :index index
458 :disp (- (* ,offset n-word-bytes)
461 (defmacro define-full-reffer+offset (name type offset lowtag scs el-type &optional translate)
465 `((:translate ,translate)))
467 (:args (object :scs (descriptor-reg))
468 (index :scs (any-reg immediate unsigned-reg)))
469 (:arg-types ,type tagged-num
470 (:constant (constant-displacement ,lowtag sb!vm:n-word-bytes ,offset)))
472 (:results (value :scs ,scs))
473 (:result-types ,el-type)
474 (:generator 3 ; pw was 5
477 (inst mov value (make-ea :dword :base object
478 :disp (- (* (+ ,offset
484 (inst mov value (make-ea :dword :base object :index index :scale 4
485 :disp (- (* (+ ,offset offset)
489 (inst mov value (make-ea :dword :base object :index index
490 :disp (- (* (+ ,offset offset)
494 (defmacro define-full-setter (name type offset lowtag scs el-type &optional translate)
498 `((:translate ,translate)))
500 (:args (object :scs (descriptor-reg))
501 (index :scs (any-reg immediate))
502 (value :scs ,scs :target result))
503 (:arg-types ,type tagged-num ,el-type)
504 (:results (result :scs ,scs))
505 (:result-types ,el-type)
506 (:generator 4 ; was 5
509 (inst mov (make-ea :dword :base object
510 :disp (- (* (+ ,offset (tn-value index))
515 (inst mov (make-ea :dword :base object :index index
516 :disp (- (* ,offset n-word-bytes) ,lowtag))
518 (move result value)))))
520 (defmacro define-full-setter+offset (name type offset lowtag scs el-type &optional translate)
524 `((:translate ,translate)))
526 (:args (object :scs (descriptor-reg))
527 (index :scs (any-reg immediate))
528 (value :scs ,scs :target result))
530 (:arg-types ,type tagged-num
531 (:constant (constant-displacement ,lowtag sb!vm:n-word-bytes ,offset)) ,el-type)
532 (:results (result :scs ,scs))
533 (:result-types ,el-type)
534 (:generator 4 ; was 5
537 (inst mov (make-ea :dword :base object
538 :disp (- (* (+ ,offset (tn-value index) offset)
543 (inst mov (make-ea :dword :base object :index index
544 :disp (- (* (+ ,offset offset)
545 n-word-bytes) ,lowtag))
547 (move result value)))))
549 ;;; helper for alien stuff.
550 (def!macro with-pinned-objects ((&rest objects) &body body)
551 "Arrange with the garbage collector that the pages occupied by
552 OBJECTS will not be moved in memory for the duration of BODY.
553 Useful for e.g. foreign calls where another thread may trigger
555 `(multiple-value-prog1
557 ,@(loop for p in objects
559 ;; There is no race here wrt to gc, because at every
560 ;; point during the execution there is a reference to
561 ;; P on the stack or in a register.
562 `(push-word-on-c-stack
563 (int-sap (sb!kernel:get-lisp-obj-address ,p))))
565 ;; If the body returned normally, we should restore the stack pointer
566 ;; for the benefit of any following code in the same function. If
567 ;; there's a non-local exit in the body, sp is garbage anyway and
568 ;; will get set appropriately from {a, the} frame pointer before it's
570 (pop-words-from-c-stack ,(length objects))))