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)
113 `(with-tls-ea (EA :base ,reg
115 :disp (make-ea-for-symbol-tls-index ,symbol))
116 (inst mov ,reg (make-ea :dword :base ,reg) :maybe-fs)))
118 (defmacro load-tl-symbol-value (reg symbol) `(load-symbol-value ,reg ,symbol))
121 (defmacro store-tl-symbol-value (reg symbol temp)
122 `(with-tls-ea (EA :base ,temp
124 :disp (make-ea-for-symbol-tls-index ,symbol))
125 (inst mov EA ,reg :maybe-fs)))
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)
133 `(with-tls-ea (EA :base ,reg
135 :disp (* 4 thread-binding-stack-pointer-slot))
136 (inst mov ,reg EA :maybe-fs))
138 `(load-symbol-value ,reg *binding-stack-pointer*))
140 (defmacro store-binding-stack-pointer (reg)
147 (with-tls-ea (EA :base eax-tn
149 :disp (* 4 thread-binding-stack-pointer-slot))
153 (with-tls-ea (EA :disp-type :constant
154 :disp (* 4 thread-binding-stack-pointer-slot))
155 (inst mov EA ,reg :maybe-fs)))
157 `(store-symbol-value ,reg *binding-stack-pointer*))
159 (defmacro load-type (target source &optional (offset 0))
161 "Loads the type bits of a pointer into target independent of
162 byte-ordering issues."
163 (once-only ((n-target target)
166 (ecase *backend-byte-order*
169 (make-ea :byte :base ,n-source :disp ,n-offset)))
172 (make-ea :byte :base ,n-source
173 :disp (+ ,n-offset (1- n-word-bytes))))))))
175 ;;;; allocation helpers
177 ;;; Allocation within alloc_region (which is thread local) can be done
178 ;;; inline. If the alloc_region is overflown allocation is done by
179 ;;; calling the C alloc() function.
181 ;;; C calls for allocation don't /seem/ to make an awful lot of
182 ;;; difference to speed. On pure consing it's about a 25%
183 ;;; gain. Guessing from historical context, it looks like inline
184 ;;; allocation was introduced before pseudo-atomic, at which time all
185 ;;; calls to alloc() would have needed a syscall to mask signals for
186 ;;; the duration. Now we have pseudoatomic there's no need for that
189 (defun allocation-dynamic-extent (alloc-tn size lowtag)
190 (inst sub esp-tn size)
191 ;; FIXME: SIZE _should_ be double-word aligned (suggested but
192 ;; unfortunately not enforced by PAD-DATA-BLOCK and
193 ;; WITH-FIXED-ALLOCATION), so that ESP is always divisible by 8 (for
194 ;; 32-bit lispobjs). In that case, this AND instruction is
195 ;; unneccessary and could be removed. If not, explain why. -- CSR,
197 (inst and esp-tn (lognot lowtag-mask))
198 (aver (not (location= alloc-tn esp-tn)))
199 (inst lea alloc-tn (make-ea :byte :base esp-tn :disp lowtag))
202 (defun allocation-notinline (alloc-tn size)
203 (let* ((alloc-tn-offset (tn-offset alloc-tn))
204 ;; C call to allocate via dispatch routines. Each
205 ;; destination has a special entry point. The size may be a
206 ;; register or a constant.
207 (tn-text (ecase alloc-tn-offset
213 (#.edi-offset "edi")))
214 (size-text (case size (8 "8_") (16 "16_") (t ""))))
215 (unless (or (eql size 8) (eql size 16))
216 (unless (and (tn-p size) (location= alloc-tn size))
217 (inst mov alloc-tn size)))
218 (inst call (make-fixup (concatenate 'string
223 (defun allocation-inline (alloc-tn size)
224 (let* ((ok (gen-label)) ;reindent after merging
226 #!+(and sb-thread win32)
227 (scratch-tns (loop for my-tn in `(,eax-tn ,ebx-tn ,edx-tn ,ecx-tn)
228 when (and (not (location= alloc-tn my-tn))
229 (or (not (tn-p size))
230 (not (location= size my-tn))))
232 (tls-prefix #!+sb-thread :fs #!-sb-thread nil)
234 (make-ea :dword :disp
235 #!+sb-thread (* n-word-bytes thread-alloc-region-slot)
236 #!-sb-thread (make-fixup "boxed_region" :foreign)
237 :scale 1)) ; thread->alloc_region.free_pointer
239 (make-ea :dword :disp
240 #!+sb-thread (* n-word-bytes (1+ thread-alloc-region-slot))
241 #!-sb-thread (make-fixup "boxed_region" :foreign 4)
242 :scale 1)) ; thread->alloc_region.end_addr
243 #!+(and sb-thread win32) (scratch-tn (pop scratch-tns))
244 #!+(and sb-thread win32) (swap-tn (pop scratch-tns)))
245 (unless (and (tn-p size) (location= alloc-tn size))
246 (inst mov alloc-tn size))
247 #!+(and sb-thread win32)
249 (inst push scratch-tn)
252 (make-ea :dword :disp
253 +win32-tib-arbitrary-field-offset+) tls-prefix)
254 (setf (ea-base free-pointer) scratch-tn
255 (ea-base end-addr) scratch-tn
257 (inst add alloc-tn free-pointer tls-prefix)
258 (inst cmp alloc-tn end-addr tls-prefix)
260 (let ((dst (ecase (tn-offset alloc-tn)
261 (#.eax-offset "alloc_overflow_eax")
262 (#.ecx-offset "alloc_overflow_ecx")
263 (#.edx-offset "alloc_overflow_edx")
264 (#.ebx-offset "alloc_overflow_ebx")
265 (#.esi-offset "alloc_overflow_esi")
266 (#.edi-offset "alloc_overflow_edi"))))
267 (inst call (make-fixup dst :foreign)))
268 (inst jmp-short done)
270 ;; Swap ALLOC-TN and FREE-POINTER
271 (cond ((and (tn-p size) (location= alloc-tn size))
272 ;; XCHG is extremely slow, use the xor swap trick
273 #!-(and sb-thread win32)
275 (inst xor alloc-tn free-pointer tls-prefix)
276 (inst xor free-pointer alloc-tn tls-prefix)
277 (inst xor alloc-tn free-pointer tls-prefix))
278 #!+(and sb-thread win32)
280 (inst mov swap-tn free-pointer tls-prefix)
281 (inst mov free-pointer alloc-tn tls-prefix)
282 (inst mov alloc-tn swap-tn)))
284 ;; It's easier if SIZE is still available.
285 (inst mov free-pointer alloc-tn tls-prefix)
286 (inst sub alloc-tn size)))
288 #!+(and sb-thread win32)
291 (inst pop scratch-tn))
295 ;;; Emit code to allocate an object with a size in bytes given by
296 ;;; SIZE. The size may be an integer or a TN. If Inline is a VOP
297 ;;; node-var then it is used to make an appropriate speed vs size
300 ;;; Allocation should only be used inside a pseudo-atomic section, which
301 ;;; should also cover subsequent initialization of the object.
303 ;;; (FIXME: so why aren't we asserting this?)
305 (defun allocation (alloc-tn size &optional inline dynamic-extent lowtag)
306 (declare (ignorable inline))
309 (allocation-dynamic-extent alloc-tn size lowtag))
310 ((or (null inline) (policy inline (>= speed space)))
311 (allocation-inline alloc-tn size))
313 (allocation-notinline alloc-tn size)))
314 (when (and lowtag (not dynamic-extent))
315 (inst lea alloc-tn (make-ea :byte :base alloc-tn :disp lowtag)))
318 ;;; Allocate an other-pointer object of fixed SIZE with a single word
319 ;;; header having the specified WIDETAG value. The result is placed in
321 (defmacro with-fixed-allocation ((result-tn widetag size &optional inline stack-allocate-p)
324 (bug "empty &body in WITH-FIXED-ALLOCATION"))
325 (once-only ((result-tn result-tn) (size size) (stack-allocate-p stack-allocate-p))
326 `(maybe-pseudo-atomic ,stack-allocate-p
327 (allocation ,result-tn (pad-data-block ,size) ,inline ,stack-allocate-p
328 other-pointer-lowtag)
329 (storew (logior (ash (1- ,size) n-widetag-bits) ,widetag)
330 ,result-tn 0 other-pointer-lowtag)
334 (defun emit-error-break (vop kind code values)
337 (inst int 3) ; i386 breakpoint instruction
339 ;; On Darwin, we need to use #x0b0f instead of int3 in order
340 ;; to generate a SIGILL instead of a SIGTRAP as darwin/x86
341 ;; doesn't seem to be reliably firing SIGTRAP
342 ;; handlers. Hopefully this will be fixed by Apple at a
346 ;; The return PC points here; note the location for the debugger.
348 (note-this-location vop :internal-error))
349 (inst byte kind) ; e.g. trap_xyyy
350 (with-adjustable-vector (vector) ; interr arguments
351 (write-var-integer code vector)
353 ;; classic CMU CL comment:
354 ;; zzzzz jrd here. tn-offset is zero for constant
356 (write-var-integer (make-sc-offset (sc-number (tn-sc tn))
357 (or (tn-offset tn) 0))
359 (inst byte (length vector))
360 (dotimes (i (length vector))
361 (inst byte (aref vector i))))))
363 (defun error-call (vop error-code &rest values)
365 "Cause an error. ERROR-CODE is the error to cause."
366 (emit-error-break vop error-trap (error-number-or-lose error-code) values))
368 (defun generate-error-code (vop error-code &rest values)
370 "Generate-Error-Code Error-code Value*
371 Emit code for an error with the specified Error-Code and context Values."
372 (assemble (*elsewhere*)
373 (let ((start-lab (gen-label)))
374 (emit-label start-lab)
375 (emit-error-break vop error-trap (error-number-or-lose error-code) values)
381 ;;; This is used to wrap operations which leave untagged memory lying
382 ;;; around. It's an operation which the AOP weenies would describe as
383 ;;; having "cross-cutting concerns", meaning it appears all over the
384 ;;; place and there's no logical single place to attach documentation.
385 ;;; grep (mostly in src/runtime) is your friend
387 ;;; KLUDGE: since the stack on the x86 is treated conservatively, it
388 ;;; does not matter whether a signal occurs during construction of a
389 ;;; dynamic-extent object, as the half-finished construction of the
390 ;;; object will not cause any difficulty. We can therefore elide
391 (defmacro maybe-pseudo-atomic (not-really-p &body forms)
394 (pseudo-atomic ,@forms)))
396 ;;; Unsafely clear pa flags so that the image can properly lose in a
399 (defmacro %clear-pseudo-atomic ()
403 '(inst mov (make-ea :dword :disp (* 4 thread-pseudo-atomic-bits-slot)) 0 :fs))
406 (defun emit-safepoint ()
407 (inst test al-tn (make-ea :byte
408 :disp (make-fixup "gc_safepoint_page" :foreign))))
411 (defmacro pseudo-atomic (&rest forms)
413 `(progn ,@forms (emit-safepoint))
415 (with-unique-names (label)
416 `(let ((,label (gen-label)))
417 (inst mov (make-ea :dword :disp (* 4 thread-pseudo-atomic-bits-slot))
420 (inst xor (make-ea :dword :disp (* 4 thread-pseudo-atomic-bits-slot))
423 ;; if PAI was set, interrupts were disabled at the same time
424 ;; using the process signal mask.
425 (inst break pending-interrupt-trap)
428 ;; In this case, when allocation thinks a GC should be done, it
429 ;; does not mark PA as interrupted, but schedules a safepoint
430 ;; trap instead. Let's take the opportunity to trigger that
431 ;; safepoint right now.
435 (defmacro pseudo-atomic (&rest forms)
436 (with-unique-names (label)
437 `(let ((,label (gen-label)))
438 (inst mov (make-ea-for-symbol-value *pseudo-atomic-bits* :dword)
441 (inst xor (make-ea-for-symbol-value *pseudo-atomic-bits* :dword)
444 ;; if PAI was set, interrupts were disabled at the same time
445 ;; using the process signal mask.
446 (inst break pending-interrupt-trap)
447 (emit-label ,label))))
449 ;;;; indexed references
451 (defmacro define-full-compare-and-swap
452 (name type offset lowtag scs el-type &optional translate)
455 ,@(when translate `((:translate ,translate)))
457 (:args (object :scs (descriptor-reg) :to :eval)
458 (index :scs (any-reg immediate unsigned-reg) :to :result)
459 (old-value :scs ,scs :target eax)
460 (new-value :scs ,scs))
461 (:arg-types ,type tagged-num ,el-type ,el-type)
462 (:temporary (:sc descriptor-reg :offset eax-offset
463 :from (:argument 2) :to :result :target value) eax)
464 (:results (value :scs ,scs))
465 (:result-types ,el-type)
468 (let ((ea (sc-case index
470 (make-ea :dword :base object
471 :disp (- (* (+ ,offset (tn-value index))
475 (make-ea :dword :base object :index index :scale 4
476 :disp (- (* ,offset n-word-bytes)
479 (make-ea :dword :base object :index index
480 :disp (- (* ,offset n-word-bytes)
482 (inst cmpxchg ea new-value :lock))
485 (defmacro define-full-reffer (name type offset lowtag scs el-type &optional translate)
489 `((:translate ,translate)))
491 (:args (object :scs (descriptor-reg))
492 (index :scs (any-reg immediate unsigned-reg)))
493 (:arg-types ,type tagged-num)
494 (:results (value :scs ,scs))
495 (:result-types ,el-type)
496 (:generator 3 ; pw was 5
499 (inst mov value (make-ea :dword :base object
500 :disp (- (* (+ ,offset (tn-value index))
504 (inst mov value (make-ea :dword :base object :index index :scale 4
505 :disp (- (* ,offset n-word-bytes)
508 (inst mov value (make-ea :dword :base object :index index
509 :disp (- (* ,offset n-word-bytes)
512 (defmacro define-full-reffer+offset (name type offset lowtag scs el-type &optional translate)
516 `((:translate ,translate)))
518 (:args (object :scs (descriptor-reg))
519 (index :scs (any-reg immediate unsigned-reg)))
520 (:arg-types ,type tagged-num
521 (:constant (constant-displacement ,lowtag sb!vm:n-word-bytes ,offset)))
523 (:results (value :scs ,scs))
524 (:result-types ,el-type)
525 (:generator 3 ; pw was 5
528 (inst mov value (make-ea :dword :base object
529 :disp (- (* (+ ,offset
535 (inst mov value (make-ea :dword :base object :index index :scale 4
536 :disp (- (* (+ ,offset offset)
540 (inst mov value (make-ea :dword :base object :index index
541 :disp (- (* (+ ,offset offset)
545 (defmacro define-full-setter (name type offset lowtag scs el-type &optional translate)
549 `((:translate ,translate)))
551 (:args (object :scs (descriptor-reg))
552 (index :scs (any-reg immediate))
553 (value :scs ,scs :target result))
554 (:arg-types ,type tagged-num ,el-type)
555 (:results (result :scs ,scs))
556 (:result-types ,el-type)
557 (:generator 4 ; was 5
560 (inst mov (make-ea :dword :base object
561 :disp (- (* (+ ,offset (tn-value index))
566 (inst mov (make-ea :dword :base object :index index
567 :disp (- (* ,offset n-word-bytes) ,lowtag))
569 (move result value)))))
571 (defmacro define-full-setter+offset (name type offset lowtag scs el-type &optional translate)
575 `((:translate ,translate)))
577 (:args (object :scs (descriptor-reg))
578 (index :scs (any-reg immediate))
579 (value :scs ,scs :target result))
581 (:arg-types ,type tagged-num
582 (:constant (constant-displacement ,lowtag sb!vm:n-word-bytes ,offset)) ,el-type)
583 (:results (result :scs ,scs))
584 (:result-types ,el-type)
585 (:generator 4 ; was 5
588 (inst mov (make-ea :dword :base object
589 :disp (- (* (+ ,offset (tn-value index) offset)
594 (inst mov (make-ea :dword :base object :index index
595 :disp (- (* (+ ,offset offset)
596 n-word-bytes) ,lowtag))
598 (move result value)))))
600 ;;; helper for alien stuff.
602 (def!macro with-pinned-objects ((&rest objects) &body body)
603 "Arrange with the garbage collector that the pages occupied by
604 OBJECTS will not be moved in memory for the duration of BODY.
605 Useful for e.g. foreign calls where another thread may trigger
608 (let ((pins (make-gensym-list (length objects)))
609 (wpo (gensym "WITH-PINNED-OBJECTS-THUNK")))
610 ;; BODY is stuffed in a function to preserve the lexical
612 `(flet ((,wpo () (progn ,@body)))
613 (declare (muffle-conditions compiler-note))
614 ;; PINS are dx-allocated in case the compiler for some
615 ;; unfathomable reason decides to allocate value-cells
616 ;; for them -- since we have DX value-cells on x86oid
617 ;; platforms this still forces them on the stack.
618 (dx-let ,(mapcar #'list pins objects)
619 (multiple-value-prog1 (,wpo)
620 ;; TOUCH-OBJECT has a VOP with an empty body: compiler
621 ;; thinks we're using the argument and doesn't flush
622 ;; the variable, but we don't have to pay any extra
623 ;; beyond that -- and MULTIPLE-VALUE-PROG1 keeps them
624 ;; live till the body has finished. *whew*
625 ,@(mapcar (lambda (pin)
626 `(touch-object ,pin))
630 ;;; Helper to hide the fact that thread access on Windows needs one more
631 ;;; instruction, needs the FS prefix in that instruction _instead_ of
632 ;;; the actual load/store, and partially hide the resulting need for a
633 ;;; temporary TN when the non-windows might have have dereferenced an EA
634 ;;; without a TN as a base.
636 (defmacro with-tls-ea ((ea-var &key base
638 (disp-type :constant)
641 "Execute BODY with various magic. BODY is expected to emit instructions.
643 In the body, EA-VAR will be an alias for an EA which BODY can use to
644 perform a thread-local load or store.
646 Within the body, :MAYBE-FS will be replaced with :FS or NIL,
647 depending on the target, and needs to be included in any instruction
648 performing an access through the EA.
650 DISP-TYPE must be :INDEX, or :CONSTANT, and DISP must be an EA/TN,
651 or an expression returning an integer, respectively.
653 BASE must be a temporary TN, except in the following situation: BASE
654 will be unused when DISP-TYPE is constant, BASE-ALREADY-LIVE-P is
655 true, _and_ we're on POSIX. This is an intentional optimization, and
656 the caller needs to take care to ignore the TN in this case, or can
659 BASE-ALREADY-LIVE-P means that at run-time, the BASE register already
660 holds an offset that we should add to instead of overwriting it.
661 The value of the BASE register is undefined following the macro invocation."
662 (check-type base-already-live-p boolean)
663 (check-type disp-type (member :index :constant))
664 #!-(and win32 sb-thread)
665 (let ((body (subst :fs :maybe-fs body)))
669 ,@(subst (if base-already-live-p
671 `(make-ea :dword :base ,base :disp ,disp)
672 ;; BASE not live and not needed, just use DISP
673 `(make-ea :dword :disp ,disp))
677 ;; need to use BASE in any case; and DISP is an EA
679 (inst ,(if base-already-live-p 'add 'mov) ,base ,disp)
680 ,@(subst `(make-ea :dword :base ,base)
683 #!+(and win32 sb-thread)
684 ;; goes through a temporary register to add the thread address into it
685 (multiple-value-bind (constant-disp ea-disp)
687 (:constant (values disp nil))
688 (:index (values 0 disp)))
691 `((inst ,(if base-already-live-p 'add 'mov) ,base ,ea-disp)))
692 (inst ,(if (or base-already-live-p ea-disp) 'add 'mov)
694 (make-ea :dword :disp +win32-tib-arbitrary-field-offset+)
696 ,@(subst `(make-ea :dword :base ,base :disp ,constant-disp)
698 (subst nil :maybe-fs body)))))