| blob | 4a5090ce9606c646d58564223dd016b09123da24 |
1 /*
2 * very-low-level utilities for runtime support
3 */
5 /*
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.
14 */
15 \f
16 #define LANGUAGE_ASSEMBLY
17 #include "sbcl.h"
18 #include "validate.h"
19 #include "genesis/closure.h"
20 #include "genesis/funcallable-instance.h"
21 #include "genesis/fdefn.h"
22 #include "genesis/static-symbols.h"
23 #include "genesis/symbol.h"
24 #include "genesis/thread.h"
26 /* Minimize conditionalization for different OS naming schemes.
27 *
28 * (As of sbcl-0.8.10, this seems no longer to be much of an issue,
29 * since everyone has converged on ELF. If this generality really
30 * turns out not to matter, perhaps it's just clutter we could get
31 * rid of? -- WHN 2004-04-18)
32 *
33 * (Except Win32, which is unlikely ever to be ELF, sorry. -- AB 2005-12-08)
34 */
35 #if defined __linux__ || defined __FreeBSD__ || defined __NetBSD__ || defined __OpenBSD__ || defined __sun
36 #define GNAME(var) var
37 #else
38 #define GNAME(var) _##var
39 #endif
41 /* Get the right type of alignment. Linux, FreeBSD and NetBSD (but not OpenBSD)
42 * want alignment in bytes.
43 *
44 * (As in the GNAME() definitions above, as of sbcl-0.8.10, this seems
45 * no longer to be much of an issue, since everyone has converged on
46 * the same value. If this generality really turns out not to
47 * matter any more, perhaps it's just clutter we could get
48 * rid of? -- WHN 2004-04-18)
49 */
50 #if defined(__linux__) || defined(__FreeBSD__) || defined(__NetBSD__) || defined(__OpenBSD__) || defined(__sun) || defined(LISP_FEATURE_WIN32)
51 #define align_4byte 4
52 #define align_8byte 8
53 #define align_16byte 16
54 #else
55 #define align_4byte 2
56 #define align_8byte 3
57 #define align_16byte 4
58 #endif
60 /*
61 * The assembler used for win32 doesn't like .type or .size directives,
62 * so we want to conditionally kill them out. So let's wrap them in macros
63 * that are defined to be no-ops on win32. Hopefully this still works on
64 * other platforms.
65 */
66 #if !defined(LISP_FEATURE_WIN32) && !defined(LISP_FEATURE_DARWIN)
67 #define TYPE(name) .type name,@function
68 #define SIZE(name) .size name,.-name
69 #else
70 #define TYPE(name)
71 #define SIZE(name)
72 #endif
74 /*
75 * x86/darwin (as of MacOS X 10.4.5) doesn't reliably file signal
76 * handlers (SIGTRAP or Mach exception handlers) for 0xCC, wo we have
77 * to use ud2 instead. ud2 is an undefined opcode, #x0b0f, or
78 * 0F 0B in low-endian notation, that causes SIGILL to fire. We check
79 * for this instruction in the SIGILL handler and if we see it, we
80 * advance the EIP by two bytes to skip over ud2 instruction and
81 * call sigtrap_handler. */
82 #if defined(LISP_FEATURE_DARWIN)
83 #define END()
84 #define TRAP ud2
85 #else
86 #define END() .end
87 #define TRAP int3
88 #endif
90 .text
91 .globl GNAME(all_threads)
92 \f
93 /*
94 * A call to call_into_c preserves esi, edi, and ebp.
95 * (The C function will preserve ebx, esi, edi, and ebp across its
96 * function call, but we trash ebx ourselves by using it to save the
97 * return Lisp address.)
98 *
99 * Return values are in eax and maybe edx for quads, or st(0) for
100 * floats.
101 *
102 * This should work for Lisp calls C calls Lisp calls C..
103 *
104 * FIXME & OAOOM: This duplicates call-out in src/compiler/x86/c-call.lisp,
105 * so if you tweak this, change that too!
106 */
107 .text
108 .align align_16byte,0x90
109 .globl GNAME(call_into_c)
110 TYPE(GNAME(call_into_c))
111 GNAME(call_into_c):
112 /* Save the return Lisp address in ebx. */
113 popl %ebx
115 /* Setup the NPX for C */
116 fstp %st(0)
117 fstp %st(0)
118 fstp %st(0)
119 fstp %st(0)
120 fstp %st(0)
121 fstp %st(0)
122 fstp %st(0)
123 fstp %st(0)
125 call *%eax # normal callout using Lisp stack
126 movl %eax,%ecx # remember integer return value
128 /* Check for a return FP value. */
129 fxam
130 fnstsw %ax
131 andl $0x4500,%eax
132 cmpl $0x4100,%eax
133 jne Lfp_rtn_value
135 /* The return value is in eax, or eax,edx? */
136 /* Set up the NPX stack for Lisp. */
137 fldz # Ensure no regs are empty.
138 fldz
139 fldz
140 fldz
141 fldz
142 fldz
143 fldz
144 fldz
146 /* Restore the return value. */
147 movl %ecx,%eax # maybe return value
149 /* Return. */
150 jmp *%ebx
152 Lfp_rtn_value:
153 /* The return result is in st(0). */
154 /* Set up the NPX stack for Lisp, placing the result in st(0). */
155 fldz # Ensure no regs are empty.
156 fldz
157 fldz
158 fldz
159 fldz
160 fldz
161 fldz
162 fxch %st(7) # Move the result back to st(0).
164 /* We don't need to restore eax, because the result is in st(0). */
166 /* Return. FIXME: It would be nice to restructure this to use RET. */
167 jmp *%ebx
169 SIZE(GNAME(call_into_c))
171 \f
172 .text
173 .globl GNAME(call_into_lisp_first_time)
174 TYPE(GNAME(call_into_lisp_first_time))
176 /* The *ALIEN-STACK* pointer is set up on the first call_into_lisp when
177 * the stack changes. We don't worry too much about saving registers
178 * here, because we never expect to return from the initial call to lisp
179 * anyway */
181 .align align_16byte,0x90
182 GNAME(call_into_lisp_first_time):
183 pushl %ebp # Save old frame pointer.
184 movl %esp,%ebp # Establish new frame.
185 #ifndef LISP_FEATURE_WIN32
186 movl %esp,ALIEN_STACK + SYMBOL_VALUE_OFFSET
187 movl GNAME(all_threads),%eax
188 /* pthread machinery takes care of this for other threads */
189 movl THREAD_CONTROL_STACK_END_OFFSET(%eax) ,%esp
190 #else
191 /* Win32 -really- doesn't like you switching stacks out from under it. */
192 movl GNAME(all_threads),%eax
193 #endif
194 jmp Lstack
195 \f
196 .text
197 .globl GNAME(call_into_lisp)
198 TYPE(GNAME(call_into_lisp))
200 /* The C conventions require that ebx, esi, edi, and ebp be preserved
201 * across function calls. */
203 .align align_16byte,0x90
204 GNAME(call_into_lisp):
205 pushl %ebp # Save old frame pointer.
206 movl %esp,%ebp # Establish new frame.
207 Lstack:
208 /* Save the NPX state */
209 fwait # Catch any pending NPX exceptions.
210 subl $108,%esp # Make room for the NPX state.
211 fnsave (%esp) # save and reset NPX
213 movl (%esp),%eax # Load NPX control word.
214 andl $0xfffff2ff,%eax # Set rounding mode to nearest.
215 orl $0x00000200,%eax # Set precision to 64 bits. (53-bit mantissa)
216 pushl %eax
217 fldcw (%esp) # Recover modes.
218 popl %eax
220 fldz # Ensure no FP regs are empty.
221 fldz
222 fldz
223 fldz
224 fldz
225 fldz
226 fldz
227 fldz
229 /* Save C regs: ebx esi edi. */
230 pushl %ebx
231 pushl %esi
232 pushl %edi
234 /* Clear descriptor regs. */
235 xorl %eax,%eax # lexenv
236 xorl %ebx,%ebx # available
237 xorl %ecx,%ecx # arg count
238 xorl %edx,%edx # first arg
239 xorl %edi,%edi # second arg
240 xorl %esi,%esi # third arg
242 /* no longer in function call */
243 movl %esp,%ebx # remember current stack
244 pushl %ebx # Save entry stack on (maybe) new stack.
246 /* Establish Lisp args. */
247 movl 8(%ebp),%eax # lexenv?
248 movl 12(%ebp),%ebx # address of arg vec
249 movl 16(%ebp),%ecx # num args
250 shll $2,%ecx # Make num args into fixnum.
251 cmpl $0,%ecx
252 je Ldone
253 movl (%ebx),%edx # arg0
254 cmpl $4,%ecx
255 je Ldone
256 movl 4(%ebx),%edi # arg1
257 cmpl $8,%ecx
258 je Ldone
259 movl 8(%ebx),%esi # arg2
260 Ldone:
261 /* Registers eax, ecx, edx, edi, and esi are now live. */
263 #ifdef LISP_FEATURE_WIN32
264 /* Establish an SEH frame. */
265 #ifdef LISP_FEATURE_SB_THREAD
266 /* FIXME: need to save BSP here. */
267 #error "need to save BSP here, but don't know how yet."
268 #else
269 pushl BINDING_STACK_POINTER + SYMBOL_VALUE_OFFSET
270 #endif
271 pushl $GNAME(exception_handler_wrapper)
272 pushl %fs:0
273 movl %esp, %fs:0
274 #endif
276 /* Alloc new frame. */
277 mov %esp,%ebx # The current sp marks start of new frame.
278 push %ebp # fp in save location S0
279 sub $8,%esp # Ensure 3 slots are allocated, one above.
280 mov %ebx,%ebp # Switch to new frame.
282 call *CLOSURE_FUN_OFFSET(%eax)
284 /* If the function returned multiple values, it will return to
285 this point. Lose them */
286 jnc LsingleValue
287 mov %ebx, %esp
288 LsingleValue:
289 /* A singled value function returns here */
291 #ifdef LISP_FEATURE_WIN32
292 /* Remove our SEH frame. */
293 popl %fs:0
294 add $8, %esp
295 #endif
297 /* Restore the stack, in case there was a stack change. */
298 popl %esp # c-sp
300 /* Restore C regs: ebx esi edi. */
301 popl %edi
302 popl %esi
303 popl %ebx
305 /* Restore the NPX state. */
306 frstor (%esp)
307 addl $108, %esp
309 popl %ebp # c-sp
310 movl %edx,%eax # c-val
311 ret
312 SIZE(GNAME(call_into_lisp))
313 \f
314 /* support for saving and restoring the NPX state from C */
315 .text
316 .globl GNAME(fpu_save)
317 TYPE(GNAME(fpu_save))
318 .align 2,0x90
319 GNAME(fpu_save):
320 movl 4(%esp),%eax
321 fnsave (%eax) # Save the NPX state. (resets NPX)
322 ret
323 SIZE(GNAME(fpu_save))
325 .globl GNAME(fpu_restore)
326 TYPE(GNAME(fpu_restore))
327 .align 2,0x90
328 GNAME(fpu_restore):
329 movl 4(%esp),%eax
330 frstor (%eax) # Restore the NPX state.
331 ret
332 SIZE(GNAME(fpu_restore))
333 \f
334 /*
335 * the undefined-function trampoline
336 */
337 .text
338 .align align_4byte,0x90
339 .globl GNAME(undefined_tramp)
340 TYPE(GNAME(undefined_tramp))
341 .byte 0, 0, 0, SIMPLE_FUN_HEADER_WIDETAG
342 GNAME(undefined_tramp):
343 TRAP
344 .byte trap_Error
345 .byte 2
346 .byte UNDEFINED_FUN_ERROR
347 .byte sc_DescriptorReg # eax in the Descriptor-reg SC
348 ret
349 SIZE(GNAME(undefined_tramp))
351 /*
352 * the closure trampoline
353 */
354 .text
355 .align align_4byte,0x90
356 .globl GNAME(closure_tramp)
357 TYPE(GNAME(closure_tramp))
358 .byte 0, 0, 0, SIMPLE_FUN_HEADER_WIDETAG
359 GNAME(closure_tramp):
360 movl FDEFN_FUN_OFFSET(%eax),%eax
361 /* FIXME: The '*' after "jmp" in the next line is from PVE's
362 * patch posted to the CMU CL mailing list Oct 6, 1999. It looks
363 * reasonable, and it certainly seems as though if CMU CL needs it,
364 * SBCL needs it too, but I haven't actually verified that it's
365 * right. It would be good to find a way to force the flow of
366 * control through here to test it. */
367 jmp *CLOSURE_FUN_OFFSET(%eax)
368 SIZE(GNAME(closure_tramp))
370 .text
371 .align align_4byte,0x90
372 .globl GNAME(funcallable_instance_tramp)
373 TYPE(GNAME(funcallable_instance_tramp))
374 GNAME(funcallable_instance_tramp):
375 movl FUNCALLABLE_INSTANCE_FUNCTION_OFFSET(%eax),%eax
376 /* KLUDGE: on this platform, whatever kind of function is in %rax
377 * now, the first word of it contains the address to jump to. */
378 jmp *CLOSURE_FUN_OFFSET(%eax)
379 SIZE(GNAME(funcallable_instance_tramp))
381 /*
382 * fun-end breakpoint magic
383 */
384 .text
385 .globl GNAME(fun_end_breakpoint_guts)
386 .align align_4byte
387 GNAME(fun_end_breakpoint_guts):
388 /* Multiple Value return */
389 jc multiple_value_return
390 /* Single value return: The eventual return will now use the
391 multiple values return convention but with a return values
392 count of one. */
393 movl %esp,%ebx # Setup ebx - the ofp.
394 subl $4,%esp # Allocate one stack slot for the return value
395 movl $4,%ecx # Setup ecx for one return value.
396 movl $(NIL),%edi # default second value
397 movl $(NIL),%esi # default third value
399 multiple_value_return:
401 .globl GNAME(fun_end_breakpoint_trap)
402 GNAME(fun_end_breakpoint_trap):
403 TRAP
404 .byte trap_FunEndBreakpoint
405 hlt # We should never return here.
407 .globl GNAME(fun_end_breakpoint_end)
408 GNAME(fun_end_breakpoint_end):
410 \f
411 .globl GNAME(do_pending_interrupt)
412 TYPE(GNAME(do_pending_interrupt))
413 .align align_4byte,0x90
414 GNAME(do_pending_interrupt):
415 TRAP
416 .byte trap_PendingInterrupt
417 ret
418 SIZE(GNAME(do_pending_interrupt))
419 \f
420 /* Allocate bytes and return the start of the allocated space
421 * in the specified destination register.
422 *
423 * In the general case the size will be in the destination register.
424 *
425 * All registers must be preserved except the destination.
426 * The C conventions will preserve ebx, esi, edi, and ebp.
427 * So only eax, ecx, and edx need special care here.
428 *
429 * ALLOC factors out the logic of calling alloc(): stack alignment, etc.
430 *
431 * DEFINE_ALLOC_TO_FOO defines an alloction routine.
432 */
434 #ifdef LISP_FEATURE_DARWIN
435 #define ALLOC(size) \
436 pushl %ebp; /* Save EBP */ \
437 movl %esp,%ebp; /* Save ESP to EBP */ \
438 pushl $0; /* Reserve space for arg */ \
439 andl $0xfffffff0,%esp; /* Align stack to 16bytes */ \
440 movl size, (%esp); /* Argument to alloc */ \
441 call GNAME(alloc); \
442 movl %ebp,%esp; /* Restore ESP from EBP */ \
443 popl %ebp; /* Restore EBP */
444 #else
445 #define ALLOC(size) \
446 pushl size; /* Argument to alloc */ \
447 call GNAME(alloc); \
448 addl $4,%esp; /* Pop argument */
449 #endif
451 #define DEFINE_ALLOC_TO_EAX(name,size) \
452 .globl GNAME(name); \
453 TYPE(GNAME(name)); \
454 .align align_4byte,0x90; \
455 GNAME(name): \
456 pushl %ecx; /* Save ECX and EDX */ \
457 pushl %edx; \
458 ALLOC(size) \
459 popl %edx; /* Restore ECX and EDX */ \
460 popl %ecx; \
461 ret; \
462 SIZE(GNAME(name))
464 #define DEFINE_ALLOC_TO_ECX(name,size) \
465 .globl GNAME(name); \
466 TYPE(GNAME(name)); \
467 .align align_4byte,0x90; \
468 GNAME(name): \
469 pushl %eax; /* Save EAX and EDX */ \
470 pushl %edx; \
471 ALLOC(size) \
472 movl %eax,%ecx; /* Result to destination */ \
473 popl %edx; \
474 popl %eax; \
475 ret; \
476 SIZE(GNAME(name))
478 #define DEFINE_ALLOC_TO_EDX(name,size) \
479 .globl GNAME(name); \
480 TYPE(GNAME(name)); \
481 .align align_4byte,0x90; \
482 GNAME(name): \
483 pushl %eax; /* Save EAX and ECX */ \
484 pushl %ecx; \
485 ALLOC(size) \
486 movl %eax,%edx; /* Restore EAX and ECX */ \
487 popl %ecx; \
488 popl %eax; \
489 ret; \
490 SIZE(GNAME(name))
492 #define DEFINE_ALLOC_TO_REG(name,reg,size) \
493 .globl GNAME(name); \
494 TYPE(GNAME(name)); \
495 .align align_4byte,0x90; \
496 GNAME(name): \
497 pushl %eax; /* Save EAX, ECX, and EDX */ \
498 pushl %ecx; \
499 pushl %edx; \
500 ALLOC(size) \
501 movl %eax,reg; /* Restore them */ \
502 popl %edx; \
503 popl %ecx; \
504 popl %eax; \
505 ret; \
506 SIZE(GNAME(name))
508 DEFINE_ALLOC_TO_EAX(alloc_to_eax,%eax)
509 DEFINE_ALLOC_TO_EAX(alloc_8_to_eax,$8)
510 DEFINE_ALLOC_TO_EAX(alloc_16_to_eax,$16)
512 DEFINE_ALLOC_TO_ECX(alloc_to_ecx,%ecx)
513 DEFINE_ALLOC_TO_ECX(alloc_8_to_ecx,$8)
514 DEFINE_ALLOC_TO_ECX(alloc_16_to_ecx,$16)
516 DEFINE_ALLOC_TO_EDX(alloc_to_edx,%edx)
517 DEFINE_ALLOC_TO_EDX(alloc_8_to_edx,$8)
518 DEFINE_ALLOC_TO_EDX(alloc_16_to_edx,$16)
520 DEFINE_ALLOC_TO_REG(alloc_to_ebx,%ebx,%ebx)
521 DEFINE_ALLOC_TO_REG(alloc_8_to_ebx,%ebx,$8)
522 DEFINE_ALLOC_TO_REG(alloc_16_to_ebx,%ebx,$16)
524 DEFINE_ALLOC_TO_REG(alloc_to_esi,%esi,%esi)
525 DEFINE_ALLOC_TO_REG(alloc_8_to_esi,%esi,$8)
526 DEFINE_ALLOC_TO_REG(alloc_16_to_esi,%esi,$16)
528 DEFINE_ALLOC_TO_REG(alloc_to_edi,%edi,%edi)
529 DEFINE_ALLOC_TO_REG(alloc_8_to_edi,%edi,$8)
530 DEFINE_ALLOC_TO_REG(alloc_16_to_edi,%edi,$16)
532 /* Called from lisp when an inline allocation overflows.
533 * Every register except the result needs to be preserved.
534 * We depend on C to preserve ebx, esi, edi, and ebp.
535 * But where necessary must save eax, ecx, edx. */
537 #ifdef LISP_FEATURE_SB_THREAD
538 #define START_REGION %fs:THREAD_ALLOC_REGION_OFFSET
539 #else
540 #define START_REGION GNAME(boxed_region)
541 #endif
543 #define ALLOC_OVERFLOW(size) \
544 /* Calculate the size for the allocation. */ \
545 subl START_REGION,size; \
546 ALLOC(size)
548 /* This routine handles an overflow with eax=crfp+size. So the
549 size=eax-crfp. */
550 .align align_4byte
551 .globl GNAME(alloc_overflow_eax)
552 TYPE(GNAME(alloc_overflow_eax))
553 GNAME(alloc_overflow_eax):
554 pushl %ecx # Save ecx
555 pushl %edx # Save edx
556 ALLOC_OVERFLOW(%eax)
557 popl %edx # Restore edx.
558 popl %ecx # Restore ecx.
559 ret
560 SIZE(GNAME(alloc_overflow_eax))
562 .align align_4byte
563 .globl GNAME(alloc_overflow_ecx)
564 TYPE(GNAME(alloc_overflow_ecx))
565 GNAME(alloc_overflow_ecx):
566 pushl %eax # Save eax
567 pushl %edx # Save edx
568 ALLOC_OVERFLOW(%ecx)
569 movl %eax,%ecx # setup the destination.
570 popl %edx # Restore edx.
571 popl %eax # Restore eax.
572 ret
573 SIZE(GNAME(alloc_overflow_ecx))
575 .align align_4byte
576 .globl GNAME(alloc_overflow_edx)
577 TYPE(GNAME(alloc_overflow_edx))
578 GNAME(alloc_overflow_edx):
579 pushl %eax # Save eax
580 pushl %ecx # Save ecx
581 ALLOC_OVERFLOW(%edx)
582 movl %eax,%edx # setup the destination.
583 popl %ecx # Restore ecx.
584 popl %eax # Restore eax.
585 ret
586 SIZE(GNAME(alloc_overflow_edx))
588 /* This routine handles an overflow with ebx=crfp+size. So the
589 size=ebx-crfp. */
590 .align align_4byte
591 .globl GNAME(alloc_overflow_ebx)
592 TYPE(GNAME(alloc_overflow_ebx))
593 GNAME(alloc_overflow_ebx):
594 pushl %eax # Save eax
595 pushl %ecx # Save ecx
596 pushl %edx # Save edx
597 ALLOC_OVERFLOW(%ebx)
598 movl %eax,%ebx # setup the destination.
599 popl %edx # Restore edx.
600 popl %ecx # Restore ecx.
601 popl %eax # Restore eax.
602 ret
603 SIZE(GNAME(alloc_overflow_ebx))
605 /* This routine handles an overflow with esi=crfp+size. So the
606 size=esi-crfp. */
607 .align align_4byte
608 .globl GNAME(alloc_overflow_esi)
609 TYPE(GNAME(alloc_overflow_esi))
610 GNAME(alloc_overflow_esi):
611 pushl %eax # Save eax
612 pushl %ecx # Save ecx
613 pushl %edx # Save edx
614 ALLOC_OVERFLOW(%esi)
615 movl %eax,%esi # setup the destination.
616 popl %edx # Restore edx.
617 popl %ecx # Restore ecx.
618 popl %eax # Restore eax.
619 ret
620 SIZE(GNAME(alloc_overflow_esi))
622 .align align_4byte
623 .globl GNAME(alloc_overflow_edi)
624 TYPE(GNAME(alloc_overflow_edi))
625 GNAME(alloc_overflow_edi):
626 pushl %eax # Save eax
627 pushl %ecx # Save ecx
628 pushl %edx # Save edx
629 ALLOC_OVERFLOW(%edi)
630 movl %eax,%edi # setup the destination.
631 popl %edx # Restore edx.
632 popl %ecx # Restore ecx.
633 popl %eax # Restore eax.
634 ret
635 SIZE(GNAME(alloc_overflow_edi))
638 #ifdef LISP_FEATURE_WIN32
639 /* The guts of the exception-handling system doesn't use
640 * frame pointers, which manages to throw off backtraces
641 * rather badly. So here we grab the (known-good) EBP
642 * and EIP from the exception context and use it to fake
643 * up a stack frame which will skip over the system SEH
644 * code. */
645 .align align_4byte
646 .globl GNAME(exception_handler_wrapper)
647 TYPE(GNAME(exception_handler_wrapper))
648 GNAME(exception_handler_wrapper):
649 /* Context layout is: */
650 /* 7 dwords before FSA. (0x1c) */
651 /* 8 dwords and 0x50 bytes in the FSA. (0x70/0x8c) */
652 /* 4 dwords segregs. (0x10/0x9c) */
653 /* 6 dwords non-stack GPRs. (0x18/0xb4) */
654 /* EBP (at 0xb4) */
655 /* EIP (at 0xb8) */
656 #define CONTEXT_EBP_OFFSET 0xb4
657 #define CONTEXT_EIP_OFFSET 0xb8
658 /* some other stuff we don't care about. */
659 pushl %ebp
660 movl 0x10(%esp), %ebp /* context */
661 pushl CONTEXT_EIP_OFFSET(%ebp)
662 pushl CONTEXT_EBP_OFFSET(%ebp)
663 movl %esp, %ebp
664 pushl 0x1c(%esp)
665 pushl 0x1c(%esp)
666 pushl 0x1c(%esp)
667 pushl 0x1c(%esp)
668 call GNAME(handle_exception)
669 lea 8(%ebp), %esp
670 popl %ebp
671 ret
672 SIZE(GNAME(exception_handler_wrapper))
673 #endif
675 #ifdef LISP_FEATURE_DARWIN
676 .align align_4byte
677 .globl GNAME(call_into_lisp_tramp)
678 TYPE(GNAME(call_into_lisp_tramp))
679 GNAME(call_into_lisp_tramp):
680 /* 1. build the stack frame from the block that's pointed to by ECX
681 2. free the block
682 3. set ECX to 0
683 4. call the function via call_into_lisp
684 */
685 pushl 0(%ecx) /* return address */
687 pushl %ebp
688 movl %esp, %ebp
690 pushl 32(%ecx) /* eflags */
691 pushl 28(%ecx) /* EAX */
692 pushl 20(%ecx) /* ECX */
693 pushl 16(%ecx) /* EDX */
694 pushl 24(%ecx) /* EBX */
695 pushl $0 /* popal is going to ignore esp */
696 pushl %ebp /* is this right?? */
697 pushl 12(%ecx) /* ESI */
698 pushl 8(%ecx) /* EDI */
699 pushl $0 /* args for call_into_lisp */
700 pushl $0
701 pushl 4(%ecx) /* function to call */
703 /* free our save block */
704 pushl %ecx /* reserve sufficient space on stack for args */
705 pushl %ecx
706 andl $0xfffffff0, %esp /* align stack */
707 movl $0x40, 4(%esp)
708 movl %ecx, (%esp)
709 call GNAME(os_invalidate)
711 /* call call_into_lisp */
712 leal -48(%ebp), %esp
713 call GNAME(call_into_lisp)
715 /* Clean up our mess */
716 leal -36(%ebp), %esp
717 popal
718 popfl
719 leave
720 ret
722 SIZE(call_into_lisp_tramp)
723 #endif
725 .align align_4byte,0x90
726 .globl GNAME(post_signal_tramp)
727 TYPE(GNAME(post_signal_tramp))
728 GNAME(post_signal_tramp):
729 /* this is notionally the second half of a function whose first half
730 * doesn't exist. This is where call_into_lisp returns when called
731 * using return_to_lisp_function */
732 addl $12,%esp /* clear call_into_lisp args from stack */
733 popal /* restore registers */
734 popfl
735 #ifdef LISP_FEATURE_DARWIN
736 /* skip two padding words */
737 addl $8,%esp
738 #endif
739 leave
740 ret
741 SIZE(GNAME(post_signal_tramp))
744 /* fast_bzero implementations and code to detect which implementation
745 * to use.
746 */
747 \f
748 .globl GNAME(fast_bzero_pointer)
749 .data
750 .align align_4byte
751 GNAME(fast_bzero_pointer):
752 /* Variable containing a pointer to the bzero function to use.
753 * Initially points to a basic function. Change this variable
754 * to fast_bzero_detect if OS supports SSE. */
755 .long GNAME(fast_bzero_base)
756 \f
757 .text
758 .align align_8byte,0x90
759 .globl GNAME(fast_bzero)
760 TYPE(GNAME(fast_bzero))
761 GNAME(fast_bzero):
762 /* Indirect function call */
763 jmp *GNAME(fast_bzero_pointer)
764 SIZE(GNAME(fast_bzero))
766 \f
767 .text
768 .align align_8byte,0x90
769 .globl GNAME(fast_bzero_detect)
770 TYPE(GNAME(fast_bzero_detect))
771 GNAME(fast_bzero_detect):
772 /* Decide whether to use SSE, MMX or REP version */
773 push %eax /* CPUID uses EAX-EDX */
774 push %ebx
775 push %ecx
776 push %edx
777 mov $1, %eax
778 cpuid
779 test $0x04000000, %edx /* SSE2 needed for MOVNTDQ */
780 jnz Lsse2
781 /* Originally there was another case here for using the
782 * MOVNTQ instruction for processors that supported MMX but
783 * not SSE2. This turned out to be a loss especially on
784 * Athlons (where this instruction is apparently microcoded
785 * somewhat slowly). So for simplicity revert to REP STOSL
786 * for all non-SSE2 processors.
787 */
788 Lbase:
789 movl $(GNAME(fast_bzero_base)), GNAME(fast_bzero_pointer)
790 jmp Lrestore
791 Lsse2:
792 movl $(GNAME(fast_bzero_sse)), GNAME(fast_bzero_pointer)
793 jmp Lrestore
795 Lrestore:
796 pop %edx
797 pop %ecx
798 pop %ebx
799 pop %eax
800 jmp *GNAME(fast_bzero_pointer)
802 SIZE(GNAME(fast_bzero_detect))
804 \f
805 .text
806 .align align_8byte,0x90
807 .globl GNAME(fast_bzero_sse)
808 TYPE(GNAME(fast_bzero_sse))
810 GNAME(fast_bzero_sse):
811 /* A fast routine for zero-filling blocks of memory that are
812 * guaranteed to start and end at a 4096-byte aligned address.
813 */
814 push %esi /* Save temporary registers */
815 push %edi
816 mov 16(%esp), %esi /* Parameter: amount of bytes to fill */
817 mov 12(%esp), %edi /* Parameter: start address */
818 shr $6, %esi /* Amount of 64-byte blocks to copy */
819 jz Lend_sse /* If none, stop */
820 movups %xmm7, -16(%esp) /* Save XMM register */
821 xorps %xmm7, %xmm7 /* Zero the XMM register */
822 jmp Lloop_sse
823 .align align_16byte
824 Lloop_sse:
826 /* Copy the 16 zeroes from xmm7 to memory, 4 times. MOVNTDQ is the
827 * non-caching double-quadword moving variant, i.e. the memory areas
828 * we're touching are not fetched into the L1 cache, since we're just
829 * going to overwrite the memory soon anyway.
830 */
831 movntdq %xmm7, 0(%edi)
832 movntdq %xmm7, 16(%edi)
833 movntdq %xmm7, 32(%edi)
834 movntdq %xmm7, 48(%edi)
836 add $64, %edi /* Advance pointer */
837 dec %esi /* Decrement 64-byte block count */
838 jnz Lloop_sse
839 movups -16(%esp), %xmm7 /* Restore the XMM register */
840 sfence /* Ensure that weakly ordered writes are flushed. */
841 Lend_sse:
842 mov 12(%esp), %esi /* Parameter: start address */
843 prefetcht0 0(%esi) /* Prefetch the start of the block into cache,
844 * since it's likely to be used immediately. */
845 pop %edi /* Restore temp registers */
846 pop %esi
847 ret
848 SIZE(GNAME(fast_bzero_sse))
850 \f
851 .text
852 .align align_8byte,0x90
853 .globl GNAME(fast_bzero_base)
854 TYPE(GNAME(fast_bzero_base))
856 GNAME(fast_bzero_base):
857 /* A fast routine for zero-filling blocks of memory that are
858 * guaranteed to start and end at a 4096-byte aligned address.
859 */
860 push %eax /* Save temporary registers */
861 push %ecx
862 push %edi
863 mov 20(%esp), %ecx /* Parameter: amount of bytes to fill */
864 mov 16(%esp), %edi /* Parameter: start address */
865 xor %eax, %eax /* Zero EAX */
866 shr $2, %ecx /* Amount of 4-byte blocks to copy */
867 jz Lend_base
869 rep
870 stosl /* Store EAX to *EDI, ECX times, incrementing
871 * EDI by 4 after each store */
873 Lend_base:
874 pop %edi /* Restore temp registers */
875 pop %ecx
876 pop %eax
877 ret
878 SIZE(GNAME(fast_bzero_base))
880 \f
881 END()