2 * very-low-level utilities for runtime support
6 * This software is part of the SBCL system. See the README file for
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.
16 #define LANGUAGE_ASSEMBLY
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.
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)
33 * (Except Win32, which is unlikely ever to be ELF, sorry. -- AB 2005-12-08)
35 #if defined __linux__ || defined __FreeBSD__ || defined __NetBSD__ || defined __OpenBSD__ || defined __sun
36 #define GNAME(var) var
38 #define GNAME(var) _##var
41 /* Get the right type of alignment. Linux, FreeBSD and NetBSD (but not OpenBSD)
42 * want alignment in bytes.
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)
50 #if defined(__linux__) || defined(__FreeBSD__) || defined(__NetBSD__) || defined(__OpenBSD__) || defined(__sun) || defined(LISP_FEATURE_WIN32)
53 #define align_16byte 16
57 #define align_16byte 4
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
66 #if !defined(LISP_FEATURE_WIN32) && !defined(LISP_FEATURE_DARWIN)
67 #define TYPE(name) .type name,@function
68 #define SIZE(name) .size name,.-name
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)
91 .globl GNAME(foreign_function_call_active)
92 .globl GNAME(all_threads)
95 * A call to call_into_c preserves esi, edi, and ebp.
96 * (The C function will preserve ebx, esi, edi, and ebp across its
97 * function call, but we trash ebx ourselves by using it to save the
98 * return Lisp address.)
100 * Return values are in eax and maybe edx for quads, or st(0) for
103 * This should work for Lisp calls C calls Lisp calls C..
106 .align align_16byte,0x90
107 .globl GNAME(call_into_c)
108 TYPE(GNAME(call_into_c))
110 movl $1,GNAME(foreign_function_call_active)
112 /* Save the return Lisp address in ebx. */
115 /* Setup the NPX for C */
125 #ifdef LISP_FEATURE_WIN32
129 #ifdef LISP_FEATURE_DARWIN
130 andl $0xfffffff0,%esp # align stack to 16-byte boundary before calling C
132 call *%eax # normal callout using Lisp stack
134 movl %eax,%ecx # remember integer return value
136 /* Check for a return FP value. */
143 /* The return value is in eax, or eax,edx? */
144 /* Set up the NPX stack for Lisp. */
145 fldz # Ensure no regs are empty.
154 /* Restore the return value. */
155 movl %ecx,%eax # maybe return value
157 movl $0,GNAME(foreign_function_call_active)
162 /* The return result is in st(0). */
163 /* Set up the NPX stack for Lisp, placing the result in st(0). */
164 fldz # Ensure no regs are empty.
171 fxch %st(7) # Move the result back to st(0).
173 /* We don't need to restore eax, because the result is in st(0). */
175 movl $0,GNAME(foreign_function_call_active)
179 SIZE(GNAME(call_into_c))
183 .globl GNAME(call_into_lisp_first_time)
184 TYPE(GNAME(call_into_lisp_first_time))
186 /* The *ALIEN-STACK* pointer is set up on the first call_into_lisp when
187 * the stack changes. We don't worry too much about saving registers
188 * here, because we never expect to return from the initial call to lisp
191 .align align_16byte,0x90
192 GNAME(call_into_lisp_first_time):
193 pushl %ebp # Save old frame pointer.
194 movl %esp,%ebp # Establish new frame.
195 #ifndef LISP_FEATURE_WIN32
196 movl %esp,ALIEN_STACK + SYMBOL_VALUE_OFFSET
197 movl GNAME(all_threads),%eax
198 movl THREAD_CONTROL_STACK_START_OFFSET(%eax) ,%esp
199 /* don't think too hard about what happens if we get interrupted
201 addl $(THREAD_CONTROL_STACK_SIZE),%esp
203 /* Win32 -really- doesn't like you switching stacks out from under it. */
204 movl GNAME(all_threads),%eax
209 .globl GNAME(call_into_lisp)
210 TYPE(GNAME(call_into_lisp))
212 /* The C conventions require that ebx, esi, edi, and ebp be preserved
213 * across function calls. */
215 .align align_16byte,0x90
216 GNAME(call_into_lisp):
217 pushl %ebp # Save old frame pointer.
218 movl %esp,%ebp # Establish new frame.
220 /* Save the NPX state */
221 fwait # Catch any pending NPX exceptions.
222 subl $108,%esp # Make room for the NPX state.
223 fnsave (%esp) # save and reset NPX
225 movl (%esp),%eax # Load NPX control word.
226 andl $0xfffff2ff,%eax # Set rounding mode to nearest.
227 orl $0x00000200,%eax # Set precision to 64 bits. (53-bit mantissa)
229 fldcw (%esp) # Recover modes.
232 fldz # Ensure no FP regs are empty.
241 /* Save C regs: ebx esi edi. */
246 /* Clear descriptor regs. */
247 xorl %eax,%eax # lexenv
248 xorl %ebx,%ebx # available
249 xorl %ecx,%ecx # arg count
250 xorl %edx,%edx # first arg
251 xorl %edi,%edi # second arg
252 xorl %esi,%esi # third arg
254 /* no longer in function call */
255 movl %eax, GNAME(foreign_function_call_active)
257 movl %esp,%ebx # remember current stack
258 pushl %ebx # Save entry stack on (maybe) new stack.
260 /* Establish Lisp args. */
261 movl 8(%ebp),%eax # lexenv?
262 movl 12(%ebp),%ebx # address of arg vec
263 movl 16(%ebp),%ecx # num args
264 shll $2,%ecx # Make num args into fixnum.
267 movl (%ebx),%edx # arg0
270 movl 4(%ebx),%edi # arg1
273 movl 8(%ebx),%esi # arg2
275 /* Registers eax, ecx, edx, edi, and esi are now live. */
277 #ifdef LISP_FEATURE_WIN32
278 /* Establish an SEH frame. */
279 #ifdef LISP_FEATURE_SB_THREAD
280 /* FIXME: need to save BSP here. */
281 #error "need to save BSP here, but don't know how yet."
283 pushl BINDING_STACK_POINTER + SYMBOL_VALUE_OFFSET
285 pushl $GNAME(exception_handler_wrapper)
290 /* Alloc new frame. */
291 mov %esp,%ebx # The current sp marks start of new frame.
292 push %ebp # fp in save location S0
293 sub $8,%esp # Ensure 3 slots are allocated, one above.
294 mov %ebx,%ebp # Switch to new frame.
296 call *CLOSURE_FUN_OFFSET(%eax)
298 /* If the function returned multiple values, it will return to
299 this point. Lose them */
303 /* A singled value function returns here */
305 #ifdef LISP_FEATURE_WIN32
306 /* Remove our SEH frame. */
311 /* Restore the stack, in case there was a stack change. */
314 /* Restore C regs: ebx esi edi. */
319 /* Restore the NPX state. */
324 movl %edx,%eax # c-val
326 SIZE(GNAME(call_into_lisp))
328 /* support for saving and restoring the NPX state from C */
330 .globl GNAME(fpu_save)
331 TYPE(GNAME(fpu_save))
335 fnsave (%eax) # Save the NPX state. (resets NPX)
337 SIZE(GNAME(fpu_save))
339 .globl GNAME(fpu_restore)
340 TYPE(GNAME(fpu_restore))
344 frstor (%eax) # Restore the NPX state.
346 SIZE(GNAME(fpu_restore))
349 * the undefined-function trampoline
352 .align align_4byte,0x90
353 .globl GNAME(undefined_tramp)
354 TYPE(GNAME(undefined_tramp))
355 .byte 0, 0, 0, SIMPLE_FUN_HEADER_WIDETAG
356 GNAME(undefined_tramp):
360 .byte UNDEFINED_FUN_ERROR
361 .byte sc_DescriptorReg # eax in the Descriptor-reg SC
363 SIZE(GNAME(undefined_tramp))
366 * the closure trampoline
369 .align align_4byte,0x90
370 .globl GNAME(closure_tramp)
371 TYPE(GNAME(closure_tramp))
372 .byte 0, 0, 0, SIMPLE_FUN_HEADER_WIDETAG
373 GNAME(closure_tramp):
374 movl FDEFN_FUN_OFFSET(%eax),%eax
375 /* FIXME: The '*' after "jmp" in the next line is from PVE's
376 * patch posted to the CMU CL mailing list Oct 6, 1999. It looks
377 * reasonable, and it certainly seems as though if CMU CL needs it,
378 * SBCL needs it too, but I haven't actually verified that it's
379 * right. It would be good to find a way to force the flow of
380 * control through here to test it. */
381 jmp *CLOSURE_FUN_OFFSET(%eax)
382 SIZE(GNAME(closure_tramp))
385 .align align_4byte,0x90
386 .globl GNAME(funcallable_instance_tramp)
387 TYPE(GNAME(funcallable_instance_tramp))
388 GNAME(funcallable_instance_tramp):
389 movl FUNCALLABLE_INSTANCE_FUNCTION_OFFSET(%eax),%eax
390 /* KLUDGE: on this platform, whatever kind of function is in %rax
391 * now, the first word of it contains the address to jump to. */
392 jmp *CLOSURE_FUN_OFFSET(%eax)
393 SIZE(GNAME(funcallable_instance_tramp))
396 * fun-end breakpoint magic
399 .globl GNAME(fun_end_breakpoint_guts)
401 GNAME(fun_end_breakpoint_guts):
402 /* Multiple Value return */
403 jc multiple_value_return
404 /* Single value return: The eventual return will now use the
405 multiple values return convention but with a return values
407 movl %esp,%ebx # Setup ebx - the ofp.
408 subl $4,%esp # Allocate one stack slot for the return value
409 movl $4,%ecx # Setup ecx for one return value.
410 movl $(NIL),%edi # default second value
411 movl $(NIL),%esi # default third value
413 multiple_value_return:
415 .globl GNAME(fun_end_breakpoint_trap)
416 GNAME(fun_end_breakpoint_trap):
418 .byte trap_FunEndBreakpoint
419 hlt # We should never return here.
421 .globl GNAME(fun_end_breakpoint_end)
422 GNAME(fun_end_breakpoint_end):
425 .globl GNAME(do_pending_interrupt)
426 TYPE(GNAME(do_pending_interrupt))
427 .align align_4byte,0x90
428 GNAME(do_pending_interrupt):
430 .byte trap_PendingInterrupt
432 SIZE(GNAME(do_pending_interrupt))
436 * Allocate bytes and return the start of the allocated space
437 * in the specified destination register.
439 * In the general case the size will be in the destination register.
441 * All registers must be preserved except the destination.
442 * The C conventions will preserve ebx, esi, edi, and ebp.
443 * So only eax, ecx, and edx need special care here.
446 .globl GNAME(alloc_to_eax)
447 TYPE(GNAME(alloc_to_eax))
448 .align align_4byte,0x90
450 pushl %ecx # Save ecx and edx as C could destroy them.
452 pushl %eax # Push the size.
454 addl $4,%esp # Pop the size arg.
455 popl %edx # Restore ecx and edx.
458 SIZE(GNAME(alloc_to_eax))
460 .globl GNAME(alloc_8_to_eax)
461 TYPE(GNAME(alloc_8_to_eax))
462 .align align_4byte,0x90
463 GNAME(alloc_8_to_eax):
464 pushl %ecx # Save ecx and edx as C could destroy them.
466 pushl $8 # Push the size.
468 addl $4,%esp # Pop the size arg.
469 popl %edx # Restore ecx and edx.
472 SIZE(GNAME(alloc_8_to_eax))
474 .globl GNAME(alloc_8_to_eax)
475 TYPE(GNAME(alloc_8_to_eax))
476 .align align_4byte,0x90
478 .globl GNAME(alloc_16_to_eax)
479 TYPE(GNAME(alloc_16_to_eax))
480 .align align_4byte,0x90
481 GNAME(alloc_16_to_eax):
482 pushl %ecx # Save ecx and edx as C could destroy them.
484 pushl $16 # Push the size.
486 addl $4,%esp # Pop the size arg.
487 popl %edx # Restore ecx and edx.
490 SIZE(GNAME(alloc_16_to_eax))
492 .globl GNAME(alloc_to_ecx)
493 TYPE(GNAME(alloc_to_ecx))
494 .align align_4byte,0x90
496 pushl %eax # Save eax and edx as C could destroy them.
498 pushl %ecx # Push the size.
500 addl $4,%esp # Pop the size arg.
501 movl %eax,%ecx # Set up the destination.
502 popl %edx # Restore eax and edx.
505 SIZE(GNAME(alloc_to_ecx))
507 .globl GNAME(alloc_8_to_ecx)
508 TYPE(GNAME(alloc_8_to_ecx))
509 .align align_4byte,0x90
510 GNAME(alloc_8_to_ecx):
511 pushl %eax # Save eax and edx as C could destroy them.
513 pushl $8 # Push the size.
515 addl $4,%esp # Pop the size arg.
516 movl %eax,%ecx # Set up the destination.
517 popl %edx # Restore eax and edx.
520 SIZE(GNAME(alloc_8_to_ecx))
522 .globl GNAME(alloc_16_to_ecx)
523 TYPE(GNAME(alloc_16_to_ecx))
524 .align align_4byte,0x90
525 GNAME(alloc_16_to_ecx):
526 pushl %eax # Save eax and edx as C could destroy them.
528 pushl $16 # Push the size.
530 addl $4,%esp # Pop the size arg.
531 movl %eax,%ecx # Set up the destination.
532 popl %edx # Restore eax and edx.
535 SIZE(GNAME(alloc_16_to_ecx))
538 .globl GNAME(alloc_to_edx)
539 TYPE(GNAME(alloc_to_edx))
540 .align align_4byte,0x90
542 pushl %eax # Save eax and ecx as C could destroy them.
544 pushl %edx # Push the size.
546 addl $4,%esp # Pop the size arg.
547 movl %eax,%edx # Set up the destination.
548 popl %ecx # Restore eax and ecx.
551 SIZE(GNAME(alloc_to_edx))
553 .globl GNAME(alloc_8_to_edx)
554 TYPE(GNAME(alloc_8_to_edx))
555 .align align_4byte,0x90
556 GNAME(alloc_8_to_edx):
557 pushl %eax # Save eax and ecx as C could destroy them.
559 pushl $8 # Push the size.
561 addl $4,%esp # Pop the size arg.
562 movl %eax,%edx # Set up the destination.
563 popl %ecx # Restore eax and ecx.
566 SIZE(GNAME(alloc_8_to_edx))
568 .globl GNAME(alloc_16_to_edx)
569 TYPE(GNAME(alloc_16_to_edx))
570 .align align_4byte,0x90
571 GNAME(alloc_16_to_edx):
572 pushl %eax # Save eax and ecx as C could destroy them.
574 pushl $16 # Push the size.
576 addl $4,%esp # Pop the size arg.
577 movl %eax,%edx # Set up the destination.
578 popl %ecx # Restore eax and ecx.
581 SIZE(GNAME(alloc_16_to_edx))
585 .globl GNAME(alloc_to_ebx)
586 TYPE(GNAME(alloc_to_ebx))
587 .align align_4byte,0x90
589 pushl %eax # Save eax, ecx, and edx as C could destroy them.
592 pushl %ebx # Push the size.
594 addl $4,%esp # Pop the size arg.
595 movl %eax,%ebx # Set up the destination.
596 popl %edx # Restore eax, ecx and edx.
600 SIZE(GNAME(alloc_to_ebx))
602 .globl GNAME(alloc_8_to_ebx)
603 TYPE(GNAME(alloc_8_to_ebx))
604 .align align_4byte,0x90
605 GNAME(alloc_8_to_ebx):
606 pushl %eax # Save eax, ecx, and edx as C could destroy them.
609 pushl $8 # Push the size.
611 addl $4,%esp # Pop the size arg.
612 movl %eax,%ebx # Set up the destination.
613 popl %edx # Restore eax, ecx and edx.
617 SIZE(GNAME(alloc_8_to_ebx))
619 .globl GNAME(alloc_16_to_ebx)
620 TYPE(GNAME(alloc_16_to_ebx))
621 .align align_4byte,0x90
622 GNAME(alloc_16_to_ebx):
623 pushl %eax # Save eax, ecx, and edx as C could destroy them.
626 pushl $16 # Push the size
628 addl $4,%esp # pop the size arg.
629 movl %eax,%ebx # setup the destination.
630 popl %edx # Restore eax, ecx and edx.
634 SIZE(GNAME(alloc_16_to_ebx))
638 .globl GNAME(alloc_to_esi)
639 TYPE(GNAME(alloc_to_esi))
640 .align align_4byte,0x90
642 pushl %eax # Save eax, ecx, and edx as C could destroy them.
645 pushl %esi # Push the size
647 addl $4,%esp # pop the size arg.
648 movl %eax,%esi # setup the destination.
649 popl %edx # Restore eax, ecx and edx.
653 SIZE(GNAME(alloc_to_esi))
655 .globl GNAME(alloc_8_to_esi)
656 TYPE(GNAME(alloc_8_to_esi))
657 .align align_4byte,0x90
658 GNAME(alloc_8_to_esi):
659 pushl %eax # Save eax, ecx, and edx as C could destroy them.
662 pushl $8 # Push the size
664 addl $4,%esp # pop the size arg.
665 movl %eax,%esi # setup the destination.
666 popl %edx # Restore eax, ecx and edx.
670 SIZE(GNAME(alloc_8_to_esi))
672 .globl GNAME(alloc_16_to_esi)
673 TYPE(GNAME(alloc_16_to_esi))
674 .align align_4byte,0x90
675 GNAME(alloc_16_to_esi):
676 pushl %eax # Save eax, ecx, and edx as C could destroy them.
679 pushl $16 # Push the size
681 addl $4,%esp # pop the size arg.
682 movl %eax,%esi # setup the destination.
683 popl %edx # Restore eax, ecx and edx.
687 SIZE(GNAME(alloc_16_to_esi))
690 .globl GNAME(alloc_to_edi)
691 TYPE(GNAME(alloc_to_edi))
692 .align align_4byte,0x90
694 pushl %eax # Save eax, ecx, and edx as C could destroy them.
697 pushl %edi # Push the size
699 addl $4,%esp # pop the size arg.
700 movl %eax,%edi # setup the destination.
701 popl %edx # Restore eax, ecx and edx.
705 SIZE(GNAME(alloc_to_edi))
707 .globl GNAME(alloc_8_to_edi)
708 TYPE(GNAME(alloc_8_to_edi))
709 .align align_4byte,0x90
710 GNAME(alloc_8_to_edi):
711 pushl %eax # Save eax, ecx, and edx as C could destroy them.
714 pushl $8 # Push the size
716 addl $4,%esp # pop the size arg.
717 movl %eax,%edi # setup the destination.
718 popl %edx # Restore eax, ecx and edx.
722 SIZE(GNAME(alloc_8_to_edi))
724 .globl GNAME(alloc_16_to_edi)
725 TYPE(GNAME(alloc_16_to_edi))
726 .align align_4byte,0x90
727 GNAME(alloc_16_to_edi):
728 pushl %eax # Save eax, ecx, and edx as C could destroy them.
731 pushl $16 # Push the size
733 addl $4,%esp # pop the size arg.
734 movl %eax,%edi # setup the destination.
735 popl %edx # Restore eax, ecx and edx.
739 SIZE(GNAME(alloc_16_to_edi))
742 /* Called from lisp when an inline allocation overflows.
743 Every register except the result needs to be preserved.
744 We depend on C to preserve ebx, esi, edi, and ebp.
745 But where necessary must save eax, ecx, edx. */
747 #ifdef LISP_FEATURE_SB_THREAD
748 #define START_REGION %fs:THREAD_ALLOC_REGION_OFFSET
750 #define START_REGION GNAME(boxed_region)
753 /* This routine handles an overflow with eax=crfp+size. So the
756 .globl GNAME(alloc_overflow_eax)
757 TYPE(GNAME(alloc_overflow_eax))
758 GNAME(alloc_overflow_eax):
759 pushl %ecx # Save ecx
760 pushl %edx # Save edx
761 /* Calculate the size for the allocation. */
762 subl START_REGION,%eax
763 pushl %eax # Push the size
765 addl $4,%esp # pop the size arg.
766 popl %edx # Restore edx.
767 popl %ecx # Restore ecx.
769 SIZE(GNAME(alloc_overflow_eax))
772 .globl GNAME(alloc_overflow_ecx)
773 TYPE(GNAME(alloc_overflow_ecx))
774 GNAME(alloc_overflow_ecx):
775 pushl %eax # Save eax
776 pushl %edx # Save edx
777 /* Calculate the size for the allocation. */
778 subl START_REGION,%ecx
779 pushl %ecx # Push the size
781 addl $4,%esp # pop the size arg.
782 movl %eax,%ecx # setup the destination.
783 popl %edx # Restore edx.
784 popl %eax # Restore eax.
786 SIZE(GNAME(alloc_overflow_ecx))
789 .globl GNAME(alloc_overflow_edx)
790 TYPE(GNAME(alloc_overflow_edx))
791 GNAME(alloc_overflow_edx):
792 pushl %eax # Save eax
793 pushl %ecx # Save ecx
794 /* Calculate the size for the allocation. */
795 subl START_REGION,%edx
796 pushl %edx # Push the size
798 addl $4,%esp # pop the size arg.
799 movl %eax,%edx # setup the destination.
800 popl %ecx # Restore ecx.
801 popl %eax # Restore eax.
803 SIZE(GNAME(alloc_overflow_edx))
805 /* This routine handles an overflow with ebx=crfp+size. So the
808 .globl GNAME(alloc_overflow_ebx)
809 TYPE(GNAME(alloc_overflow_ebx))
810 GNAME(alloc_overflow_ebx):
811 pushl %eax # Save eax
812 pushl %ecx # Save ecx
813 pushl %edx # Save edx
814 /* Calculate the size for the allocation. */
815 subl START_REGION,%ebx
816 pushl %ebx # Push the size
818 addl $4,%esp # pop the size arg.
819 movl %eax,%ebx # setup the destination.
820 popl %edx # Restore edx.
821 popl %ecx # Restore ecx.
822 popl %eax # Restore eax.
824 SIZE(GNAME(alloc_overflow_ebx))
826 /* This routine handles an overflow with esi=crfp+size. So the
829 .globl GNAME(alloc_overflow_esi)
830 TYPE(GNAME(alloc_overflow_esi))
831 GNAME(alloc_overflow_esi):
832 pushl %eax # Save eax
833 pushl %ecx # Save ecx
834 pushl %edx # Save edx
835 /* Calculate the size for the allocation. */
836 subl START_REGION,%esi
837 pushl %esi # Push the size
839 addl $4,%esp # pop the size arg.
840 movl %eax,%esi # setup the destination.
841 popl %edx # Restore edx.
842 popl %ecx # Restore ecx.
843 popl %eax # Restore eax.
845 SIZE(GNAME(alloc_overflow_esi))
848 .globl GNAME(alloc_overflow_edi)
849 TYPE(GNAME(alloc_overflow_edi))
850 GNAME(alloc_overflow_edi):
851 pushl %eax # Save eax
852 pushl %ecx # Save ecx
853 pushl %edx # Save edx
854 /* Calculate the size for the allocation. */
855 subl START_REGION,%edi
856 pushl %edi # Push the size
858 addl $4,%esp # pop the size arg.
859 movl %eax,%edi # setup the destination.
860 popl %edx # Restore edx.
861 popl %ecx # Restore ecx.
862 popl %eax # Restore eax.
864 SIZE(GNAME(alloc_overflow_edi))
867 #ifdef LISP_FEATURE_WIN32
868 /* The guts of the exception-handling system doesn't use
869 * frame pointers, which manages to throw off backtraces
870 * rather badly. So here we grab the (known-good) EBP
871 * and EIP from the exception context and use it to fake
872 * up a stack frame which will skip over the system SEH
875 .globl GNAME(exception_handler_wrapper)
876 TYPE(GNAME(exception_handler_wrapper))
877 GNAME(exception_handler_wrapper):
878 /* Context layout is: */
879 /* 7 dwords before FSA. (0x1c) */
880 /* 8 dwords and 0x50 bytes in the FSA. (0x70/0x8c) */
881 /* 4 dwords segregs. (0x10/0x9c) */
882 /* 6 dwords non-stack GPRs. (0x18/0xb4) */
885 #define CONTEXT_EBP_OFFSET 0xb4
886 #define CONTEXT_EIP_OFFSET 0xb8
887 /* some other stuff we don't care about. */
889 movl 0x10(%esp), %ebp /* context */
890 pushl CONTEXT_EIP_OFFSET(%ebp)
891 pushl CONTEXT_EBP_OFFSET(%ebp)
897 call GNAME(handle_exception)
901 SIZE(GNAME(exception_handler_wrapper))
904 #ifdef LISP_FEATURE_DARWIN
906 .globl GNAME(call_into_lisp_tramp)
907 TYPE(GNAME(call_into_lisp_tramp))
908 GNAME(call_into_lisp_tramp):
909 /* 1. build the stack frame from the block that's pointed to by ECX
912 4. call the function via call_into_lisp
914 pushl 0(%ecx) /* return address */
919 pushl 32(%ecx) /* eflags */
920 pushl 28(%ecx) /* EAX */
921 pushl 20(%ecx) /* ECX */
922 pushl 16(%ecx) /* EDX */
923 pushl 24(%ecx) /* EBX */
924 pushl $0 /* popal is going to ignore esp */
925 pushl %ebp /* is this right?? */
926 pushl 12(%ecx) /* ESI */
927 pushl 8(%ecx) /* EDI */
928 pushl $0 /* args for call_into_lisp */
930 pushl 4(%ecx) /* function to call */
932 /* free our save block */
933 pushl %ecx /* reserve sufficient space on stack for args */
935 andl $0xfffffff0, %esp /* align stack */
938 call GNAME(os_invalidate)
940 /* call call_into_lisp */
942 call GNAME(call_into_lisp)
944 /* Clean up our mess */
951 SIZE(call_into_lisp_tramp)
954 .align align_4byte,0x90
955 .globl GNAME(post_signal_tramp)
956 TYPE(GNAME(post_signal_tramp))
957 GNAME(post_signal_tramp):
958 /* this is notionally the second half of a function whose first half
959 * doesn't exist. This is where call_into_lisp returns when called
960 * using return_to_lisp_function */
961 addl $12,%esp /* clear call_into_lisp args from stack */
962 popal /* restore registers */
964 #ifdef LISP_FEATURE_DARWIN
965 /* skip two padding words */
970 SIZE(GNAME(post_signal_tramp))
973 /* fast_bzero implementations and code to detect which implementation
977 .globl GNAME(fast_bzero_pointer)
980 GNAME(fast_bzero_pointer):
981 /* Variable containing a pointer to the bzero function to use.
982 * Initially points to a basic function. Change this variable
983 * to fast_bzero_detect if OS supports SSE. */
984 .long GNAME(fast_bzero_base)
987 .align align_8byte,0x90
988 .globl GNAME(fast_bzero)
989 TYPE(GNAME(fast_bzero))
991 /* Indirect function call */
992 jmp *GNAME(fast_bzero_pointer)
993 SIZE(GNAME(fast_bzero))
997 .align align_8byte,0x90
998 .globl GNAME(fast_bzero_detect)
999 TYPE(GNAME(fast_bzero_detect))
1000 GNAME(fast_bzero_detect):
1001 /* Decide whether to use SSE, MMX or REP version */
1002 push %eax /* CPUID uses EAX-EDX */
1008 test $0x04000000, %edx /* SSE2 needed for MOVNTDQ */
1010 /* Originally there was another case here for using the
1011 * MOVNTQ instruction for processors that supported MMX but
1012 * not SSE2. This turned out to be a loss especially on
1013 * Athlons (where this instruction is apparently microcoded
1014 * somewhat slowly). So for simplicity revert to REP STOSL
1015 * for all non-SSE2 processors.
1018 movl $(GNAME(fast_bzero_base)), GNAME(fast_bzero_pointer)
1021 movl $(GNAME(fast_bzero_sse)), GNAME(fast_bzero_pointer)
1029 jmp *GNAME(fast_bzero_pointer)
1031 SIZE(GNAME(fast_bzero_detect))
1035 .align align_8byte,0x90
1036 .globl GNAME(fast_bzero_sse)
1037 TYPE(GNAME(fast_bzero_sse))
1039 GNAME(fast_bzero_sse):
1040 /* A fast routine for zero-filling blocks of memory that are
1041 * guaranteed to start and end at a 4096-byte aligned address.
1043 push %esi /* Save temporary registers */
1045 mov 16(%esp), %esi /* Parameter: amount of bytes to fill */
1046 mov 12(%esp), %edi /* Parameter: start address */
1047 shr $6, %esi /* Amount of 64-byte blocks to copy */
1048 jz Lend_sse /* If none, stop */
1049 movups %xmm7, -16(%esp) /* Save XMM register */
1050 xorps %xmm7, %xmm7 /* Zero the XMM register */
1055 /* Copy the 16 zeroes from xmm7 to memory, 4 times. MOVNTDQ is the
1056 * non-caching double-quadword moving variant, i.e. the memory areas
1057 * we're touching are not fetched into the L1 cache, since we're just
1058 * going to overwrite the memory soon anyway.
1060 movntdq %xmm7, 0(%edi)
1061 movntdq %xmm7, 16(%edi)
1062 movntdq %xmm7, 32(%edi)
1063 movntdq %xmm7, 48(%edi)
1065 add $64, %edi /* Advance pointer */
1066 dec %esi /* Decrement 64-byte block count */
1068 movups -16(%esp), %xmm7 /* Restore the XMM register */
1069 sfence /* Ensure that weakly ordered writes are flushed. */
1071 mov 12(%esp), %esi /* Parameter: start address */
1072 prefetcht0 0(%esi) /* Prefetch the start of the block into cache,
1073 * since it's likely to be used immediately. */
1074 pop %edi /* Restore temp registers */
1077 SIZE(GNAME(fast_bzero_sse))
1081 .align align_8byte,0x90
1082 .globl GNAME(fast_bzero_base)
1083 TYPE(GNAME(fast_bzero_base))
1085 GNAME(fast_bzero_base):
1086 /* A fast routine for zero-filling blocks of memory that are
1087 * guaranteed to start and end at a 4096-byte aligned address.
1089 push %eax /* Save temporary registers */
1092 mov 20(%esp), %ecx /* Parameter: amount of bytes to fill */
1093 mov 16(%esp), %edi /* Parameter: start address */
1094 xor %eax, %eax /* Zero EAX */
1095 shr $2, %ecx /* Amount of 4-byte blocks to copy */
1097 cld /* Set direction of STOSL to increment */
1100 stosl /* Store EAX to *EDI, ECX times, incrementing
1101 * EDI by 4 after each store */
1104 pop %edi /* Restore temp registers */
1108 SIZE(GNAME(fast_bzero_base))