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(all_threads)
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.)
99 * Return values are in eax and maybe edx for quads, or st(0) for
102 * This should work for Lisp calls C calls Lisp calls C..
104 * FIXME & OAOOM: This duplicates call-out in src/compiler/x86/c-call.lisp,
105 * so if you tweak this, change that too!
108 .align align_16byte,0x90
109 .globl GNAME(call_into_c)
110 TYPE(GNAME(call_into_c))
112 /* Save the return Lisp address in ebx. */
115 /* Setup the NPX for C */
125 call *%eax # normal callout using Lisp stack
126 movl %eax,%ecx # remember integer return value
128 /* Check for a return FP 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.
146 /* Restore the return value. */
147 movl %ecx,%eax # maybe return 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.
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. */
169 SIZE(GNAME(call_into_c))
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
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 movl THREAD_CONTROL_STACK_START_OFFSET(%eax) ,%esp
189 /* don't think too hard about what happens if we get interrupted
191 addl $(THREAD_CONTROL_STACK_SIZE),%esp
193 /* Win32 -really- doesn't like you switching stacks out from under it. */
194 movl GNAME(all_threads),%eax
199 .globl GNAME(call_into_lisp)
200 TYPE(GNAME(call_into_lisp))
202 /* The C conventions require that ebx, esi, edi, and ebp be preserved
203 * across function calls. */
205 .align align_16byte,0x90
206 GNAME(call_into_lisp):
207 pushl %ebp # Save old frame pointer.
208 movl %esp,%ebp # Establish new frame.
210 /* Save the NPX state */
211 fwait # Catch any pending NPX exceptions.
212 subl $108,%esp # Make room for the NPX state.
213 fnsave (%esp) # save and reset NPX
215 movl (%esp),%eax # Load NPX control word.
216 andl $0xfffff2ff,%eax # Set rounding mode to nearest.
217 orl $0x00000200,%eax # Set precision to 64 bits. (53-bit mantissa)
219 fldcw (%esp) # Recover modes.
222 fldz # Ensure no FP regs are empty.
231 /* Save C regs: ebx esi edi. */
236 /* Clear descriptor regs. */
237 xorl %eax,%eax # lexenv
238 xorl %ebx,%ebx # available
239 xorl %ecx,%ecx # arg count
240 xorl %edx,%edx # first arg
241 xorl %edi,%edi # second arg
242 xorl %esi,%esi # third arg
244 /* no longer in function call */
245 movl %esp,%ebx # remember current stack
246 pushl %ebx # Save entry stack on (maybe) new stack.
248 /* Establish Lisp args. */
249 movl 8(%ebp),%eax # lexenv?
250 movl 12(%ebp),%ebx # address of arg vec
251 movl 16(%ebp),%ecx # num args
252 shll $2,%ecx # Make num args into fixnum.
255 movl (%ebx),%edx # arg0
258 movl 4(%ebx),%edi # arg1
261 movl 8(%ebx),%esi # arg2
263 /* Registers eax, ecx, edx, edi, and esi are now live. */
265 #ifdef LISP_FEATURE_WIN32
266 /* Establish an SEH frame. */
267 #ifdef LISP_FEATURE_SB_THREAD
268 /* FIXME: need to save BSP here. */
269 #error "need to save BSP here, but don't know how yet."
271 pushl BINDING_STACK_POINTER + SYMBOL_VALUE_OFFSET
273 pushl $GNAME(exception_handler_wrapper)
278 /* Alloc new frame. */
279 mov %esp,%ebx # The current sp marks start of new frame.
280 push %ebp # fp in save location S0
281 sub $8,%esp # Ensure 3 slots are allocated, one above.
282 mov %ebx,%ebp # Switch to new frame.
284 call *CLOSURE_FUN_OFFSET(%eax)
286 /* If the function returned multiple values, it will return to
287 this point. Lose them */
291 /* A singled value function returns here */
293 #ifdef LISP_FEATURE_WIN32
294 /* Remove our SEH frame. */
299 /* Restore the stack, in case there was a stack change. */
302 /* Restore C regs: ebx esi edi. */
307 /* Restore the NPX state. */
312 movl %edx,%eax # c-val
314 SIZE(GNAME(call_into_lisp))
316 /* support for saving and restoring the NPX state from C */
318 .globl GNAME(fpu_save)
319 TYPE(GNAME(fpu_save))
323 fnsave (%eax) # Save the NPX state. (resets NPX)
325 SIZE(GNAME(fpu_save))
327 .globl GNAME(fpu_restore)
328 TYPE(GNAME(fpu_restore))
332 frstor (%eax) # Restore the NPX state.
334 SIZE(GNAME(fpu_restore))
337 * the undefined-function trampoline
340 .align align_4byte,0x90
341 .globl GNAME(undefined_tramp)
342 TYPE(GNAME(undefined_tramp))
343 .byte 0, 0, 0, SIMPLE_FUN_HEADER_WIDETAG
344 GNAME(undefined_tramp):
348 .byte UNDEFINED_FUN_ERROR
349 .byte sc_DescriptorReg # eax in the Descriptor-reg SC
351 SIZE(GNAME(undefined_tramp))
354 * the closure trampoline
357 .align align_4byte,0x90
358 .globl GNAME(closure_tramp)
359 TYPE(GNAME(closure_tramp))
360 .byte 0, 0, 0, SIMPLE_FUN_HEADER_WIDETAG
361 GNAME(closure_tramp):
362 movl FDEFN_FUN_OFFSET(%eax),%eax
363 /* FIXME: The '*' after "jmp" in the next line is from PVE's
364 * patch posted to the CMU CL mailing list Oct 6, 1999. It looks
365 * reasonable, and it certainly seems as though if CMU CL needs it,
366 * SBCL needs it too, but I haven't actually verified that it's
367 * right. It would be good to find a way to force the flow of
368 * control through here to test it. */
369 jmp *CLOSURE_FUN_OFFSET(%eax)
370 SIZE(GNAME(closure_tramp))
373 .align align_4byte,0x90
374 .globl GNAME(funcallable_instance_tramp)
375 TYPE(GNAME(funcallable_instance_tramp))
376 GNAME(funcallable_instance_tramp):
377 movl FUNCALLABLE_INSTANCE_FUNCTION_OFFSET(%eax),%eax
378 /* KLUDGE: on this platform, whatever kind of function is in %rax
379 * now, the first word of it contains the address to jump to. */
380 jmp *CLOSURE_FUN_OFFSET(%eax)
381 SIZE(GNAME(funcallable_instance_tramp))
384 * fun-end breakpoint magic
387 .globl GNAME(fun_end_breakpoint_guts)
389 GNAME(fun_end_breakpoint_guts):
390 /* Multiple Value return */
391 jc multiple_value_return
392 /* Single value return: The eventual return will now use the
393 multiple values return convention but with a return values
395 movl %esp,%ebx # Setup ebx - the ofp.
396 subl $4,%esp # Allocate one stack slot for the return value
397 movl $4,%ecx # Setup ecx for one return value.
398 movl $(NIL),%edi # default second value
399 movl $(NIL),%esi # default third value
401 multiple_value_return:
403 .globl GNAME(fun_end_breakpoint_trap)
404 GNAME(fun_end_breakpoint_trap):
406 .byte trap_FunEndBreakpoint
407 hlt # We should never return here.
409 .globl GNAME(fun_end_breakpoint_end)
410 GNAME(fun_end_breakpoint_end):
413 .globl GNAME(do_pending_interrupt)
414 TYPE(GNAME(do_pending_interrupt))
415 .align align_4byte,0x90
416 GNAME(do_pending_interrupt):
418 .byte trap_PendingInterrupt
420 SIZE(GNAME(do_pending_interrupt))
422 /* Allocate bytes and return the start of the allocated space
423 * in the specified destination register.
425 * In the general case the size will be in the destination register.
427 * All registers must be preserved except the destination.
428 * The C conventions will preserve ebx, esi, edi, and ebp.
429 * So only eax, ecx, and edx need special care here.
431 * ALLOC factors out the logic of calling alloc(): stack alignment, etc.
433 * DEFINE_ALLOC_TO_FOO defines an alloction routine.
436 #ifdef LISP_FEATURE_DARWIN
437 #define ALLOC(size) \
438 pushl %ebp; /* Save EBP */ \
439 movl %esp,%ebp; /* Save ESP to EBP */ \
440 andl $0xfffffff0,%esp; /* Align stack */ \
441 pushl $0; /* Padding */ \
442 pushl size; /* Argument to alloc */ \
444 movl %ebp,%esp; /* Restore ESP from EBP */ \
445 popl %ebp; /* Restore EBP */
447 #define ALLOC(size) \
448 pushl size; /* Argument to alloc */ \
450 addl $4,%esp; /* Pop argument */
453 #define DEFINE_ALLOC_TO_EAX(name,size) \
454 .globl GNAME(name); \
456 .align align_4byte,0x90; \
458 pushl %ecx; /* Save ECX and EDX */ \
461 popl %edx; /* Restore ECX and EDX */ \
466 #define DEFINE_ALLOC_TO_ECX(name,size) \
467 .globl GNAME(name); \
469 .align align_4byte,0x90; \
471 pushl %eax; /* Save EAX and EDX */ \
474 movl %eax,%ecx; /* Result to destination */ \
480 #define DEFINE_ALLOC_TO_EDX(name,size) \
481 .globl GNAME(name); \
483 .align align_4byte,0x90; \
485 pushl %eax; /* Save EAX and ECX */ \
488 movl %eax,%edx; /* Restore EAX and ECX */ \
494 #define DEFINE_ALLOC_TO_REG(name,reg,size) \
495 .globl GNAME(name); \
497 .align align_4byte,0x90; \
499 pushl %eax; /* Save EAX, ECX, and EDX */ \
503 movl %eax,reg; /* Restore them */ \
510 DEFINE_ALLOC_TO_EAX(alloc_to_eax,%eax)
511 DEFINE_ALLOC_TO_EAX(alloc_8_to_eax,$8)
512 DEFINE_ALLOC_TO_EAX(alloc_16_to_eax,$16)
514 DEFINE_ALLOC_TO_ECX(alloc_to_ecx,%ecx)
515 DEFINE_ALLOC_TO_ECX(alloc_8_to_ecx,$8)
516 DEFINE_ALLOC_TO_ECX(alloc_16_to_ecx,$16)
518 DEFINE_ALLOC_TO_EDX(alloc_to_edx,%edx)
519 DEFINE_ALLOC_TO_EDX(alloc_8_to_edx,$8)
520 DEFINE_ALLOC_TO_EDX(alloc_16_to_edx,$16)
522 DEFINE_ALLOC_TO_REG(alloc_to_ebx,%ebx,%ebx)
523 DEFINE_ALLOC_TO_REG(alloc_8_to_ebx,%ebx,$8)
524 DEFINE_ALLOC_TO_REG(alloc_16_to_ebx,%ebx,$16)
526 DEFINE_ALLOC_TO_REG(alloc_to_esi,%esi,%esi)
527 DEFINE_ALLOC_TO_REG(alloc_8_to_esi,%esi,$8)
528 DEFINE_ALLOC_TO_REG(alloc_16_to_esi,%esi,$16)
530 DEFINE_ALLOC_TO_REG(alloc_to_edi,%edi,%edi)
531 DEFINE_ALLOC_TO_REG(alloc_8_to_edi,%edi,$8)
532 DEFINE_ALLOC_TO_REG(alloc_16_to_edi,%edi,$16)
534 /* Called from lisp when an inline allocation overflows.
535 * Every register except the result needs to be preserved.
536 * We depend on C to preserve ebx, esi, edi, and ebp.
537 * But where necessary must save eax, ecx, edx. */
539 #ifdef LISP_FEATURE_SB_THREAD
540 #define START_REGION %fs:THREAD_ALLOC_REGION_OFFSET
542 #define START_REGION GNAME(boxed_region)
545 #define ALLOC_OVERFLOW(size) \
546 /* Calculate the size for the allocation. */ \
547 subl START_REGION,size; \
550 /* This routine handles an overflow with eax=crfp+size. So the
553 .globl GNAME(alloc_overflow_eax)
554 TYPE(GNAME(alloc_overflow_eax))
555 GNAME(alloc_overflow_eax):
556 pushl %ecx # Save ecx
557 pushl %edx # Save edx
559 popl %edx # Restore edx.
560 popl %ecx # Restore ecx.
562 SIZE(GNAME(alloc_overflow_eax))
565 .globl GNAME(alloc_overflow_ecx)
566 TYPE(GNAME(alloc_overflow_ecx))
567 GNAME(alloc_overflow_ecx):
568 pushl %eax # Save eax
569 pushl %edx # Save edx
571 movl %eax,%ecx # setup the destination.
572 popl %edx # Restore edx.
573 popl %eax # Restore eax.
575 SIZE(GNAME(alloc_overflow_ecx))
578 .globl GNAME(alloc_overflow_edx)
579 TYPE(GNAME(alloc_overflow_edx))
580 GNAME(alloc_overflow_edx):
581 pushl %eax # Save eax
582 pushl %ecx # Save ecx
584 movl %eax,%edx # setup the destination.
585 popl %ecx # Restore ecx.
586 popl %eax # Restore eax.
588 SIZE(GNAME(alloc_overflow_edx))
590 /* This routine handles an overflow with ebx=crfp+size. So the
593 .globl GNAME(alloc_overflow_ebx)
594 TYPE(GNAME(alloc_overflow_ebx))
595 GNAME(alloc_overflow_ebx):
596 pushl %eax # Save eax
597 pushl %ecx # Save ecx
598 pushl %edx # Save edx
600 movl %eax,%ebx # setup the destination.
601 popl %edx # Restore edx.
602 popl %ecx # Restore ecx.
603 popl %eax # Restore eax.
605 SIZE(GNAME(alloc_overflow_ebx))
607 /* This routine handles an overflow with esi=crfp+size. So the
610 .globl GNAME(alloc_overflow_esi)
611 TYPE(GNAME(alloc_overflow_esi))
612 GNAME(alloc_overflow_esi):
613 pushl %eax # Save eax
614 pushl %ecx # Save ecx
615 pushl %edx # Save edx
617 movl %eax,%esi # setup the destination.
618 popl %edx # Restore edx.
619 popl %ecx # Restore ecx.
620 popl %eax # Restore eax.
622 SIZE(GNAME(alloc_overflow_esi))
625 .globl GNAME(alloc_overflow_edi)
626 TYPE(GNAME(alloc_overflow_edi))
627 GNAME(alloc_overflow_edi):
628 pushl %eax # Save eax
629 pushl %ecx # Save ecx
630 pushl %edx # Save edx
632 movl %eax,%edi # setup the destination.
633 popl %edx # Restore edx.
634 popl %ecx # Restore ecx.
635 popl %eax # Restore eax.
637 SIZE(GNAME(alloc_overflow_edi))
640 #ifdef LISP_FEATURE_WIN32
641 /* The guts of the exception-handling system doesn't use
642 * frame pointers, which manages to throw off backtraces
643 * rather badly. So here we grab the (known-good) EBP
644 * and EIP from the exception context and use it to fake
645 * up a stack frame which will skip over the system SEH
648 .globl GNAME(exception_handler_wrapper)
649 TYPE(GNAME(exception_handler_wrapper))
650 GNAME(exception_handler_wrapper):
651 /* Context layout is: */
652 /* 7 dwords before FSA. (0x1c) */
653 /* 8 dwords and 0x50 bytes in the FSA. (0x70/0x8c) */
654 /* 4 dwords segregs. (0x10/0x9c) */
655 /* 6 dwords non-stack GPRs. (0x18/0xb4) */
658 #define CONTEXT_EBP_OFFSET 0xb4
659 #define CONTEXT_EIP_OFFSET 0xb8
660 /* some other stuff we don't care about. */
662 movl 0x10(%esp), %ebp /* context */
663 pushl CONTEXT_EIP_OFFSET(%ebp)
664 pushl CONTEXT_EBP_OFFSET(%ebp)
670 call GNAME(handle_exception)
674 SIZE(GNAME(exception_handler_wrapper))
677 #ifdef LISP_FEATURE_DARWIN
679 .globl GNAME(call_into_lisp_tramp)
680 TYPE(GNAME(call_into_lisp_tramp))
681 GNAME(call_into_lisp_tramp):
682 /* 1. build the stack frame from the block that's pointed to by ECX
685 4. call the function via call_into_lisp
687 pushl 0(%ecx) /* return address */
692 pushl 32(%ecx) /* eflags */
693 pushl 28(%ecx) /* EAX */
694 pushl 20(%ecx) /* ECX */
695 pushl 16(%ecx) /* EDX */
696 pushl 24(%ecx) /* EBX */
697 pushl $0 /* popal is going to ignore esp */
698 pushl %ebp /* is this right?? */
699 pushl 12(%ecx) /* ESI */
700 pushl 8(%ecx) /* EDI */
701 pushl $0 /* args for call_into_lisp */
703 pushl 4(%ecx) /* function to call */
705 /* free our save block */
706 pushl %ecx /* reserve sufficient space on stack for args */
708 andl $0xfffffff0, %esp /* align stack */
711 call GNAME(os_invalidate)
713 /* call call_into_lisp */
715 call GNAME(call_into_lisp)
717 /* Clean up our mess */
724 SIZE(call_into_lisp_tramp)
727 .align align_4byte,0x90
728 .globl GNAME(post_signal_tramp)
729 TYPE(GNAME(post_signal_tramp))
730 GNAME(post_signal_tramp):
731 /* this is notionally the second half of a function whose first half
732 * doesn't exist. This is where call_into_lisp returns when called
733 * using return_to_lisp_function */
734 addl $12,%esp /* clear call_into_lisp args from stack */
735 popal /* restore registers */
737 #ifdef LISP_FEATURE_DARWIN
738 /* skip two padding words */
743 SIZE(GNAME(post_signal_tramp))
746 /* fast_bzero implementations and code to detect which implementation
750 .globl GNAME(fast_bzero_pointer)
753 GNAME(fast_bzero_pointer):
754 /* Variable containing a pointer to the bzero function to use.
755 * Initially points to a basic function. Change this variable
756 * to fast_bzero_detect if OS supports SSE. */
757 .long GNAME(fast_bzero_base)
760 .align align_8byte,0x90
761 .globl GNAME(fast_bzero)
762 TYPE(GNAME(fast_bzero))
764 /* Indirect function call */
765 jmp *GNAME(fast_bzero_pointer)
766 SIZE(GNAME(fast_bzero))
770 .align align_8byte,0x90
771 .globl GNAME(fast_bzero_detect)
772 TYPE(GNAME(fast_bzero_detect))
773 GNAME(fast_bzero_detect):
774 /* Decide whether to use SSE, MMX or REP version */
775 push %eax /* CPUID uses EAX-EDX */
781 test $0x04000000, %edx /* SSE2 needed for MOVNTDQ */
783 /* Originally there was another case here for using the
784 * MOVNTQ instruction for processors that supported MMX but
785 * not SSE2. This turned out to be a loss especially on
786 * Athlons (where this instruction is apparently microcoded
787 * somewhat slowly). So for simplicity revert to REP STOSL
788 * for all non-SSE2 processors.
791 movl $(GNAME(fast_bzero_base)), GNAME(fast_bzero_pointer)
794 movl $(GNAME(fast_bzero_sse)), GNAME(fast_bzero_pointer)
802 jmp *GNAME(fast_bzero_pointer)
804 SIZE(GNAME(fast_bzero_detect))
808 .align align_8byte,0x90
809 .globl GNAME(fast_bzero_sse)
810 TYPE(GNAME(fast_bzero_sse))
812 GNAME(fast_bzero_sse):
813 /* A fast routine for zero-filling blocks of memory that are
814 * guaranteed to start and end at a 4096-byte aligned address.
816 push %esi /* Save temporary registers */
818 mov 16(%esp), %esi /* Parameter: amount of bytes to fill */
819 mov 12(%esp), %edi /* Parameter: start address */
820 shr $6, %esi /* Amount of 64-byte blocks to copy */
821 jz Lend_sse /* If none, stop */
822 movups %xmm7, -16(%esp) /* Save XMM register */
823 xorps %xmm7, %xmm7 /* Zero the XMM register */
828 /* Copy the 16 zeroes from xmm7 to memory, 4 times. MOVNTDQ is the
829 * non-caching double-quadword moving variant, i.e. the memory areas
830 * we're touching are not fetched into the L1 cache, since we're just
831 * going to overwrite the memory soon anyway.
833 movntdq %xmm7, 0(%edi)
834 movntdq %xmm7, 16(%edi)
835 movntdq %xmm7, 32(%edi)
836 movntdq %xmm7, 48(%edi)
838 add $64, %edi /* Advance pointer */
839 dec %esi /* Decrement 64-byte block count */
841 movups -16(%esp), %xmm7 /* Restore the XMM register */
842 sfence /* Ensure that weakly ordered writes are flushed. */
844 mov 12(%esp), %esi /* Parameter: start address */
845 prefetcht0 0(%esi) /* Prefetch the start of the block into cache,
846 * since it's likely to be used immediately. */
847 pop %edi /* Restore temp registers */
850 SIZE(GNAME(fast_bzero_sse))
854 .align align_8byte,0x90
855 .globl GNAME(fast_bzero_base)
856 TYPE(GNAME(fast_bzero_base))
858 GNAME(fast_bzero_base):
859 /* A fast routine for zero-filling blocks of memory that are
860 * guaranteed to start and end at a 4096-byte aligned address.
862 push %eax /* Save temporary registers */
865 mov 20(%esp), %ecx /* Parameter: amount of bytes to fill */
866 mov 16(%esp), %edi /* Parameter: start address */
867 xor %eax, %eax /* Zero EAX */
868 shr $2, %ecx /* Amount of 4-byte blocks to copy */
872 stosl /* Store EAX to *EDI, ECX times, incrementing
873 * EDI by 4 after each store */
876 pop %edi /* Restore temp registers */
880 SIZE(GNAME(fast_bzero_base))