/* * very-low-level utilities for runtime support */ /* * This software is part of the SBCL system. See the README file for * more information. * * This software is derived from the CMU CL system, which was * written at Carnegie Mellon University and released into the * public domain. The software is in the public domain and is * provided with absolutely no warranty. See the COPYING and CREDITS * files for more information. */ #define LANGUAGE_ASSEMBLY #include "sbcl.h" #include "validate.h" #include "genesis/closure.h" #include "genesis/funcallable-instance.h" #include "genesis/fdefn.h" #include "genesis/static-symbols.h" #include "genesis/symbol.h" #include "genesis/thread.h" /* Minimize conditionalization for different OS naming schemes. * * (As of sbcl-0.8.10, this seems no longer to be much of an issue, * since everyone has converged on ELF. If this generality really * turns out not to matter, perhaps it's just clutter we could get * rid of? -- WHN 2004-04-18) * * (Except Win32, which is unlikely ever to be ELF, sorry. -- AB 2005-12-08) */ #if defined __linux__ || defined __FreeBSD__ || defined __NetBSD__ || defined __OpenBSD__ || defined __sun #define GNAME(var) var #else #define GNAME(var) _##var #endif /* Get the right type of alignment. Linux, FreeBSD and NetBSD (but not OpenBSD) * want alignment in bytes. * * (As in the GNAME() definitions above, as of sbcl-0.8.10, this seems * no longer to be much of an issue, since everyone has converged on * the same value. If this generality really turns out not to * matter any more, perhaps it's just clutter we could get * rid of? -- WHN 2004-04-18) */ #if defined(__linux__) || defined(__FreeBSD__) || defined(__NetBSD__) || defined(__OpenBSD__) || defined(__sun) || defined(LISP_FEATURE_WIN32) #define align_4byte 4 #define align_8byte 8 #define align_16byte 16 #else #define align_4byte 2 #define align_8byte 3 #define align_16byte 4 #endif /* * The assembler used for win32 doesn't like .type or .size directives, * so we want to conditionally kill them out. So let's wrap them in macros * that are defined to be no-ops on win32. Hopefully this still works on * other platforms. */ #if !defined(LISP_FEATURE_WIN32) && !defined(LISP_FEATURE_DARWIN) #define TYPE(name) .type name,@function #define SIZE(name) .size name,.-name #else #define TYPE(name) #define SIZE(name) #endif /* * x86/darwin (as of MacOS X 10.4.5) doesn't reliably file signal * handlers (SIGTRAP or Mach exception handlers) for 0xCC, wo we have * to use ud2 instead. ud2 is an undefined opcode, #x0b0f, or * 0F 0B in low-endian notation, that causes SIGILL to fire. We check * for this instruction in the SIGILL handler and if we see it, we * advance the EIP by two bytes to skip over ud2 instruction and * call sigtrap_handler. */ #if defined(LISP_FEATURE_UD2_BREAKPOINTS) #define END() #define TRAP ud2 #else #define END() .end #define TRAP int3 #endif .text .globl GNAME(all_threads) /* * A call to call_into_c preserves esi, edi, and ebp. * (The C function will preserve ebx, esi, edi, and ebp across its * function call, but we trash ebx ourselves by using it to save the * return Lisp address.) * * Return values are in eax and maybe edx for quads, or st(0) for * floats. * * This should work for Lisp calls C calls Lisp calls C.. * * FIXME & OAOOM: This duplicates call-out in src/compiler/x86/c-call.lisp, * so if you tweak this, change that too! */ .text .align align_16byte,0x90 .globl GNAME(call_into_c) TYPE(GNAME(call_into_c)) GNAME(call_into_c): /* Save the return Lisp address in ebx. */ popl %ebx /* Setup the NPX for C */ fstp %st(0) fstp %st(0) fstp %st(0) fstp %st(0) fstp %st(0) fstp %st(0) fstp %st(0) fstp %st(0) call *%eax # normal callout using Lisp stack movl %eax,%ecx # remember integer return value /* Check for a return FP value. */ fxam fnstsw %ax andl $0x4500,%eax cmpl $0x4100,%eax jne Lfp_rtn_value /* The return value is in eax, or eax,edx? */ /* Set up the NPX stack for Lisp. */ fldz # Ensure no regs are empty. fldz fldz fldz fldz fldz fldz fldz /* Restore the return value. */ movl %ecx,%eax # maybe return value /* Return. */ jmp *%ebx Lfp_rtn_value: /* The return result is in st(0). */ /* Set up the NPX stack for Lisp, placing the result in st(0). */ fldz # Ensure no regs are empty. fldz fldz fldz fldz fldz fldz fxch %st(7) # Move the result back to st(0). /* We don't need to restore eax, because the result is in st(0). */ /* Return. FIXME: It would be nice to restructure this to use RET. */ jmp *%ebx SIZE(GNAME(call_into_c)) .text .globl GNAME(call_into_lisp_first_time) TYPE(GNAME(call_into_lisp_first_time)) /* We don't worry too much about saving registers * here, because we never expect to return from the initial call to lisp * anyway */ .align align_16byte,0x90 GNAME(call_into_lisp_first_time): pushl %ebp # Save old frame pointer. movl %esp,%ebp # Establish new frame. #ifndef LISP_FEATURE_WIN32 movl GNAME(all_threads),%eax /* pthread machinery takes care of this for other threads */ movl THREAD_CONTROL_STACK_END_OFFSET(%eax) ,%esp #else /* Win32 -really- doesn't like you switching stacks out from under it. */ movl GNAME(all_threads),%eax #endif jmp Lstack .text .globl GNAME(call_into_lisp) TYPE(GNAME(call_into_lisp)) /* The C conventions require that ebx, esi, edi, and ebp be preserved * across function calls. */ .align align_16byte,0x90 GNAME(call_into_lisp): pushl %ebp # Save old frame pointer. movl %esp,%ebp # Establish new frame. Lstack: /* Save the NPX state */ fwait # Catch any pending NPX exceptions. subl $108,%esp # Make room for the NPX state. fnsave (%esp) # save and reset NPX movl (%esp),%eax # Load NPX control word. andl $0xfffff2ff,%eax # Set rounding mode to nearest. orl $0x00000200,%eax # Set precision to 64 bits. (53-bit mantissa) pushl %eax fldcw (%esp) # Recover modes. popl %eax fldz # Ensure no FP regs are empty. fldz fldz fldz fldz fldz fldz fldz /* Save C regs: ebx esi edi. */ pushl %ebx pushl %esi pushl %edi /* Clear descriptor regs. */ xorl %eax,%eax # lexenv xorl %ebx,%ebx # available xorl %ecx,%ecx # arg count xorl %edx,%edx # first arg xorl %edi,%edi # second arg xorl %esi,%esi # third arg /* no longer in function call */ movl %esp,%ebx # remember current stack pushl %ebx # Save entry stack on (maybe) new stack. /* Establish Lisp args. */ movl 8(%ebp),%eax # lexenv? movl 12(%ebp),%ebx # address of arg vec movl 16(%ebp),%ecx # num args shll $2,%ecx # Make num args into fixnum. cmpl $0,%ecx je Ldone movl (%ebx),%edx # arg0 cmpl $4,%ecx je Ldone movl 4(%ebx),%edi # arg1 cmpl $8,%ecx je Ldone movl 8(%ebx),%esi # arg2 Ldone: /* Registers eax, ecx, edx, edi, and esi are now live. */ #ifdef LISP_FEATURE_WIN32 /* Establish an SEH frame. */ #ifdef LISP_FEATURE_SB_THREAD /* FIXME: need to save BSP here. */ #error "need to save BSP here, but don't know how yet." #else pushl BINDING_STACK_POINTER + SYMBOL_VALUE_OFFSET #endif pushl $GNAME(exception_handler_wrapper) pushl %fs:0 movl %esp, %fs:0 #endif /* Alloc new frame. */ push %ebp # Dummy for return address push %ebp # fp in save location S1 mov %esp,%ebp # The current sp marks start of new frame. sub $4,%esp # Ensure 3 slots are allocated, two above. call *CLOSURE_FUN_OFFSET(%eax) /* If the function returned multiple values, it will return to this point. Lose them */ jnc LsingleValue mov %ebx, %esp LsingleValue: /* A singled value function returns here */ #ifdef LISP_FEATURE_WIN32 /* Remove our SEH frame. */ popl %fs:0 add $8, %esp #endif /* Restore the stack, in case there was a stack change. */ popl %esp # c-sp /* Restore C regs: ebx esi edi. */ popl %edi popl %esi popl %ebx /* Restore the NPX state. */ frstor (%esp) addl $108, %esp popl %ebp # c-sp movl %edx,%eax # c-val ret SIZE(GNAME(call_into_lisp)) /* support for saving and restoring the NPX state from C */ .text .globl GNAME(fpu_save) TYPE(GNAME(fpu_save)) .align 2,0x90 GNAME(fpu_save): movl 4(%esp),%eax fnsave (%eax) # Save the NPX state. (resets NPX) ret SIZE(GNAME(fpu_save)) .globl GNAME(fpu_restore) TYPE(GNAME(fpu_restore)) .align 2,0x90 GNAME(fpu_restore): movl 4(%esp),%eax frstor (%eax) # Restore the NPX state. ret SIZE(GNAME(fpu_restore)) /* * the undefined-function trampoline */ .text .align align_16byte,0x90 .globl GNAME(undefined_tramp) TYPE(GNAME(undefined_tramp)) .byte 0, 0, 0, SIMPLE_FUN_HEADER_WIDETAG GNAME(undefined_tramp): pop 4(%ebp) # Save return PC for backtrace. TRAP .byte trap_Error .byte 2 .byte UNDEFINED_FUN_ERROR .byte sc_DescriptorReg # eax in the Descriptor-reg SC ret SIZE(GNAME(undefined_tramp)) /* * the closure trampoline */ .text .align align_16byte,0x90 .globl GNAME(closure_tramp) TYPE(GNAME(closure_tramp)) .byte 0, 0, 0, SIMPLE_FUN_HEADER_WIDETAG GNAME(closure_tramp): movl FDEFN_FUN_OFFSET(%eax),%eax /* FIXME: The '*' after "jmp" in the next line is from PVE's * patch posted to the CMU CL mailing list Oct 6, 1999. It looks * reasonable, and it certainly seems as though if CMU CL needs it, * SBCL needs it too, but I haven't actually verified that it's * right. It would be good to find a way to force the flow of * control through here to test it. */ jmp *CLOSURE_FUN_OFFSET(%eax) SIZE(GNAME(closure_tramp)) .text .align align_16byte,0x90 .globl GNAME(funcallable_instance_tramp) TYPE(GNAME(funcallable_instance_tramp)) GNAME(funcallable_instance_tramp): movl FUNCALLABLE_INSTANCE_FUNCTION_OFFSET(%eax),%eax /* KLUDGE: on this platform, whatever kind of function is in %rax * now, the first word of it contains the address to jump to. */ jmp *CLOSURE_FUN_OFFSET(%eax) SIZE(GNAME(funcallable_instance_tramp)) /* * fun-end breakpoint magic */ /* * For an explanation of the magic involved in function-end * breakpoints, see the implementation in ppc-assem.S. */ .text .globl GNAME(fun_end_breakpoint_guts) .align align_16byte GNAME(fun_end_breakpoint_guts): /* Multiple Value return */ jc multiple_value_return /* Single value return: The eventual return will now use the multiple values return convention but with a return values count of one. */ movl %esp,%ebx # Setup ebx - the ofp. subl $4,%esp # Allocate one stack slot for the return value movl $4,%ecx # Setup ecx for one return value. movl $(NIL),%edi # default second value movl $(NIL),%esi # default third value multiple_value_return: .globl GNAME(fun_end_breakpoint_trap) GNAME(fun_end_breakpoint_trap): TRAP .byte trap_FunEndBreakpoint hlt # We should never return here. .globl GNAME(fun_end_breakpoint_end) GNAME(fun_end_breakpoint_end): .globl GNAME(do_pending_interrupt) TYPE(GNAME(do_pending_interrupt)) .align align_16byte,0x90 GNAME(do_pending_interrupt): TRAP .byte trap_PendingInterrupt ret SIZE(GNAME(do_pending_interrupt)) /* Allocate bytes and return the start of the allocated space * in the specified destination register. * * In the general case the size will be in the destination register. * * All registers must be preserved except the destination. * The C conventions will preserve ebx, esi, edi, and ebp. * So only eax, ecx, and edx need special care here. * * ALLOC factors out the logic of calling alloc(): stack alignment, etc. * * DEFINE_ALLOC_TO_FOO defines an alloction routine. */ #ifdef LISP_FEATURE_DARWIN #define ALLOC(size) \ pushl %ebp; /* Save EBP */ \ movl %esp,%ebp; /* Save ESP to EBP */ \ pushl $0; /* Reserve space for arg */ \ andl $0xfffffff0,%esp; /* Align stack to 16bytes */ \ movl size, (%esp); /* Argument to alloc */ \ call GNAME(alloc); \ movl %ebp,%esp; /* Restore ESP from EBP */ \ popl %ebp; /* Restore EBP */ #else #define ALLOC(size) \ pushl size; /* Argument to alloc */ \ call GNAME(alloc); \ addl $4,%esp; /* Pop argument */ #endif #define DEFINE_ALLOC_TO_EAX(name,size) \ .globl GNAME(name); \ TYPE(GNAME(name)); \ .align align_16byte,0x90; \ GNAME(name): \ pushl %ecx; /* Save ECX and EDX */ \ pushl %edx; \ ALLOC(size) \ popl %edx; /* Restore ECX and EDX */ \ popl %ecx; \ ret; \ SIZE(GNAME(name)) #define DEFINE_ALLOC_TO_ECX(name,size) \ .globl GNAME(name); \ TYPE(GNAME(name)); \ .align align_16byte,0x90; \ GNAME(name): \ pushl %eax; /* Save EAX and EDX */ \ pushl %edx; \ ALLOC(size) \ movl %eax,%ecx; /* Result to destination */ \ popl %edx; \ popl %eax; \ ret; \ SIZE(GNAME(name)) #define DEFINE_ALLOC_TO_EDX(name,size) \ .globl GNAME(name); \ TYPE(GNAME(name)); \ .align align_16byte,0x90; \ GNAME(name): \ pushl %eax; /* Save EAX and ECX */ \ pushl %ecx; \ ALLOC(size) \ movl %eax,%edx; /* Restore EAX and ECX */ \ popl %ecx; \ popl %eax; \ ret; \ SIZE(GNAME(name)) #define DEFINE_ALLOC_TO_REG(name,reg,size) \ .globl GNAME(name); \ TYPE(GNAME(name)); \ .align align_16byte,0x90; \ GNAME(name): \ pushl %eax; /* Save EAX, ECX, and EDX */ \ pushl %ecx; \ pushl %edx; \ ALLOC(size) \ movl %eax,reg; /* Restore them */ \ popl %edx; \ popl %ecx; \ popl %eax; \ ret; \ SIZE(GNAME(name)) DEFINE_ALLOC_TO_EAX(alloc_to_eax,%eax) DEFINE_ALLOC_TO_EAX(alloc_8_to_eax,$8) DEFINE_ALLOC_TO_EAX(alloc_16_to_eax,$16) DEFINE_ALLOC_TO_ECX(alloc_to_ecx,%ecx) DEFINE_ALLOC_TO_ECX(alloc_8_to_ecx,$8) DEFINE_ALLOC_TO_ECX(alloc_16_to_ecx,$16) DEFINE_ALLOC_TO_EDX(alloc_to_edx,%edx) DEFINE_ALLOC_TO_EDX(alloc_8_to_edx,$8) DEFINE_ALLOC_TO_EDX(alloc_16_to_edx,$16) DEFINE_ALLOC_TO_REG(alloc_to_ebx,%ebx,%ebx) DEFINE_ALLOC_TO_REG(alloc_8_to_ebx,%ebx,$8) DEFINE_ALLOC_TO_REG(alloc_16_to_ebx,%ebx,$16) DEFINE_ALLOC_TO_REG(alloc_to_esi,%esi,%esi) DEFINE_ALLOC_TO_REG(alloc_8_to_esi,%esi,$8) DEFINE_ALLOC_TO_REG(alloc_16_to_esi,%esi,$16) DEFINE_ALLOC_TO_REG(alloc_to_edi,%edi,%edi) DEFINE_ALLOC_TO_REG(alloc_8_to_edi,%edi,$8) DEFINE_ALLOC_TO_REG(alloc_16_to_edi,%edi,$16) /* Called from lisp when an inline allocation overflows. * Every register except the result needs to be preserved. * We depend on C to preserve ebx, esi, edi, and ebp. * But where necessary must save eax, ecx, edx. */ #ifdef LISP_FEATURE_SB_THREAD #define START_REGION %fs:THREAD_ALLOC_REGION_OFFSET #else #define START_REGION GNAME(boxed_region) #endif #define ALLOC_OVERFLOW(size) \ /* Calculate the size for the allocation. */ \ subl START_REGION,size; \ ALLOC(size) /* This routine handles an overflow with eax=crfp+size. So the size=eax-crfp. */ .align align_16byte .globl GNAME(alloc_overflow_eax) TYPE(GNAME(alloc_overflow_eax)) GNAME(alloc_overflow_eax): pushl %ecx # Save ecx pushl %edx # Save edx ALLOC_OVERFLOW(%eax) popl %edx # Restore edx. popl %ecx # Restore ecx. ret SIZE(GNAME(alloc_overflow_eax)) .align align_16byte .globl GNAME(alloc_overflow_ecx) TYPE(GNAME(alloc_overflow_ecx)) GNAME(alloc_overflow_ecx): pushl %eax # Save eax pushl %edx # Save edx ALLOC_OVERFLOW(%ecx) movl %eax,%ecx # setup the destination. popl %edx # Restore edx. popl %eax # Restore eax. ret SIZE(GNAME(alloc_overflow_ecx)) .align align_16byte .globl GNAME(alloc_overflow_edx) TYPE(GNAME(alloc_overflow_edx)) GNAME(alloc_overflow_edx): pushl %eax # Save eax pushl %ecx # Save ecx ALLOC_OVERFLOW(%edx) movl %eax,%edx # setup the destination. popl %ecx # Restore ecx. popl %eax # Restore eax. ret SIZE(GNAME(alloc_overflow_edx)) /* This routine handles an overflow with ebx=crfp+size. So the size=ebx-crfp. */ .align align_16byte .globl GNAME(alloc_overflow_ebx) TYPE(GNAME(alloc_overflow_ebx)) GNAME(alloc_overflow_ebx): pushl %eax # Save eax pushl %ecx # Save ecx pushl %edx # Save edx ALLOC_OVERFLOW(%ebx) movl %eax,%ebx # setup the destination. popl %edx # Restore edx. popl %ecx # Restore ecx. popl %eax # Restore eax. ret SIZE(GNAME(alloc_overflow_ebx)) /* This routine handles an overflow with esi=crfp+size. So the size=esi-crfp. */ .align align_16byte .globl GNAME(alloc_overflow_esi) TYPE(GNAME(alloc_overflow_esi)) GNAME(alloc_overflow_esi): pushl %eax # Save eax pushl %ecx # Save ecx pushl %edx # Save edx ALLOC_OVERFLOW(%esi) movl %eax,%esi # setup the destination. popl %edx # Restore edx. popl %ecx # Restore ecx. popl %eax # Restore eax. ret SIZE(GNAME(alloc_overflow_esi)) .align align_16byte .globl GNAME(alloc_overflow_edi) TYPE(GNAME(alloc_overflow_edi)) GNAME(alloc_overflow_edi): pushl %eax # Save eax pushl %ecx # Save ecx pushl %edx # Save edx ALLOC_OVERFLOW(%edi) movl %eax,%edi # setup the destination. popl %edx # Restore edx. popl %ecx # Restore ecx. popl %eax # Restore eax. ret SIZE(GNAME(alloc_overflow_edi)) #ifdef LISP_FEATURE_WIN32 /* The guts of the exception-handling system doesn't use * frame pointers, which manages to throw off backtraces * rather badly. So here we grab the (known-good) EBP * and EIP from the exception context and use it to fake * up a stack frame which will skip over the system SEH * code. */ .align align_16byte .globl GNAME(exception_handler_wrapper) TYPE(GNAME(exception_handler_wrapper)) GNAME(exception_handler_wrapper): /* Context layout is: */ /* 7 dwords before FSA. (0x1c) */ /* 8 dwords and 0x50 bytes in the FSA. (0x70/0x8c) */ /* 4 dwords segregs. (0x10/0x9c) */ /* 6 dwords non-stack GPRs. (0x18/0xb4) */ /* EBP (at 0xb4) */ /* EIP (at 0xb8) */ #define CONTEXT_EBP_OFFSET 0xb4 #define CONTEXT_EIP_OFFSET 0xb8 /* some other stuff we don't care about. */ pushl %ebp movl 0x10(%esp), %ebp /* context */ pushl CONTEXT_EIP_OFFSET(%ebp) pushl CONTEXT_EBP_OFFSET(%ebp) movl %esp, %ebp pushl 0x1c(%esp) pushl 0x1c(%esp) pushl 0x1c(%esp) pushl 0x1c(%esp) call GNAME(handle_exception) lea 8(%ebp), %esp popl %ebp ret SIZE(GNAME(exception_handler_wrapper)) #endif #ifdef LISP_FEATURE_DARWIN .align align_16byte .globl GNAME(call_into_lisp_tramp) TYPE(GNAME(call_into_lisp_tramp)) GNAME(call_into_lisp_tramp): /* 1. build the stack frame from the block that's pointed to by ECX 2. free the block 3. set ECX to 0 4. call the function via call_into_lisp */ pushl 0(%ecx) /* return address */ pushl %ebp movl %esp, %ebp pushl 32(%ecx) /* eflags */ pushl 28(%ecx) /* EAX */ pushl 20(%ecx) /* ECX */ pushl 16(%ecx) /* EDX */ pushl 24(%ecx) /* EBX */ pushl $0 /* popal is going to ignore esp */ pushl %ebp /* is this right?? */ pushl 12(%ecx) /* ESI */ pushl 8(%ecx) /* EDI */ pushl $0 /* args for call_into_lisp */ pushl $0 pushl 4(%ecx) /* function to call */ /* free our save block */ pushl %ecx /* reserve sufficient space on stack for args */ pushl %ecx andl $0xfffffff0, %esp /* align stack */ movl $0x40, 4(%esp) movl %ecx, (%esp) call GNAME(os_invalidate) /* call call_into_lisp */ leal -48(%ebp), %esp call GNAME(call_into_lisp) /* Clean up our mess */ leal -36(%ebp), %esp popal popfl leave ret SIZE(call_into_lisp_tramp) #endif .align align_16byte,0x90 .globl GNAME(post_signal_tramp) TYPE(GNAME(post_signal_tramp)) GNAME(post_signal_tramp): /* this is notionally the second half of a function whose first half * doesn't exist. This is where call_into_lisp returns when called * using return_to_lisp_function */ addl $12,%esp /* clear call_into_lisp args from stack */ popal /* restore registers */ popfl #ifdef LISP_FEATURE_DARWIN /* skip two padding words */ addl $8,%esp #endif leave ret SIZE(GNAME(post_signal_tramp)) /* fast_bzero implementations and code to detect which implementation * to use. */ .globl GNAME(fast_bzero_pointer) .data .align align_16byte GNAME(fast_bzero_pointer): /* Variable containing a pointer to the bzero function to use. * Initially points to a basic function. Change this variable * to fast_bzero_detect if OS supports SSE. */ .long GNAME(fast_bzero_base) .text .align align_16byte,0x90 .globl GNAME(fast_bzero) TYPE(GNAME(fast_bzero)) GNAME(fast_bzero): /* Indirect function call */ jmp *GNAME(fast_bzero_pointer) SIZE(GNAME(fast_bzero)) .text .align align_16byte,0x90 .globl GNAME(fast_bzero_detect) TYPE(GNAME(fast_bzero_detect)) GNAME(fast_bzero_detect): /* Decide whether to use SSE, MMX or REP version */ push %eax /* CPUID uses EAX-EDX */ push %ebx push %ecx push %edx mov $1, %eax cpuid test $0x04000000, %edx /* SSE2 needed for MOVNTDQ */ jnz Lsse2 /* Originally there was another case here for using the * MOVNTQ instruction for processors that supported MMX but * not SSE2. This turned out to be a loss especially on * Athlons (where this instruction is apparently microcoded * somewhat slowly). So for simplicity revert to REP STOSL * for all non-SSE2 processors. */ Lbase: movl $(GNAME(fast_bzero_base)), GNAME(fast_bzero_pointer) jmp Lrestore Lsse2: movl $(GNAME(fast_bzero_sse)), GNAME(fast_bzero_pointer) jmp Lrestore Lrestore: pop %edx pop %ecx pop %ebx pop %eax jmp *GNAME(fast_bzero_pointer) SIZE(GNAME(fast_bzero_detect)) .text .align align_16byte,0x90 .globl GNAME(fast_bzero_sse) TYPE(GNAME(fast_bzero_sse)) GNAME(fast_bzero_sse): /* A fast routine for zero-filling blocks of memory that are * guaranteed to start and end at a 4096-byte aligned address. */ push %esi /* Save temporary registers */ push %edi mov 16(%esp), %esi /* Parameter: amount of bytes to fill */ mov 12(%esp), %edi /* Parameter: start address */ shr $6, %esi /* Amount of 64-byte blocks to copy */ jz Lend_sse /* If none, stop */ movups %xmm7, -16(%esp) /* Save XMM register */ xorps %xmm7, %xmm7 /* Zero the XMM register */ jmp Lloop_sse .align align_16byte Lloop_sse: /* Copy the 16 zeroes from xmm7 to memory, 4 times. MOVNTDQ is the * non-caching double-quadword moving variant, i.e. the memory areas * we're touching are not fetched into the L1 cache, since we're just * going to overwrite the memory soon anyway. */ movntdq %xmm7, 0(%edi) movntdq %xmm7, 16(%edi) movntdq %xmm7, 32(%edi) movntdq %xmm7, 48(%edi) add $64, %edi /* Advance pointer */ dec %esi /* Decrement 64-byte block count */ jnz Lloop_sse movups -16(%esp), %xmm7 /* Restore the XMM register */ sfence /* Ensure that weakly ordered writes are flushed. */ Lend_sse: mov 12(%esp), %esi /* Parameter: start address */ prefetcht0 0(%esi) /* Prefetch the start of the block into cache, * since it's likely to be used immediately. */ pop %edi /* Restore temp registers */ pop %esi ret SIZE(GNAME(fast_bzero_sse)) .text .align align_16byte,0x90 .globl GNAME(fast_bzero_base) TYPE(GNAME(fast_bzero_base)) GNAME(fast_bzero_base): /* A fast routine for zero-filling blocks of memory that are * guaranteed to start and end at a 4096-byte aligned address. */ push %eax /* Save temporary registers */ push %ecx push %edi mov 20(%esp), %ecx /* Parameter: amount of bytes to fill */ mov 16(%esp), %edi /* Parameter: start address */ xor %eax, %eax /* Zero EAX */ shr $2, %ecx /* Amount of 4-byte blocks to copy */ jz Lend_base rep stosl /* Store EAX to *EDI, ECX times, incrementing * EDI by 4 after each store */ Lend_base: pop %edi /* Restore temp registers */ pop %ecx pop %eax ret SIZE(GNAME(fast_bzero_base)) /* When LISP_FEATURE_C_STACK_IS_CONTROL_STACK, we cannot safely scrub * the control stack from C, largely due to not knowing where the * active stack frame ends. On such platforms, we reimplement the * core scrubbing logic in assembly, in this case here: */ .text .align align_16byte,0x90 .globl GNAME(arch_scrub_control_stack) TYPE(GNAME(arch_scrub_control_stack)) GNAME(arch_scrub_control_stack): /* We are passed three parameters: * A (struct thread *) at [ESP+4], * the address of the guard page at [ESP+8], and * the address of the hard guard page at [ESP+12]. * We may trash EAX, ECX, and EDX with impunity. * [ESP] is our return address, [ESP-4] is the first * stack slot to scrub. */ /* We start by setting up our scrub pointer in EAX, our * guard page upper bound in ECX, and our hard guard * page upper bound in EDX. */ lea -4(%esp), %eax mov GNAME(os_vm_page_size),%edx mov %edx, %ecx add 8(%esp), %ecx add 12(%esp), %edx /* We need to do a memory operation relative to the * thread pointer, so put it in %ecx and our guard * page upper bound in 4(%esp). */ xchg 4(%esp), %ecx /* Now we begin our main scrub loop. */ ascs_outer_loop: /* If we're about to scrub the hard guard page, exit. */ cmp %edx, %eax jae ascs_check_guard_page cmp 12(%esp), %eax ja ascs_finished ascs_check_guard_page: /* If we're about to scrub the guard page, and the guard * page is protected, exit. */ cmp 4(%esp), %eax jae ascs_clear_loop cmp 8(%esp), %eax jbe ascs_clear_loop cmpl $(NIL), THREAD_CONTROL_STACK_GUARD_PAGE_PROTECTED_OFFSET(%ecx) jne ascs_finished /* Clear memory backwards to the start of the (4KiB) page */ ascs_clear_loop: movl $0, (%eax) test $0xfff, %eax lea -4(%eax), %eax jnz ascs_clear_loop /* If we're about to hit the hard guard page, exit. */ cmp %edx, %eax jae ascs_finished /* If the next (previous?) 4KiB page contains a non-zero * word, continue scrubbing. */ ascs_check_loop: testl $-1, (%eax) jnz ascs_outer_loop test $0xfff, %eax lea -4(%eax), %eax jnz ascs_check_loop ascs_finished: ret SIZE(GNAME(arch_scrub_control_stack)) END()