#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"
* that are defined to be no-ops on win32. Hopefully this still works on
* other platforms.
*/
-#ifndef LISP_FEATURE_WIN32
+#if !defined(LISP_FEATURE_WIN32) && !defined(LISP_FEATURE_DARWIN)
#define TYPE(name) .type name,@function
#define SIZE(name) .size name,.-name
#else
#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_DARWIN)
+#define END()
+#define TRAP ud2
+#else
+#define END() .end
+#define TRAP int3
+#endif
+
.text
- .global GNAME(foreign_function_call_active)
- .global GNAME(all_threads)
+ .globl GNAME(all_threads)
\f
/*
* A call to call_into_c preserves esi, edi, and ebp.
* 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
- .global GNAME(call_into_c)
+ .globl GNAME(call_into_c)
TYPE(GNAME(call_into_c))
GNAME(call_into_c):
- movl $1,GNAME(foreign_function_call_active)
-
/* Save the return Lisp address in ebx. */
popl %ebx
cld
#endif
+#ifdef LISP_FEATURE_DARWIN
+ andl $0xfffffff0,%esp # align stack to 16-byte boundary before calling C
+#endif
call *%eax # normal callout using Lisp stack
movl %eax,%ecx # remember integer return value
/* Restore the return value. */
movl %ecx,%eax # maybe return value
- movl $0,GNAME(foreign_function_call_active)
/* Return. */
jmp *%ebx
/* We don't need to restore eax, because the result is in st(0). */
- movl $0,GNAME(foreign_function_call_active)
/* Return. */
jmp *%ebx
\f
.text
- .global GNAME(call_into_lisp_first_time)
+ .globl GNAME(call_into_lisp_first_time)
TYPE(GNAME(call_into_lisp_first_time))
/* The *ALIEN-STACK* pointer is set up on the first call_into_lisp when
movl THREAD_CONTROL_STACK_START_OFFSET(%eax) ,%esp
/* don't think too hard about what happens if we get interrupted
* here */
- addl $THREAD_CONTROL_STACK_SIZE-4,%esp
+ addl $(THREAD_CONTROL_STACK_SIZE),%esp
#else
/* Win32 -really- doesn't like you switching stacks out from under it. */
movl GNAME(all_threads),%eax
jmp Lstack
\f
.text
- .global GNAME(call_into_lisp)
+ .globl GNAME(call_into_lisp)
TYPE(GNAME(call_into_lisp))
/* The C conventions require that ebx, esi, edi, and ebp be preserved
xorl %esi,%esi # third arg
/* no longer in function call */
- movl %eax, GNAME(foreign_function_call_active)
-
movl %esp,%ebx # remember current stack
pushl %ebx # Save entry stack on (maybe) new stack.
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. */
mov %esp,%ebx # The current sp marks start of new frame.
push %ebp # fp in save location S0
/* 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
\f
/* support for saving and restoring the NPX state from C */
.text
- .global GNAME(fpu_save)
+ .globl GNAME(fpu_save)
TYPE(GNAME(fpu_save))
.align 2,0x90
GNAME(fpu_save):
ret
SIZE(GNAME(fpu_save))
- .global GNAME(fpu_restore)
+ .globl GNAME(fpu_restore)
TYPE(GNAME(fpu_restore))
.align 2,0x90
GNAME(fpu_restore):
*/
.text
.align align_4byte,0x90
- .global GNAME(undefined_tramp)
+ .globl GNAME(undefined_tramp)
TYPE(GNAME(undefined_tramp))
.byte 0, 0, 0, SIMPLE_FUN_HEADER_WIDETAG
GNAME(undefined_tramp):
- int3
+ TRAP
.byte trap_Error
.byte 2
.byte UNDEFINED_FUN_ERROR
*/
.text
.align align_4byte,0x90
- .global GNAME(closure_tramp)
+ .globl GNAME(closure_tramp)
TYPE(GNAME(closure_tramp))
.byte 0, 0, 0, SIMPLE_FUN_HEADER_WIDETAG
GNAME(closure_tramp):
jmp *CLOSURE_FUN_OFFSET(%eax)
SIZE(GNAME(closure_tramp))
+ .text
+ .align align_4byte,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
*/
.text
- .global GNAME(fun_end_breakpoint_guts)
+ .globl GNAME(fun_end_breakpoint_guts)
.align align_4byte
GNAME(fun_end_breakpoint_guts):
/* Multiple Value return */
- jmp 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
+ movl $(NIL),%edi # default second value
+ movl $(NIL),%esi # default third value
multiple_value_return:
- .global GNAME(fun_end_breakpoint_trap)
+ .globl GNAME(fun_end_breakpoint_trap)
GNAME(fun_end_breakpoint_trap):
- int3
+ TRAP
.byte trap_FunEndBreakpoint
hlt # We should never return here.
- .global GNAME(fun_end_breakpoint_end)
+ .globl GNAME(fun_end_breakpoint_end)
GNAME(fun_end_breakpoint_end):
\f
- .global GNAME(do_pending_interrupt)
+ .globl GNAME(do_pending_interrupt)
TYPE(GNAME(do_pending_interrupt))
.align align_4byte,0x90
GNAME(do_pending_interrupt):
- int3
+ TRAP
.byte trap_PendingInterrupt
ret
SIZE(GNAME(do_pending_interrupt))
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_4byte
+ .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_4byte
+ .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_4byte,0x90
.globl GNAME(post_signal_tramp)
TYPE(GNAME(post_signal_tramp))
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))
-#ifdef LISP_FEATURE_WIN32
- /*
- * This is part of the funky magic for exception handling on win32.
- * see sigtrap_emulator() in win32-os.c for details.
- */
- .global GNAME(sigtrap_trampoline)
-GNAME(sigtrap_trampoline):
- pushl %eax
- pushl %ebp
- movl %esp, %ebp
- call GNAME(sigtrap_wrapper)
- pop %eax
- pop %eax
- int3
- .byte trap_ContextRestore
- hlt # We should never return here.
- /*
- * This is part of the funky magic for exception handling on win32.
- * see handle_exception() in win32-os.c for details.
- */
- .global GNAME(exception_trampoline)
-GNAME(exception_trampoline):
- pushl %eax
- pushl %ebp
- movl %esp, %ebp
- call GNAME(handle_win32_exception_wrapper)
- pop %eax
- pop %eax
- int3
- .byte trap_ContextRestore
- hlt # We should never return here.
-#endif
-
- .end
+ /* fast_bzero implementations and code to detect which implementation
+ * to use.
+ */
+\f
+ .globl GNAME(fast_bzero_pointer)
+ .data
+ .align align_4byte
+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)
+\f
+ .text
+ .align align_8byte,0x90
+ .globl GNAME(fast_bzero)
+ TYPE(GNAME(fast_bzero))
+GNAME(fast_bzero):
+ /* Indirect function call */
+ jmp *GNAME(fast_bzero_pointer)
+ SIZE(GNAME(fast_bzero))
+
+\f
+ .text
+ .align align_8byte,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))
+
+\f
+ .text
+ .align align_8byte,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))
+
+\f
+ .text
+ .align align_8byte,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
+ cld /* Set direction of STOSL to increment */
+
+ 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))
+
+\f
+ END()
+
\ No newline at end of file
projects / sbcl.git / blobdiff