X-Git-Url: http://repo.macrolet.net/gitweb/?a=blobdiff_plain;f=src%2Fruntime%2Fx86-assem.S;h=d6dd52bc62826aaedf756b03c780df0dd9307619;hb=d1942842cebf682ea4161792b2920eec7ccfe179;hp=9470533fbb24860fae70a5b79dc6df2ec40ad0c5;hpb=f94339428a65d2002f73086bcc6022885f656c1f;p=sbcl.git diff --git a/src/runtime/x86-assem.S b/src/runtime/x86-assem.S index 9470533..d6dd52b 100644 --- a/src/runtime/x86-assem.S +++ b/src/runtime/x86-assem.S @@ -79,7 +79,7 @@ * 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) +#if defined(LISP_FEATURE_UD2_BREAKPOINTS) #define END() #define TRAP ud2 #else @@ -122,15 +122,12 @@ GNAME(call_into_c): fstp %st(0) fstp %st(0) - cld # clear out DF: Darwin, Solaris and Win32 at - # least need this, and it should not hurt others - call *%eax # normal callout using Lisp stack movl %eax,%ecx # remember integer return value /* Check for a return FP value. */ fxam - fnstsw %eax + fnstsw %ax andl $0x4500,%eax cmpl $0x4100,%eax jne Lfp_rtn_value @@ -176,8 +173,7 @@ Lfp_rtn_value: .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 - * the stack changes. We don't worry too much about saving registers +/* We don't worry too much about saving registers * here, because we never expect to return from the initial call to lisp * anyway */ @@ -186,12 +182,9 @@ GNAME(call_into_lisp_first_time): pushl %ebp # Save old frame pointer. movl %esp,%ebp # Establish new frame. #ifndef LISP_FEATURE_WIN32 - movl %esp,ALIEN_STACK + SYMBOL_VALUE_OFFSET movl GNAME(all_threads),%eax - 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),%esp + /* 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 @@ -279,10 +272,10 @@ Ldone: #endif /* Alloc new frame. */ - mov %esp,%ebx # The current sp marks start of new frame. - push %ebp # fp in save location S0 - sub $8,%esp # Ensure 3 slots are allocated, one above. - mov %ebx,%ebp # Switch to 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) @@ -340,11 +333,12 @@ GNAME(fpu_restore): * the undefined-function trampoline */ .text - .align align_4byte,0x90 + .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 @@ -357,7 +351,7 @@ GNAME(undefined_tramp): * the closure trampoline */ .text - .align align_4byte,0x90 + .align align_16byte,0x90 .globl GNAME(closure_tramp) TYPE(GNAME(closure_tramp)) .byte 0, 0, 0, SIMPLE_FUN_HEADER_WIDETAG @@ -373,7 +367,7 @@ GNAME(closure_tramp): SIZE(GNAME(closure_tramp)) .text - .align align_4byte,0x90 + .align align_16byte,0x90 .globl GNAME(funcallable_instance_tramp) TYPE(GNAME(funcallable_instance_tramp)) GNAME(funcallable_instance_tramp): @@ -386,9 +380,15 @@ 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_4byte + .align align_16byte GNAME(fun_end_breakpoint_guts): /* Multiple Value return */ jc multiple_value_return @@ -415,7 +415,7 @@ GNAME(fun_end_breakpoint_end): .globl GNAME(do_pending_interrupt) TYPE(GNAME(do_pending_interrupt)) - .align align_4byte,0x90 + .align align_16byte,0x90 GNAME(do_pending_interrupt): TRAP .byte trap_PendingInterrupt @@ -440,17 +440,15 @@ GNAME(do_pending_interrupt): #define ALLOC(size) \ pushl %ebp; /* Save EBP */ \ movl %esp,%ebp; /* Save ESP to EBP */ \ - andl $0xfffffff0,%esp; /* Align stack */ \ - pushl $0; /* Padding */ \ - pushl size; /* Argument to alloc */ \ - cld; /* Clear DF */ \ + 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 */ \ - cld; /* Clear DF */ \ call GNAME(alloc); \ addl $4,%esp; /* Pop argument */ #endif @@ -458,7 +456,7 @@ GNAME(do_pending_interrupt): #define DEFINE_ALLOC_TO_EAX(name,size) \ .globl GNAME(name); \ TYPE(GNAME(name)); \ - .align align_4byte,0x90; \ + .align align_16byte,0x90; \ GNAME(name): \ pushl %ecx; /* Save ECX and EDX */ \ pushl %edx; \ @@ -471,7 +469,7 @@ GNAME(name): \ #define DEFINE_ALLOC_TO_ECX(name,size) \ .globl GNAME(name); \ TYPE(GNAME(name)); \ - .align align_4byte,0x90; \ + .align align_16byte,0x90; \ GNAME(name): \ pushl %eax; /* Save EAX and EDX */ \ pushl %edx; \ @@ -485,7 +483,7 @@ GNAME(name): \ #define DEFINE_ALLOC_TO_EDX(name,size) \ .globl GNAME(name); \ TYPE(GNAME(name)); \ - .align align_4byte,0x90; \ + .align align_16byte,0x90; \ GNAME(name): \ pushl %eax; /* Save EAX and ECX */ \ pushl %ecx; \ @@ -499,7 +497,7 @@ GNAME(name): \ #define DEFINE_ALLOC_TO_REG(name,reg,size) \ .globl GNAME(name); \ TYPE(GNAME(name)); \ - .align align_4byte,0x90; \ + .align align_16byte,0x90; \ GNAME(name): \ pushl %eax; /* Save EAX, ECX, and EDX */ \ pushl %ecx; \ @@ -554,7 +552,7 @@ DEFINE_ALLOC_TO_REG(alloc_16_to_edi,%edi,$16) /* This routine handles an overflow with eax=crfp+size. So the size=eax-crfp. */ - .align align_4byte + .align align_16byte .globl GNAME(alloc_overflow_eax) TYPE(GNAME(alloc_overflow_eax)) GNAME(alloc_overflow_eax): @@ -566,7 +564,7 @@ GNAME(alloc_overflow_eax): ret SIZE(GNAME(alloc_overflow_eax)) - .align align_4byte + .align align_16byte .globl GNAME(alloc_overflow_ecx) TYPE(GNAME(alloc_overflow_ecx)) GNAME(alloc_overflow_ecx): @@ -579,7 +577,7 @@ GNAME(alloc_overflow_ecx): ret SIZE(GNAME(alloc_overflow_ecx)) - .align align_4byte + .align align_16byte .globl GNAME(alloc_overflow_edx) TYPE(GNAME(alloc_overflow_edx)) GNAME(alloc_overflow_edx): @@ -594,7 +592,7 @@ GNAME(alloc_overflow_edx): /* This routine handles an overflow with ebx=crfp+size. So the size=ebx-crfp. */ - .align align_4byte + .align align_16byte .globl GNAME(alloc_overflow_ebx) TYPE(GNAME(alloc_overflow_ebx)) GNAME(alloc_overflow_ebx): @@ -611,7 +609,7 @@ GNAME(alloc_overflow_ebx): /* This routine handles an overflow with esi=crfp+size. So the size=esi-crfp. */ - .align align_4byte + .align align_16byte .globl GNAME(alloc_overflow_esi) TYPE(GNAME(alloc_overflow_esi)) GNAME(alloc_overflow_esi): @@ -626,7 +624,7 @@ GNAME(alloc_overflow_esi): ret SIZE(GNAME(alloc_overflow_esi)) - .align align_4byte + .align align_16byte .globl GNAME(alloc_overflow_edi) TYPE(GNAME(alloc_overflow_edi)) GNAME(alloc_overflow_edi): @@ -649,7 +647,7 @@ GNAME(alloc_overflow_edi): * 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 + .align align_16byte .globl GNAME(exception_handler_wrapper) TYPE(GNAME(exception_handler_wrapper)) GNAME(exception_handler_wrapper): @@ -680,7 +678,7 @@ GNAME(exception_handler_wrapper): #endif #ifdef LISP_FEATURE_DARWIN - .align align_4byte + .align align_16byte .globl GNAME(call_into_lisp_tramp) TYPE(GNAME(call_into_lisp_tramp)) GNAME(call_into_lisp_tramp): @@ -729,7 +727,7 @@ GNAME(call_into_lisp_tramp): SIZE(call_into_lisp_tramp) #endif - .align align_4byte,0x90 + .align align_16byte,0x90 .globl GNAME(post_signal_tramp) TYPE(GNAME(post_signal_tramp)) GNAME(post_signal_tramp): @@ -754,7 +752,7 @@ GNAME(post_signal_tramp): .globl GNAME(fast_bzero_pointer) .data - .align align_4byte + .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 @@ -762,7 +760,7 @@ GNAME(fast_bzero_pointer): .long GNAME(fast_bzero_base) .text - .align align_8byte,0x90 + .align align_16byte,0x90 .globl GNAME(fast_bzero) TYPE(GNAME(fast_bzero)) GNAME(fast_bzero): @@ -772,7 +770,7 @@ GNAME(fast_bzero): .text - .align align_8byte,0x90 + .align align_16byte,0x90 .globl GNAME(fast_bzero_detect) TYPE(GNAME(fast_bzero_detect)) GNAME(fast_bzero_detect): @@ -810,7 +808,7 @@ Lrestore: .text - .align align_8byte,0x90 + .align align_16byte,0x90 .globl GNAME(fast_bzero_sse) TYPE(GNAME(fast_bzero_sse)) @@ -856,7 +854,7 @@ Lend_sse: .text - .align align_8byte,0x90 + .align align_16byte,0x90 .globl GNAME(fast_bzero_base) TYPE(GNAME(fast_bzero_base)) @@ -872,7 +870,6 @@ GNAME(fast_bzero_base): 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 @@ -884,7 +881,81 @@ Lend_base: 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() - \ No newline at end of file