patchDotGNU Portable.NET - Patches: patch #7620, Added support for the windows...

 
 

You are not allowed to post comments on this tracker with your current authentication level.

patch #7620: Added support for the windows stack guard page in libjit.

Submitter:  Patrick van Beem <pvbeem>
Submitted:  Wed 21 Sep 2011 05:12:51 AM UTC
   
 
Category:  None Priority:  5 - Normal
Status:  None Privacy:  Public
Assigned to:  None Open/Closed:  Open

Wed 21 Sep 2011 05:12:51 AM UTC, original submission:  

In windows, the allocation of physical memory for the stack is controlled by a guard page on the last not allocated page of the stack. Functions that use more then 4K of stack, should access every page of the stack they need in the correct order, to ensure the allocation of physical memory for the complete stack they use. Failing to do so, results in an access violation exception. Libjit did not do this. The changes below implement this. See also:
http://blogs.technet.com/b/markrussinovich/archive/2009/07/08/3261309.aspx

The fixes are below. I couldn't make a diff, because my previous patches are already applied in this code, but not in the main trunk...




Fix for 32 bit (jit-rules-x86.c), chages are on and after the #ifdef WIN32 block:


void *_jit_gen_prolog(jit_gencode_t gen, jit_function_t func, void *buf)
{
        unsigned char prolog[JIT_PROLOG_SIZE];
        unsigned char *inst = prolog;
        int reg;

        /* Push ebp onto the stack */
        x86_push_reg(inst, X86_EBP);

        /* Initialize EBP for the current frame */
        x86_mov_reg_reg(inst, X86_EBP, X86_ESP, sizeof(void *));

        /* Allocate space for the local variable frame */
        if(func->builder->frame_size > 0)
        {
#ifdef WIN32
                /* For windows: make sure we hit the guard page on the stack. Call the routine below.
                 * I coudn't get the compiler to omit the prologue on a function, so I created the
                 * function in-line in this function. This causes problems in unrolling the stack when
                 * an exceptions occurs in this code. However, it's very unlikely that will happen.
                 */
                if ( func->builder->frame_size >= 4096 )
                {
                        void* entrypoint;
                        __asm__ (
                                "jmp    .L3\n"
                                ".chkstk:\n\t"
                                "push    %%ecx\n\t"
                                "mov     %%esp,%%ecx\n\t"
                                "add     $8,%%ecx\n"
                                ".L1:\t"
                                "cmp     $0x1000,%%eax\n\t"
                                "jb      .L2\n\t"
                                "sub     $0x1000,%%ecx\n\t"
                                "or      $0,(%%ecx)\n\t"
                                "sub     $0x1000,%%eax\n\t"
                                "jmp     .L1\n"
                                ".L2:\t"
                                "sub     %%eax,%%ecx\n\t"
                                "or      $0,(%%ecx)\n\t"
                                "mov     %%esp,%%eax\n\t"
                                "mov     %%ecx,%%esp\n\t"
                                "mov     (%%eax),%%ecx\n\t"
                                "mov     4(%%eax),%%eax\n\t"
                                "jmp     %%eax\n\t"
                                ".L3:\t"
                                "lea     .chkstk,%0\n\t"
                                : "=r"(entrypoint)
                        );
                        x86_mov_reg_imm(inst, X86_EAX, func->builder->frame_size);
                        /* We generate in a temporary buffer, so relative calls can't be used (as is). Use long calls.
                         * ECX is caller-saved, so we can use that one here.
                         */
                        x86_mov_reg_imm(inst, X86_ECX, entrypoint);
                        x86_call_reg(inst, X86_ECX);
                }
                else
#endif
                {
                        x86_alu_reg_imm(inst, X86_SUB, X86_ESP, (int)(func->builder->frame_size));
                }
        }

        /* Save registers that we need to preserve */
        for(reg = 0; reg <= 7; ++reg)
        {
                if(jit_reg_is_used(gen->touched, reg) &&
                   (_jit_reg_info[reg].flags & JIT_REG_CALL_USED) == 0)
                {
                        x86_push_reg(inst, _jit_reg_info[reg].cpu_reg);
                }
        }

        /* Copy the prolog into place and return the adjusted entry position */
        reg = (int)(inst - prolog);
        jit_memcpy(((unsigned char *)buf) + JIT_PROLOG_SIZE - reg, prolog, reg);
        return (void *)(((unsigned char *)buf) + JIT_PROLOG_SIZE - reg);
}





Fix for 64 bit (jit-rules-x86-64.c), changes are in the block with the inline assembler within the #ifdef WIN64:


void *
_jit_gen_prolog(jit_gencode_t gen, jit_function_t func, void *buf)
{
        unsigned char prolog[JIT_PROLOG_SIZE];
        unsigned char *inst = prolog;
        int reg;
        int frame_size = 0;
        int regs_to_save = _jit_count_regs_saved(gen);

#ifdef WIN64
        /* Bail out if there is insufficient space for the prolog unwind information. */
        #define UNWIND_CODE_COUNT 8
        _jit_cache_check_space(&gen->posn, sizeof(RUNTIME_FUNCTION) + sizeof(UNWIND_INFO) + sizeof(UNWIND_CODE) * (UNWIND_CODE_COUNT - 1) + 4 );

        PRUNTIME_FUNCTION win64_runtime_function;
        PUNWIND_INFO win64_unwind_info;
        int unwindslots = 0;

        /* Create the WIN64 unwind information */
        func->win64_runtime_function_base = buf;
        win64_runtime_function = (PRUNTIME_FUNCTION)(((jit_nuint)gen->posn.ptr + 3) & ~(jit_nuint)3);
        func->win64_runtime_function = win64_runtime_function;
        win64_runtime_function->EndAddress = (ULONG)((void*)gen->posn.ptr - buf);
        gen->posn.ptr = (unsigned char*)win64_runtime_function + sizeof(RUNTIME_FUNCTION);
        win64_unwind_info = (PUNWIND_INFO)(gen->posn.ptr + sizeof(UNWIND_CODE) * UNWIND_CODE_COUNT);
        gen->posn.ptr += sizeof(UNWIND_INFO) + sizeof(UNWIND_CODE) * (UNWIND_CODE_COUNT - 1);
#endif

        /* Push ebp onto the stack */
        x86_64_push_reg_size(inst, X86_64_RBP, 8);

#ifdef WIN64
        /* Register this operation code in the prolog unwind table. */
        unwindslots++;
        win64_unwind_info = (void*)win64_unwind_info - sizeof(UNWIND_CODE);
        win64_unwind_info->UnwindCode[0].CodeOffset = (ULONG)(inst - prolog);
        win64_unwind_info->UnwindCode[0].UnwindOp = UWOP_PUSH_NONVOL;
        win64_unwind_info->UnwindCode[0].OpInfo = 5;
#endif

        /* Initialize EBP for the current frame */
        x86_64_mov_reg_reg_size(inst, X86_64_RBP, X86_64_RSP, 8);

        /* Allocate space for the local variable frame */
        if(func->builder->frame_size > 0)
        {
                /* Make sure that the framesize is a multiple of 8 bytes */
                frame_size = (func->builder->frame_size + 0x7) & ~0x7;
        }

        /* add the register save area to the initial frame size */
        frame_size += (regs_to_save << 3);

#ifdef JIT_USE_PARAM_AREA
        /* Add the param area to the frame_size if the additional offset
           doesnt cause the offsets in the register saves become 4 bytes */
        if(func->builder->param_area_size > 0 &&
           (func->builder->param_area_size <= 0x50 || regs_to_save == 0))
        {
                frame_size += func->builder->param_area_size;
        }
#endif /* JIT_USE_PARAM_AREA */

        /* Make sure that the framesize is a multiple of 16 bytes */
        /* so that the final RSP will be alligned on a 16byte boundary. */
        frame_size = (frame_size + 0xf) & ~0xf;

        if(frame_size > 0)
        {
#ifdef WIN64
                /* For windows: make sure we hit the guard page on the stack. Call the routine below.
                 * I coudn't get the compiler to omit the prologue on a function, so I created the
                 * function in-line in this function. This causes problems in unrolling the stack when
                 * an exceptions occurs in this code. However, it's very unlikely that will happen.
                 */
                if ( frame_size >= 4096 )
                {
                        void* entrypoint;
                        __asm__ (
                                "jmp    .L3\n"
                                ".chkstk:\n\t"
                                "pushq   %%rcx\n\t"
                                "pushq   %%rax\n\t"
                                "cmpq    $0x1000, %%rax\n\t"
                                "leaq    0x18(%%rsp), %%rcx\n\t"
                                "jb      .L2\n"
                                ".L1:\t"
                                "subq    $0x1000,%%rcx\n\t"
                                "orq     $0,(%%rcx)\n\t"
                                "subq    $0x1000,%%rax\n\t"
                                "cmpq    $0x1000,%%rax\n\t"
                                "ja      .L1\n"
                                ".L2:\t"
                                "subq    %%rax, %%rcx\n\t"
                                "orq     $0,(%%rcx)\n\t"
                                "popq    %%rax\n\t"
                                "popq    %%rcx\n\t"
                                "ret\n"
                                ".L3:\t"
                                "leaq    .chkstk,%0\n\t"
                                : "=r"(entrypoint)
                        );
                        x86_64_mov_reg_imm_size(inst, X86_64_RAX, frame_size, 8);
                        /* We generate in a temporary buffer, so relative calls can't be used (as is). Use long calls. */
                        x86_64_mov_reg_imm_size(inst, X86_64_SCRATCH, (jit_nint)entrypoint, 8);
                        x86_64_call_reg(inst, X86_64_SCRATCH);
                }
#endif
                x86_64_sub_reg_imm_size(inst, X86_64_RSP, frame_size, 8);
#ifdef WIN64
                /* Register this operation code in the prolog unwind table. */
                if ( frame_size <= 128 )
                {
                        unwindslots++;
                        win64_unwind_info = (void*)win64_unwind_info - sizeof(UNWIND_CODE);
                        win64_unwind_info->UnwindCode[0].CodeOffset = (ULONG)(inst - prolog);
                        win64_unwind_info->UnwindCode[0].UnwindOp = UWOP_ALLOC_SMALL;
                        win64_unwind_info->UnwindCode[0].OpInfo = (frame_size / 8) - 1;
                }
                else
                {
                        unwindslots++;
                        win64_unwind_info = (void*)win64_unwind_info - sizeof(UNWIND_CODE);
                        win64_unwind_info->UnwindCode[0].FrameOffset = frame_size / 8;
                        unwindslots++;
                        win64_unwind_info = (void*)win64_unwind_info - sizeof(UNWIND_CODE);
                        win64_unwind_info->UnwindCode[0].CodeOffset = (ULONG)(inst - prolog);
                        win64_unwind_info->UnwindCode[0].UnwindOp = UWOP_ALLOC_LARGE;
                        win64_unwind_info->UnwindCode[0].OpInfo = 0;
                }
#endif
        }

        if(regs_to_save > 0)
        {
                int current_offset;
#ifdef JIT_USE_PARAM_AREA
                if(func->builder->param_area_size > 0 &&
                   func->builder->param_area_size <= 0x50)
                {
                        current_offset = func->builder->param_area_size;
                }
                else
#endif /* JIT_USE_PARAM_AREA */
                {
                        current_offset = 0;
                }

                /* Save registers that we need to preserve */
                for(reg = 0; reg <= 14; ++reg)
                {
                        if(jit_reg_is_used(gen->touched, reg) &&
                           (_jit_reg_info[reg].flags & JIT_REG_CALL_USED) == 0)
                        {
                                x86_64_mov_membase_reg_size(inst, X86_64_RSP, current_offset,
                                                                                        _jit_reg_info[reg].cpu_reg, 8);
                                current_offset += 8;
                        }
                }
        }
#ifdef JIT_USE_PARAM_AREA
        if(func->builder->param_area_size > 0x50 && regs_to_save > 0)
        {
                x86_64_sub_reg_imm_size(inst, X86_64_RSP, func->builder->param_area_size, 8);
#ifdef WIN64
                /* Register this operation code in the prolog unwind table. */
                unwindslots++;
                win64_unwind_info = (void*)win64_unwind_info - sizeof(UNWIND_CODE);
                win64_unwind_info->UnwindCode[0].FrameOffset = func->builder->param_area_size / 8;
                unwindslots++;
                win64_unwind_info = (void*)win64_unwind_info - sizeof(UNWIND_CODE);
                win64_unwind_info->UnwindCode[0].CodeOffset = (ULONG)(inst - prolog);
                win64_unwind_info->UnwindCode[0].UnwindOp = UWOP_ALLOC_LARGE;
                win64_unwind_info->UnwindCode[0].OpInfo = 0;
#endif // WIN64
        }
#endif /* JIT_USE_PARAM_AREA */

        /* Copy the prolog into place and return the adjusted entry position */
        reg = (int)(inst - prolog);
        jit_memcpy(((unsigned char *)buf) + JIT_PROLOG_SIZE - reg, prolog, reg);

#ifdef WIN64
        /* Register stack unwinding information that is ready to be used by RtlAddFunctionTable */
        win64_runtime_function->BeginAddress = JIT_PROLOG_SIZE - reg;
        win64_unwind_info->Version = 1;
        win64_unwind_info->Flags = 0;
        win64_unwind_info->CountOfCodes = unwindslots;
        win64_unwind_info->FrameRegister = 0;
        win64_unwind_info->FrameOffset = 0;
        win64_unwind_info->SizeOfProlog = reg;
        win64_runtime_function->UnwindData = (ULONG)((void*)win64_unwind_info - buf);
#endif

        return (void *)(((unsigned char *)buf) + JIT_PROLOG_SIZE - reg);
}

void
_jit_gen_epilog(jit_gencode_t gen, jit_function_t func)
{
        unsigned char *inst;
        int reg;
        int current_offset;
        jit_int *fixup;
        jit_int *next;

        /* Bail out if there is insufficient space for the epilog */
        _jit_cache_check_space(&gen->posn, 256); // Was 48, 76 needed for FAA, 256 to be safe...

        inst = gen->posn.ptr;

        /* Perform fixups on any blocks that jump to the epilog */
        fixup = (jit_int *)(gen->epilog_fixup);
        while(fixup != 0)
        {
                if(DEBUG_FIXUPS)
                {
                        fprintf(stderr, "Fixup Address: %lx, Value: %x\n",
                                        (jit_nint)fixup, fixup[0]);
                }
                next = (jit_int *)_JIT_CALC_NEXT_FIXUP(fixup, fixup[0]);
                fixup[0] = (jit_int)(((jit_nint)inst) - ((jit_nint)fixup) - 4);
                fixup = next;
        }
        gen->epilog_fixup = 0;

        /* Perform fixups on any alloca calls */
        fixup = (jit_int *)(gen->alloca_fixup);
        while (fixup != 0)
        {
                next = (jit_int *)_JIT_CALC_NEXT_FIXUP(fixup, fixup[0]);
                fixup[0] = func->builder->param_area_size;
                if(DEBUG_FIXUPS)
                {
                        fprintf(stderr, "Fixup Param Area Size: %lx, Value: %x\n",
                                        (jit_nint)fixup, fixup[0]);
                }
                fixup = next;
        }
        gen->alloca_fixup = 0;

        /* Restore the used callee saved registers */
        if(gen->stack_changed)
        {
                int frame_size = func->builder->frame_size;
                int regs_saved = _jit_count_regs_saved(gen);

                /* add the register save area to the initial frame size */
                frame_size += (regs_saved << 3);

                /* Make sure that the framesize is a multiple of 16 bytes */
                /* so that the final RSP will be alligned on a 16byte boundary. */
                frame_size = (frame_size + 0xf) & ~0xf;

                current_offset = -frame_size;

                for(reg = 0; reg <= 14; ++reg)
                {
                        if(jit_reg_is_used(gen->touched, reg) &&
                           (_jit_reg_info[reg].flags & JIT_REG_CALL_USED) == 0)
                        {
                                x86_64_mov_reg_membase_size(inst, _jit_reg_info[reg].cpu_reg,
                                                            X86_64_RBP, current_offset, 8);
                                current_offset += 8;
                        }
                }
        }
        else
        {
#ifdef JIT_USE_PARAM_AREA
                if(func->builder->param_area_size > 0)
                {
                        current_offset = func->builder->param_area_size;
                }
                else
                {
                        current_offset = 0;
                }
#else /* !JIT_USE_PARAM_AREA */
                current_offset = 0;
#endif /* !JIT_USE_PARAM_AREA */
                for(reg = 0; reg <= 14; ++reg)
                {
                        if(jit_reg_is_used(gen->touched, reg) &&
                           (_jit_reg_info[reg].flags & JIT_REG_CALL_USED) == 0)
                        {
                                x86_64_mov_reg_membase_size(inst, _jit_reg_info[reg].cpu_reg,
                                                            X86_64_RSP, current_offset, 8);
                                current_offset += 8;
                        }
                }
        }

        /* Restore stackpointer and frame register */
        x86_64_mov_reg_reg_size(inst, X86_64_RSP, X86_64_RBP, 8);
        x86_64_pop_reg_size(inst, X86_64_RBP, 8);

        /* and return */
        x86_64_ret(inst);

        gen->posn.ptr = inst;
}


Patrick van Beem <pvbeem>

 

(Note: upload size limit is set to 16384 kB, after insertion of the required escape characters.)

No files currently attached

 

Depends on the following items: None found

Items that depend on this one: None found

 

Carbon-Copy List
  • -email is unavailable- added by pvbeem (Submitted the item)
  •  

    There are 0 votes so far. Votes easily highlight which items people would like to see resolved in priority, independently of the priority of the item set by tracker managers.

     

    No changes have been made to this item

    Back to the top

    Powered by Savane 3.14-f13d.
    Corresponding source code