My main language is C/C++ and so I'm used to seeing function signatures such as the following:

void Func(void);
int Func(void);
void Func(double, int, char*);

which would be written as follows in Visual Basic:

Private Sub Func()
Private Function Func() As Integer
Private Sub Func(ByVal arg1 As Double, ByVal arg2 As Integer, ByVal arg3 As String)

A few months ago, I devised a way in x86_64 assembler and C++ of dynamically allocating a separate stack to use for a single invocation of a function, and in the end I even got it working with exception handling. But it only worked for functions that don't take an argument and don't return a value.

Since then, I've tried to come up with a universal solution that will work for all function signatures (even those with supernumerary parameters, or that return a very big struct by value), and I've got it working for System V x86_64.

Here's what I currently have, the assembler is inline among the C++ code:

https://godbolt.org/z/W4x7vbE7q

And here it is copy-pasted:

#include <cassert> // assert
#include <cstddef> // size_t
#include <memory> // unique_ptr
#include <utility> // forward

thread_local char *p_original, *p_replacement;
thread_local void (*f)(void);
thread_local char *bottom_of_stack;

extern "C" {
void Assembler_set_bottom_of_stack (void) noexcept;
void Assembler_set_stack_pointer_and_invoke(void) noexcept;
}

__asm("Assembler_set_bottom_of_stack: \n"
".intel_syntax noprefix \n"
" mov r10, rsp \n"
" add r10, 16 \n" // +8 return addr, +8 to be safe
" mov QWORD PTR fs:bottom_of_stack@tpoff, r10 \n"
" ret \n"
".att_syntax");

template<typename R, typename... Params>
class Invoker {

Invoker(char *const arg_p, R(*const arg_f)(Params...)) noexcept
{
p_replacement = arg_p; // sets a thread_local variable
f = reinterpret_cast<void (*)(void)>(arg_f); // sets a thread_local variable
}

public:
R operator()(Params... args) // This could be static function but I like operator()
{
Assembler_set_bottom_of_stack();
R (*const funcptr)(Params...) = reinterpret_cast<R(*)(Params...)>(Assembler_set_stack_pointer_and_invoke);
return funcptr( std::forward<Params>(args)... );
}

friend class Stacker;
};

class Stacker {
char *p;
std::unique_ptr<char[]> mystack;

public:

Stacker(std::size_t const len) noexcept(false) // might throw bad_alloc
{
assert( len >= 128u );

mystack.reset( new char[len] );
p = mystack.get() + len - 16u;
}

Stacker(char *const arg, std::size_t const len) noexcept
{
assert( nullptr != arg );
assert( len >= 128u );

p = arg + len - 16u;
}

template<typename R, typename... Params>
Invoker<R,Params...> operator()( R(*const arg)(Params...) )
{
return Invoker<R,Params...>(this->p, arg);
}
};

/* In the following function written in x86_64 assembler using
the System V calling convention, we can only clobber r10
and r11 because all of the other caller-saved registers
must be preserved for the 'jmp' to the target function. */

__asm("Assembler_set_stack_pointer_and_invoke:\n"
".intel_syntax noprefix \n"
// Step 1: Save the original stack pointer
" mov QWORD PTR fs:p_original@tpoff, rsp \n"
// Step 2: Retrieve the replacement stack pointer
" push r15 \n" // save to restore later
" mov r15, rsp \n" // pointer to the r15 we just pushed onto stack
" add r15, 8 \n" // sets 'r15' to top of old stack
" mov r10, QWORD PTR fs:p_replacement@tpoff \n" // sets 'r10' to top of new stack
" mov rax, QWORD PTR fs:bottom_of_stack@tpoff \n" // sets 'rax' to bottom of old stack
// Right now: R15 is the top of the old stack
// R10 is the top of the new stack
// RAX is the bottom of the old stack
// We want to do:
// while ( rax != r15 ) *r10-- = *rax--;
// Step 3: Copy the old stack to the new stack (it might contain supernumerary arguments or a big return struct)
" jmp cond \n" // Jump to condition of 'while' loop
"loop: \n" // ----<----<----<----<----
" mov r11, qword ptr [rax] \n" // |
" mov qword ptr [r10], r11 \n" // ^
" sub r10, 1 \n" // | Loop
" sub rax, 1 \n" // |
"cond: \n" // ^
" cmp rax, r15 \n" // |
" jne loop \n" // ---->---->---->---->----
" pop r15 \n" // restore original value
// Step 4: Change the stack pointer to the new stack ============================================ " mov rsp, r10 \n" // ================================================= new stack
// Step 5: Set the return address to after the 'jmp' instruction
" lea r10, [Label_Jump_Back] \n"
" add rsp, 8 \n" // This line and the next line replace the return address on the stack
" push r10 \n" // This line and the previous line replace the return address on the stack
// Step 5: Invoke the function
" jmp QWORD PTR fs:f@tpoff \n" // --- Invoke the function!
"Label_Jump_Back: \n"
// Note: The label has already been popped off the stack by the callee
// Step 9: Restore the original stack pointer
" mov rsp, QWORD PTR fs:p_original@tpoff \n"
// Step 10: Jump back to the original address
" ret \n"
".att_syntax");

// =================== And now the test code ==============================================
#include <iostream> // cout, endl
using std::cout, std::endl;

struct VeryBigStruct {
double a[3];
int b[3];
double c[3];
int d[3];
double e[3];
int f[3];
};

VeryBigStruct Func2(int a1, int a2, int a3, int a4, int a5, int a6, int a7, int a8, int a9, int a10)
{ VeryBigStruct vbs;
vbs.f[2] = a1+a2+a3+a4+a5+a6+a7+a8+a9+a10;
return vbs;
}

int main(void)
{ cout << "first line in main\n";

cout << "Retval: " << Func2(1,2,3,4,5,6,7,8,9,10).f[2] << endl;

cout << "Retval: " << Stacker(1048576000u)(Func2)(1,2,3,4,5,6,7,8,9,10)..f[2] << endl;

cout << "last line in main\n";
}

On Friday, July 14, 2023 at 1:12:54 PM UTC+1, Frederick Virchanza Gotham wrote:
>
> A few months ago, I devised a way in x86_64 assembler and C++ of dynamically allocating a separate stack to use for a single invocation of a function, and in the end I even got it working with exception handling. But it only worked for functions that don't take an argument and don't return a value..
>
> Since then, I've tried to come up with a universal solution that will work for all function signatures (even those with supernumerary parameters, or that return a very big struct by value), and I've got it working for System V x86_64.

I'm trying to get it working with exception handling, but it's segfaulting. I've tried copying the entire stack but it's still segfaulting inside libgcc inside the function "uw_update_context_1" when it tries to copy the context struct. Anybody know what's wrong with the following?

https://godbolt.org/z/vv7hTGWdr

And here it is copy-pasted:

#include <cassert> // assert
#include <cstddef> // size_t
#include <climits> // ULONG_LONG_MAX
#include <cstdlib> // strtoull
#include <cstring> // strstr
#include <cstdint> // UINTPTR_MAX
#include <memory> // unique_ptr
#include <utility> // forward
#include <exception> // exception_ptr, current_exception
#include <type_traits> // is_rvalue_reference, is_trivially_destructible

#include <iostream> // cout, endl ============================= REMOVE THIS
using std::cerr, std::cout, std::endl;

#include <unistd.h> // lseek, read, close
#include <fcntl.h> // open

char *GetStackBottom(void) noexcept
{ static thread_local int fd = -1;

if ( fd < 0 )
{
fd = ::open("/proc/thread-self/maps", O_RDONLY);

if ( fd < 0 ) return nullptr;

// std::atexit( [](){ ::close(fd); } ); - Won't work
}

if ( 0 != ::lseek(fd, 0, SEEK_SET) ) return nullptr;

static char buf[8192u];
if ( ::read(fd, buf, 8191u) < 32 ) return nullptr;
buf[8191u] = '\0';

char *vm = std::strstr(buf, "[stack]\n");
if ( nullptr == vm ) return nullptr;

while ( '\n' != *vm ) --vm;

++vm;

while ( '-' != *vm ) ++vm;

char *vm2 = ++vm;

while ( ' ' != *vm2 ) ++vm2;

*vm2 = '\0';

static_assert( ULONG_LONG_MAX >= UINTPTR_MAX );
long long unsigned const addr = std::strtoull(vm,nullptr,16u);

return reinterpret_cast<char*>(addr);
}

thread_local char *p_original, *p_replacement;
thread_local void (*f)(void), (*g)(void);
thread_local char *bottom_of_stack;
thread_local std::exception_ptr e;

extern "C" {
void Assembler_set_bottom_of_stack (void) noexcept;
void Assembler_set_stack_pointer_and_invoke(void) noexcept;
}

__asm("Assembler_set_bottom_of_stack: \n"
".intel_syntax noprefix \n"
" mov r10, rsp \n"
" add r10, 16 \n" // +8 return addr, +8 to be safe
" mov QWORD PTR fs:bottom_of_stack@tpoff, r10 \n"
" ret \n"
".att_syntax");

template<typename T> requires std::is_trivially_destructible_v<std::remove_cvref_t<T> >
std::remove_cvref_t<T> dummy_prvalue(void) noexcept
{ typedef std::remove_cvref_t<T> TT;
void (*const tmp)(void) = [](){};
TT (*const funcptr)(void) = reinterpret_cast<TT(*)(void)>(tmp);
return funcptr(); // guaranteed elision of move/copy operations here
}

template<typename R, typename... Params>
class Invoker {

static R exception_capable(Params... args) noexcept
{
cerr << "Entered exception_capable\n";

R (*const funcptr)(Params...) = reinterpret_cast<R(*)(Params...)>(g);

try
{
cerr << "Entered try-block\n";
return funcptr( std::forward<Params>(args)... );
}
catch (...)
{
cerr << "Exception thrown!\n";
e = std::current_exception();
}

return dummy_prvalue<R>();
}

Invoker(char *const arg_p, R(*const arg_f)(Params...)) noexcept
{
p_replacement = arg_p; // sets a thread_local variable
g = reinterpret_cast<void (*)(void)>(arg_f); // sets a thread_local variable
f = reinterpret_cast<void (*)(void)>(exception_capable); // sets a thread_local variable
}

public:
R operator()(Params... args) noexcept(false) // This could be static function but I like operator()
{
//Assembler_set_bottom_of_stack();
//cout << "\nBottom of stack: " << (void*)bottom_of_stack << " (my own assembler)\n";
bottom_of_stack = GetStackBottom() - 8u;
//cout << "Bottom of stack: " << (void*)bottom_of_stack << " (thread-self/maps)\n";
e = nullptr;
R (*const funcptr)(Params...) = reinterpret_cast<R(*)(Params...)>(Assembler_set_stack_pointer_and_invoke);
R retval = funcptr( std::forward<Params>(args)... );
if ( nullptr != e )
{
cerr << "About to rethrow!\n";
std::rethrow_exception(e);
}
return retval;
}

friend class Stacker;
};

class Stacker {
char *p;
std::unique_ptr<char[]> mystack;

public:

Stacker(std::size_t const len) noexcept(false) // might throw bad_alloc
{
assert( len >= 128u );

mystack.reset( new char[len] );
p = mystack.get() + len - 16u;
}

Stacker(char *const arg, std::size_t const len) noexcept
{
assert( nullptr != arg );
assert( len >= 128u );

p = arg + len - 16u;
}

template<typename R, typename... Params>
Invoker<R,Params...> operator()( R(*const arg)(Params...) ) noexcept
{
return Invoker<R,Params...>(this->p, arg);
}
};

/* In the following function written in x86_64 assembler using
the System V calling convention, we can only clobber r10
and r11 because all of the other caller-saved registers
must be preserved for the 'jmp' to the target function. */

__asm("Assembler_set_stack_pointer_and_invoke:\n"
".intel_syntax noprefix \n"
// Step 1: Save the original stack pointer
" mov QWORD PTR fs:p_original@tpoff, rsp \n"
// Step 2: Retrieve the replacement stack pointer
" push r15 \n" // save to restore later
" mov r15, rsp \n" // pointer to the r15 we just pushed onto stack
" add r15, 8 \n" // sets 'r15' to top of old stack
" mov r10, QWORD PTR fs:p_replacement@tpoff \n" // sets 'r10' to top of new stack
" mov rax, QWORD PTR fs:bottom_of_stack@tpoff \n" // sets 'rax' to bottom of old stack
// Right now: R15 is the top of the old stack
// R10 is the top of the new stack
// RAX is the bottom of the old stack
// We want to do:
// while ( rax != r15 ) *r10-- = *rax--;
// Step 3: Copy the old stack to the new stack (it might contain supernumerary arguments or a big return struct)
" jmp cond \n" // Jump to condition of 'while' loop
"loop: \n" // ----<----<----<----<----
" mov r11, qword ptr [rax] \n" // |
" mov qword ptr [r10], r11 \n" // ^
" sub r10, 1 \n" // | Loop
" sub rax, 1 \n" // |
"cond: \n" // ^
" cmp rax, r15 \n" // |
" jne loop \n" // ---->---->---->---->----
" pop r15 \n" // restore original value
// Step 4: Change the stack pointer to the new stack ============================================ " mov rsp, r10 \n" // ================================================= new stack
// Step 5: Set the return address to after the 'jmp' instruction
" lea r10, [Label_Jump_Back] \n"
" add rsp, 8 \n" // This line and the next line replace the return address on the stack
" push r10 \n" // This line and the previous line replace the return address on the stack
// Step 6: Invoke the function
" jmp QWORD PTR fs:f@tpoff \n" // --- Invoke the function!
"Label_Jump_Back: \n"
// Note: The label has already been popped off the stack by the callee
// Step 7: Restore the original stack pointer
" mov rsp, QWORD PTR fs:p_original@tpoff \n"
// Step 8: Jump back to the original address
" ret \n"
".att_syntax");