Flummi/amxmodx
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Merge pull request #24 from alliedmodders/hamsandwich-stkalign16

Browse Source 
Align stack on 16 byte boundary in hamsandwich trampolines.
This commit is contained in:
Scott Ehlert 2014-05-31 17:19:47 -05:00
commit 0c1be7ca7b
1 changed files with 81 additions and 140 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

221
dlls/hamsandwich/Trampolines.h
View File

@ -57,6 +57,7 @@
#include <stdlib.h> // memalign #include <stdlib.h> // memalign
#include <stdio.h> #include <stdio.h>
#include <am-utility.h>
namespace Trampolines namespace Trampolines
{ {
@ -72,28 +73,32 @@ namespace Trampolines
namespace Bytecode namespace Bytecode
{ {
/** /**
* Prologue for a void function * Prologue for a function
* Clobbers EBX and EAX
*/ */
const unsigned char codeVoidPrologue[] = { const unsigned char codePrologue[] = {
0x55, // push ebp 0x55, // push ebp
0x89, 0xE5, // mov ebp, esp 0x89, 0xE5, // mov ebp, esp
0x50, // push eax
}; };
/** /**
* Prologue for a function that returns * Align stack on 16 byte boundary
* Clobbers EBX, EAX too but not after call */
const unsigned char codeAlignStack16[] = {
0x83, 0xE4, 0xF0, // and esp, 0xFFFFFFF0
};
/**
* Allocate stack space (8-bit) by adding to ESP
*/ */
const unsigned char codeReturnPrologue[] = { const unsigned char codeAllocStack[] = {
0x55, // push ebp 0x83, 0xEC, 0xFF, // sub esp, 0xFF
0x89, 0xE5, // mov ebp, esp
};
const unsigned char codeThisReturnPrologue[] = {
0x55, // push ebp
0x89, 0xE5, // mov ebp, esp
}; };
/**
* Offset of codeAllocStack to modify at runtime
* to contain amount of stack space to allocate.
*/
const unsigned int codeAllocStackReplace = 2;
/** /**
* Takes a paramter from the trampoline's stack * Takes a paramter from the trampoline's stack
@ -167,36 +172,19 @@ namespace Trampolines
const unsigned int codeFreeStackReplace = 2; const unsigned int codeFreeStackReplace = 2;
/** /**
* Epilogue of a simple return function * Epilogue of a simple function
*/ */
const unsigned char codeReturnEpilogue[] = { const unsigned char codeEpilogue[] = {
0x89, 0xEC, // mov esp, ebp
0x5D, // pop ebp 0x5D, // pop ebp
0xC3 // ret 0xC3 // ret
}; };
const unsigned char codeReturnEpilogueN[] = { const unsigned char codeEpilogueN[] = {
0x89, 0xEC, // mov esp, ebp
0x5D, // pop ebp 0x5D, // pop ebp
0xC2, 0xCD, 0xAB // retn 0xABCD 0xC2, 0xCD, 0xAB // retn 0xABCD
}; };
const int codeReturnEpilogueNReplace = 2; const int codeEpilogueNReplace = 4;
/**
* Epilogue of a void return function
*/
const unsigned char codeVoidEpilogue[] = {
0x58, // pop eax
0x5D, // pop ebp
0xC3 // ret
};
const unsigned char codeVoidEpilogueN[] = {
0x58, // pop eax
0x5D, // pop ebp
0xC2, 0xCD, 0xAB // retn 0xABCD
};
const int codeVoidEpilogueNReplace = 3;
const unsigned char codeBreakpoint[] = { const unsigned char codeBreakpoint[] = {
0xCC // int 3 0xCC // int 3
@ -272,65 +260,50 @@ namespace Trampolines
}; };
/** /**
* Adds the "return prologue", pushes registers and prepares stack * Adds the prologue, pushes registers, prepares the stack
*/ */
void ReturnPrologue() void Prologue()
{ {
Append(&::Trampolines::Bytecode::codeReturnPrologue[0],sizeof(::Trampolines::Bytecode::codeReturnPrologue)); Append(&::Trampolines::Bytecode::codePrologue[0],sizeof(::Trampolines::Bytecode::codePrologue));
m_paramstart=0;
m_thiscall=0;
};
void ThisReturnPrologue()
{
this->ReturnPrologue();
m_thiscall=1;
};
/**
* Adds the void prologue pushes registers, prepares the stack
*/
void VoidPrologue()
{
Append(&::Trampolines::Bytecode::codeVoidPrologue[0],sizeof(::Trampolines::Bytecode::codeVoidPrologue));
m_paramstart=0; m_paramstart=0;
m_thiscall=0; m_thiscall=0;
}; };
/** /**
* Flags this trampoline as a thiscall trampoline, and prepares the void prologue. * Flags this trampoline as a thiscall trampoline, and prepares the prologue.
*/ */
void ThisVoidPrologue() void ThisPrologue()
{ {
this->VoidPrologue(); this->Prologue();
m_thiscall=1; m_thiscall=1;
}; };
/** /**
* Epilogue for a returning function pops registers but does not free any more of the stack! * Epilogue for a function pops registers but does not free any more of the stack!
*/ */
void ReturnEpilogue() void Epilogue()
{ {
Append(&::Trampolines::Bytecode::codeReturnEpilogue[0],sizeof(::Trampolines::Bytecode::codeReturnEpilogue)); Append(&::Trampolines::Bytecode::codeEpilogue[0],sizeof(::Trampolines::Bytecode::codeEpilogue));
}; };
/** /**
* Epilogue that also frees it's estimated stack usage. Useful for stdcall/thiscall/fastcall. * Epilogue that also frees it's estimated stack usage. Useful for stdcall/thiscall/fastcall.
*/ */
void ReturnEpilogueAndFree() void EpilogueAndFree()
{ {
this->ReturnEpilogue(m_mystack); this->Epilogue(m_mystack);
}; };
/** /**
* Return epilogue. Pops registers, and frees given amount of data from the stack. * Epilogue. Pops registers, and frees given amount of data from the stack.
* *
* @param howmuch How many bytes to free from the stack. * @param howmuch How many bytes to free from the stack.
*/ */
void ReturnEpilogue(int howmuch) void Epilogue(int howmuch)
{ {
unsigned char code[sizeof(::Trampolines::Bytecode::codeEpilogueN)];
unsigned char code[sizeof(::Trampolines::Bytecode::codeReturnEpilogueN)]; memcpy(&code[0],&::Trampolines::Bytecode::codeEpilogueN[0],sizeof(::Trampolines::Bytecode::codeEpilogueN));
memcpy(&code[0],&::Trampolines::Bytecode::codeReturnEpilogueN[0],sizeof(::Trampolines::Bytecode::codeReturnEpilogueN));
unsigned char *c=&code[0]; unsigned char *c=&code[0];
@ -343,58 +316,48 @@ namespace Trampolines
bi.i=howmuch; bi.i=howmuch;
c+=::Trampolines::Bytecode::codeReturnEpilogueNReplace; c+=::Trampolines::Bytecode::codeEpilogueNReplace;
*c++=bi.b[0]; *c++=bi.b[0];
*c++=bi.b[1]; *c++=bi.b[1];
Append(&code[0],sizeof(::Trampolines::Bytecode::codeReturnEpilogueN)); Append(&code[0],sizeof(::Trampolines::Bytecode::codeEpilogueN));
//Append(&::Trampolines::Bytecode::codeReturnEpilogueN[0],sizeof(::Trampolines::Bytecode::codeReturnEpilogueN));
}; };
/** /**
* Void epilogue, pops registers and frees the estimated stack usage of the trampoline. * Aligns stack on 16 byte boundary for functions that use aligned SSE instructions.
* This also allocates extra stack space to allow the specified number of slots to be used
* for function paramaters that will be pushed onto the stack.
*/ */
void VoidEpilogueAndFree() void AlignStack16(int slots)
{ {
this->VoidEpilogue(m_mystack); const size_t stackNeeded = slots * sizeof(void *);
}; const size_t stackReserve = ke::Align(stackNeeded, 16);
/** const size_t stackExtra = stackReserve - stackNeeded;
* Void epilogue, pops registers, nothing else done with stack.
*/ // Stack space should fit in a byte
void VoidEpilogue() assert(stackExtra <= 0xFF);
const size_t codeAlignStackSize = sizeof(::Trampolines::Bytecode::codeAlignStack16);
const size_t codeAllocStackSize = sizeof(::Trampolines::Bytecode::codeAllocStack);
unsigned char code[codeAlignStackSize + codeAllocStackSize];
memcpy(&code[0], &::Trampolines::Bytecode::codeAlignStack16[0], codeAlignStackSize);
if (stackExtra > 0)
{ {
Append(&::Trampolines::Bytecode::codeVoidEpilogue[0],sizeof(::Trampolines::Bytecode::codeVoidEpilogue)); unsigned char *c = &code[codeAlignStackSize];
}; memcpy(c, &::Trampolines::Bytecode::codeAllocStack[0], codeAllocStackSize);
/**
* Void epilogue, pops registers, frees given amount of data off of the stack. c += ::Trampolines::Bytecode::codeAllocStackReplace;
* *c = (unsigned char)stackExtra;
* @param howmuch How many bytes to free from the stack.
*/ Append(&code[0], codeAlignStackSize + codeAllocStackSize);
void VoidEpilogue(int howmuch) }
else
{ {
Append(&code[0], codeAlignStackSize);
unsigned char code[sizeof(::Trampolines::Bytecode::codeVoidEpilogueN)]; }
}
memcpy(&code[0],&::Trampolines::Bytecode::codeVoidEpilogueN[0],sizeof(::Trampolines::Bytecode::codeVoidEpilogueN));
unsigned char *c=&code[0];
union
{
int i;
unsigned char b[4];
} bi;
bi.i=howmuch;
c+=::Trampolines::Bytecode::codeVoidEpilogueNReplace;
*c++=bi.b[0];
*c++=bi.b[1];
Append(&code[0],sizeof(::Trampolines::Bytecode::codeVoidEpilogueN));
Append(&::Trampolines::Bytecode::codeVoidEpilogueN[0],sizeof(::Trampolines::Bytecode::codeVoidEpilogueN));
};
/** /**
* Pushes the "this" pointer onto the callee stack. Pushes ECX for MSVC, and param0 on GCC. * Pushes the "this" pointer onto the callee stack. Pushes ECX for MSVC, and param0 on GCC.
@ -645,27 +608,15 @@ inline void *CreateGenericTrampoline(bool thiscall, bool voidcall, bool retbuf,
{ {
Trampolines::TrampolineMaker tramp; Trampolines::TrampolineMaker tramp;
if (voidcall)
{
if (thiscall) if (thiscall)
{ {
tramp.ThisVoidPrologue(); tramp.ThisPrologue();
tramp.AlignStack16(paramcount + 2); // Param count + this ptr + extra ptr
} }
else else
{ {
tramp.VoidPrologue(); tramp.Prologue();
} tramp.AlignStack16(paramcount + 1); // Param count + extra ptr
}
else
{
if (thiscall)
{
tramp.ThisReturnPrologue();
}
else
{
tramp.ReturnPrologue();
}
} }
while (paramcount) while (paramcount)
@ -679,31 +630,21 @@ inline void *CreateGenericTrampoline(bool thiscall, bool voidcall, bool retbuf,
tramp.PushNum(reinterpret_cast<int>(extraptr)); tramp.PushNum(reinterpret_cast<int>(extraptr));
tramp.Call(callee); tramp.Call(callee);
tramp.FreeTargetStack(); tramp.FreeTargetStack();
if (voidcall)
{
#if defined(_WIN32) #if defined(_WIN32)
tramp.VoidEpilogueAndFree(); tramp.EpilogueAndFree();
#elif defined(__linux__) || defined(__APPLE__) #elif defined(__linux__) || defined(__APPLE__)
if (retbuf) if (retbuf)
{ {
tramp.VoidEpilogue(4); tramp.Epilogue(4);
} }
else else
{ {
tramp.VoidEpilogue(); tramp.Epilogue();
} }
#endif #endif
}
else
{
#if defined(_WIN32)
tramp.ReturnEpilogueAndFree();
#elif defined(__linux__) || defined(__APPLE__)
tramp.ReturnEpilogue();
#endif
}
return tramp.Finish(NULL);
return tramp.Finish(NULL);
}; };