Annotated x64 Disassembly
The following very simple function illustrates the x64 calling convention.
int Simple(int i, int j)
{
return i*5 + j + 3;
}
This compiles to code like this:
01001080 lea eax,[rdx+rcx*4] ; eax = rdx+rcx*4
01001083 lea eax,[rcx+rax+0x3] ; eax = rcx+rax+3
01001087 ret
The i and j parameters are passed in the ecx and edx registers, respectively. Since there are only two parameters, the routine does not use the stack at all.
The particular code generated exploits three tricks, one of which is specific to the x64:
The lea operation can be used to perform a series of simple arithmetic operations as a single operation. The first instruction stores j+i*4 in eax, and the second instruction adds i+3 to the result, for a total of j+i*5+3.
Many operations, such as addition and multiplication, can be done with extra precision, and then truncated to the correct precision. In this instance, the code uses 64-bit addition and multiplication. We can safely truncate the result to 32 bits.
On the x64, any operation that outputs to a 32-bit register automatically zero-extends the result. In this case, outputting to eax has the effect of truncating the result to 32 bits.
Return values are passed in the rax register. In this case, the result is already in the rax register, so the function returns.
Next we consider a more complicated function to demonstrate typical x64 disassembly:
HRESULT Meaningless(IDispatch *pdisp, DISPID dispid, BOOL fUnique, LPCWSTR pszExe)
{
IQueryAssociations *pqa;
HRESULT hr = AssocCreate(CLSID_QueryAssociations, IID_IQueryAssociations, (void**)&pqa);
if (SUCCEEDED(hr)) {
hr = pqa->Init(ASSOCF_INIT_BYEXENAME, pszExe, NULL, NULL);
if (SUCCEEDED(hr)) {
WCHAR wszName[MAX_PATH];
DWORD cchName = MAX_PATH;
hr = pqa->GetString(0, ASSOCSTR_FRIENDLYAPPNAME, NULL, wszName, &cchName);
if (SUCCEEDED(hr)) {
VARIANTARG rgvarg[2] = { 0 };
V_VT(&rgvarg[0]) = VT_BSTR;
V_BSTR(&rgvarg[0]) = SysAllocString(wszName);
if (V_BSTR(&rgvarg[0])) {
DISPPARAMS dp;
LONG lUnique = InterlockedIncrement(&lCounter);
V_VT(&rgvarg[1]) = VT_I4;
V_I4(&rgvarg[1]) = fUnique ? lUnique : 0;
dp.rgvarg = rgvarg;
dp.cArgs = 2;
dp.rgdispidNamedArgs = NULL;
dp.cNamedArgs = 0;
hr = pdisp->Invoke(dispid, IID_NULL, 0, DISPATCH_METHOD, &dp, NULL, NULL, NULL);
VariantClear(&rgvarg[0]);
VariantClear(&rgvarg[1]);
} else {
hr = E_OUTOFMEMORY;
}
}
}
pqa->Release();
}
return hr;
}
We'll go through this function and the equivalent assembly line by line.
When entered, the function's parameters are stored as follows:
rcx = pdisp.
rdx = dispid.
r8 = fUnique.
r9 = pszExe.
Recall that the first four parameters are passed in registers. Since this function has only four parameters, none are passed on the stack.
The assembly begins as follows:
Meaningless:
010010e0 push rbx ; save
010010e1 push rsi ; save
010010e2 push rdi ; save
010010e3 push r12d ; save
010010e5 push r13d ; save
010010e7 push r14d ; save
010010e9 push r15d ; save
010010eb sub rsp,0x2c0 ; reserve stack
010010f2 mov rbx,r9 ; rbx = pszExe
010010f5 mov r12d,r8d ; r12 = fUnique (zero-extend)
010010f8 mov r13d,edx ; r13 = dispid (zero-extend)
010010fb mov rsi,rcx ; rsi = pdisp
The function begins by saving nonvolatile registers, and then reserving stack space for local variables. It then saves parameters in nonvolatile registers. Note that the destination of the middle two mov instructions is a 32-bit register, so they are implicitly zero-extended to 64 bits.
IQueryAssociations *pqa;
HRESULT hr = AssocCreate(CLSID_QueryAssociations, IID_IQueryAssociations, (void**)&pqa);
The first parameter to AssocCreate is a 128-bit CLSID passed by value. Since this doesn't fit in a 64-bit register, the CLSID is copied to the stack, and a pointer to the stack location is passed instead.
010010fe movdqu xmm0,oword ptr [CLSID_QueryAssociations (01001060)]
01001106 movdqu oword ptr [rsp+0x60],xmm0 ; temp buffer for first parameter
0100110c lea r8,[rsp+0x58] ; arg3 = &pqa
01001111 lea rdx,[IID_IQueryAssociations (01001070)] ; arg2 = &IID_IQueryAssociations
01001118 lea rcx,[rsp+0x60] ; arg1 = &temporary
0100111d call qword ptr [_imp_AssocCreate (01001028)] ; call
The movdqu instruction transfers 128-bits values to and from xmmn registers. In this instance, the assembly code uses it to copy the CLSID to the stack. The pointer to the CLSID is passed in r8. The other two arguments are passed in rcx and rdx.
if (SUCCEEDED(hr)) {
01001123 test eax,eax
01001125 jl ReturnEAX (01001281)
The code checks to see if the return value is a success.
hr = pqa->Init(ASSOCF_INIT_BYEXENAME, pszExe, NULL, NULL);
0100112b mov rcx,[rsp+0x58] ; arg1 = pqa
01001130 mov rax,[rcx] ; rax = pqa.vtbl
01001133 xor r14d,r14d ; r14 = 0
01001136 mov [rsp+0x20],r14 ; arg5 = 0
0100113b xor r9d,r9d ; arg4 = 0
0100113e mov r8,rbx ; arg3 = pszExe
01001141 mov r15d,0x2 ; r15 = 2 (for later)
01001147 mov edx,r15d ; arg2 = 2 (ASSOCF_INIT_BY_EXENAME)
0100114a call qword ptr [rax+0x18] ; call Init method
This is an indirect function call using a C++ vtable. The this pointer is passed in rcx as the first parameter. The first three parameters are passed in registers, while the final parameter is passed on the stack. The function reserves 16 bytes for the parameters passed in registers, so the fifth parameter begins at rsp+0x20.
if (SUCCEEDED(hr)) {
0100114d mov ebx,eax ; ebx = hr
0100114f test ebx,ebx ; FAILED?
01001151 jl ReleasePQA (01001274) ; jump if so
The assembly-language code saves the result in ebx, and checks to see if it's a success code.
WCHAR wszName[MAX_PATH];
DWORD cchName = MAX_PATH;
hr = pqa->GetString(0, ASSOCSTR_FRIENDLYAPPNAME, NULL, wszName, &cchName);
if (SUCCEEDED(hr)) {
01001157 mov dword ptr [rsp+0x50],0x104 ; cchName = MAX_PATH
0100115f mov rcx,[rsp+0x58] ; arg1 = pqa
01001164 mov rax,[rcx] ; rax = pqa.vtbl
01001167 lea rdx,[rsp+0x50] ; rdx = &cchName
0100116c mov [rsp+0x28],rdx ; arg6 = cchName
01001171 lea rdx,[rsp+0xb0] ; rdx = &wszName[0]
01001179 mov [rsp+0x20],rdx ; arg5 = &wszName[0]
0100117e xor r9d,r9d ; arg4 = 0
01001181 mov r8d,0x4 ; arg3 = 4 (ASSOCSTR_FRIENDLYNAME)
01001187 xor edx,edx ; arg2 = 0
01001189 call qword ptr [rax+0x20] ; call GetString method
0100118c mov ebx,eax ; ebx = hr
0100118e test ebx,ebx ; FAILED?
01001190 jl ReleasePQA (01001274) ; jump if so
Once again, we set up the parameters and call a function, then test the return value for success.
VARIANTARG rgvarg[2] = { 0 };
01001196 lea rdi,[rsp+0x82] ; rdi = &rgvarg
0100119e xor eax,eax ; rax = 0
010011a0 mov ecx,0x2e ; rcx = sizeof(rgvarg)
010011a5 rep stosb ; Zero it out
The idiomatic method for zeroing out a buffer on x64 is the same as x86.
V_VT(&rgvarg[0]) = VT_BSTR;
V_BSTR(&rgvarg[0]) = SysAllocString(wszName);
if (V_BSTR(&rgvarg[0])) {
010011a7 mov word ptr [rsp+0x80],0x8 ; V_VT(&rgvarg[0]) = VT_BSTR
010011b1 lea rcx,[rsp+0xb0] ; arg1 = &wszName[0]
010011b9 call qword ptr [_imp_SysAllocString (01001010)] ; call
010011bf mov [rsp+0x88],rax ; V_BSTR(&rgvarg[0]) = result
010011c7 test rax,rax ; anything allocated?
010011ca je OutOfMemory (0100126f) ; jump if failed
DISPPARAMS dp;
LONG lUnique = InterlockedIncrement(&lCounter);
010011d0 lea rax,[lCounter (01002000)]
010011d7 mov ecx,0x1
010011dc lock xadd [rax],ecx ; interlocked exchange and add
010011e0 add ecx,0x1
InterlockedIncrement compiles directly to machine code. The lock xadd instruction performs an atomic exchange and add. The final result is stored in ecx.
V_VT(&rgvarg[1]) = VT_I4;
V_I4(&rgvarg[1]) = fUnique ? lUnique : 0;
010011e3 mov word ptr [rsp+0x98],0x3 ; V_VT(&rgvarg[1]) = VT_I4;
010011ed mov eax,r14d ; rax = 0 (r14d is still zero)
010011f0 test r12d,r12d ; fUnique set?
010011f3 cmovne eax,ecx ; if so, then set rax=lCounter
010011f6 mov [rsp+0xa0],eax ; V_I4(&rgvarg[1]) = ...
Since x64 supports the cmov instruction, the ?: construct can be compiled without using a jump.
dp.rgvarg = rgvarg;
dp.cArgs = 2;
dp.rgdispidNamedArgs = NULL;
dp.cNamedArgs = 0;
010011fd lea rax,[rsp+0x80] ; rax = &rgvarg[0]
01001205 mov [rsp+0x60],rax ; dp.rgvarg = rgvarg
0100120a mov [rsp+0x70],r15d ; dp.cArgs = 2 (r15 is still 2)
0100120f mov [rsp+0x68],r14 ; dp.rgdispidNamedArgs = NULL
01001214 mov [rsp+0x74],r14d ; dp.cNamedArgs = 0
This code initializes the rest of the members of DISPPARAMS. Note that the compiler reuses the space on the stack previously used by the CLSID.
hr = pdisp->Invoke(dispid, IID_NULL, 0, DISPATCH_METHOD, &dp, NULL, NULL, NULL);
01001219 mov rax,[rsi] ; rax = pdisp.vtbl
0100121c mov [rsp+0x40],r14 ; arg9 = 0
01001221 mov [rsp+0x38],r14 ; arg8 = 0
01001226 mov [rsp+0x30],r14 ; arg7 = 0
0100122b lea rcx,[rsp+0x60] ; rcx = &dp
01001230 mov [rsp+0x28],rcx ; arg6 = &dp
01001235 mov word ptr [rsp+0x20],0x1 ; arg5 = 1 (DISPATCH_METHOD)
0100123c xor r9d,r9d ; arg4 = 0
0100123f lea r8,[GUID_NULL (01001080)] ; arg3 = &IID_NULL
01001246 mov edx,r13d ; arg2 = dispid
01001249 mov rcx,rsi ; arg1 = pdisp
0100124c call qword ptr [rax+0x30] ; call Invoke method
0100124f mov ebx,eax ; hr = result
The code then sets up the parameters and calls the Invoke method.
VariantClear(&rgvarg[0]);
VariantClear(&rgvarg[1]);
01001251 lea rcx,[rsp+0x80] ; arg1 = &rgvarg[0]
01001259 call qword ptr [_imp_VariantClear (01001018)]
0100125f lea rcx,[rsp+0x98] ; arg1 = &rgvarg[1]
01001267 call qword ptr [_imp_VariantClear (01001018)]
0100126d jmp ReleasePQA (01001274)
The code finishes up the current branch of the conditional, and skips over the else branch.
} else {
hr = E_OUTOFMEMORY;
}
}
OutOfMemory:
0100126f mov ebx,0x8007000e ; hr = E_OUTOFMEMORY
pqa->Release();
ReleasePQA:
01001274 mov rcx,[rsp+0x58] ; arg1 = pqa
01001279 mov rax,[rcx] ; rax = pqa.vtbl
0100127c call qword ptr [rax+0x10] ; release
The else branch.
return hr;
}
0100127f mov eax,ebx ; rax = hr (for return value)
ReturnEAX:
01001281 add rsp,0x2c0 ; clean up the stack
01001288 pop r15d ; restore
0100128a pop r14d ; restore
0100128c pop r13d ; restore
0100128e pop r12d ; restore
01001290 pop rdi ; restore
01001291 pop rsi ; restore
01001292 pop rbx ; restore
01001293 ret ; return (do not pop arguments)
The return value is stored in rax, and then the non-volatile registers are restored before returning.
See also
Feedback
Coming soon: Throughout 2024 we will be phasing out GitHub Issues as the feedback mechanism for content and replacing it with a new feedback system. For more information see: https://aka.ms/ContentUserFeedback.
Submit and view feedback for