[LLVMdev] About JIT by LLVM 2.9 or later

[Eli Friedman]

2011/11/1 空明流转 <wuye9036 at gmail.com>:
> Hello guys,
>
>      Thanks for your help when you are busing.
>      I am working on an open source project. It supports shader language and
> I want JIT feature, so LLVM is used.
>      But now I find the ABI & Calling Convention did not co-work with MSVC.
> For example, following code I have:
>
> struct float4 { float x, y, z, w; };
> struct float4x4 { float4 x, y, z, w; };
>
> float4 fetch_vs( float4x4* mat ){ return mat->y; }
>
> Caller:
>
> // ...
> float4x4 mat; // Initialized
> float4 ret = fetch(mat); // fetch is JITed by LLVM
> float4 ret_vs = fetch_vs(mat)
> // ...
>
> Callee(LLVM):
>
> %vec4 = type { float, float, float, float }
> %mat44 = type { %vec4, %vec4, %vec4, %vec4 }
> define %vec4 @fetch( %mat44* %m ) {
>        %matval = load %mat44* %m
>        %v2 = extractvalue %mat44 %matval, 2
>        ret %vec4 %v2
>  }
>
>      But if it is implemented by LLVM and called the JIT-ed function in
> MSVC, the program will be crashed.
>      I traced into the implementations, ASMs are:
>
>     Caller:
>
>         float4x4 f;
>         float4 b = fetch(&f);
>     // Calling function. first address is a temporary result generated by
> caller. And secondary is the &f.
>     013C1428  lea         eax,[ebp-48h]
>     013C142B  push        eax
>     013C142C  lea         ecx,[ebp-138h]
>     013C1432  push        ecx
>     013C1433  call        fetch (13C11D6h)
>     013C1438  add         esp,8
>
>     // Copy result to another temporary vairable.
>     013C143B  mov         edx,dword ptr [eax]
>     013C143D  mov         dword ptr [ebp-150h],edx
>     013C1443  mov         ecx,dword ptr [eax+4]
>     013C1446  mov         dword ptr [ebp-14Ch],ecx
>     013C144C  mov         edx,dword ptr [eax+8]
>     013C144F  mov         dword ptr [ebp-148h],edx
>     013C1455  mov         eax,dword ptr [eax+0Ch]
>     013C1458  mov         dword ptr [ebp-144h],eax
>     013C145E  mov         ecx,dword ptr [ebp-150h]
>
>     // Copy secondary temporary to variable 'b'
>     013C1464  mov         dword ptr [ebp-60h],ecx
>     013C1467  mov         edx,dword ptr [ebp-14Ch]
>     013C146D  mov         dword ptr [ebp-5Ch],edx
>     013C1470  mov         eax,dword ptr [ebp-148h]
>     013C1476  mov         dword ptr [ebp-58h],eax
>     013C1479  mov         ecx,dword ptr [ebp-144h]
>     013C147F  mov         dword ptr [ebp-54h],ecx
>
>     Callee( 'fetch_vs' MSVC ):
>
>       float4 __cdecl fetch( float4x4* mat ){
>
> // Stack protection
> 002C13B0  push        ebp
> 002C13B1  mov         ebp,esp
> 002C13B3  sub         esp,0C0h
> 002C13B9  push        ebx
> 002C13BA  push        esi
> 002C13BB  push        edi
> 002C13BC  lea         edi,[ebp-0C0h]
> 002C13C2  mov         ecx,30h
> 002C13C7  mov         eax,0CCCCCCCCh
> 002C13CC  rep stos    dword ptr es:[edi]
>         return mat->y;
>
> // Copy to address of first temporary variable.
> 002C13CE  mov         eax,dword ptr [mat]
> 002C13D1  add         eax,10h
> 002C13D4  mov         ecx,dword ptr [ebp+8]
> 002C13D7  mov         edx,dword ptr [eax]
> 002C13D9  mov         dword ptr [ecx],edx
> 002C13DB  mov         edx,dword ptr [eax+4]
> 002C13DE  mov         dword ptr [ecx+4],edx
> 002C13E1  mov         edx,dword ptr [eax+8]
> 002C13E4  mov         dword ptr [ecx+8],edx
> 002C13E7  mov         eax,dword ptr [eax+0Ch]
> 002C13EA  mov         dword ptr [ecx+0Ch],eax
> 002C13ED  mov         eax,dword ptr [ebp+8]
>     }
> 002C13F0  pop         edi
> 002C13F1  pop         esi
> 002C13F2  pop         ebx
> 002C13F3  mov         esp,ebp
> 002C13F5  pop         ebp
> 002C13F6  ret
>
>     Callee( 'fetch' LLVM ):
>
> 010B0010  mov         eax,dword ptr [esp+4]
> 010B0014  mov         ecx,dword ptr [esp+8]
> 010B0018  movss       xmm0,dword ptr [ecx+1Ch]
> 010B001D  movss       dword ptr [eax+0Ch],xmm0
> 010B0022  movss       xmm0,dword ptr [ecx+18h]
> 010B0027  movss       dword ptr [eax+8],xmm0
> 010B002C  movss       xmm0,dword ptr [ecx+10h]
> 010B0031  movss       xmm1,dword ptr [ecx+14h]
> 010B0036  movss       dword ptr [eax+4],xmm1
> 010B003B  movss       dword ptr [eax],xmm0
> 010B003F  ret         4                                           // There
> are nothing push/pop in function and stack space is managed by caller, so
> why ret 4 is generated?
>
> It will cause the stack unbalance before and after function call. the other
> code look like well.
>
> Following is my questions:
>
> 1. Why it generates "ret 4" but not "ret" without modify esp?

>From what I can find, that follows the Mingw calling convention (which
is generally the same as cdecl, but not quite the same when returning
a struct).  Looks like it's an oversight; I'd suggest filing a bug.

> 2. Does it doesn't support Microsoft Visual C++ compiler directly ?
>
> 3. I want to integrate LLVM JIT in MSVC, is it possible? what I will do,
> such as generate a adapter function ? Any porting document for it?
> 4. Does Clang support MS's call convention? If it does, how it work? I
> traced into it, but code is too large.

People have used LLVM with MSVC; making your calls across compilers
simpler generally helps.

> 5. Does it support Mingw directly ?

LLVM support for mingw is generally more mature than the MSVC support.

> 6. Does x64 work if your solution is applied ?

The Windows 64-bit calling convention should work... it hasn't gotten
much testing as far as I know, though.

-Eli