Re: Loader Optimization

[Alexandre Julliard]

Robert Shearman  writes:
> ChangeLog:
> Optimize for the case where a DLL with no path is requested and it is
> already loaded. This change is correct since RtlDosSearchPath_U did not
> change the path being looked at - if libname contains no path then
> file_part will be the same as libname.

Yes, but the current code checks for the full name first and your
patch doesn't, so it will change the behavior if we have two modules
with the same base name. This will need some test cases.

Loader Optimization

[Robert Shearman]

Alexandre Julliard wrote:
> Robert Shearman  writes:
> > ChangeLog:
> > Optimize for the case where a DLL with no path is requested and it is
> > already loaded. This change is correct since RtlDosSearchPath_U did not
> > change the path being looked at - if libname contains no path then
> > file_part will be the same as libname.
>
> Yes, but the current code checks for the full name first and your
> patch doesn't, so it will change the behavior if we have two modules
> with the same base name. This will need some test cases.

What sort of tests do you want? I don't think I'll be able to come up 
with anything that can be put into the Wine test framework.

Loader Optimization

[Alexandre Julliard]

Robert Shearman  writes:
> What sort of tests do you want? I don't think I'll be able to come up
> with anything that can be put into the Wine test framework.

Agreed, it's probably not possible to put that in the test framework
since it will need native dlls. All we really need is a small test app
that we can run on Windows to determine the proper behavior.

Loader Optimization

[Rob Shearman]

Alexandre Julliard wrote:
> Robert Shearman  writes:
> > What sort of tests do you want? I don't think I'll be able to come up
> > with anything that can be put into the Wine test framework.
>
> Agreed, it's probably not possible to put that in the test framework
> since it will need native dlls. All we really need is a small test app
> that we can run on Windows to determine the proper behavior.

I performed two tests. One was a control to make sure the assumptions I 
made were correct and one was to verify the behaviour of the loader when 
it comes to finding matching modules.

Preparations for both tests:
1. Windows XP SP2 was used for the tests.
2. A test DLL was build that just consisted of the following code:
BOOL WINAPI DllMain(
  HINSTANCE hinstDLL,
  DWORD fdwReason,
  LPVOID lpvReserved)
{
    switch (fdwReason)
    {
    case DLL_PROCESS_ATTACH:
        OutputDebugString("Root\n"); / OutputDebugString("System\n");
        DisableThreadLibraryCalls(hinstDLL);
        break;
    }
}
3. Two versions of the DLL were built with different debug strings to be 
output depending on where the DLL was located (this was to ensure the 
correct DLL was being loaded for the later tests during the setup of the 
environment.
4. One DLL (to be known as "Root") was copied to C:\ (which had the 
debug string of "Root").
5. Another DLL (to be known as "System") was copied to 
%SystemRoot%\system32.
6. Both DLLs had the same name of "LoaderTest.dll"

Test 1:
Purpose - this test is a control to see whether the assumption was 
correct that using SetDllDirectory would cause the root directory 
version of the DLL to be returned from the last DLL loader.
The following test code will be used:
    HMODULE hSystem = NULL;
    HMODULE hRoot = NULL;
    HMODULE hModule = NULL;
//    hSystem = LoadLibrary("LoaderTest.dll");
    hRoot = LoadLibraryEx("C:\LoaderTest.dll", NULL, 
LOAD_WITH_ALTERED_SEARCH_PATH);
    SetDllDirectory("C:\");
    hModule = LoadLibrary("LoaderTest.dll");
    if (hModule == hSystem)
        printf("Last module load resolved to system directory version of 
the DLL\n");
    else if (hModule == hRoot)
        printf("Last module load resolved to root directory version of 
the DLL\n");
    else
        printf("Last module load resolved to neither of the two previous 
DLL loads\n");

Result - "Last module load resolved to root directory version of the DLL"
Conclusion - the test succeeded in verifying that without the system 
directory version of the DLL loaded then the last DLL load will return 
C:\LoaderTest.dll.

Test 2:
Purpose - this test will see whether loading a DLL by specifying just 
its basename causes it to be matched using its full name or its base name.
The following test code will be used:
    HMODULE hSystem = NULL;
    HMODULE hRoot = NULL;
    HMODULE hModule = NULL;
    hSystem = LoadLibrary("LoaderTest.dll");
    hRoot = LoadLibraryEx("C:\LoaderTest.dll", NULL, 
LOAD_WITH_ALTERED_SEARCH_PATH);
    SetDllDirectory("C:\");
    hModule = LoadLibrary("LoaderTest.dll");
    if (hModule == hSystem)
        printf("Last module load resolved to system directory version of 
the DLL\n");
    else if (hModule == hRoot)
        printf("Last module load resolved to root directory version of 
the DLL\n");
    else
        printf("Last module load resolved to neither of the two previous 
DLL loads\n");

Result - "Last module load resolved to system directory version of the DLL"
Conclusion - The result shows that the last module load returns the same 
module as the first (system directory) module load. This can be because 
of two reasons. The first could be that using a DLL with no path always 
causes the system directory version of the DLL to be used, but we 
discounted this possibility with Test 1 above. The second reason is that 
it could be preferring the first loaded module. However, since the full 
DLL names are unique and matching any base DLL name would cause bad 
behaviour, it must be concluded that it is indeed matching on the base 
DLL name.

I believe that this proves that my patch is correct. If I have overlook 
something I am happy to perform further tests.

Rob