On Wed, Mar 15, 2017 at 09:31:45AM +0100, Raffaello Giulietti wrote:

Suppose my code directly accesses "my.dll" which, in turn, depends on "her.dll" which my code does not need directly and suppose that library is located in folder "<path-to-her-folder>". How can I enforce the UFFI to perform the search there?

On Linux, I use a shell script "gopharo" to run the VM: #!/bin/sh PHAROVMPATH=$(dirname `readlink -f "$0"`) LD_LIBRARY_PATH="$PHAROVMPATH" exec "$PHAROVMPATH/pharo" $@ LD_LIBRARY_PATH is customizable. In my case, I place 3rd party files like libsqlite.so, libunqlite.so together with the standard VM files. Pierce

Not that well on macOS. Yes it works but you want to launch Pharo from thr Dock and not via the command line. But yes put the DLLs in with the VM files should work in all main OSs Mark On 17/03/2017 08:44, raffaello.giulietti@lifeware.ch wrote:

Hi Pierce,

sure, this works on any Unix-like system.

On 17/03/17 03:37, Pierce Ng wrote:

...

On Wed, Mar 15, 2017 at 09:31:45AM +0100, Raffaello Giulietti wrote:

...

Suppose my code directly accesses "my.dll" which, in turn, depends on "her.dll" which my code does not need directly and suppose that library is located in folder "<path-to-her-folder>". How can I enforce the UFFI to perform the search there?

On Linux, I use a shell script "gopharo" to run the VM:

#!/bin/sh PHAROVMPATH=$(dirname `readlink -f "$0"`) LD_LIBRARY_PATH="$PHAROVMPATH" exec "$PHAROVMPATH/pharo" $@

LD_LIBRARY_PATH is customizable. In my case, I place 3rd party files like libsqlite.so, libunqlite.so together with the standard VM files.

Pierce

-- Mark

Hi Ben, my understanding is that SetDllDirectory only affects the search path of libraries that are loaded at *run-time* with LoadLibrary. What I'm asking for is a mechanism that works at *load-time*, when a library I'm accessing directly depends on another one which I'm not targeting directly. But I'll try whether SetDllDirectory also affects load-time searches. Anyway, I hope future versions of the UFFI to offer setting the search paths more easily. Greetings Raffaello On 2017-03-15 07:22, Ben Coman wrote:

On Tue, Mar 14, 2017 at 11:23 PM, Raffaello Giulietti <raffaello.giulietti@lifeware.ch <mailto:raffaello.giulietti@lifeware.ch>> wrote:

Hi,

say I want to directly use a library that depends on other libraries in different folders. How does UFFI proceed to find these other libraries?

Is there a way to specify the search folders in UFFI?

I'm on Windows and in Pharo5/6.

Greetings Raffaello

I assume you mean specifying the search path outside of Pharo. Perhaps call SetDllDirectory via Pharo FFI. (YMMV disclaimer, my knowledge of this is a few minutes searching)

SetDllDirectory function https://msdn.microsoft.com/en-us/library/windows/desktop/ms686203(v=vs.85).a...

LoadLibrary function https://msdn.microsoft.com/en-us/library/windows/desktop/ms684175(v=vs.85).a...

Dynamic-Link Library Search Order https://msdn.microsoft.com/en-us/library/windows/desktop/ms682586(v=vs.85).a...

cheers -ben

On 2017-03-15 13:51, Ben Coman wrote:

On Wed, Mar 15, 2017 at 4:42 PM, Raffaello Giulietti <raffaello.giulietti@lifeware.ch <mailto:raffaello.giulietti@lifeware.ch>> wrote:

Hi Ben,

my understanding is that SetDllDirectory only affects the search path of libraries that are loaded at *run-time* with LoadLibrary.

What I'm asking for is a mechanism that works at *load-time*, when a library I'm accessing directly depends on another one which I'm not targeting directly.

I don't quite follow. I'm not intimate with Windows dynamic loading, so these questions are an opportunity for me to learn...

You can't mean when Pharo loads, because it wouldn't know of OTHER.DLL ?

Do you mean that LoadLibrary is not called explicitly by YOUR.DLL before its uses the function from OTHER.DLL?

During the build of YOUR.DLL, you can specify that it should depend on OTHER.LIB (kind of binary which lists function headers and static data declarations for OTHER.DLL). Linking resolution is then done at load-time of YOUR.DLL, not by invoking LoadLibrary("OTHER.DLL"). When YOUR.DLL is loaded by UFFI at run-time (here by means of LoadLibrary(), I guess), Windows (not the UFFI) also attempts to load OTHER.DLL because of the linking information gathered during the build. Notice that the linking information does *not* include a specific path for OTHER.DLL. (You don't have to mention standard *.LIBs like Kernel32.lib because they are included by default. But of course, Windows knows where to find the corresponding .dll.) OTHER.DLL is searched using the strategy described in the MSDN page you mention. My expectation is/was that UFFI has/had an API for specifying the search paths where to find OTHER.DLL in a platform-independent way. In other words, I would expect the UFFI to do the call of SetDllDirectory() or whatever is needed on Windows for me or to use other mechanisms on other platforms.

Do you mean you link OTHER.DLL to YOUR.DLL at compile time? But still, YOUR.DLL it not loaded until called via Pharo FFI, so neither is OTHER.DLL, and Dynamic-Link Library Search Order [1] says... "If a DLL with the same module name is already loaded in memory, the system uses the loaded DLL, no matter which directory it is in. The system does not search for the DLL."

So maybe before calling any of YOUR.DLL, do a SetDllDirectory followed by a LoadLibrary(OTHER.DLL) and later when you call into YOUR.DLL, OTHER.DLL is already loaded.

Yes, this might work but is rather contrived.

Otherwise it would seems(?) your options are either Manifests or Dynamic-Link Library Redirection, both done outside of Pharo.

What if neither YOUR.DLL nor OTHER.DLL are under my control? The only thing I know is where the .dll and the corresponding .lib + .h are. Thanks for your interest Raffaello

But I'll try whether SetDllDirectory also affects load-time searches.

Anyway, I hope future versions of the UFFI to offer setting the search paths more easily.

Greetings Raffaello

Hi Esteban, I understand this is the current status of the UFFI, so I can certainly use the workarounds discussed below. But I hope UFFI will soon offer a more "declarative" way to specify the search folders, kind of LD_LIBRARY_PATH mechanism but in the UFFI. Greetings Raffaello On 2017-03-15 14:52, Esteban Lorenzano wrote:

Hi,

UFFI cannot do what you want. If I understand well, you have:

Pharo -> libA.dll -> libB.dll

Pharo does a LoadLibrary(libA.dll), but has no control on how libA calls libB… and is not possible for us to influence it more than the predefined platform ways.

One posible workaround (just possible, no idea if it will work) is to perform a load of libB before is required by libB. Then hopefully the loader will understand is same library and will answer same handler (but not sure, because if it discriminates by path… then you’re in the same position). Anyway, if you want to try it, it would be something like this:

DynamicLoader loadLibrary: 'path/to/libB.dll’. DynamicLoader loadLibrary: 'path/to/libA.dll’.

then you can try to execute functions of libA…

cheers, Esteban

...

On 15 Mar 2017, at 14:32, Raffaello Giulietti <raffaello.giulietti@lifeware.ch> wrote:

On 2017-03-15 13:51, Ben Coman wrote:

...

On Wed, Mar 15, 2017 at 4:42 PM, Raffaello Giulietti <raffaello.giulietti@lifeware.ch <mailto:raffaello.giulietti@lifeware.ch>> wrote:

Hi Ben,

my understanding is that SetDllDirectory only affects the search path of libraries that are loaded at *run-time* with LoadLibrary.

What I'm asking for is a mechanism that works at *load-time*, when a library I'm accessing directly depends on another one which I'm not targeting directly.

I don't quite follow. I'm not intimate with Windows dynamic loading, so these questions are an opportunity for me to learn...

You can't mean when Pharo loads, because it wouldn't know of OTHER.DLL ?

Do you mean that LoadLibrary is not called explicitly by YOUR.DLL before its uses the function from OTHER.DLL?

During the build of YOUR.DLL, you can specify that it should depend on OTHER.LIB (kind of binary which lists function headers and static data declarations for OTHER.DLL). Linking resolution is then done at load-time of YOUR.DLL, not by invoking LoadLibrary("OTHER.DLL").

When YOUR.DLL is loaded by UFFI at run-time (here by means of LoadLibrary(), I guess), Windows (not the UFFI) also attempts to load OTHER.DLL because of the linking information gathered during the build. Notice that the linking information does *not* include a specific path for OTHER.DLL.

(You don't have to mention standard *.LIBs like Kernel32.lib because they are included by default. But of course, Windows knows where to find the corresponding .dll.)

OTHER.DLL is searched using the strategy described in the MSDN page you mention. My expectation is/was that UFFI has/had an API for specifying the search paths where to find OTHER.DLL in a platform-independent way. In other words, I would expect the UFFI to do the call of SetDllDirectory() or whatever is needed on Windows for me or to use other mechanisms on other platforms.

...

Do you mean you link OTHER.DLL to YOUR.DLL at compile time? But still, YOUR.DLL it not loaded until called via Pharo FFI, so neither is OTHER.DLL, and Dynamic-Link Library Search Order [1] says... "If a DLL with the same module name is already loaded in memory, the system uses the loaded DLL, no matter which directory it is in. The system does not search for the DLL."

So maybe before calling any of YOUR.DLL, do a SetDllDirectory followed by a LoadLibrary(OTHER.DLL) and later when you call into YOUR.DLL, OTHER.DLL is already loaded.

Yes, this might work but is rather contrived.

...

Otherwise it would seems(?) your options are either Manifests or Dynamic-Link Library Redirection, both done outside of Pharo.

What if neither YOUR.DLL nor OTHER.DLL are under my control? The only thing I know is where the .dll and the corresponding .lib + .h are.

Thanks for your interest Raffaello

...

But I'll try whether SetDllDirectory also affects load-time searches.

Anyway, I hope future versions of the UFFI to offer setting the search paths more easily.

Greetings Raffaello

On 2017-03-15 15:20, Esteban Lorenzano wrote:

...

On 15 Mar 2017, at 15:08, Raffaello Giulietti <raffaello.giulietti@lifeware.ch> wrote:

Hi Esteban,

I understand this is the current status of the UFFI, so I can certainly use the workarounds discussed below.

But I hope UFFI will soon offer a more "declarative" way to specify the search folders, kind of LD_LIBRARY_PATH mechanism but in the UFFI.

that will NOT solve the problem of dependencies of libraries indirectly references. There is no way to do that (indirect reference solving) on an executing program that I know. So is not “current status of UFFI”: it will never provide that.

Do you mean that SetDllDirectory() has no influence when invoked between Pharo's start and the LoadLibrary() call done by the UFFI?

All that, unless someone come and says: "look, Esteban is like this”.

I spent last three weeks (not full time of course) trying to do exactly that for our current linux vm, to force libgit2 to use the version of libssh2 we want (and not the version on system)… and I failed. I solved some cases, but not all. So at the end I opted to change zeroconf to define LD_LIBRARY_PATH before calling pharo. Only way I found I can force the lookup paths.

cheers, Esteban

...

Greetings Raffaello

On 2017-03-15 14:52, Esteban Lorenzano wrote:

...

Hi,

UFFI cannot do what you want. If I understand well, you have:

Pharo -> libA.dll -> libB.dll

Pharo does a LoadLibrary(libA.dll), but has no control on how libA calls libB… and is not possible for us to influence it more than the predefined platform ways.

One posible workaround (just possible, no idea if it will work) is to perform a load of libB before is required by libB. Then hopefully the loader will understand is same library and will answer same handler (but not sure, because if it discriminates by path… then you’re in the same position). Anyway, if you want to try it, it would be something like this:

DynamicLoader loadLibrary: 'path/to/libB.dll’. DynamicLoader loadLibrary: 'path/to/libA.dll’.

then you can try to execute functions of libA…

cheers, Esteban

...

On 15 Mar 2017, at 14:32, Raffaello Giulietti <raffaello.giulietti@lifeware.ch> wrote:

On 2017-03-15 13:51, Ben Coman wrote:

...

On Wed, Mar 15, 2017 at 4:42 PM, Raffaello Giulietti <raffaello.giulietti@lifeware.ch <mailto:raffaello.giulietti@lifeware.ch>> wrote:

Hi Ben,

my understanding is that SetDllDirectory only affects the search path of libraries that are loaded at *run-time* with LoadLibrary.

What I'm asking for is a mechanism that works at *load-time*, when a library I'm accessing directly depends on another one which I'm not targeting directly.

I don't quite follow. I'm not intimate with Windows dynamic loading, so these questions are an opportunity for me to learn...

You can't mean when Pharo loads, because it wouldn't know of OTHER.DLL ?

Do you mean that LoadLibrary is not called explicitly by YOUR.DLL before its uses the function from OTHER.DLL?

During the build of YOUR.DLL, you can specify that it should depend on OTHER.LIB (kind of binary which lists function headers and static data declarations for OTHER.DLL). Linking resolution is then done at load-time of YOUR.DLL, not by invoking LoadLibrary("OTHER.DLL").

When YOUR.DLL is loaded by UFFI at run-time (here by means of LoadLibrary(), I guess), Windows (not the UFFI) also attempts to load OTHER.DLL because of the linking information gathered during the build. Notice that the linking information does *not* include a specific path for OTHER.DLL.

(You don't have to mention standard *.LIBs like Kernel32.lib because they are included by default. But of course, Windows knows where to find the corresponding .dll.)

OTHER.DLL is searched using the strategy described in the MSDN page you mention. My expectation is/was that UFFI has/had an API for specifying the search paths where to find OTHER.DLL in a platform-independent way. In other words, I would expect the UFFI to do the call of SetDllDirectory() or whatever is needed on Windows for me or to use other mechanisms on other platforms.

...

Do you mean you link OTHER.DLL to YOUR.DLL at compile time? But still, YOUR.DLL it not loaded until called via Pharo FFI, so neither is OTHER.DLL, and Dynamic-Link Library Search Order [1] says... "If a DLL with the same module name is already loaded in memory, the system uses the loaded DLL, no matter which directory it is in. The system does not search for the DLL."

So maybe before calling any of YOUR.DLL, do a SetDllDirectory followed by a LoadLibrary(OTHER.DLL) and later when you call into YOUR.DLL, OTHER.DLL is already loaded.

Yes, this might work but is rather contrived.

...

Otherwise it would seems(?) your options are either Manifests or Dynamic-Link Library Redirection, both done outside of Pharo.

What if neither YOUR.DLL nor OTHER.DLL are under my control? The only thing I know is where the .dll and the corresponding .lib + .h are.

Thanks for your interest Raffaello

...

But I'll try whether SetDllDirectory also affects load-time searches.

Anyway, I hope future versions of the UFFI to offer setting the search paths more easily.

Greetings Raffaello

Please note the moment you use a FFI of any language you step to a C territory. VMs tend not to touch the code because the code is already in machine code format. There is zero reason for UFFI to worry about the dependencies of a library loaded, this must happen at the library level that UFFI does not focus on. What you describe here is standard practice because the vast majority of libraries depend on other libraries. In the end UFFI or any FFI in any language is not there to replace proper C coding but rather to allow you to utilize C libraries. You cannot hope to achieve performance comparable to C if you try to use a language that is not C to do standard things like dependency management. If the library fail to load it wont be because of UFFI it will be because the library is buggy at handling its dependencies. That means that the library will be buggy even if you use it from C directly. Even you do not know what you doing at C level, you going to crash Pharo countless times anyway and the least of your worries will be handling dependencies which far more obvious than some other nasty C bugs. I think its a great idea that UFFI does zero hand holding. This way it gives the clear message that if you do not understand C then you should not be using the UFFI in the first place. On Wed, Mar 15, 2017 at 4:51 PM Esteban Lorenzano <estebanlm@gmail.com> wrote:

...

On 15 Mar 2017, at 15:44, Raffaello Giulietti < raffaello.giulietti@lifeware.ch> wrote:

On 2017-03-15 15:20, Esteban Lorenzano wrote:

...

...

On 15 Mar 2017, at 15:08, Raffaello Giulietti <

raffaello.giulietti@lifeware.ch> wrote:

...

...

Hi Esteban,

I understand this is the current status of the UFFI, so I can

certainly use the workarounds discussed below.

...

But I hope UFFI will soon offer a more "declarative" way to specify

the search folders, kind of LD_LIBRARY_PATH mechanism but in the UFFI.

that will NOT solve the problem of dependencies of libraries indirectly references. There is no way to do that (indirect reference solving) on an executing program that I know. So is not “current status of UFFI”: it will never provide that.

Do you mean that SetDllDirectory() has no influence when invoked between Pharo's start and the LoadLibrary() call done by the UFFI?

I really have no idea. You can try to add it to your app.

Esteban

...

...

All that, unless someone come and says: "look, Esteban is like this”.

I spent last three weeks (not full time of course) trying to do exactly

that for our current linux vm, to force libgit2 to use the version of libssh2 we want (and not the version on system)… and I failed. I solved some cases, but not all. So at the end I opted to change zeroconf to define LD_LIBRARY_PATH before calling pharo. Only way I found I can force the lookup paths.

...

...

cheers, Esteban

...

Greetings Raffaello

On 2017-03-15 14:52, Esteban Lorenzano wrote:

...

Hi,

UFFI cannot do what you want. If I understand well, you have:

Pharo -> libA.dll -> libB.dll

Pharo does a LoadLibrary(libA.dll), but has no control on how libA

calls libB… and is not possible for us to influence it more than the predefined platform ways.

...

...

...

One posible workaround (just possible, no idea if it will work) is to

perform a load of libB before is required by libB. Then hopefully the loader will understand is same library and will answer same handler (but not sure, because if it discriminates by path… then you’re in the same position). Anyway, if you want to try it, it would be something like this:

...

DynamicLoader loadLibrary: 'path/to/libB.dll’. DynamicLoader loadLibrary: 'path/to/libA.dll’.

then you can try to execute functions of libA…

cheers, Esteban

...

On 15 Mar 2017, at 14:32, Raffaello Giulietti <

raffaello.giulietti@lifeware.ch> wrote:

...

...

On 2017-03-15 13:51, Ben Coman wrote:

...

On Wed, Mar 15, 2017 at 4:42 PM, Raffaello Giulietti <raffaello.giulietti@lifeware.ch <mailto:raffaello.giulietti@lifeware.ch>> wrote:

Hi Ben,

my understanding is that SetDllDirectory only affects the search path of libraries that are loaded at *run-time* with LoadLibrary.

What I'm asking for is a mechanism that works at *load-time*, when

a

...

...

library I'm accessing directly depends on another one which I'm not targeting directly.

I don't quite follow. I'm not intimate with Windows dynamic loading, so these questions are an opportunity for me to learn...

You can't mean when Pharo loads, because it wouldn't know of OTHER.DLL ?

Do you mean that LoadLibrary is not called explicitly by YOUR.DLL before its uses the function from OTHER.DLL?

During the build of YOUR.DLL, you can specify that it should depend on OTHER.LIB (kind of binary which lists function headers and static data declarations for OTHER.DLL). Linking resolution is then done at load-time of YOUR.DLL, not by invoking LoadLibrary("OTHER.DLL").

When YOUR.DLL is loaded by UFFI at run-time (here by means of LoadLibrary(), I guess), Windows (not the UFFI) also attempts to load OTHER.DLL because of the linking information gathered during the build. Notice that the linking information does *not* include a specific path for OTHER.DLL.

(You don't have to mention standard *.LIBs like Kernel32.lib because they are included by default. But of course, Windows knows where to find the corresponding .dll.)

OTHER.DLL is searched using the strategy described in the MSDN page you mention. My expectation is/was that UFFI has/had an API for specifying the search paths where to find OTHER.DLL in a platform-independent way. In other words, I would expect the UFFI to do the call of SetDllDirectory() or whatever is needed on Windows for me or to use other mechanisms on other platforms.

...

Do you mean you link OTHER.DLL to YOUR.DLL at compile time? But still, YOUR.DLL it not loaded until called via Pharo FFI, so neither is OTHER.DLL, and Dynamic-Link Library Search Order [1] says... "If a DLL with the same module name is already loaded in memory, the system uses the loaded DLL, no matter which directory it is in. The system does not search for the DLL."

So maybe before calling any of YOUR.DLL, do a SetDllDirectory followed by a LoadLibrary(OTHER.DLL) and later when you call into YOUR.DLL, OTHER.DLL is already loaded.

Yes, this might work but is rather contrived.

...

Otherwise it would seems(?) your options are either Manifests or Dynamic-Link Library Redirection, both done outside of Pharo.

What if neither YOUR.DLL nor OTHER.DLL are under my control? The only thing I know is where the .dll and the corresponding .lib + .h are.

Thanks for your interest Raffaello

...

But I'll try whether SetDllDirectory also affects load-time searches.

Anyway, I hope future versions of the UFFI to offer setting the search paths more easily.

Greetings Raffaello

Yes the calling process can alter where these indirect DLLs are loaded from and so should be able to be done in UFFI and directly from Smalltalk The OS will load the DLL specified then it runs DllMain and doing that causes the OS to load any DLLs that it calls MSDN https://msdn.microsoft.com/en-us/library/7d83bc18.aspx says With both implicit and explicit linking, Windows first searches for "known DLLs", such as Kernel32.dll and User32.dll. Windows then searches for the DLLs in the following sequence: The directory where the executable module for the current process is located. The current directory. The Windows system directory. The GetSystemDirectory function retrieves the path of this directory. The Windows directory. The GetWindowsDirectory function retrieves the path of this directory. The directories listed in the PATH environment variable. This can be changed using the call LoadLibraryEx instead of LoadLibrary Flags can be set to change the directories used https://msdn.microsoft.com/en-us/library/windows/desktop/ms684179(v=vs.85).a... Also see AddDllDirectory https://msdn.microsoft.com/en-us/library/windows/desktop/hh310513(v=vs.85).a... Mark On 15/03/2017 17:22, Raffaello Giulietti wrote:

Hi Dimitris,

I don't get your point about proper C coding, C performance or Pharo crashes. It is not the focus of this thread.

How dependencies between libraries are resolved at load-time is dictated by the OS. If the OS offers a working mechanism to specify folders that the OS uses to search for libraries at load-time, well, I think that Pharo, hence the UFFI, should offer it as well and in a more fashionable format. UFFI is quite elegant in many respects but I wouldn't call it a minimalist design (zero hand holding, if I understand you correctly). But I'm happy the extra elegance is there because it makes life more comfortable.

Again, I still have to experiment with SetDllDirectory() in the case of Windows, so I'm not in a position to say something authoritative yet.

On 2017-03-15 17:41, Dimitris Chloupis wrote:

...

Please note the moment you use a FFI of any language you step to a C territory. VMs tend not to touch the code because the code is already in machine code format.

There is zero reason for UFFI to worry about the dependencies of a library loaded, this must happen at the library level that UFFI does not focus on.

What you describe here is standard practice because the vast majority of libraries depend on other libraries. In the end UFFI or any FFI in any language is not there to replace proper C coding but rather to allow you to utilize C libraries. You cannot hope to achieve performance comparable to C if you try to use a language that is not C to do standard things like dependency management.

If the library fail to load it wont be because of UFFI it will be because the library is buggy at handling its dependencies. That means that the library will be buggy even if you use it from C directly.

Even you do not know what you doing at C level, you going to crash Pharo countless times anyway and the least of your worries will be handling dependencies which far more obvious than some other nasty C bugs.

I think its a great idea that UFFI does zero hand holding. This way it gives the clear message that if you do not understand C then you should not be using the UFFI in the first place.

On Wed, Mar 15, 2017 at 4:51 PM Esteban Lorenzano <estebanlm@gmail.com <mailto:estebanlm@gmail.com>> wrote:

> On 15 Mar 2017, at 15:44, Raffaello Giulietti <raffaello.giulietti@lifeware.ch <mailto:raffaello.giulietti@lifeware.ch>> wrote: > > On 2017-03-15 15:20, Esteban Lorenzano wrote: >> >>> On 15 Mar 2017, at 15:08, Raffaello Giulietti <raffaello.giulietti@lifeware.ch <mailto:raffaello.giulietti@lifeware.ch>> wrote: >>> >>> Hi Esteban, >>> >>> I understand this is the current status of the UFFI, so I can certainly use the workarounds discussed below. >>> >>> But I hope UFFI will soon offer a more "declarative" way to specify the search folders, kind of LD_LIBRARY_PATH mechanism but in the UFFI. >> >> that will NOT solve the problem of dependencies of libraries indirectly references. >> There is no way to do that (indirect reference solving) on an executing program that I know. >> So is not “current status of UFFI”: it will never provide that. >> > > Do you mean that SetDllDirectory() has no influence when invoked between Pharo's start and the LoadLibrary() call done by the UFFI?

I really have no idea. You can try to add it to your app.

Esteban

> > > > >> All that, unless someone come and says: "look, Esteban is like this”. >> >> I spent last three weeks (not full time of course) trying to do exactly that for our current linux vm, to force libgit2 to use the version of libssh2 we want (and not the version on system)… and I failed. I solved some cases, but not all. So at the end I opted to change zeroconf to define LD_LIBRARY_PATH before calling pharo. Only way I found I can force the lookup paths. >> >> cheers, >> Esteban >> >>> >>> >>> >>> Greetings >>> Raffaello >>> >>> >>> >>> >>> On 2017-03-15 14:52, Esteban Lorenzano wrote: >>>> Hi, >>>> >>>> UFFI cannot do what you want. >>>> If I understand well, you have: >>>> >>>> Pharo -> libA.dll -> libB.dll >>>> >>>> Pharo does a LoadLibrary(libA.dll), but has no control on how libA calls libB… and is not possible for us to influence it more than the predefined platform ways. >>>> >>>> One posible workaround (just possible, no idea if it will work) is to perform a load of libB before is required by libB. Then hopefully the loader will understand is same library and will answer same handler (but not sure, because if it discriminates by path… then you’re in the same position). Anyway, if you want to try it, it would be something like this: >>>> >>>> DynamicLoader loadLibrary: 'path/to/libB.dll’. >>>> DynamicLoader loadLibrary: 'path/to/libA.dll’. >>>> >>>> then you can try to execute functions of libA… >>>> >>>> cheers, >>>> Esteban >>>> >>>> >>>>> On 15 Mar 2017, at 14:32, Raffaello Giulietti <raffaello.giulietti@lifeware.ch <mailto:raffaello.giulietti@lifeware.ch>> wrote: >>>>> >>>>> On 2017-03-15 13:51, Ben Coman wrote: >>>>>> >>>>>> >>>>>> On Wed, Mar 15, 2017 at 4:42 PM, Raffaello Giulietti >>>>>> <raffaello.giulietti@lifeware.ch <mailto:raffaello.giulietti@lifeware.ch> >>>>>> <mailto:raffaello.giulietti@lifeware.ch <mailto:raffaello.giulietti@lifeware.ch>>> wrote: >>>>>> >>>>>> Hi Ben, >>>>>> >>>>>> my understanding is that SetDllDirectory only affects the search >>>>>> path of libraries that are loaded at *run-time* with LoadLibrary. >>>>>> >>>>>> What I'm asking for is a mechanism that works at *load-time*, when a >>>>>> library I'm accessing directly depends on another one which I'm not >>>>>> targeting directly. >>>>>> >>>>>> >>>>>> I don't quite follow. I'm not intimate with Windows dynamic loading, so >>>>>> these questions are an opportunity for me to learn... >>>>>> >>>>>> You can't mean when Pharo loads, because it wouldn't know of OTHER.DLL ? >>>>>> >>>>>> Do you mean that LoadLibrary is not called explicitly by YOUR.DLL before >>>>>> its uses the function from OTHER.DLL? >>>>>> >>>>> >>>>> During the build of YOUR.DLL, you can specify that it should depend on OTHER.LIB (kind of binary which lists function headers and static data declarations for OTHER.DLL). Linking resolution is then done at load-time of YOUR.DLL, not by invoking LoadLibrary("OTHER.DLL"). >>>>> >>>>> When YOUR.DLL is loaded by UFFI at run-time (here by means of LoadLibrary(), I guess), Windows (not the UFFI) also attempts to load OTHER.DLL because of the linking information gathered during the build. Notice that the linking information does *not* include a specific path for OTHER.DLL. >>>>> >>>>> (You don't have to mention standard *.LIBs like Kernel32.lib because they are included by default. But of course, Windows knows where to find the corresponding .dll.) >>>>> >>>>> OTHER.DLL is searched using the strategy described in the MSDN page you mention. My expectation is/was that UFFI has/had an API for specifying the search paths where to find OTHER.DLL in a platform-independent way. In other words, I would expect the UFFI to do the call of SetDllDirectory() or whatever is needed on Windows for me or to use other mechanisms on other platforms. >>>>> >>>>> >>>>> >>>>>> Do you mean you link OTHER.DLL to YOUR.DLL at compile time? >>>>>> But still, YOUR.DLL it not loaded until called via Pharo FFI, so neither >>>>>> is OTHER.DLL, >>>>>> and Dynamic-Link Library Search Order [1] says... "If a DLL with the >>>>>> same module name >>>>>> is already loaded in memory, the system uses the loaded DLL, no matter >>>>>> which directory it is in. >>>>>> The system does not search for the DLL." >>>>>> >>>>>> So maybe before calling any of YOUR.DLL, >>>>>> do a SetDllDirectory followed by a LoadLibrary(OTHER.DLL) >>>>>> and later when you call into YOUR.DLL, >>>>>> OTHER.DLL is already loaded. >>>>>> >>>>> >>>>> Yes, this might work but is rather contrived. >>>>> >>>>> >>>>> >>>>>> Otherwise it would seems(?) your options are either Manifests >>>>>> or Dynamic-Link Library Redirection, both done outside of Pharo. >>>>>> >>>>> >>>>> What if neither YOUR.DLL nor OTHER.DLL are under my control? The only thing I know is where the .dll and the corresponding .lib + .h are. >>>>> >>>>> >>>>> >>>>> Thanks for your interest >>>>> Raffaello >>>>> >>>>> >>>>> >>>>> >>>>>> But I'll try whether SetDllDirectory also affects load-time searches. >>>>>> >>>>>> Anyway, I hope future versions of the UFFI to offer setting the >>>>>> search paths more easily. >>>>>> >>>>>> >>>>>> >>>>>> Greetings >>>>>> Raffaello >>>>>> >>>>> >>>>> >>>> >>>> >>> >>> >> >> > >

-- Mark

My point was that anything you do in Pharo will lower performance quite a bit, but it wont be noticable anyway unless its in a big loop or something. Second that Pharo cannot debug such issues which is a huge deal when you work with OS stuff and OS calls. You are far better doing this outside UFFI and straight inside a C IDE because you will have an army of tools that will make your life a ton easier. Hence why I talked about performance, C coding and Pharo crashes. In my case I had to write and test the code in C to port it to UFFI because it was impossible to resolve the bugs from inside Pharo. Also crashing Pharo all the time got really annoying and I managed to corrupt a couple of images too in the process. Of course debugging C code from a C IDE is quite similar to how you work with Pharo. I also got an experience with loading libraries because I worked with Shared memory mapped files which is the technology OS uses to load DLLs. Its really low level stuff that come directly from OS kernel. It was a ton of fun. Also C is easier because obviously there is a ton of docs and all you have to do is copy paste. In case of UFFI you have to first figure out the C code and then how to convert it to UFFI so its double the effort for little gain. Its much better thing to just wrap C code and minimize the usage of UFFI. The benefits of Pharo are lost anyway as soon as you drop so low level. On Thu, Mar 16, 2017 at 3:05 PM <raffaello.giulietti@lifeware.ch> wrote:

Indeed, experimenting using some combination of AddDllDirectory(), SetDllDirectory() and LoadLibraryEx() was in my plan for next week.

Thanks

On 16/03/17 11:59, Mark Bestley wrote:

...

Yes the calling process can alter where these indirect DLLs are loaded from and so should be able to be done in UFFI and directly from Smalltalk

The OS will load the DLL specified then it runs DllMain and doing that causes the OS to load any DLLs that it calls

MSDN https://msdn.microsoft.com/en-us/library/7d83bc18.aspx says

With both implicit and explicit linking, Windows first searches for "known DLLs", such as Kernel32.dll and User32.dll. Windows then searches for the DLLs in the following sequence: The directory where the executable module for the current process is located. The current directory. The Windows system directory. The GetSystemDirectory function retrieves the path of this directory. The Windows directory. The GetWindowsDirectory function retrieves the path of this directory. The directories listed in the PATH environment variable.

This can be changed using the call LoadLibraryEx instead of LoadLibrary Flags can be set to change the directories used

https://msdn.microsoft.com/en-us/library/windows/desktop/ms684179(v=vs.85).a...

...

Also see AddDllDirectory

https://msdn.microsoft.com/en-us/library/windows/desktop/hh310513(v=vs.85).a...

...

Mark

On 15/03/2017 17:22, Raffaello Giulietti wrote:

...

Hi Dimitris,

I don't get your point about proper C coding, C performance or Pharo crashes. It is not the focus of this thread.

How dependencies between libraries are resolved at load-time is dictated by the OS. If the OS offers a working mechanism to specify folders that the OS uses to search for libraries at load-time, well, I think that Pharo, hence the UFFI, should offer it as well and in a more fashionable format. UFFI is quite elegant in many respects but I wouldn't call it a minimalist design (zero hand holding, if I understand you correctly). But I'm happy the extra elegance is there because it makes life more comfortable.

Again, I still have to experiment with SetDllDirectory() in the case of Windows, so I'm not in a position to say something authoritative yet.

On 2017-03-15 17:41, Dimitris Chloupis wrote:

...

Please note the moment you use a FFI of any language you step to a C territory. VMs tend not to touch the code because the code is already in machine code format.

There is zero reason for UFFI to worry about the dependencies of a library loaded, this must happen at the library level that UFFI does not focus on.

What you describe here is standard practice because the vast majority of libraries depend on other libraries. In the end UFFI or any FFI in any language is not there to replace proper C coding but rather to allow you to utilize C libraries. You cannot hope to achieve performance comparable to C if you try to use a language that is not C to do standard things like dependency management.

If the library fail to load it wont be because of UFFI it will be because the library is buggy at handling its dependencies. That means that the library will be buggy even if you use it from C directly.

Even you do not know what you doing at C level, you going to crash Pharo countless times anyway and the least of your worries will be handling dependencies which far more obvious than some other nasty C bugs.

I think its a great idea that UFFI does zero hand holding. This way it gives the clear message that if you do not understand C then you should not be using the UFFI in the first place.

On Wed, Mar 15, 2017 at 4:51 PM Esteban Lorenzano <estebanlm@gmail.com <mailto:estebanlm@gmail.com>> wrote:

> On 15 Mar 2017, at 15:44, Raffaello Giulietti <raffaello.giulietti@lifeware.ch <mailto:raffaello.giulietti@lifeware.ch>> wrote: > > On 2017-03-15 15:20, Esteban Lorenzano wrote: >> >>> On 15 Mar 2017, at 15:08, Raffaello Giulietti <raffaello.giulietti@lifeware.ch <mailto:raffaello.giulietti@lifeware.ch>> wrote: >>> >>> Hi Esteban, >>> >>> I understand this is the current status of the UFFI, so I can certainly use the workarounds discussed below. >>> >>> But I hope UFFI will soon offer a more "declarative" way to specify the search folders, kind of LD_LIBRARY_PATH mechanism but in the UFFI. >> >> that will NOT solve the problem of dependencies of libraries indirectly references. >> There is no way to do that (indirect reference solving) on an executing program that I know. >> So is not “current status of UFFI”: it will never provide that. >> > > Do you mean that SetDllDirectory() has no influence when invoked between Pharo's start and the LoadLibrary() call done by the UFFI?

I really have no idea. You can try to add it to your app.

Esteban

> > > > >> All that, unless someone come and says: "look, Esteban is like this”. >> >> I spent last three weeks (not full time of course) trying to do exactly that for our current linux vm, to force libgit2 to use the version of libssh2 we want (and not the version on system)… and I failed. I solved some cases, but not all. So at the end I opted to change zeroconf to define LD_LIBRARY_PATH before calling pharo. Only way I found I can force the lookup paths. >> >> cheers, >> Esteban >> >>> >>> >>> >>> Greetings >>> Raffaello >>> >>> >>> >>> >>> On 2017-03-15 14:52, Esteban Lorenzano wrote: >>>> Hi, >>>> >>>> UFFI cannot do what you want. >>>> If I understand well, you have: >>>> >>>> Pharo -> libA.dll -> libB.dll >>>> >>>> Pharo does a LoadLibrary(libA.dll), but has no control on how libA calls libB… and is not possible for us to influence it more than the predefined platform ways. >>>> >>>> One posible workaround (just possible, no idea if it will work) is to perform a load of libB before is required by libB. Then hopefully the loader will understand is same library and will answer same handler (but not sure, because if it discriminates by path… then you’re in the same position). Anyway, if you want to try it, it would be something like this: >>>> >>>> DynamicLoader loadLibrary: 'path/to/libB.dll’. >>>> DynamicLoader loadLibrary: 'path/to/libA.dll’. >>>> >>>> then you can try to execute functions of libA… >>>> >>>> cheers, >>>> Esteban >>>> >>>> >>>>> On 15 Mar 2017, at 14:32, Raffaello Giulietti <raffaello.giulietti@lifeware.ch <mailto:raffaello.giulietti@lifeware.ch>> wrote: >>>>> >>>>> On 2017-03-15 13:51, Ben Coman wrote: >>>>>> >>>>>> >>>>>> On Wed, Mar 15, 2017 at 4:42 PM, Raffaello Giulietti >>>>>> <raffaello.giulietti@lifeware.ch <mailto:raffaello.giulietti@lifeware.ch> >>>>>> <mailto:raffaello.giulietti@lifeware.ch <mailto:raffaello.giulietti@lifeware.ch>>> wrote: >>>>>> >>>>>> Hi Ben, >>>>>> >>>>>> my understanding is that SetDllDirectory only affects the search >>>>>> path of libraries that are loaded at *run-time* with LoadLibrary. >>>>>> >>>>>> What I'm asking for is a mechanism that works at *load-time*, when a >>>>>> library I'm accessing directly depends on another one which I'm not >>>>>> targeting directly. >>>>>> >>>>>> >>>>>> I don't quite follow. I'm not intimate with Windows dynamic loading, so >>>>>> these questions are an opportunity for me to learn... >>>>>> >>>>>> You can't mean when Pharo loads, because it wouldn't know of OTHER.DLL ? >>>>>> >>>>>> Do you mean that LoadLibrary is not called explicitly by YOUR.DLL before >>>>>> its uses the function from OTHER.DLL? >>>>>> >>>>> >>>>> During the build of YOUR.DLL, you can specify that it should depend on OTHER.LIB (kind of binary which lists function headers and static data declarations for OTHER.DLL). Linking resolution is then done at load-time of YOUR.DLL, not by invoking LoadLibrary("OTHER.DLL"). >>>>> >>>>> When YOUR.DLL is loaded by UFFI at run-time (here by means of LoadLibrary(), I guess), Windows (not the UFFI) also attempts to load OTHER.DLL because of the linking information gathered during the build. Notice that the linking information does *not* include a specific path for OTHER.DLL. >>>>> >>>>> (You don't have to mention standard *.LIBs like Kernel32.lib because they are included by default. But of course, Windows knows where to find the corresponding .dll.) >>>>> >>>>> OTHER.DLL is searched using the strategy described in the MSDN page you mention. My expectation is/was that UFFI has/had an API for specifying the search paths where to find OTHER.DLL in a platform-independent way. In other words, I would expect the UFFI to do the call of SetDllDirectory() or whatever is needed on Windows for me or to use other mechanisms on other platforms. >>>>> >>>>> >>>>> >>>>>> Do you mean you link OTHER.DLL to YOUR.DLL at compile time? >>>>>> But still, YOUR.DLL it not loaded until called via Pharo FFI, so neither >>>>>> is OTHER.DLL, >>>>>> and Dynamic-Link Library Search Order [1] says... "If a DLL with the >>>>>> same module name >>>>>> is already loaded in memory, the system uses the loaded DLL, no matter >>>>>> which directory it is in. >>>>>> The system does not search for the DLL." >>>>>> >>>>>> So maybe before calling any of YOUR.DLL, >>>>>> do a SetDllDirectory followed by a LoadLibrary(OTHER.DLL) >>>>>> and later when you call into YOUR.DLL, >>>>>> OTHER.DLL is already loaded. >>>>>> >>>>> >>>>> Yes, this might work but is rather contrived. >>>>> >>>>> >>>>> >>>>>> Otherwise it would seems(?) your options are either

Manifests

...

...

...

>>>>>> or Dynamic-Link Library Redirection, both done outside of Pharo. >>>>>> >>>>> >>>>> What if neither YOUR.DLL nor OTHER.DLL are under my control? The only thing I know is where the .dll and the corresponding .lib + .h are. >>>>> >>>>> >>>>> >>>>> Thanks for your interest >>>>> Raffaello >>>>> >>>>> >>>>> >>>>> >>>>>> But I'll try whether SetDllDirectory also affects load-time searches. >>>>>> >>>>>> Anyway, I hope future versions of the UFFI to offer setting the >>>>>> search paths more easily. >>>>>> >>>>>> >>>>>> >>>>>> Greetings >>>>>> Raffaello >>>>>> >>>>> >>>>> >>>> >>>> >>> >>> >> >> > >

In general I would agree do in C or C++ to cut down the interface and pass as little data between the two languages as possible. So hiding of C++ objects and exceptions and the others listed in you last paragraph is reasonable However the OS functions for loading need to be controlled outside the function called and so control should be in the UFFI layer. Also calling DLLs from Pharo will have a different environment from DLL called from a C main() call so can't be debugged in the C IDE. (My experience in this comes from DLLs called from Excel so a similar case, you end up running the excel executable under the C debugger and set C breakpoints and even then that differs from production use) Mark On 16/03/2017 16:19, raffaello.giulietti@lifeware.ch wrote:

Hi Dimitris,

I think I understand what you mean.

You are probably saying that, when facing multiple libraries that need to be glued together, it is far better to implement a C "gluing" library and then targeting that one from the UFFI rather than trying to do the gluing code using Smalltalk and the UFFI directly, right?

I can only agree with you that, except for the most simplest cases, a gluing C bridge library is far better and simpler to get right.

Personally, I've got no problems in using the UFFI for now, except for C varargs (the ... in a C function header), for invoking functions through pointers rather than by name and for indirectly handling C++ exceptions. For that I had to resort to a thin gluing C library written just for that purpose.

Greetings Raffaello

On 16/03/17 15:48, Dimitris Chloupis wrote:

...

My point was that anything you do in Pharo will lower performance quite a bit, but it wont be noticable anyway unless its in a big loop or something.

Second that Pharo cannot debug such issues which is a huge deal when you work with OS stuff and OS calls.

You are far better doing this outside UFFI and straight inside a C IDE because you will have an army of tools that will make your life a ton easier.

Hence why I talked about performance, C coding and Pharo crashes.

In my case I had to write and test the code in C to port it to UFFI because it was impossible to resolve the bugs from inside Pharo. Also crashing Pharo all the time got really annoying and I managed to corrupt a couple of images too in the process.

Of course debugging C code from a C IDE is quite similar to how you work with Pharo.

I also got an experience with loading libraries because I worked with Shared memory mapped files which is the technology OS uses to load DLLs. Its really low level stuff that come directly from OS kernel. It was a ton of fun.

Also C is easier because obviously there is a ton of docs and all you have to do is copy paste. In case of UFFI you have to first figure out the C code and then how to convert it to UFFI so its double the effort for little gain. Its much better thing to just wrap C code and minimize the usage of UFFI.

The benefits of Pharo are lost anyway as soon as you drop so low level.

On Thu, Mar 16, 2017 at 3:05 PM <raffaello.giulietti@lifeware.ch <mailto:raffaello.giulietti@lifeware.ch>> wrote:

Indeed, experimenting using some combination of AddDllDirectory(), SetDllDirectory() and LoadLibraryEx() was in my plan for next week.

Thanks

On 16/03/17 11:59, Mark Bestley wrote: > Yes the calling process can alter where these indirect DLLs are loaded > from and so should be able to be done in UFFI and directly from Smalltalk > > The OS will load the DLL specified then it runs DllMain and doing that > causes the OS to load any DLLs that it calls > > MSDN https://msdn.microsoft.com/en-us/library/7d83bc18.aspx says > > With both implicit and explicit linking, Windows first searches for > "known DLLs", such as Kernel32.dll and User32.dll. Windows then searches > for the DLLs in the following sequence: > The directory where the executable module for the current process is > located. > The current directory. > The Windows system directory. The GetSystemDirectory function retrieves > the path of this directory. > The Windows directory. The GetWindowsDirectory function retrieves the > path of this directory. > The directories listed in the PATH environment variable. > > > This can be changed using the call LoadLibraryEx instead of LoadLibrary > Flags can be set to change the directories used > > https://msdn.microsoft.com/en-us/library/windows/desktop/ms684179(v=vs.85).a... > > > > Also see AddDllDirectory > https://msdn.microsoft.com/en-us/library/windows/desktop/hh310513(v=vs.85).a... > > > Mark > > On 15/03/2017 17:22, Raffaello Giulietti wrote: >> Hi Dimitris, >> >> I don't get your point about proper C coding, C performance or Pharo >> crashes. It is not the focus of this thread. >> >> How dependencies between libraries are resolved at load-time is dictated >> by the OS. If the OS offers a working mechanism to specify folders that >> the OS uses to search for libraries at load-time, well, I think that >> Pharo, hence the UFFI, should offer it as well and in a more fashionable >> format. >> UFFI is quite elegant in many respects but I wouldn't call it a >> minimalist design (zero hand holding, if I understand you correctly). >> But I'm happy the extra elegance is there because it makes life more >> comfortable. >> >> Again, I still have to experiment with SetDllDirectory() in the case of >> Windows, so I'm not in a position to say something authoritative yet. >> >> >> >> >> >> >> >> On 2017-03-15 17:41, Dimitris Chloupis wrote: >>> Please note the moment you use a FFI of any language you step to a C >>> territory. VMs tend not to touch the code because the code is already >>> in machine code format. >>> >>> There is zero reason for UFFI to worry about the dependencies of a >>> library loaded, this must happen at the library level that UFFI does >>> not focus on. >>> >>> What you describe here is standard practice because the vast majority >>> of libraries depend on other libraries. In the end UFFI or any FFI in >>> any language is not there to replace proper C coding but rather to >>> allow you to utilize C libraries. You cannot hope to achieve >>> performance comparable to C if you try to use a language that is not C >>> to do standard things like dependency management. >>> >>> If the library fail to load it wont be because of UFFI it will be >>> because the library is buggy at handling its dependencies. That means >>> that the library will be buggy even if you use it from C directly. >>> >>> Even you do not know what you doing at C level, you going to crash >>> Pharo countless times anyway and the least of your worries will be >>> handling dependencies which far more obvious than some other nasty C >>> bugs. >>> >>> I think its a great idea that UFFI does zero hand holding. This way it >>> gives the clear message that if you do not understand C then you >>> should not be using the UFFI in the first place. >>> >>> On Wed, Mar 15, 2017 at 4:51 PM Esteban Lorenzano >>> <estebanlm@gmail.com <mailto:estebanlm@gmail.com> >>> <mailto:estebanlm@gmail.com <mailto:estebanlm@gmail.com>>> wrote: >>> >>> >>> > On 15 Mar 2017, at 15:44, Raffaello Giulietti >>> <raffaello.giulietti@lifeware.ch <mailto:raffaello.giulietti@lifeware.ch> >>> <mailto:raffaello.giulietti@lifeware.ch <mailto:raffaello.giulietti@lifeware.ch>>> >>> wrote: >>> > >>> > On 2017-03-15 15:20, Esteban Lorenzano wrote: >>> >> >>> >>> On 15 Mar 2017, at 15:08, Raffaello Giulietti >>> <raffaello.giulietti@lifeware.ch <mailto:raffaello.giulietti@lifeware.ch> >>> <mailto:raffaello.giulietti@lifeware.ch <mailto:raffaello.giulietti@lifeware.ch>>> >>> wrote: >>> >>> >>> >>> Hi Esteban, >>> >>> >>> >>> I understand this is the current status of the UFFI, so I can >>> certainly use the workarounds discussed below. >>> >>> >>> >>> But I hope UFFI will soon offer a more "declarative" way to >>> specify the search folders, kind of LD_LIBRARY_PATH mechanism but >>> in the UFFI. >>> >> >>> >> that will NOT solve the problem of dependencies of libraries >>> indirectly references. >>> >> There is no way to do that (indirect reference solving) on an >>> executing program that I know. >>> >> So is not “current status of UFFI”: it will never provide that. >>> >> >>> > >>> > Do you mean that SetDllDirectory() has no influence when invoked >>> between Pharo's start and the LoadLibrary() call done by the UFFI? >>> >>> I really have no idea. You can try to add it to your app. >>> >>> Esteban >>> >>> > >>> > >>> > >>> > >>> >> All that, unless someone come and says: "look, Esteban is like >>> this”. >>> >> >>> >> I spent last three weeks (not full time of course) trying to do >>> exactly that for our current linux vm, to force libgit2 to use the >>> version of libssh2 we want (and not the version on system)… and I >>> failed. I solved some cases, but not all. So at the end I opted to >>> change zeroconf to define LD_LIBRARY_PATH before calling pharo. >>> Only way I found I can force the lookup paths. >>> >> >>> >> cheers, >>> >> Esteban >>> >> >>> >>> >>> >>> >>> >>> >>> >>> Greetings >>> >>> Raffaello >>> >>> >>> >>> >>> >>> >>> >>> >>> >>> On 2017-03-15 14:52, Esteban Lorenzano wrote: >>> >>>> Hi, >>> >>>> >>> >>>> UFFI cannot do what you want. >>> >>>> If I understand well, you have: >>> >>>> >>> >>>> Pharo -> libA.dll -> libB.dll >>> >>>> >>> >>>> Pharo does a LoadLibrary(libA.dll), but has no control on how >>> libA calls libB… and is not possible for us to influence it more >>> than the predefined platform ways. >>> >>>> >>> >>>> One posible workaround (just possible, no idea if it will >>> work) is to perform a load of libB before is required by libB. >>> Then hopefully the loader will understand is same library and will >>> answer same handler (but not sure, because if it discriminates by >>> path… then you’re in the same position). Anyway, if you want to >>> try it, it would be something like this: >>> >>>> >>> >>>> DynamicLoader loadLibrary: 'path/to/libB.dll’. >>> >>>> DynamicLoader loadLibrary: 'path/to/libA.dll’. >>> >>>> >>> >>>> then you can try to execute functions of libA… >>> >>>> >>> >>>> cheers, >>> >>>> Esteban >>> >>>> >>> >>>> >>> >>>>> On 15 Mar 2017, at 14:32, Raffaello Giulietti >>> <raffaello.giulietti@lifeware.ch <mailto:raffaello.giulietti@lifeware.ch> >>> <mailto:raffaello.giulietti@lifeware.ch <mailto:raffaello.giulietti@lifeware.ch>>> >>> wrote: >>> >>>>> >>> >>>>> On 2017-03-15 13:51, Ben Coman wrote: >>> >>>>>> >>> >>>>>> >>> >>>>>> On Wed, Mar 15, 2017 at 4:42 PM, Raffaello Giulietti >>> >>>>>> >>> <raffaello.giulietti@lifeware.ch <mailto:raffaello.giulietti@lifeware.ch> >>> <mailto:raffaello.giulietti@lifeware.ch <mailto:raffaello.giulietti@lifeware.ch>> >>> >>>>>> >>> <mailto:raffaello.giulietti@lifeware.ch <mailto:raffaello.giulietti@lifeware.ch> >>> <mailto:raffaello.giulietti@lifeware.ch <mailto:raffaello.giulietti@lifeware.ch>>>> >>> wrote: >>> >>>>>> >>> >>>>>> Hi Ben, >>> >>>>>> >>> >>>>>> my understanding is that SetDllDirectory only affects the >>> search >>> >>>>>> path of libraries that are loaded at *run-time* with >>> LoadLibrary. >>> >>>>>> >>> >>>>>> What I'm asking for is a mechanism that works at >>> *load-time*, when a >>> >>>>>> library I'm accessing directly depends on another one >>> which I'm not >>> >>>>>> targeting directly. >>> >>>>>> >>> >>>>>> >>> >>>>>> I don't quite follow. I'm not intimate with Windows >>> dynamic loading, so >>> >>>>>> these questions are an opportunity for me to learn... >>> >>>>>> >>> >>>>>> You can't mean when Pharo loads, because it wouldn't know >>> of OTHER.DLL ? >>> >>>>>> >>> >>>>>> Do you mean that LoadLibrary is not called explicitly by >>> YOUR.DLL before >>> >>>>>> its uses the function from OTHER.DLL? >>> >>>>>> >>> >>>>> >>> >>>>> During the build of YOUR.DLL, you can specify that it should >>> depend on OTHER.LIB (kind of binary which lists function headers >>> and static data declarations for OTHER.DLL). Linking resolution is >>> then done at load-time of YOUR.DLL, not by invoking >>> LoadLibrary("OTHER.DLL"). >>> >>>>> >>> >>>>> When YOUR.DLL is loaded by UFFI at run-time (here by means >>> of LoadLibrary(), I guess), Windows (not the UFFI) also attempts >>> to load OTHER.DLL because of the linking information gathered >>> during the build. Notice that the linking information does *not* >>> include a specific path for OTHER.DLL. >>> >>>>> >>> >>>>> (You don't have to mention standard *.LIBs like Kernel32.lib >>> because they are included by default. But of course, Windows knows >>> where to find the corresponding .dll.) >>> >>>>> >>> >>>>> OTHER.DLL is searched using the strategy described in the >>> MSDN page you mention. My expectation is/was that UFFI has/had an >>> API for specifying the search paths where to find OTHER.DLL in a >>> platform-independent way. In other words, I would expect the UFFI >>> to do the call of SetDllDirectory() or whatever is needed on >>> Windows for me or to use other mechanisms on other platforms. >>> >>>>> >>> >>>>> >>> >>>>> >>> >>>>>> Do you mean you link OTHER.DLL to YOUR.DLL at compile time? >>> >>>>>> But still, YOUR.DLL it not loaded until called via Pharo >>> FFI, so neither >>> >>>>>> is OTHER.DLL, >>> >>>>>> and Dynamic-Link Library Search Order [1] says... "If a DLL >>> with the >>> >>>>>> same module name >>> >>>>>> is already loaded in memory, the system uses the loaded >>> DLL, no matter >>> >>>>>> which directory it is in. >>> >>>>>> The system does not search for the DLL." >>> >>>>>> >>> >>>>>> So maybe before calling any of YOUR.DLL, >>> >>>>>> do a SetDllDirectory followed by a LoadLibrary(OTHER.DLL) >>> >>>>>> and later when you call into YOUR.DLL, >>> >>>>>> OTHER.DLL is already loaded. >>> >>>>>> >>> >>>>> >>> >>>>> Yes, this might work but is rather contrived. >>> >>>>> >>> >>>>> >>> >>>>> >>> >>>>>> Otherwise it would seems(?) your options are either Manifests >>> >>>>>> or Dynamic-Link Library Redirection, both done outside of >>> Pharo. >>> >>>>>> >>> >>>>> >>> >>>>> What if neither YOUR.DLL nor OTHER.DLL are under my control? >>> The only thing I know is where the .dll and the corresponding .lib >>> + .h are. >>> >>>>> >>> >>>>> >>> >>>>> >>> >>>>> Thanks for your interest >>> >>>>> Raffaello >>> >>>>> >>> >>>>> >>> >>>>> >>> >>>>> >>> >>>>>> But I'll try whether SetDllDirectory also affects >>> load-time searches. >>> >>>>>> >>> >>>>>> Anyway, I hope future versions of the UFFI to offer >>> setting the >>> >>>>>> search paths more easily. >>> >>>>>> >>> >>>>>> >>> >>>>>> >>> >>>>>> Greetings >>> >>>>>> Raffaello >>> >>>>>> >>> >>>>> >>> >>>>> >>> >>>> >>> >>>> >>> >>> >>> >>> >>> >> >>> >> >>> > >>> > >>> >>> >> > >

-- Mark

It is true that you can kill an image or two doing this. A reason why I added this little thing https://github.com/Pharophile/HOImageSaver and use it quite often. I also got a debug and assert VM to build for Windows, so I have the full symbols table available. Running Pharo and, say CodeBlocks IDE side by side, and connecting GDB in CodeBlocks through attach to process using the Pharo PID works pretty well to figure out what is going on. I am also using Snagit a lot to capture screenshots with values etc for future reference. Using a remote Pharo image from another image that will stay untouched may be a better way to go in the future but will not really work if one has a breakpoint triggered in the remote one anyway,.. But I found the process of wrapping a library pretty much painless compared to the sea of wrapping code one gets from Swig for example. Once we will have wrapped a couple external libraries with UFFI I expect some patterns to emerge and our ability to wrap things to be enhanced drastically, to the point that we will see incredible growth in terms of capabilities. We will then really up our game on how to load code from a central repo, a la npmjs, or mavencentral, or gems, or pip. Because it sucks big time in a couple of occasions. With any tech, there is always this hype cycle with the disullusionment after the "peak of excitement" and that is needed to reach the productivity. I would urge all Pharo users to keep pushing and not let go of the vision, because there is really nothing like Pharo around. I just got remembered of how good it was yesterday: a given job stopped processing its input because the underlying format of the file changed under its feet. It was a matter of 30 minutes between diagnostic and fix, and I was able to deliver the output for a critical meeting by the client. There is zero way I would have been able to inspect, diagnose, and fix that fast with something else. Using Pharo makes someone a kind of maverick but the vista is quite good. Phil On Thu, Mar 16, 2017 at 3:48 PM, Dimitris Chloupis <kilon.alios@gmail.com> wrote:

My point was that anything you do in Pharo will lower performance quite a bit, but it wont be noticable anyway unless its in a big loop or something.

Second that Pharo cannot debug such issues which is a huge deal when you work with OS stuff and OS calls.

You are far better doing this outside UFFI and straight inside a C IDE because you will have an army of tools that will make your life a ton easier.

Hence why I talked about performance, C coding and Pharo crashes.

In my case I had to write and test the code in C to port it to UFFI because it was impossible to resolve the bugs from inside Pharo. Also crashing Pharo all the time got really annoying and I managed to corrupt a couple of images too in the process.

Of course debugging C code from a C IDE is quite similar to how you work with Pharo.

I also got an experience with loading libraries because I worked with Shared memory mapped files which is the technology OS uses to load DLLs. Its really low level stuff that come directly from OS kernel. It was a ton of fun.

Also C is easier because obviously there is a ton of docs and all you have to do is copy paste. In case of UFFI you have to first figure out the C code and then how to convert it to UFFI so its double the effort for little gain. Its much better thing to just wrap C code and minimize the usage of UFFI.

The benefits of Pharo are lost anyway as soon as you drop so low level.

On Thu, Mar 16, 2017 at 3:05 PM <raffaello.giulietti@lifeware.ch> wrote:

...

Indeed, experimenting using some combination of AddDllDirectory(), SetDllDirectory() and LoadLibraryEx() was in my plan for next week.

Thanks

On 16/03/17 11:59, Mark Bestley wrote:

...

Yes the calling process can alter where these indirect DLLs are loaded from and so should be able to be done in UFFI and directly from Smalltalk

The OS will load the DLL specified then it runs DllMain and doing that causes the OS to load any DLLs that it calls

MSDN https://msdn.microsoft.com/en-us/library/7d83bc18.aspx says

With both implicit and explicit linking, Windows first searches for "known DLLs", such as Kernel32.dll and User32.dll. Windows then searches for the DLLs in the following sequence: The directory where the executable module for the current process is located. The current directory. The Windows system directory. The GetSystemDirectory function retrieves the path of this directory. The Windows directory. The GetWindowsDirectory function retrieves the path of this directory. The directories listed in the PATH environment variable.

This can be changed using the call LoadLibraryEx instead of LoadLibrary Flags can be set to change the directories used

https://msdn.microsoft.com/en-us/library/windows/desktop/ ms684179(v=vs.85).aspx

Also see AddDllDirectory https://msdn.microsoft.com/en-us/library/windows/desktop/ hh310513(v=vs.85).aspx

Mark

On 15/03/2017 17:22, Raffaello Giulietti wrote:

...

Hi Dimitris,

I don't get your point about proper C coding, C performance or Pharo crashes. It is not the focus of this thread.

How dependencies between libraries are resolved at load-time is dictated by the OS. If the OS offers a working mechanism to specify folders that the OS uses to search for libraries at load-time, well, I think that Pharo, hence the UFFI, should offer it as well and in a more fashionable format. UFFI is quite elegant in many respects but I wouldn't call it a minimalist design (zero hand holding, if I understand you correctly). But I'm happy the extra elegance is there because it makes life more comfortable.

Again, I still have to experiment with SetDllDirectory() in the case of Windows, so I'm not in a position to say something authoritative yet.

On 2017-03-15 17:41, Dimitris Chloupis wrote:

...

Please note the moment you use a FFI of any language you step to a C territory. VMs tend not to touch the code because the code is already in machine code format.

There is zero reason for UFFI to worry about the dependencies of a library loaded, this must happen at the library level that UFFI does not focus on.

What you describe here is standard practice because the vast majority of libraries depend on other libraries. In the end UFFI or any FFI in any language is not there to replace proper C coding but rather to allow you to utilize C libraries. You cannot hope to achieve performance comparable to C if you try to use a language that is not C to do standard things like dependency management.

If the library fail to load it wont be because of UFFI it will be because the library is buggy at handling its dependencies. That means that the library will be buggy even if you use it from C directly.

Even you do not know what you doing at C level, you going to crash Pharo countless times anyway and the least of your worries will be handling dependencies which far more obvious than some other nasty C bugs.

I think its a great idea that UFFI does zero hand holding. This way it gives the clear message that if you do not understand C then you should not be using the UFFI in the first place.

On Wed, Mar 15, 2017 at 4:51 PM Esteban Lorenzano <estebanlm@gmail.com <mailto:estebanlm@gmail.com>> wrote:

> On 15 Mar 2017, at 15:44, Raffaello Giulietti <raffaello.giulietti@lifeware.ch <mailto:raffaello.giulietti@lifeware.ch>> wrote: > > On 2017-03-15 15:20, Esteban Lorenzano wrote: >> >>> On 15 Mar 2017, at 15:08, Raffaello Giulietti <raffaello.giulietti@lifeware.ch <mailto:raffaello.giulietti@lifeware.ch>> wrote: >>> >>> Hi Esteban, >>> >>> I understand this is the current status of the UFFI, so I can certainly use the workarounds discussed below. >>> >>> But I hope UFFI will soon offer a more "declarative" way to specify the search folders, kind of LD_LIBRARY_PATH mechanism but in the UFFI. >> >> that will NOT solve the problem of dependencies of libraries indirectly references. >> There is no way to do that (indirect reference solving) on an executing program that I know. >> So is not “current status of UFFI”: it will never provide that. >> > > Do you mean that SetDllDirectory() has no influence when invoked between Pharo's start and the LoadLibrary() call done by the UFFI?

I really have no idea. You can try to add it to your app.

Esteban

> > > > >> All that, unless someone come and says: "look, Esteban is like this”. >> >> I spent last three weeks (not full time of course) trying to do exactly that for our current linux vm, to force libgit2 to use the version of libssh2 we want (and not the version on system)… and I failed. I solved some cases, but not all. So at the end I opted to change zeroconf to define LD_LIBRARY_PATH before calling pharo. Only way I found I can force the lookup paths. >> >> cheers, >> Esteban >> >>> >>> >>> >>> Greetings >>> Raffaello >>> >>> >>> >>> >>> On 2017-03-15 14:52, Esteban Lorenzano wrote: >>>> Hi, >>>> >>>> UFFI cannot do what you want. >>>> If I understand well, you have: >>>> >>>> Pharo -> libA.dll -> libB.dll >>>> >>>> Pharo does a LoadLibrary(libA.dll), but has no control on how libA calls libB… and is not possible for us to influence it more than the predefined platform ways. >>>> >>>> One posible workaround (just possible, no idea if it will work) is to perform a load of libB before is required by libB. Then hopefully the loader will understand is same library and will answer same handler (but not sure, because if it discriminates by path… then you’re in the same position). Anyway, if you want to try it, it would be something like this: >>>> >>>> DynamicLoader loadLibrary: 'path/to/libB.dll’. >>>> DynamicLoader loadLibrary: 'path/to/libA.dll’. >>>> >>>> then you can try to execute functions of libA… >>>> >>>> cheers, >>>> Esteban >>>> >>>> >>>>> On 15 Mar 2017, at 14:32, Raffaello Giulietti <raffaello.giulietti@lifeware.ch <mailto:raffaello.giulietti@lifeware.ch>> wrote: >>>>> >>>>> On 2017-03-15 13:51, Ben Coman wrote: >>>>>> >>>>>> >>>>>> On Wed, Mar 15, 2017 at 4:42 PM, Raffaello Giulietti >>>>>> <raffaello.giulietti@lifeware.ch <mailto:raffaello.giulietti@lifeware.ch> >>>>>> <mailto:raffaello.giulietti@lifeware.ch <mailto:raffaello.giulietti@lifeware.ch>>> wrote: >>>>>> >>>>>> Hi Ben, >>>>>> >>>>>> my understanding is that SetDllDirectory only affects the search >>>>>> path of libraries that are loaded at *run-time* with LoadLibrary. >>>>>> >>>>>> What I'm asking for is a mechanism that works at *load-time*, when a >>>>>> library I'm accessing directly depends on another one which I'm not >>>>>> targeting directly. >>>>>> >>>>>> >>>>>> I don't quite follow. I'm not intimate with Windows dynamic loading, so >>>>>> these questions are an opportunity for me to learn... >>>>>> >>>>>> You can't mean when Pharo loads, because it wouldn't know of OTHER.DLL ? >>>>>> >>>>>> Do you mean that LoadLibrary is not called explicitly by YOUR.DLL before >>>>>> its uses the function from OTHER.DLL? >>>>>> >>>>> >>>>> During the build of YOUR.DLL, you can specify that it should depend on OTHER.LIB (kind of binary which lists function headers and static data declarations for OTHER.DLL). Linking resolution is then done at load-time of YOUR.DLL, not by invoking LoadLibrary("OTHER.DLL"). >>>>> >>>>> When YOUR.DLL is loaded by UFFI at run-time (here by means of LoadLibrary(), I guess), Windows (not the UFFI) also attempts to load OTHER.DLL because of the linking information gathered during the build. Notice that the linking information does *not* include a specific path for OTHER.DLL. >>>>> >>>>> (You don't have to mention standard *.LIBs like Kernel32.lib because they are included by default. But of course, Windows knows where to find the corresponding .dll.) >>>>> >>>>> OTHER.DLL is searched using the strategy described in the MSDN page you mention. My expectation is/was that UFFI has/had an API for specifying the search paths where to find OTHER.DLL in a platform-independent way. In other words, I would expect the UFFI to do the call of SetDllDirectory() or whatever is needed on Windows for me or to use other mechanisms on other platforms. >>>>> >>>>> >>>>> >>>>>> Do you mean you link OTHER.DLL to YOUR.DLL at compile time? >>>>>> But still, YOUR.DLL it not loaded until called via Pharo FFI, so neither >>>>>> is OTHER.DLL, >>>>>> and Dynamic-Link Library Search Order [1] says... "If a DLL with the >>>>>> same module name >>>>>> is already loaded in memory, the system uses the loaded DLL, no matter >>>>>> which directory it is in. >>>>>> The system does not search for the DLL." >>>>>> >>>>>> So maybe before calling any of YOUR.DLL, >>>>>> do a SetDllDirectory followed by a LoadLibrary(OTHER.DLL) >>>>>> and later when you call into YOUR.DLL, >>>>>> OTHER.DLL is already loaded. >>>>>> >>>>> >>>>> Yes, this might work but is rather contrived. >>>>> >>>>> >>>>> >>>>>> Otherwise it would seems(?) your options are either Manifests >>>>>> or Dynamic-Link Library Redirection, both done outside of Pharo. >>>>>> >>>>> >>>>> What if neither YOUR.DLL nor OTHER.DLL are under my control? The only thing I know is where the .dll and the corresponding .lib + .h are. >>>>> >>>>> >>>>> >>>>> Thanks for your interest >>>>> Raffaello >>>>> >>>>> >>>>> >>>>> >>>>>> But I'll try whether SetDllDirectory also affects load-time searches. >>>>>> >>>>>> Anyway, I hope future versions of the UFFI to offer setting the >>>>>> search paths more easily. >>>>>> >>>>>> >>>>>> >>>>>> Greetings >>>>>> Raffaello >>>>>> >>>>> >>>>> >>>> >>>> >>> >>> >> >> > >

Why all the hype about python? why not more flexible languages like JavaScript ES6, or Perl5 that are capable of preserving string to object literal isomorphism out of the box? Of course Dynamic Languages are fine for this, that's essentially what Objective C is; C + FakeTalk. I dream of a future where we can throw out the silly C processor macros and replace them with a real script engine; eg write c in a php file and have it generate code at preprocessor time, and if you need help at runtime, load the php runtime, and interact with it. This can be done in any language, For SmallTalk it's something like objective c but objective c is halfway there, it didn't replace c preprocessor with inline smalltalk. But yeah, hopefully we see more elegant compiletime-runtime interop in the future. On Mar 17, 2017 6:21 AM, "Dimitris Chloupis" <kilon.alios@gmail.com> wrote:

With any tech, there is always this hype cycle with the disullusionment

...

after the "peak of excitement" and that is needed to reach the productivity.

I would urge all Pharo users to keep pushing and not let go of the vision, because there is really nothing like Pharo around.

The one things I love about Pharo is Live Coding. The ability to change the code while it executes without all this recompile-restart nonsense.

I also was under the assumption you were when it came to live coding that none is doing better than Smalltalk and Lisp. So just for fun I went on to experiment how easy it would be to re implement live coding in C++ and Python. I was not expecting much if anything.

With Python it was super easy, I only need a main loop, wrap the loop inside an exception that in case of an error would catch the error , not stop the execution but rather using the import tools that come with Python will import the module. Worked like a charm.

I use this to develop a blender addon.

But the really shock was how easy it was to do this with C++. I thought that memory leaks would make live coding with C++ impossible , I was dead wrong. Apparently OS exceptions can capture even crashes, they wont allow the executable to crash. Instead of python modules I use DLLs on windows , .dylibs on MacOS and .so on Ubuntu. Executable is a tiny loop reloading the dlls and providing to them a pointer to a very large array that contains the entire memory of the app, DLLs handle the memory to share it between objects etc.

https://github.com/kilon/LiveCPP

Next step is integrating iPython and also python debugger and testing unit testing with live coding.

So yes definetly learn Pharo because it has many hidden gems , if you are newcomer. But if you miss Pharo , like me, when you code in other language, despair not, its very easy to recreate at least live coding.

Tx phil for these cool words. We are pushing and we need inputs too because from time to time I have the impression that we are focusing a bit too much on our own system but this is natural :) Stef On Fri, 17 Mar 2017 07:17:38 +0100, phil@highoctane.be <phil@highoctane.be> wrote:

It is true that you can kill an image or two doing this.

A reason why I added this little thing https://github.com/Pharophile/HOImageSaver and use it quite often.

I also got a debug and assert VM to build for Windows, so I have the full symbols table available.

Running Pharo and, say CodeBlocks IDE side by side, and connecting GDB in CodeBlocks through attach to process using the Pharo PID works pretty well to figure out what is going on.

I am also using Snagit a lot to capture screenshots with values etc for future reference.

Using a remote Pharo image from another image that will stay untouched may be a better way to go in the future but will not really work if one has a breakpoint triggered in the remote one >anyway,..

But I found the process of wrapping a library pretty much painless compared to the sea of wrapping code one gets from Swig for example.

Once we will have wrapped a couple external libraries with UFFI I expect some patterns to emerge and our ability to wrap things to be enhanced drastically, to the point that we will see >incredible growth in terms of capabilities.

We will then really up our game on how to load code from a central repo, a la npmjs, or mavencentral, or gems, or pip. Because it sucks big time in a couple of occasions.

With any tech, there is always this hype cycle with the disullusionment after the "peak of excitement" and that is needed to reach the productivity. I would urge all Pharo users to keep pushing and not let go of the vision, because there is really nothing like Pharo around.

I just got remembered of how good it was yesterday: a given job stopped processing its input because the underlying format of the file changed under its feet. It was a matter of 30 minutes >between diagnostic and fix, and I was able to deliver the output for a critical meeting by the client. There is zero way I would have been able to inspect, diagnose, and fix that fast with something >else. Using Pharo makes someone a kind of maverick but the vista is quite good.

Phil

On Thu, Mar 16, 2017 at 3:48 PM, Dimitris Chloupis <kilon.alios@gmail.com> wrote:

...

My point was that anything you do in Pharo will lower performance quite a bit, but it wont be noticable anyway unless its in a big loop or something. Second that Pharo cannot debug such issues which is a huge deal when you work with OS stuff and OS calls. You are far better doing this outside UFFI and straight inside a C IDE because you will have an army of tools that will make your life a ton easier. Hence why I talked about performance, C coding and Pharo crashes. In my case I had to write and test the code in C to port it to UFFI because it was impossible to resolve the bugs from inside Pharo. Also crashing Pharo all the time got really annoying and I >>managed to corrupt a couple of images too in the process. Of course debugging C code from a C IDE is quite similar to how you work with Pharo. I also got an experience with loading libraries because I worked with Shared memory mapped files which is the technology OS uses to load DLLs. Its really low level stuff that come directly >>from OS kernel. It was a ton of fun. Also C is easier because obviously there is a ton of docs and all you have to do is copy paste. In case of UFFI you have to first figure out the C code and then how to convert it to UFFI so its >>double the effort for little gain. Its much better thing to just wrap C code and minimize the usage of UFFI. The benefits of Pharo are lost anyway as soon as you drop so low level. On Thu, Mar 16, 2017 at 3:05 PM <raffaello.giulietti@lifeware.ch> wrote:

...

Indeed, experimenting using some combination of AddDllDirectory(), SetDllDirectory() and LoadLibraryEx() was in my plan for next week.

Thanks

On 16/03/17 11:59, Mark Bestley wrote:

...

Yes the calling process can alter where these indirect DLLs are loaded from and so should be able to be done in UFFI and directly from Smalltalk

The OS will load the DLL specified then it runs DllMain and doing that causes the OS to load any DLLs that it calls

MSDN https://msdn.microsoft.com/en-us/library/7d83bc18.aspx says

With both implicit and explicit linking, Windows first searches for "known DLLs", such as Kernel32.dll and User32.dll. Windows then searches for the DLLs in the following sequence: The directory where the executable module for the current process is located. The current directory. The Windows system directory. The GetSystemDirectory function retrieves the path of this directory. The Windows directory. The GetWindowsDirectory function retrieves the path of this directory. The directories listed in the PATH environment variable.

This can be changed using the call LoadLibraryEx instead of LoadLibrary Flags can be set to change the directories used

https://msdn.microsoft.com/en-us/library/windows/desktop/ms684179(v=vs.85).a...

Also see AddDllDirectory https://msdn.microsoft.com/en-us/library/windows/desktop/hh310513(v=vs.85).a...

Mark

On 15/03/2017 17:22, Raffaello Giulietti wrote:

...

Hi Dimitris,

I don't get your point about proper C coding, C performance or Pharo crashes. It is not the focus of this thread.

How dependencies between libraries are resolved at load-time is dictated by the OS. If the OS offers a working mechanism to specify folders that the OS uses to search for libraries at load-time, well, I think that Pharo, hence the UFFI, should offer it as well and in a more fashionable format. UFFI is quite elegant in many respects but I wouldn't call it a minimalist design (zero hand holding, if I understand you correctly). But I'm happy the extra elegance is there because it makes life more comfortable.

Again, I still have to experiment with SetDllDirectory() in the case of Windows, so I'm not in a position to say something authoritative yet.

On 2017-03-15 17:41, Dimitris Chloupis wrote:

...

Please note the moment you use a FFI of any language you step to a C territory. VMs tend not to touch the code because the code is already in machine code format.

There is zero reason for UFFI to worry about the dependencies of a library loaded, this must happen at the library level that UFFI does not focus on.

What you describe here is standard practice because the vast majority of libraries depend on other libraries. In the end UFFI or any FFI in any language is not there to replace proper C coding but rather to allow you to utilize C libraries. You cannot hope to achieve performance comparable to C if you try to use a language that is not C to do standard things like dependency management.

If the library fail to load it wont be because of UFFI it will be because the library is buggy at handling its dependencies. That means that the library will be buggy even if you use it from C directly.

Even you do not know what you doing at C level, you going to crash Pharo countless times anyway and the least of your worries will be handling dependencies which far more obvious than some other nasty C bugs.

I think its a great idea that UFFI does zero hand holding. This way it gives the clear message that if you do not understand C then you should not be using the UFFI in the first place.

On Wed, Mar 15, 2017 at 4:51 PM Esteban Lorenzano <estebanlm@gmail.com <mailto:estebanlm@gmail.com>> wrote:

> On 15 Mar 2017, at 15:44, Raffaello Giulietti <raffaello.giulietti@lifeware.ch <mailto:raffaello.giulietti@lifeware.ch>> wrote: > > On 2017-03-15 15:20, Esteban Lorenzano wrote: >> >>> On 15 Mar 2017, at 15:08, Raffaello Giulietti <raffaello.giulietti@lifeware.ch <mailto:raffaello.giulietti@lifeware.ch>> wrote: >>> >>> Hi Esteban, >>> >>> I understand this is the current status of the UFFI, so I can certainly use the workarounds discussed below. >>> >>> But I hope UFFI will soon offer a more "declarative" way to specify the search folders, kind of LD_LIBRARY_PATH mechanism but in the UFFI. >> >> that will NOT solve the problem of dependencies of libraries indirectly references. >> There is no way to do that (indirect reference solving) on an executing program that I know. >> So is not “current status of UFFI”: it will never provide that. >> > > Do you mean that SetDllDirectory() has no influence when invoked between Pharo's start and the LoadLibrary() call done by the UFFI?

I really have no idea. You can try to add it to your app.

Esteban

> > > > >> All that, unless someone come and says: "look, Esteban is like this”. >> >> I spent last three weeks (not full time of course) trying to do exactly that for our current linux vm, to force libgit2 to use the version of libssh2 we want (and not the version on system)… and I failed. I solved some cases, but not all. So at the end I opted to change zeroconf to define LD_LIBRARY_PATH before calling pharo. Only way I found I can force the lookup paths. >> >> cheers, >> Esteban >> >>> >>> >>> >>> Greetings >>> Raffaello >>> >>> >>> >>> >>> On 2017-03-15 14:52, Esteban Lorenzano wrote: >>>> Hi, >>>> >>>> UFFI cannot do what you want. >>>> If I understand well, you have: >>>> >>>> Pharo -> libA.dll -> libB.dll >>>> >>>> Pharo does a LoadLibrary(libA.dll), but has no control on how libA calls libB… and is not possible for us to influence it more than the predefined platform ways. >>>> >>>> One posible workaround (just possible, no idea if it will work) is to perform a load of libB before is required by libB. Then hopefully the loader will understand is same library and will answer same handler (but not sure, because if it discriminates by path… then you’re in the same position). Anyway, if you want to try it, it would be something like this: >>>> >>>> DynamicLoader loadLibrary: 'path/to/libB.dll’. >>>> DynamicLoader loadLibrary: 'path/to/libA.dll’. >>>> >>>> then you can try to execute functions of libA… >>>> >>>> cheers, >>>> Esteban >>>> >>>> >>>>> On 15 Mar 2017, at 14:32, Raffaello Giulietti <raffaello.giulietti@lifeware.ch <mailto:raffaello.giulietti@lifeware.ch>> wrote: >>>>> >>>>> On 2017-03-15 13:51, Ben Coman wrote: >>>>>> >>>>>> >>>>>> On Wed, Mar 15, 2017 at 4:42 PM, Raffaello Giulietti >>>>>> <raffaello.giulietti@lifeware.ch <mailto:raffaello.giulietti@lifeware.ch> >>>>>> <mailto:raffaello.giulietti@lifeware.ch <mailto:raffaello.giulietti@lifeware.ch>>> wrote: >>>>>> >>>>>> Hi Ben, >>>>>> >>>>>> my understanding is that SetDllDirectory only affects the search >>>>>> path of libraries that are loaded at *run-time* with LoadLibrary. >>>>>> >>>>>> What I'm asking for is a mechanism that works at *load-time*, when a >>>>>> library I'm accessing directly depends on another one which I'm not >>>>>> targeting directly. >>>>>> >>>>>> >>>>>> I don't quite follow. I'm not intimate with Windows dynamic loading, so >>>>>> these questions are an opportunity for me to learn... >>>>>> >>>>>> You can't mean when Pharo loads, because it wouldn't know of OTHER.DLL ? >>>>>> >>>>>> Do you mean that LoadLibrary is not called explicitly by YOUR.DLL before >>>>>> its uses the function from OTHER.DLL? >>>>>> >>>>> >>>>> During the build of YOUR.DLL, you can specify that it should depend on OTHER.LIB (kind of binary which lists function headers and static data declarations for OTHER.DLL). Linking resolution is then done at load-time of YOUR.DLL, not by invoking LoadLibrary("OTHER.DLL"). >>>>> >>>>> When YOUR.DLL is loaded by UFFI at run-time (here by means of LoadLibrary(), I guess), Windows (not the UFFI) also attempts to load OTHER.DLL because of the linking information gathered during the build. Notice that the linking information does *not* include a specific path for OTHER.DLL. >>>>> >>>>> (You don't have to mention standard *.LIBs like Kernel32.lib because they are included by default. But of course, Windows knows where to find the corresponding .dll.) >>>>> >>>>> OTHER.DLL is searched using the strategy described in the MSDN page you mention. My expectation is/was that UFFI has/had an API for specifying the search paths where to find OTHER.DLL in a platform-independent way. In other words, I would expect the UFFI to do the call of SetDllDirectory() or whatever is needed on Windows for me or to use other mechanisms on other platforms. >>>>> >>>>> >>>>> >>>>>> Do you mean you link OTHER.DLL to YOUR.DLL at compile time? >>>>>> But still, YOUR.DLL it not loaded until called via Pharo FFI, so neither >>>>>> is OTHER.DLL, >>>>>> and Dynamic-Link Library Search Order [1] says... "If a DLL with the >>>>>> same module name >>>>>> is already loaded in memory, the system uses the loaded DLL, no matter >>>>>> which directory it is in. >>>>>> The system does not search for the DLL." >>>>>> >>>>>> So maybe before calling any of YOUR.DLL, >>>>>> do a SetDllDirectory followed by a LoadLibrary(OTHER.DLL) >>>>>> and later when you call into YOUR.DLL, >>>>>> OTHER.DLL is already loaded. >>>>>> >>>>> >>>>> Yes, this might work but is rather contrived. >>>>> >>>>> >>>>> >>>>>> Otherwise it would seems(?) your options are either Manifests >>>>>> or Dynamic-Link Library Redirection, both done outside of Pharo. >>>>>> >>>>> >>>>> What if neither YOUR.DLL nor OTHER.DLL are under my control? The only thing I know is where the .dll and the corresponding .lib + .h are. >>>>> >>>>> >>>>> >>>>> Thanks for your interest >>>>> Raffaello >>>>> >>>>> >>>>> >>>>> >>>>>> But I'll try whether SetDllDirectory also affects load-time searches. >>>>>> >>>>>> Anyway, I hope future versions of the UFFI to offer setting the >>>>>> search paths more easily. >>>>>> >>>>>> >>>>>> >>>>>> Greetings >>>>>> Raffaello >>>>>> >>>>> >>>>> >>>> >>>> >>> >>> >> >> > >

-- Using Opera's mail client: http://www.opera.com/mail/