On Sun, Jul 26, 2015 at 3:35 PM, Tom Kent wrote:

On Sun, Jul 26, 2015 at 8:10 AM, Andrey Semashev < andrey.semashev@gmail.com> wrote:

...

On Sat, May 23, 2015 at 4:40 PM, Hartmut Kaiser wrote:

...

...

All tests for Boost.Log (and I suspect any other built libraries) on

...

...

14 are failing because of this error:

LINK : fatal error LNK1101: incorrect MSPDB140.DLL version; recheck installation of this product

My workaround was to edit the 'Path' env variable in one of the temp files generated by bjam (b2_msvc_14.0_vcvarsall_amd64.cmd in your temp

...

not to point to the x64 cross compiler, but to the native one. For me

MSVC- directory) this

...

involved changing

C:\Program Files (x86)\Microsoft Visual Studio 14.0\VC\BIN\amd64_x86

To

C:\Program Files (x86)\Microsoft Visual Studio 14.0\VC\BIN\amd64

I encountered the same problem with VC14 RTM on my VM. This was a clean install of VC14 RTM, there were no CTPs or RCs installed on this machine. I removed the b2_msvc_14.0_vcvarsall*.cmd scripts and ran b2 with address-model=64 to build my tests so that these scripts are re-created from scratch, so it looks like there is a problem with how these scripts are generated by b2.

I found 3 instances of mspdb140.dll on my system in these locations:

C:\Program Files (x86)\Microsoft Visual Studio 14.0\Common7\IDE C:\Program Files (x86)\Microsoft Visual Studio 14.0\VC\bin C:\Program Files (x86)\Microsoft Visual Studio 14.0\VC\bin\amd64

The last two are for 32 and 64-bit native compilers. The first one, I assume, is for use by IDE. The problem is that the first path is listed in b2_msvc_14.0_vcvarsall_x86_amd64.cmd in the "SET Path=..." line prior to the compiler-specific paths. Removing it from the Path solved the problem.

Could Boost.Build be adjusted so that the cmd files are generated correctly in the first place?

Yes! I've been struggling with this all weekend trying to figure out what was going wrong. I've don multiple re-installs of my VM thinking that I might have gotten something corrupted. Your results seem to match what I was seeing exactly.

After experimenting a bit more, I've been able to exactly re-produce what Andrey saw. I would consider this something that would be a blocking bug for the 1.59.0 release, as it prevents any user from successfully building 64-bit binaries of boost with msvc-14.0....at least without the user going into their temp directory and editing a file that was created by boost build. Looking at the boost-build code, this batch file is created in build/src/tools/msvc.jam:744. This file will create the temp file in question by grabbing the output of a run of the visual studio environment setup routine (i.e. C:\Program Files (x86)\Microsoft Visual Studio 14.0\vc\vcvarsall.bat amd64). However, for the msvc-14.0 / address-model=64 case, it seems that it is combining the 32 bit and 64 bit variables into one causing this issue (and other variables that shouldn't be set by the amd64 run, such as FrameworkVersion32 and FrameworkVersion64...just running the above command only sets FrameworkVersion64). Is it possible that build is running the 32 bit variable file, then running the 64 bit instance? Tom

On Tue, Jul 28, 2015 at 1:49 AM, Vladimir Prus wrote:

Hi Tom,

On 28-Jul-15 3:44 AM, Tom Kent wrote:

Looking at the boost-build code, this batch file is created in

...

build/src/tools/msvc.jam:744. This file will create the temp file in question by grabbing the output of a run of the visual studio environment setup routine (i.e. C:\Program Files (x86)\Microsoft Visual Studio 14.0\vc\vcvarsall.bat amd64). However, for the msvc-14.0 / address-model=64 case, it seems that it is combining the 32 bit and 64 bit variables into one causing this issue (and other variables that shouldn't be set by the amd64 run, such as FrameworkVersion32 and FrameworkVersion64...just running the above command only sets FrameworkVersion64).

Is it possible that build is running the 32 bit variable file, then running the 64 bit instance?

local original-vars = [ SPLIT_BY_CHARACTERS [ SHELL set ] : "\n" ] ; local new-vars = [ SPLIT_BY_CHARACTERS [ SHELL "$(setup-script) $(setup-options)>nul && set" ] : "\n" ] ; local diff-vars = [ set.difference $(new-vars) : $(original-vars) ] ;

So we should be capturing the changes that the particular setup script does to environment; it's not obvious why this would start to fail with VS 14 - adding

ECHO $(diff-vars) ;

after the above lines might be illuminating. I'm installing VS 14 now, but it might take a day till I get to debug.

- Volodya

Does anyone know if any progress has been made on this? I think that the 1.59 release is just around the corner, and I would consider this a show-stopper as a user can't do a build for msvc-14.0 with address-model=64. Tom

What are the steps to reproduce this? 64-bit msvc-14 builds seem to be OK for me, John. On 07/08/2015 16:56, Andrey Semashev wrote:

On 07.08.2015 18:37, Tom Kent wrote:

...

On Tue, Jul 28, 2015 at 1:49 AM, Vladimir Prus wrote:

...

Hi Tom,

On 28-Jul-15 3:44 AM, Tom Kent wrote:

Looking at the boost-build code, this batch file is created in

...

build/src/tools/msvc.jam:744. This file will create the temp file in question by grabbing the output of a run of the visual studio environment setup routine (i.e. C:\Program Files (x86)\Microsoft Visual Studio 14.0\vc\vcvarsall.bat amd64). However, for the msvc-14.0 / address-model=64 case, it seems that it is combining the 32 bit and 64 bit variables into one causing this issue (and other variables that shouldn't be set by the amd64 run, such as FrameworkVersion32 and FrameworkVersion64...just running the above command only sets FrameworkVersion64).

Is it possible that build is running the 32 bit variable file, then running the 64 bit instance?

local original-vars = [ SPLIT_BY_CHARACTERS [ SHELL set ] : "\n" ] ; local new-vars = [ SPLIT_BY_CHARACTERS [ SHELL "$(setup-script) $(setup-options)>nul && set" ] : "\n" ] ; local diff-vars = [ set.difference $(new-vars) : $(original-vars) ] ;

So we should be capturing the changes that the particular setup script does to environment; it's not obvious why this would start to fail with VS 14 - adding

ECHO $(diff-vars) ;

after the above lines might be illuminating. I'm installing VS 14 now, but it might take a day till I get to debug.

- Volodya

Does anyone know if any progress has been made on this? I think that the 1.59 release is just around the corner, and I would consider this a show-stopper as a user can't do a build for msvc-14.0 with address-model=64.

+1, I agree with the 'show-stopper' here.

_______________________________________________ Unsubscribe & other changes: http://lists.boost.org/mailman/listinfo.cgi/boost

On 07/08/2015 17:24, Andrey Semashev wrote:

On 07.08.2015 19:09, John Maddock wrote:

...

What are the steps to reproduce this? 64-bit msvc-14 builds seem to be OK for me,

I described my steps at the beginning of the thread:

http://boost.2283326.n4.nabble.com/build-Problems-with-VC14-and-MSPDB140-DLL...

Did you remove the b2_msvc_14.0_vcvarsall* files in your temp directory?

Got it, can confirm it's an issue when building with 32-bit bjam (which uses the cross compiler), if you build with 64-bit bjam then it uses the native 64-bit compiler, and everything is OK. IMO we should find a way to fix this. John.

On Fri, Aug 7, 2015 at 5:21 PM, Andrey Semashev wrote:

On 07.08.2015 21:21, John Maddock wrote:

...

On 07/08/2015 17:24, Andrey Semashev wrote:

...

On 07.08.2015 19:09, John Maddock wrote:

...

What are the steps to reproduce this? 64-bit msvc-14 builds seem to be OK for me,

I described my steps at the beginning of the thread:

http://boost.2283326.n4.nabble.com/build-Problems-with-VC14-and-MSPDB140-DLL...

Did you remove the b2_msvc_14.0_vcvarsall* files in your temp directory?

Got it, can confirm it's an issue when building with 32-bit bjam (which uses the cross compiler), if you build with 64-bit bjam then it uses the native 64-bit compiler, and everything is OK.

I thought there's only one way to build bjam - by running tools/build/src/engine/build.bat. That's how I've always been building it anyway, and it always produced a 32-bit binary. How do you build a 64-bit bjam?

This problem will occur (with msvc-14.0) when using the bootstrap.bat file specified in the getting started directions.

On Fri, Aug 7, 2015 at 11:21 AM, John Maddock wrote:

On 07/08/2015 17:24, Andrey Semashev wrote:

...

On 07.08.2015 19:09, John Maddock wrote:

...

What are the steps to reproduce this? 64-bit msvc-14 builds seem to be OK for me,

I described my steps at the beginning of the thread:

http://boost.2283326.n4.nabble.com/build-Problems-with-VC14-and-MSPDB140-DLL...

Did you remove the b2_msvc_14.0_vcvarsall* files in your temp directory?

Got it, can confirm it's an issue when building with 32-bit bjam (which uses the cross compiler), if you build with 64-bit bjam then it uses the native 64-bit compiler, and everything is OK.

IMO we should find a way to fix this. John.

Everyone -- What is the status of this issue? -- Marshall

On Tue, Jul 28, 2015 at 9:49 AM, Vladimir Prus wrote:

Hi Tom,

On 28-Jul-15 3:44 AM, Tom Kent wrote:

...

Looking at the boost-build code, this batch file is created in build/src/tools/msvc.jam:744. This file will create the temp file in question by grabbing the output of a run of the visual studio environment setup routine (i.e. C:\Program Files (x86)\Microsoft Visual Studio 14.0\vc\vcvarsall.bat amd64). However, for the msvc-14.0 / address-model=64 case, it seems that it is combining the 32 bit and 64 bit variables into one causing this issue (and other variables that shouldn't be set by the amd64 run, such as FrameworkVersion32 and FrameworkVersion64...just running the above command only sets FrameworkVersion64).

Is it possible that build is running the 32 bit variable file, then running the 64 bit instance?

local original-vars = [ SPLIT_BY_CHARACTERS [ SHELL set ] : "\n" ] ; local new-vars = [ SPLIT_BY_CHARACTERS [ SHELL "$(setup-script) $(setup-options)>nul && set" ] : "\n" ] ; local diff-vars = [ set.difference $(new-vars) : $(original-vars) ] ;

So we should be capturing the changes that the particular setup script does to environment; it's not obvious why this would start to fail with VS 14 - adding

ECHO $(diff-vars) ;

after the above lines might be illuminating. I'm installing VS 14 now, but it might take a day till I get to debug.

I have tried to add this echo and indeed it outputs the broken Path already. Also, I tried to print %PATH% from the console (VS2015 x86 x64 Cross Tools Command Prompt) and it also has the %DevEnvDir% path first there. I inspected the vcvarsx86_amd64.bat script which sets up the environment and I can see that in line 62 it indeed inserts %DevEnvDir% before the paths to the compilers, so we have our culprit. The question is what do we do about it?

On Mon, Aug 10, 2015, Andrey Semashev wrote:

I created a Connect ticket about it: https://connect.microsoft.com/VisualStudio/feedback/details/1651822/incorrec...

Are bugs being deleted? I see it on Google's cache: http://tinyurl.com/nvedtmd But the Connect URL now says "Page Not Found". Glen

On 11.08.2015 01:18, Stephan T. Lavavej wrote:

[Andrey Semashev]

...

I created a Connect ticket about it: https://connect.microsoft.com/VisualStudio/feedback/details/1651822/incorrec...

[Glen Fernandes]

...

Are bugs being deleted? I see it on Google's cache: http://tinyurl.com/nvedtmd But the Connect URL now says "Page Not Found".

The link is also inaccessible to me, but it was successfully replicated to DevDiv#1209792 in TFS and remains active there.

I noticed that the Connect bug disappeared yesterday too and wrote a Connect feedback about it. I got no answer, but the link is accessible now.

Mere moments ago, quoth I:

On 11/08/2015 06:35, Andrey Semashev wrote:

...

I have tried to add this echo and indeed it outputs the broken Path already. Also, I tried to print %PATH% from the console (VS2015 x86 x64 Cross Tools Command Prompt) and it also has the %DevEnvDir% path first there.

I inspected the vcvarsx86_amd64.bat script which sets up the environment and I can see that in line 62 it indeed inserts %DevEnvDir% before the paths to the compilers, so we have our culprit. The question is what do we do about it?

One question is why Boost is trying to load the cross tools rather than the native x64 compiler.

While it's not unheard of to want to build 64-bit code from a 32-bit build machine, the vast majority of cases will be the reverse, so the native compilers should be preferred unless explicitly requested otherwise.

(Recent stats say that 92% of new Windows machines are 64-bit; older machines have a higher percentage of 32-bit but it's very likely that developers, at least, will have 64-bit machines by now.)

Ok, examining msvc.jam I think I've found the issue. It defaults to "x86_amd64" (cross compiler), and then swaps this to "amd64" (native compiler) if the environment variable PROCESSOR_ARCHITECTURE is set to AMD64. However this will only be the case if the process executing the jam file is also 64-bit. A 32-bit process running under WOW64 will see "x86" for that environment variable, just like it would on 32-bit Windows. So, either Boost.Build itself needs to be a 64-bit process, or it needs to be more clever at detecting a 64-bit architecture even as a 32-bit process. The simplest way of doing the latter is to read PROCESSOR_ARCHITEW6432, which will be unset for native 32 or 64-bit and set to AMD64 on 32-bit WOW64. (ie. the jamfile should assume native 64-bit if either of PROCESSOR_ARCHITECTURE or PROCESSOR_ARCHITEW6432 are set to AMD64.)

On Tue, Aug 11, 2015 at 3:35 AM, Gavin Lambert wrote:

Mere moments ago, quoth I:

...

Ok, examining msvc.jam I think I've found the issue. It defaults to "x86_amd64" (cross compiler), and then swaps this to "amd64" (native compiler) if the environment variable PROCESSOR_ARCHITECTURE is set to AMD64.

However this will only be the case if the process executing the jam file is also 64-bit. A 32-bit process running under WOW64 will see "x86" for that environment variable, just like it would on 32-bit Windows.

So, either Boost.Build itself needs to be a 64-bit process, or it needs to be more clever at detecting a 64-bit architecture even as a 32-bit process. The simplest way of doing the latter is to read PROCESSOR_ARCHITEW6432, which will be unset for native 32 or 64-bit and set to AMD64 on 32-bit WOW64. (ie. the jamfile should assume native 64-bit if either of PROCESSOR_ARCHITECTURE or PROCESSOR_ARCHITEW6432 are set to AMD64.)

PR for the latter workaround: https://github.com/boostorg/build/pull/87

Thanks Gavin. I have tried this change and it fixes the problem (both in develop and RC2), so we can use it at least as a workaround for 1.59. I can understand the current Boost.Build behavior though. Normally you build binaries for the same architecture the build tool is built for, which is 32-bit x86 in case of 32-bit bjam. When you build for another architecture, like amd64, this is a cross build, and it is logical to use a cross-compiler for that. This approach also makes it possible to build 64-bit binaries on a 32-bit machine, although in case of x86 I doubt this is a highly demanded feature. I wonder if we should be building a 64-bit bjam on a 64-bit Windows by default though. It's not a change for 1.59 but still worth considering for future releases.

On Tue, Aug 11, 2015 at 8:47 AM, Gavin Lambert wrote:

On 11/08/2015 16:55, Andrey Semashev wrote:

...

I wonder if we should be building a 64-bit bjam on a 64-bit Windows by default though. It's not a change for 1.59 but still worth considering for future releases.

That would be another way to solve it, yes. Although then the question would be whether it should use the native x86 compiler or the amd64_x86 cross-compiler. (It is actually possible in some Windows versions to remove the ability to run 32-bit code, although I would be quite surprised if anyone actually does that, especially in a development environment.)

Note that currently a 64-bit bjam *won't* use the cross-compiler; I think it just operates on a "by default we're running x86" assumption rather than trying to follow the native architecture like you suggested above. Comments near the code I patched seem to support this. :)

I'm not the author of that code but I suspect the reason for that is that the amd64_x86 cross-compiler was not shipped until VS2015; there was only the native x86 compiler. I think it is safe to keep using it (the x86 compiler) since we know every 64-bit Windows is able to run 32-bit programs. The other way around is not true though - not every Windows that manifests itself as 32-bit is able to run 64-bit programs.