Re: gfortran static linking under OS X (was: Re: )
On Sat, 20 Mar 2021, Iain Sandoe wrote: Daniel Feenberg wrote: On 3/19/21, Iain Sandoe wrote: On Fri, 19 Mar 2021, Tobias Burnus wrote: you should be able to work around this without changing the compiler or rebuilding it, %rename lib liborig *lib: -lquadmath -lm %(libgcc) %(liborig) ... I managed to install 10.2 and found that the error message goes away. No need to mess with the .spec file.A "static" binary can be produced with: gfortran taxsim.for -static-libgfortran -static-libgcc in version 10.2. however, I have only one Mac, so I can't be 100% sure the binary will port to another machine, but I am hopeful. Thank you all for your attention.. Daniel Feenberg NBER
Re: gfortran static linking under OS X (was: Re: )
On 3/19/21, Iain Sandoe wrote: >> On Fri, 19 Mar 2021, Tobias Burnus wrote: > > you should be able to work around this without changing the compiler or > rebuilding it, > > find > > /path/to/compiler/install/lib/libgfortran.spec > > make a copy of that (for backup only) > > the file contains something like: > > # > # This spec file is read by gfortran when linking. > # It is used to specify the libraries we need to link in, in the right > # order. > # > > %rename lib liborig > *lib: -lquadmath -lm %(libgcc) %(liborig) > > change the last line line to : > > *lib: %{!static-libgfortran: -lquadmath } %{static-libgfortran: > libquadmath.a%s} %(libgcc) %(liborig) > > === > > and try your link again > Tried that, the error message is: ld: library not found for -lSystem Indeed there is no libsystem.a anywhere on the computer. Any ideas? Daniel Feenberg
Re: gfortran static linking under OS X (was: Re: )
Daniel Feenberg wrote: On Fri, 19 Mar 2021, Tobias Burnus wrote: This seems to be a OS X issue ? and I have no idea about OS X, but I found the following: https://github.com/fxcoudert/gfortran-for-macOS/issues/12 It is certainly an OS X issue. Actually, it’s a gfortran issue. There is no problem in Linux, FreeBSD or Windows with the same compiler. As I understand it, OS X doesn't allow true dynamic linking, but it does allow a compiler/linker to produce an executable binary which only requires OS supplied libraries at execution time, and which includes all the compiler specific libraries. OSX doesn’t allow a completely statically-linked user space application. but it does allow libgfortran and libquadmath to be statically linked - the issue is that there’s no spec to deal with it. Apparently the problem arises because the authors of libquadm don't want users to use it in statically linked binaries, no such issue, there is a static libquadmath.a installed ... you should be able to work around this without changing the compiler or rebuilding it, find /path/to/compiler/install/lib/libgfortran.spec make a copy of that (for backup only) the file contains something like: # # This spec file is read by gfortran when linking. # It is used to specify the libraries we need to link in, in the right # order. # %rename lib liborig *lib: -lquadmath -lm %(libgcc) %(liborig) change the last line line to : *lib: %{!static-libgfortran: -lquadmath } %{static-libgfortran: libquadmath.a%s} %(libgcc) %(liborig) === and try your link again look at the linked object with “otool -Lv executable” and you should see that it only refers to libSystem.dylib. HTH Iain
Re: gfortran static linking under OS X (was: Re: )
On Fri, 19 Mar 2021, Tobias Burnus wrote: Hi, I am not sure whether it helps, but I want to point out that libm is the math library which is on Linux usually GLIBC and I assume on OS X it is provided by the OS vendor. Additionally, that libm is linked dynamically. This seems to be a OS X issue ? and I have no idea about OS X, but I found the following: https://github.com/fxcoudert/gfortran-for-macOS/issues/12 It is certainly an OS X issue. There is no problem in Linux, FreeBSD or Windows with the same compiler. As I understand it, OS X doesn't allow true dynamic linking, but it does allow a compiler/linker to produce an executable binary which only requires OS supplied libraries at execution time, and which includes all the compiler specific libraries. So I wouldn't actually expect libm to be statically linked. Apparently the problem arises because the authors of libquadm don't want users to use it in statically linked binaries, and did something to prevent that from happening. Unfortunately, it means that specialized knowledge is required to statically link gfortran programs, even if they don't use quad precision math. Daniel Feenberg which suggests that there is a mismatch of the XCode / MacOS X SDK version. As work around, it seems to work to set MACOSX_DEPLOYMENT_TARGET and/or to ensure that you have the same xcode version as the one used by gfortran, e.g. by "xcode-select --install". In some other thread, the suggestion was to use -L/usr/lib but I assume that won't help. Looking at https://github.com/xianyi/OpenBLAS/issues/3032 ? it seems as if you can use 'gfortran -v taxsim.for' to see the SDK version used when building GCC. At least that example had some |string like: ||-syslibroot /Library/Developer/CommandLineTools/SDKs/MacOSX11.1.sdk/| |If you had installed 11.0 instead, ||MACOSX_DEPLOYMENT_TARGET=11.0 would work.| |I hope it helps. If not, you need to find someone else as I have no idea about OS X.| |Good luck!| |Tobias | On 19.03.21 21:22, Daniel Feenberg via Fortran wrote: Is there a way to make a statically linked binary with fortran in OS X? For much of the past year I have been using: gfortran taxsim.for -static-libgfortran -static-libgcc but since January I only get the error message; ld: library not found for -lm. collect2: error: ld returned 1 exit status? This is OS X 11,2,3 Big Sur and fortran version 6.3.0. I need static linking because my users are not developers and do not have Xcode or gcc installed. This is free software. I have seen postings from 2015 suggesting that I rename libquadmath.0.dylib, which I did try but which did not help. Of course I have no need for lquad precision variables, which I understand is the source of the problem. Daniel Feenberg NBER
Re: gfortran static linking under OS X (was: Re: )
Hi Daniel, Tobias Burnus wrote: I am not sure whether it helps, but I want to point out that libm is the math library which is on Linux usually GLIBC and I assume on OS X it is provided by the OS vendor. actually part of libSystem (but, yes, provided by the vendor) On 19.03.21 21:22, Daniel Feenberg via Fortran wrote: Is there a way to make a statically linked binary with fortran in OS X? For much of the past year I have been using: gfortran taxsim.for -static-libgfortran -static-libgcc OK - that should work - modulo the quadmath issue (if present) - .. but that is probably solvable (with some changes to the link spec). but since January I only get the error message; ld: library not found for -lm. collect2: error: ld returned 1 exit status? lm has not been needed on macOS for a [very] long time (for many releases it was simply a convenience symlink to libSystem.dylib, for the sake of OSS code that appends ‘-lm’). There doesn’t seem to be an issue with gcc-11 (master, development) or 10.2.1 (upcoming 10.3) on macOS 11 - will see if I can fire up a copy of gcc6.5 there ... Will have a think about how to fake the libm too .. not so easy these days. This is OS X 11,2,3 Big Sur and fortran version 6.3.0. This is an old version of Fortran on a very new version of macOS, at present the first supported GCC version for macOS 11 is the upcoming 10.3 release (although homebrew no doubt has a preview courtesy of FX). Is there any way you would be able to update to a newer (and eventually supported) Fortran version ? I need static linking because my users are not developers and do not have Xcode or gcc installed. understood. This is free software. I have seen postings from 2015 suggesting that I rename libquadmath.0.dylib, which I did try but which did not help. Of course I have no need for lquad precision variables, which I understand is the source of the problem. not from what you posted - it’s the absence of “libm.dylib” that results in the message. I realise that this mail contains no solutions - but will try to reproduce the issue over the weekend. cheers Iain
gfortran static linking under OS X (was: Re: )
Hi, I am not sure whether it helps, but I want to point out that libm is the math library which is on Linux usually GLIBC and I assume on OS X it is provided by the OS vendor. Additionally, that libm is linked dynamically. This seems to be a OS X issue – and I have no idea about OS X, but I found the following: https://github.com/fxcoudert/gfortran-for-macOS/issues/12 which suggests that there is a mismatch of the XCode / MacOS X SDK version. As work around, it seems to work to set MACOSX_DEPLOYMENT_TARGET and/or to ensure that you have the same xcode version as the one used by gfortran, e.g. by "xcode-select --install". In some other thread, the suggestion was to use -L/usr/lib but I assume that won't help. Looking at https://github.com/xianyi/OpenBLAS/issues/3032 – it seems as if you can use 'gfortran -v taxsim.for' to see the SDK version used when building GCC. At least that example had some |string like: ||-syslibroot /Library/Developer/CommandLineTools/SDKs/MacOSX11.1.sdk/| |If you had installed 11.0 instead, ||MACOSX_DEPLOYMENT_TARGET=11.0 would work.| |I hope it helps. If not, you need to find someone else as I have no idea about OS X.| |Good luck!| |Tobias | On 19.03.21 21:22, Daniel Feenberg via Fortran wrote: Is there a way to make a statically linked binary with fortran in OS X? For much of the past year I have been using: gfortran taxsim.for -static-libgfortran -static-libgcc but since January I only get the error message; ld: library not found for -lm. collect2: error: ld returned 1 exit status? This is OS X 11,2,3 Big Sur and fortran version 6.3.0. I need static linking because my users are not developers and do not have Xcode or gcc installed. This is free software. I have seen postings from 2015 suggesting that I rename libquadmath.0.dylib, which I did try but which did not help. Of course I have no need for lquad precision variables, which I understand is the source of the problem. Daniel Feenberg NBER