Bug#897006: petsc FTBFS on mips: f951: out of memory

[Adrian Bunk]

On Fri, Apr 27, 2018 at 09:48:17AM +0800, Drew Parsons wrote:
> Source: petsc
> Version: 3.9.0+dfsg1-1exp4
> Severity: normal
> 
> PETSc 3.9 Fails To Build From Source on mips. PETSC 3.8 continues to build fine.
> 
> Build logs are at
>   https://buildd.debian.org/status/logs.php?pkg=petsc&arch=mips
> e.g.
>   https://buildd.debian.org/status/fetch.php?pkg=petsc&arch=mips&ver=3.9.0%2Bdfsg1-1exp4&stamp=1524789882&raw=0
> 
> The build log reports
> -====
> ...
>          CXX mips-linux-gnu-real-debug/obj/tao/linesearch/interface/dlregis_taolinesearch.o
>          CXX mips-linux-gnu-real-debug/obj/tao/linesearch/impls/gpcglinesearch/gpcglinesearch.o
>          CXX mips-linux-gnu-real-debug/obj/tao/linesearch/interface/ftn-auto/taolinesearchf.o
>          CXX mips-linux-gnu-real-debug/obj/tao/linesearch/interface/ftn-custom/ztaolinesearchf.o
>          CXX mips-linux-gnu-real-debug/obj/tao/linesearch/interface/taolinesearch.o
>          CXX mips-linux-gnu-real-debug/obj/tao/leastsquares/impls/pounders/gqt.o
>          CXX mips-linux-gnu-real-debug/obj/tao/leastsquares/impls/pounders/pounders.o
>           FC mips-linux-gnu-real-debug/obj/mat/f90-mod/petscmatmod.o
>           FC mips-linux-gnu-real-debug/obj/dm/f90-mod/petscdmmod.o
> 
> f951: out of memory allocating 2097152 bytes after a total of 1844178944 bytes
> make[4]: *** [gmakefile:164: mips-linux-gnu-real-debug/obj/dm/f90-mod/petscdmmod.o] Error 1
> -====
> 
> Looks like the fortran module for petscdm has gotten bigger than it
> used to be.
> 
> Nevertheless, on amd64 the 3.9 library is 15539008 bytes,  while 3.8 was 14727872,
> not much smaller.  The amd64 3.9 debug library is 86882368 bytes where 3.8
> was 84142848.
> 
> On mips the 3.8 debug library is 68622188 bytes, smaller than amd64.
> Is this out-of-memory error a problem with the mips buildds? Are there
> buildd memory constraints we're hitting against? Or is there a problem with
> fortran95 on mips?
>...

mips and mipsel are the only architectures where we have only 2 GB 
address space for userspace, so when the compiler needs more than
2 GB of memory you are getting this error.

On other 32bit buildds you have 3 GB or even 4 GB address space,
mips/mipsel is the canary for huge memory usage.

gcc sometimes has huge memory usage compiling C++ hitting this 2 GB 
limit, it is possible that there are also problems with Fortran code.

It is very annoying that the buildd log hides the compiler flags,
if you fix that I'll take a look regarding a workaround.

cu
Adrian

-- 

       "Is there not promise of rain?" Ling Tan asked suddenly out
        of the darkness. There had been need of rain for many days.
       "Only a promise," Lao Er said.
                                       Pearl S. Buck - Dragon Seed