[Rd] Byte-compilation failure on different architectures / low-memory systems

Tomas Kalibera tom@@@k@liber@ @ending from gm@il@com
Sat Jun 9 10:50:18 CEST 2018

It turned out slightly more complicated, the fallback to AST is was 
already in place, but memory available to malloc got exhausted via heap 
expansion. Messages starting "Error: compilation failed" mean there was 
an error while compiling, including out of memory, and that the AST will 
be used - so these are worth looking at but not fatal. The error message 
"cannot allocate buffer" was the real problem.

We have addressed that with Luke in R-devel and R-patched (more general 
fix might come later to R-devel only). Now there is a GC on the fallback 
path which should release some memory from R heap to be usable by malloc 
again. Also, the compiler needs a bit less memory when analyzing similar 
code patterns as those causing trouble in fBasics (solves the fBasics 
problems with "ulimit -v 1000000")


On 06/05/2018 08:59 PM, Dirk Eddelbuettel wrote:
> On 4 June 2018 at 20:06, Tomas Kalibera wrote:
> | thanks for the report. Access to the test system is not necessary, the
> | memory requirements of the byte-code compiler are usually
> | platform-independent and specifically with this package I can reproduce
> | they are very high. We'll have a look what we can do, certainly there
> | should at least be a way to recover and use the uncompiled version when
> | memory allocation fails, this is already done by the JIT but not when
> | compiling during installation. Before R or the package is patched, the
> | only workaround for memory constrained systems is probably to disable
> | byte-compilation of this package, as I read you are doing already.
> Yes. And as a shortcut, we just turned it off unconditionally, ie on all
> build architectures.  Worked fine as per
>     https://buildd.debian.org/status/package.php?p=fbasics
> it has been built everywhere where we have R 3.5.0 (some 20 or so platforms).
> The fix you suggest sounds ideal: if possible recover, and maybe WARN.
> Dirk

More information about the R-devel mailing list