run4flat / perl_nvcc Goto Github PK

The fragment of code in sub linker, around line 152 in ExtUtils::nvcc::Backend, that composes "-Xlinker=..." parameter shouln't use unfiltered @other_args, but should translate GCC convention of passing options to linker via -Wl,<opt1> -Wl,<opt2> to nvcc convention of passing linker oprions via -Xlinker=<opt1>,<opt2>.

The code currently reads:

    # Set up the flags for the compiler arguments:
    unshift @nvcc_args, ("-Xlinker=" . join ',', @other_args)
        if @other_args;

It should instead be something like the following

    # Set up the flags for the compiler arguments:
    unshift @nvcc_args, ("-Xlinker=" . join ',', map { s/^-Wl,// ? $_ : () } @other_args)
        if @other_args;

Though perhaps it shoul be refactored in separate function, e.g. process_linker_args, or at least a separate variable, e.g. @linker_args/

A long-standing issue in CUDA::Minimal was recently solved with a simple redefinition of dNOOP, in accordance with the definition currently used in Perl's bleeding branch. See run4flat/perl-CUDA-Minima@ff3b9900cf20b4b02edbc3fa2b16931e38f0d57a. It would be easy to have EU::nvcc patch c source code with this redefinition for offending versions of Perl, and such behavior would be in line with the stated purpose of EU::nvcc.