Re: [PATCH] nvptx newlib port

[Jeff Johnston <jjohnstn at redhat dot com>]

Looks fine.  I might suggest you add a check in configure.host that issues
an error or warning message if the elix-level is not set to 1 and
--enable-global-io-streams is not set
(if what you said above is required).  Other than that, I can check it in
shortly.

-- Jeff J.

On Fri, Mar 30, 2018 at 4:22 PM, Cesar Philippidis <cesar@codesourcery.com>
wrote:

> The attached patch adds a nvptx port to newlib. At present, the nvptx
> target is comprised of a family of Nvidia GPUs with compute capability
> sm_30 and above. The primary use of newlib on nvptx targets is for use
> with OpenACC and OpenMP GPU offloading with GCC. For the past couple of
> years, Mentor Graphics has been maintaining a newlib port for nvptx here
> <https://github.com/MentorEmbedded/nvptx-newlib>.
>
> There are a couple of quirks with nvptx. Unlike traditional processors,
> nvptx targets a text-based virtual instruction set known as PTX. For
> starters, PTX does not support "circular references" where an array is
> initialized with a value consisting of an offset to itself. This happens
> in newlib/libc/reent/impure.c:impure_data. To get around that problem,
> newlib needs to be configured with --enable-newlib-global-stdio-streams.
> In practice that's fine because those stdio stream are only used for
> standalone (i.e., non-GPU offloading) testing of GCC. The important
> debug functions, printf, putchar and puts, have been remapped to PTX's
> vprintf. Besides, our original port in github was initializing those
> file streams to NULL.
>
> Because Nvidia GPUs are primarily used as offload accelerators, we
> didn't spend much time emulating systems calls for the nvptx target. As
> such, the nvptx target needs to be configured with
> --enable-newlib-elix-level=1. However, I did have to implement some
> dummy system calls write, exit, close, lseek and read. You may noticed
> that this patch also overrides assert. That's because the PTX JIT chokes
> with the call to fiprintf in assert_func. We also overrode the memory
> management functions to utilize CUDA's malloc.
>
> Beyond that, the changes in this port are relatively non-invasive. I
> compressed the patch because the regenerated configure scripts and
> Makefiles made it roughly 260k. The nvptx port first appeared in GCC 5.
> The instructions on how to build a GCC toolchain with nvptx offloading
> can be found here <https://gcc.gnu.org/wiki/Offloading>. Just be sure to
> configure newlib with the aforementioned
> --enable-newlib-global-stdio-streams --enable-newlib-elix-level=1
> options. We got around those options in the github version of newlib by
> hacking the generic newlib code with nvptx-specific workarounds.
>
> Cesar
>