Re: [PATCH 6/n] OpenMP 4.0 offloading infrastructure: option handling

[Ilya Verbin <iverbin@gmail.com>]

2015-09-21 18:15 GMT+03:00 Thomas Schwinge <thomas@codesourcery.com>:
> (<https://gcc.gnu.org/PR67300>, "--foffload* undocumented", has recently
> been filed.)
>
> (In the following, "intelmic" is short for
> "x86_64-intelmicemul-linux-gnu", and "nvptx" is short for "nvptx-none".)
>
> What is the syntax to use for building both intelmic and nvptx offloading
> code?  I understand we allow for separate -foffload=intelmic
> -foffload=nvptx options.  Do we also intend to allow
> -foffload=intelmic,nvptx or -foffload=intelmic:nvptx?
>
> And then, we allow for specifying offloading compiler options with
> -foffload=intelmic=[...] and -foffload=nvptx=[...]; do we also intend to
> allow -foffload=intelmic,nvptx=[...] (do the options apply to nvptx only,
> or to both intelmic and nvptx?), and/or
> -foffload=intelmic=[...],nvptx=[...], and/or
> -foffload=intelmic:nvptx=[...] (which "looks a bit like" the options
> ought to apply to nvptx only -- or to both intelmic and nvptx?), and/or
> -foffload=intelmic=[...]:nvptx=[...]?

The plan was:

1. -foffload=intelmic,nvptx=[...]  <- apply options to both intelmic,nvptx.
   Just like -foffload=[...] applies to both targets (if configured so).
2. -foffload=intelmic=[...],nvptx=[...]  <- is not allowed.
3. To apply different options to different targets, one should pass:
   -foffload=intelmic=[...] -foffload=nvptx=[...].

>       3612        /* Check that GCC is configured to support the offload target.  */
>       3613        c = OFFLOAD_TARGETS;
>       3614        while (c)
>       3615          {
>       3616            n = strchr (c, ',');
>       3617            if (n == NULL)
>       3618              n = strchr (c, '\0');
>       3619
>       3620            if (next - cur == n - c && strncmp (target, c, n - c) == 0)
>       3621              break;
>       3622
>       3623            c = *n ? n + 1 : NULL;
>       3624          }
>       3625
>       3626        if (!c)
>       3627          fatal_error (input_location,
>       3628                       "GCC is not configured to support %s as offload target",
>       3629                       target);
>
> So, this code will not do the right thing when configured with
> --enable-offload-targets=intelmic,nvptx (thus,
> OFFLOAD_TARGETS=intelmic:nvptx): using -foffload=nvptx will then result
> in "xgcc: fatal error: GCC is not configured to support nvptx as offload
> target".
>
> If I'm understanding the following code correctly, this supports the idea
> that the intention has been for -foffload=[targets]=[options] to separate
> the targets by commas, and separate the options by spaces -- is that
> correct?

Yes, targets are separated by commas, options are the whole string after the
equal sign, spaces inside are allowed.

  -- Ilya