On 2022-02-04T13:09:29+0100, Tom de Vries via Gcc <g...@gcc.gnu.org> wrote:
On 2/4/22 08:21, Thomas Schwinge wrote:
On 2022-02-03T13:35:55+0000, "vries at gcc dot gnu.org via Gcc-bugs"
<gcc-b...@gcc.gnu.org> wrote:
I've tested this using (recommended) driver 470.94 on boards:
while iterating over dimensions { -mptx=3.1 , -mptx=6.3 } x { GOMP_NVPTX_JIT=-O0,
<default> }.
Do you use separate (nvptx-none offload target only?) builds for
different '-mptx' variants (likewise: '-misa'), or have you hacked up the
multilib configuration?
Neither, I'm using --target_board=unix/foffload= for that.
ACK, I see. So these flags then only affect GCC/nvptx code generation
for the actual user code (here: GCC libgomp test cases), but for the
GCC/nvptx target libraries (such as: libc, libm, libgfortran, libgomp --
the latter especially relevant for OpenMP), it uses PTX code from one of
the two "pre-compiled" GCC/nvptx multilibs: default or '-mptx=3.1'.
Meaning, one can't just use such a flag for "completely building code"
for a specific configuration. Random example,
'-foffload-options=nvptx-none=-march=sm_75': as GCC/nvptx target
libraries aren't being built for '-march=sm_75' multilib,
'-foffload-options=nvptx-none=-march=sm_75' uses the default multilib,
which isn't '-march=sm_75'.
('gcc/config/nvptx/t-nvptx:MULTILIB_OPTIONS'
etc., I suppose?) Should we add a few representative configurations to
be built by default? And/or, should we have a way to 'configure' per
user needs (I suppose: '--with-multilib-list=[...]', as supported for a
few other targets?)? (I see there's also a new
'--with-multilib-generator=[...]', haven't looked in detail.) No matter
which way: again, combinatorial explosion is a problem, of course...
As far as I know, the gcc build doesn't finish when switching default to
higher than sm_35, so there's little point to go to a multilib setup at
this point. But once we fix that, we could reconsider, otherwise,
things are likely to regress again.
As far as I remember, several issues have been fixed. Still waiting for
Roger's "middle-end: Support ABIs that pass FP values as wider integers"
or something similar, but that PR104489 issue is being worked around by
"Limit HFmode support to mexperimental", if I got that right.
Now I'm not suggesting we should now enable all or any random GCC/nvptx
multilibs, to get all these variants of GCC/nvptx target libraries built;
especially also given that GCC/nvptx code generation currently doesn't
make too much use of the new capabilities.
However, we do have a specific request that a customer would like to be
able to change at GCC 'configure' time the GCC/nvptx default multilib
(including that being used for building corresponding GCC/nvptx target
libraries).
Per 'gcc/doc/install.texi', I do see that some GCC targets allow for
GCC 'configure'-time '--with-multilib-list=[...]', or
'--with-multilib-generator=[...]', and I suppose we could be doing
something similar? But before starting implementing, I'd like your
input, as you'll be the one to approve in the end. And/or, maybe you've
already made up your own ideas about that?