On 11/12/2015 03:59 PM, Alexander Monakov wrote:
On Thu, 12 Nov 2015, Bernd Schmidt wrote:
I've run it through make -k check-c regtesting. These are new fails, all
mysterious:
These would have to be investigated first.
Any specific suggestions? The PTX code emitted from GCC differs only in
prologue/epilogue, so whatever's broken... I think is unlikely due to this
change. I can give it another try after upgrading CUDA driver and cuda-gdb
from 7.0 to latest.
Yeah, load it into cuda-gdb, that may help show what's happening.
+ fprintf (file, "\tmul%s.u32 %%fstmp1, %%fstmp0, %d;\n",
+ bits == 64 ? ".wide" : "", bits);
Use a shift.
I think mul is acceptable here: PTX JIT is handling it properly, according to
what I saw while investigating in cuda-gdb. If I used a shift, I'd also have
to introduce another instruction for a widening integer conversion in the
64-bit case. Do you insist?
Nah, it's fine.
This is crt0.s, which is linked in only for single-threaded testing with
-mmainkernel; for OpenMP, the intention is to handle it in the file that
implements libgomp_nvptx_main.
Yeah, that's what I meant. It might be nice to see that too if it
already exists.
Bernd
