https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97203
--- Comment #3 from Tom de Vries <vries at gcc dot gnu.org> ---
[ Note, this is with GOMP_NVPTX_JIT=-O0. ]
In sinf, we have:
...
45: return -__kernel_cosf(y[0],y[1]);
...
which translates to:
...
.loc 1 45 12
ld.f32 %r67,[%frame+4];
ld.f32 %r65,[%frame];
{
.param .f32 %value_in;
.param .f32 %out_arg1;
st.param.f32 [%out_arg1],%r65;
.param .f32 %out_arg2;
st.param.f32 [%out_arg2],%r67;
call (%value_in),__kernel_cosf,(%out_arg1,%out_arg2);
ld.param.f32 %r68,[%value_in];
}
.loc 1 45 11
neg.f32 %r37,%r68;
...
If I place (using GOMP_NVPTX_PTXRW) a trap before the first load:
...
.loc 1 45 12
+trap
ld.f32 %r67,[%frame+4];
...
I get:
...
libgomp: cuCtxSynchronize error: an illegal instruction was encountered
...
If I place it after the first load, I get:
...
libgomp: cuCtxSynchronize error: an illegal memory access was encountered
...
