On 16/11/2022 11:42, Tobias Burnus wrote:
This is a part of a patch by Andrew (hi!) - namely that part that only
adds the
__builtin_gcn_kernarg_ptr. More is planned, see below.
The short term benefit of this patch is to permit replacing hardcoded
numbers
by a builtin – like in libgomp (see patch) or in newlib (not submitted):
--- a/newlib/libc/sys/amdgcn/write.c
+++ b/newlib/libc/sys/amdgcn/write.c
@@ -59,1 +59,5 @@ _READ_WRITE_RETURN_TYPE write (int fd, const void
*buf, size_t count)
+#if defined(__has_builtin) && __has_builtin(__builtin_gcn_kernarg_ptr)
+ register void **kernargs = __builtin_gcn_kernarg_ptr ();
+#else
register void **kernargs asm("s8");
+#endif
It would also replace the 'asm("s8")' in reverse offload (GCN) patch, i.e.
https://gcc.gnu.org/pipermail/gcc-patches/2022-September/602339.html
However, this patch is only the very first step. Next one is to add
several additional builtins, namely those that are required for newlib,
i.e. newlib/libc/machine/amdgcn/mlock.c (sbrk) and
newlib/libc/machine/amdgcn/getreent.c (__getreent) use some additional
hard-coded value for heap and stack memory.
And at some point - but only after newlib has been updated -
we can think of making stack variables non-private.
That's a general goal - and in any case required for reverse
offload to be able to transfer between the host and on-device stack
variables.
* * *
Regarding the patch: Besides the obvious change (addition of the builtin),
the change to DEFAULT memory space is required to avoid a memory-space
conversion
ICE when using the new builtin. The gcn_oacc_dim_size change is mainly just
picked from Andrew's patch as it seems to be reasonable. In terms of the
libgomp
testsuite, I did not spot anything except that the -O2 run now does no
longer fail
with "libgomp: target function wasn't mapped" for
libgomp.oacc-fortran/kernels-map-1.f90 - but I am not sure it is related
or not.
In any case, the libgomp testsuite shows no fails (but the usual fails)
with the attached patch.
OK for mainline?
OK.
Andrew
Tobias
PS: The plan is to have at least all builtins in GCC and use them in
newlib by at
the end of this year (i.e. in newlib's end of year snapshot - aka as annual
release).
PPS: I wonder whether
[Patch] libgomp/gcn: Prepare for reverse-offload callback handling
https://gcc.gnu.org/pipermail/gcc-patches/2022-September/602339.html
would be okay after this patch - with the asm("s8") replaced by the
builtin - or not.
The code itself would be fine, but it is unreachable until
GOMP_OFFLOAD_get_num_devices accepts reverse offload and the latter depends
on the support for non-private stack variables.
