[Bug c++/83046] ICE in nvptx offloading, C++ compilation of libgomp.oacc-c-c++-common/gang-static-2.c

2017-12-28 Thread vries at gcc dot gnu.org 
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83046

--- Comment #16 from Tom de Vries  ---
Created attachment 42977
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=42977=edit
Updated patch, with both openacc and openmp testcase

[Bug c++/83046] ICE in nvptx offloading, C++ compilation of libgomp.oacc-c-c++-common/gang-static-2.c

2017-12-28 Thread vries at gcc dot gnu.org 
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83046

Tom de Vries  changed:

   What|Removed |Added

   Keywords||openmp
 CC||jakub at gcc dot gnu.org

--- Comment #15 from Tom de Vries  ---
Reproduced with openmp as well.

[Bug c++/83046] ICE in nvptx offloading, C++ compilation of libgomp.oacc-c-c++-common/gang-static-2.c

2017-12-28 Thread vries at gcc dot gnu.org 
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83046

--- Comment #14 from Tom de Vries  ---
(In reply to Tom de Vries from comment #13)
> Created attachment 42974 [details]
> Tentative patch with testcase
> 
> Currently doing:
> - bootstrap and reg test on x86_64, and
> - build & libgomp regtest on x86_64 with nvptx accelerator.

Done, no issues found.

[Bug c++/83046] ICE in nvptx offloading, C++ compilation of libgomp.oacc-c-c++-common/gang-static-2.c

2017-12-27 Thread vries at gcc dot gnu.org 
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83046

--- Comment #13 from Tom de Vries  ---
Created attachment 42974
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=42974=edit
Tentative patch with testcase

Currently doing:
- bootstrap and reg test on x86_64, and
- build & libgomp regtest on x86_64 with nvptx accelerator.

[Bug c++/83046] ICE in nvptx offloading, C++ compilation of libgomp.oacc-c-c++-common/gang-static-2.c

2017-12-27 Thread vries at gcc dot gnu.org 
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83046

--- Comment #12 from Tom de Vries  ---
Tentative patch:
...
diff --git a/gcc/lto-cgraph.c b/gcc/lto-cgraph.c
index ed3df15b143..de5655abbb6 100644
--- a/gcc/lto-cgraph.c
+++ b/gcc/lto-cgraph.c
@@ -,6 +,16 @@ output_offload_tables (void)
   struct lto_simple_output_block *ob
 = lto_create_simple_output_block (LTO_section_offload_table);

+  for (unsigned i = 0; i < vec_safe_length (offload_funcs);)
+{
+  if (!cgraph_node::get((*offload_funcs)[i]))
+   {
+ offload_funcs->ordered_remove (i);
+ continue;
+   }
+  i++;
+}
+
   for (unsigned i = 0; i < vec_safe_length (offload_funcs); i++)
 {
   streamer_write_enum (ob->main_stream, LTO_symtab_tags,
...

[Bug c++/83046] ICE in nvptx offloading, C++ compilation of libgomp.oacc-c-c++-common/gang-static-2.c

2017-12-27 Thread vries at gcc dot gnu.org 
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83046

--- Comment #11 from Tom de Vries  ---
Rewriting the test-case like this triggers the ICE without -Wreturn-type being
active:
...
/* { dg-do run { target openacc_nvidia_accel_selected } } */
/* This code uses nvptx inline assembly guarded with acc_on_device, which is
   not optimized away at -O0, and then confuses the target assembler.
   { dg-skip-if "" { *-*-* } { "-O0" } { "" } } */

#include 
#include 

#define N 100

#define GANG_ID(I)  \
  (acc_on_device (acc_device_nvidia)\
   ? ({unsigned __r;\
   __asm__ volatile ("mov.u32 %0,%%ctaid.x;" : "=r" (__r)); \
   __r; }) : (I))

int
main ()
{
  int a[N];
  int i, x;
  int c;

  c = 1;
#pragma acc parallel loop gang (static:*) num_gangs (10)
  for (i = 0; i < 100; i++)
a[i] = GANG_ID (i);

  if (c)
__builtin_unreachable ();

#pragma acc parallel loop gang (static:1) num_gangs (10)
  for (i = 0; i < 100; i++)
a[i] = GANG_ID (i);


  return 0;
}
...

[Bug c++/83046] ICE in nvptx offloading, C++ compilation of libgomp.oacc-c-c++-common/gang-static-2.c

2017-12-27 Thread vries at gcc dot gnu.org 
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83046

Tom de Vries  changed:

   What|Removed |Added

 Target||nvptx

--- Comment #10 from Tom de Vries  ---
(In reply to Martin Liška from comment #8)
> (In reply to Martin Liška from comment #7)
> > Created attachment 42667 [details]
> > Untested patch
> 
> Can you please Thomas test it and send to mailing list (and commit).
> Thanks.

I've tested and committed the work-around, which re-enables the test-case.

Of course the ICE still needs to be fixed.

[Bug c++/83046] ICE in nvptx offloading, C++ compilation of libgomp.oacc-c-c++-common/gang-static-2.c

2017-12-26 Thread vries at gcc dot gnu.org 
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83046

--- Comment #9 from Tom de Vries  ---
Author: vries
Date: Wed Dec 27 07:50:04 2017
New Revision: 256008

URL: https://gcc.gnu.org/viewcvs?rev=256008=gcc=rev
Log:
Workaround PR83046 in gang-static-2.c

2017-12-27  Tom de Vries  

PR c++/83046
* testsuite/libgomp.oacc-c-c++-common/gang-static-2.c (test_static)
(test_nonstatic): Fix return type to workaround PR83046.

Modified:
trunk/libgomp/ChangeLog
trunk/libgomp/testsuite/libgomp.oacc-c-c++-common/gang-static-2.c

[Bug c++/83046] ICE in nvptx offloading, C++ compilation of libgomp.oacc-c-c++-common/gang-static-2.c

2017-11-21 Thread marxin at gcc dot gnu.org 
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83046

Martin Liška  changed:

   What|Removed |Added

 Status|WAITING |NEW

--- Comment #8 from Martin Liška  ---
(In reply to Martin Liška from comment #7)
> Created attachment 42667 [details]
> Untested patch

Can you please Thomas test it and send to mailing list (and commit).
Thanks.

[Bug c++/83046] ICE in nvptx offloading, C++ compilation of libgomp.oacc-c-c++-common/gang-static-2.c

2017-11-21 Thread marxin at gcc dot gnu.org 
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83046

--- Comment #7 from Martin Liška  ---
Created attachment 42667
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=42667=edit
Untested patch

[Bug c++/83046] ICE in nvptx offloading, C++ compilation of libgomp.oacc-c-c++-common/gang-static-2.c

2017-11-20 Thread tschwinge at gcc dot gnu.org 
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83046

--- Comment #6 from Thomas Schwinge  ---
(In reply to Martin Liška from comment #5)
> Thanks for instructions, but apparently does not work for me:
> 
> make check-target-libgomp

>   === libgomp Summary ===
> 
> # of untested testcases   2

Oh, sorry, that must be the tests being skipped when you don't have a Nvidia
GPU.  Please remove the "dg-do run { target openacc_nvidia_accel_selected }"
directive from the test case file.

[Bug c++/83046] ICE in nvptx offloading, C++ compilation of libgomp.oacc-c-c++-common/gang-static-2.c

2017-11-20 Thread marxin at gcc dot gnu.org 
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83046

--- Comment #5 from Martin Liška  ---
Thanks for instructions, but apparently does not work for me:

make check-target-libgomp
RUNTESTFLAGS='--target_board=unix\{-foffload=disable\}
{c,c++}.exp=gang-static-2.c'
make[1]: Entering directory
'/home/marxin/Programming/gcc/objdir2/x86_64-pc-linux-gnu/libgomp'
Making check in testsuite
make[2]: Entering directory
'/home/marxin/Programming/gcc/objdir2/x86_64-pc-linux-gnu/libgomp/testsuite'
make  check-DEJAGNU
make[3]: Entering directory
'/home/marxin/Programming/gcc/objdir2/x86_64-pc-linux-gnu/libgomp/testsuite'
srcdir='../../../../libgomp/testsuite'; export srcdir; \
EXPECT=expect; export EXPECT; \
runtest=runtest; \
if /bin/sh -c "$runtest --version" > /dev/null 2>&1; then \
  exit_status=0; l='libgomp'; for tool in $l; do \
if $runtest  --tool $tool --srcdir $srcdir
--target_board=unix\{-foffload=disable\} {c,c++}.exp=gang-static-2.c; \
then :; else exit_status=1; fi; \
  done; \
else echo "WARNING: could not find \`runtest'" 1>&2; :;\
fi; \
exit $exit_status
Test run by marxin on Mon Nov 20 15:03:00 2017
Native configuration is x86_64-pc-linux-gnu

=== libgomp tests ===

Schedule of variations:
unix/-foffload=disable

Running target unix/-foffload=disable
Using /usr/share/dejagnu/baseboards/unix.exp as board description file for
target.
Using /usr/share/dejagnu/config/unix.exp as generic interface file for target.
Using ../../../../libgomp/testsuite/config/default.exp as
tool-and-target-specific interface file.
Running ../../../../libgomp/testsuite/libgomp.c/c.exp ...
Running ../../../../libgomp/testsuite/libgomp.c++/c++.exp ...
Running ../../../../libgomp/testsuite/libgomp.hsa.c/c.exp ...
Running ../../../../libgomp/testsuite/libgomp.oacc-c/c.exp ...
Running ../../../../libgomp/testsuite/libgomp.oacc-c++/c++.exp ...

=== libgomp Summary ===

# of untested testcases 2
# of unsupported tests  2
make[3]: Leaving directory
'/home/marxin/Programming/gcc/objdir2/x86_64-pc-linux-gnu/libgomp/testsuite'
make[2]: Leaving directory
'/home/marxin/Programming/gcc/objdir2/x86_64-pc-linux-gnu/libgomp/testsuite'
make[2]: Entering directory
'/home/marxin/Programming/gcc/objdir2/x86_64-pc-linux-gnu/libgomp'
true  DO=all multi-do # make
:
:
:
make[2]: Leaving directory
'/home/marxin/Programming/gcc/objdir2/x86_64-pc-linux-gnu/libgomp'
make[1]: Leaving directory
'/home/marxin/Programming/gcc/objdir2/x86_64-pc-linux-gnu/libgomp'

$ ./gcc/xgcc -v
Using built-in specs.
COLLECT_GCC=./gcc/xgcc
OFFLOAD_TARGET_NAMES=nvptx-none
Target: x86_64-pc-linux-gnu
Configured with: ../configure --enable-languages=c,c++,fortran
--disable-multilib --prefix=/home/marxin/bin/gcc --disable-bootstrap
--enable-offload-targets=nvptx-none=/no/where
Thread model: posix
gcc version 8.0.0 20171120 (experimental) (GCC)

[Bug c++/83046] ICE in nvptx offloading, C++ compilation of libgomp.oacc-c-c++-common/gang-static-2.c

2017-11-20 Thread tschwinge at gcc dot gnu.org 
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83046

--- Comment #4 from Thomas Schwinge  ---
(In reply to Martin Liška from comment #3)
> (In reply to Thomas Schwinge from comment #0)
> > ... starting with r254437 "Instrument function exit with
> > __builtin_unreachable in C++".

> > Obviously, that test case has a "-Wreturn-type mismatch" (which is not
> > diagnosed for C++; see reduced PR83045), and fixing that cures this nvptx
> > offloading ICE.
> 
> Can you please provide steps to reproduce this?

You'll either have to build an offloading compiler, but -- much simpler! -- I
assume it will be sufficient to resolve the issue seen in the
"-foffload=disable" case:

Configure GCC as usual, but add "--enable-offload-targets=nvptx-none=/no/where"
(to generally enable offloading, but without providing an actual nvptx
offloading compiler), build, and then run:

$ make check-target-libgomp
RUNTESTFLAGS='--target_board=unix\{-foffload=disable\}
{c,c++}.exp=gang-static-2.c'
[...]
Running target unix/-foffload=disable
Using /usr/share/dejagnu/baseboards/unix.exp as board description file for
target.
Using /usr/share/dejagnu/config/unix.exp as generic interface file for
target.
Using ../../../../source-gcc/libgomp/testsuite/config/default.exp as
tool-and-target-specific interface file.
Running ../../../../source-gcc/libgomp/testsuite/libgomp.c/c.exp ...
Running ../../../../source-gcc/libgomp/testsuite/libgomp.c++/c++.exp ...
Running ../../../../source-gcc/libgomp/testsuite/libgomp.hsa.c/c.exp ...
Running ../../../../source-gcc/libgomp/testsuite/libgomp.oacc-c/c.exp ...
FAIL: libgomp.oacc-c/../libgomp.oacc-c-c++-common/gang-static-2.c
-DACC_DEVICE_TYPE_nvidia=1 -DACC_MEM_SHARED=0  -O2  execution test
Running ../../../../source-gcc/libgomp/testsuite/libgomp.oacc-c++/c++.exp
...
FAIL: libgomp.oacc-c++/../libgomp.oacc-c-c++-common/gang-static-2.c
-DACC_DEVICE_TYPE_nvidia=1 -DACC_MEM_SHARED=0  -O2  (test for excess errors)

=== libgomp Summary ===

# of expected passes1
# of unexpected failures2
# of unresolved testcases   1
# of unsupported tests  4
[...]

Compilation should either succeed (as it still does for C, and did before for
C++), or terminate with a suitable diagnostic (the issue discussed in PR83045
"-Wreturn-type regression in C++").

For "c.exp" the "execution test" FAILs: "libgomp: target function wasn't
mapped" (expected in this special configuration), but for c++.exp you'll see
compilation FAIL with "excess errors":

/tmp/ccBzwk9H.o:(.gnu.offload_funcs+0x0): undefined reference to
`main._omp_fn.7'
/tmp/ccBzwk9H.o:(.gnu.offload_funcs+0x8): undefined reference to
`main._omp_fn.6'
/tmp/ccBzwk9H.o:(.gnu.offload_funcs+0x10): undefined reference to
`main._omp_fn.5'
/tmp/ccBzwk9H.o:(.gnu.offload_funcs+0x18): undefined reference to
`main._omp_fn.4'
/tmp/ccBzwk9H.o:(.gnu.offload_funcs+0x20): undefined reference to
`main._omp_fn.3'
/tmp/ccBzwk9H.o:(.gnu.offload_funcs+0x28): undefined reference to
`main._omp_fn.2'
/tmp/ccBzwk9H.o:(.gnu.offload_funcs+0x30): undefined reference to
`main._omp_fn.1'
collect2: error: ld returned 1 exit status
compiler exited with status 1

(I'll probably not be able to look into that myself in the next few weeks.)

[Bug c++/83046] ICE in nvptx offloading, C++ compilation of libgomp.oacc-c-c++-common/gang-static-2.c

2017-11-20 Thread marxin at gcc dot gnu.org 
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83046

--- Comment #3 from Martin Liška  ---
(In reply to Thomas Schwinge from comment #0)
> ... starting with r254437 "Instrument function exit with
> __builtin_unreachable in C++".
> 
> Program received signal SIGSEGV, Segmentation fault.
> input_offload_tables (do_force_output=true) at
> [...]/source-gcc/gcc/lto-cgraph.c:1942
> 1942varpool_node::get (var_decl)->force_output = 1;
> (gdb) bt
> #0  input_offload_tables (do_force_output=true) at
> [...]/source-gcc/gcc/lto-cgraph.c:1942
> #1  0x005a0927 in read_cgraph_and_symbols (fnames= out>, nfiles=) at [...]/source-gcc/gcc/lto/lto.c:2863
> #2  lto_main () at [...]/source-gcc/gcc/lto/lto.c:3314
> #3  0x00a7f63f in compile_file () at
> [...]/source-gcc/gcc/toplev.c:455
> #4  0x0056ef20 in do_compile () at
> [...]/source-gcc/gcc/toplev.c:2059
> #5  toplev::main (this=this@entry=0x7fffcfb0, argc=argc@entry=15,
> argv=0x17908e0, argv@entry=0x7fffd0b8) at
> [...]/source-gcc/gcc/toplev.c:2194
> #6  0x00571457 in main (argc=15, argv=0x7fffd0b8) at
> [...]/source-gcc/gcc/main.c:39
> 
> Obviously, that test case has a "-Wreturn-type mismatch" (which is not
> diagnosed for C++; see reduced PR83045), and fixing that cures this nvptx
> offloading ICE.

Hello.

Can you please provide steps to reproduce this?

[Bug c++/83046] ICE in nvptx offloading, C++ compilation of libgomp.oacc-c-c++-common/gang-static-2.c

2017-11-19 Thread tschwinge at gcc dot gnu.org 
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83046

--- Comment #2 from Thomas Schwinge  ---
Also, for "-foffload=disable -x c++" we now run into:

/tmp/ccAXdy3B.o:(.gnu.offload_funcs+0x0): undefined reference to
`main._omp_fn.7'
/tmp/ccAXdy3B.o:(.gnu.offload_funcs+0x8): undefined reference to
`main._omp_fn.6'
/tmp/ccAXdy3B.o:(.gnu.offload_funcs+0x10): undefined reference to
`main._omp_fn.5'
/tmp/ccAXdy3B.o:(.gnu.offload_funcs+0x18): undefined reference to
`main._omp_fn.4'
/tmp/ccAXdy3B.o:(.gnu.offload_funcs+0x20): undefined reference to
`main._omp_fn.3'
/tmp/ccAXdy3B.o:(.gnu.offload_funcs+0x28): undefined reference to
`main._omp_fn.2'
/tmp/ccAXdy3B.o:(.gnu.offload_funcs+0x30): undefined reference to
`main._omp_fn.1'
collect2: error: ld returned 1 exit status

..., which also suggests some C++ front end/middle end offloading
inconsistency.

[Bug c++/83046] ICE in nvptx offloading, C++ compilation of libgomp.oacc-c-c++-common/gang-static-2.c

2017-11-19 Thread tschwinge at gcc dot gnu.org 
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83046

Thomas Schwinge  changed:

   What|Removed |Added

 Status|UNCONFIRMED |WAITING
   Last reconfirmed||2017-11-19
 Ever confirmed|0   |1

--- Comment #1 from Thomas Schwinge  ---
I suggest we first analyze PR83045.