PGI 20.1+ crash on test/future and when_all(one_fut).wait()

Moved from pull request #181:

Recently released PGI 20.1 gets an internal compiler error (ICE) when building test/future.cpp. The failure mode looks like this:

"/usr/local/pkg/gcc/8.3.0/include/c++/8.3.0/tuple", line 640: internal error:
assertion failed: missing default rescan info (exprutil.c, line 4731
in get_expr_rescan_info)

  _TMC<_UElements...>::template
  ^

This appears to be independent of which libstdc++ version is backing PGI (seen with all of GCC 7, 8 and 9), but appears to be specific to PGI v20.1. On our side, the problem affects versions back to at least 2019.9.0 and up through the current tip of develop @ 95500de. It may or may not be specific to x86_64. Although it involves the same header, this appears to be a distinct issue from issue ~~#290~~ (although it might involve a flawed workaround recently deployed by PGI to address that problem).

Here is a minimal program that reproduces the compiler crash back to at least UPC++ 2019.9.0:

#include <upcxx/future.hpp>

using namespace upcxx;

future<int> fib(int i) {
  return make_future(i+7);
}


int main() {
  future<int> ans0 = fib(5);

  //int x = ans0.wait(); // works
  int x = when_all(ans0).wait(); // crash

  return 0;
}

Demonstration on dirac:

$ module switch upcxx upcxx/2019.9.0/gcc-9.2.0
$ upcxx -g -c future-crash.cpp 
$ module switch pgi pgi/19.10
$ module switch upcxx upcxx/2019.9.0/pgi-19.10
$ upcxx -g -c future-crash.cpp
$ module switch pgi pgi/20.1
$ module switch upcxx upcxx/2019.9.0/pgi-20.1
$ upcxx -g -c future-crash.cpp               
"/usr/local/pkg/gcc/8.3.0/include/c++/8.3.0/tuple", line 640: internal error:
          assertion failed: missing default rescan info (exprutil.c, line 4731
          in get_expr_rescan_info)

                  _TMC<_UElements...>::template
                  ^

1 catastrophic error detected in the compilation of "future-crash.cpp".
Compilation aborted.
pgc++-Fatal-/usr/local/pkg/pgi/linux86-64-llvm/20.1/bin/pggpp1 TERMINATED by signal 6
Arguments to /usr/local/pkg/pgi/linux86-64-llvm/20.1/bin/pggpp1
/usr/local/pkg/pgi/linux86-64-llvm/20.1/bin/pggpp1 --llalign --no_warnings -Dunix -D__unix -D__unix__ -Dlinux -D__linux -D__linux__ -D__NO_MATH_INLINES -D__LP64__ -D__x86_64 -D__x86_64__ -D__LONG_MAX__=9223372036854775807L '-D__SIZE_TYPE__=unsigned long int' '-D__PTRDIFF_TYPE__=long int' -D__amd64 -D__amd64__ -D__k8 -D__k8__ -D__MMX__ -D__SSE__ -D__SSE2__ -D__SSE3__ -D__SSSE3__ -D__SSE4_1__ -D__SSE4_2__ -D__POPCNT__ -D__FXSR__ -D__PGI -D_GNU_SOURCE -D_PGCG_SOURCE -I/usr/local/pkg/upcxx-dirac/pgi-20.1/stable-2019.9.0/gasnet.debug/include -I/usr/local/pkg/upcxx-dirac/pgi-20.1/stable-2019.9.0/gasnet.debug/include/ibv-conduit -I/usr/local/pkg/upcxx-dirac/pgi-20.1/stable-2019.9.0/upcxx.debug.gasnet_seq.ibv/include -I/usr/local/pkg/openmpi-4.0.2/pgi-20.1/include -I- -I/usr/local/pkg/pgi/linux86-64-llvm/20.1/include-gcc70 -I/usr/local/pkg/pgi/linux86-64-llvm/20.1/include/nvomp -I/usr/local/pkg/pgi/linux86-64-llvm/20.1/include -I/usr/local/pkg/gcc/8.3.0/include/c++/8.3.0 -I/usr/local/pkg/gcc/8.3.0/include/c++/8.3.0/x86_64-pc-linux-gnu -I/usr/local/pkg/gcc/8.3.0/include/c++/8.3.0/backward -I/usr/local/pkg/gcc/8.3.0/lib/gcc/x86_64-pc-linux-gnu/8.3.0/include -I/usr/local/include -I/usr/local/pkg/gcc/8.3.0/include -I/usr/local/pkg/gcc/8.3.0/lib/gcc/x86_64-pc-linux-gnu/8.3.0/include-fixed -I/usr/include -I/usr/local/pkg/pgi/linux86-64-llvm/20.1/include/nvomp -I/usr/local/pkg/pgi/linux86-64-llvm/20.1/include-gcc70 -I/usr/local/pkg/pgi/linux86-64-llvm/20.1/include -I/usr/local/pkg/gcc/8.3.0/lib/gcc/x86_64-pc-linux-gnu/8.3.0/include -I/usr/local/include -I/usr/local/pkg/gcc/8.3.0/include -I/usr/local/pkg/gcc/8.3.0/lib/gcc/x86_64-pc-linux-gnu/8.3.0/include-fixed -I/usr/include -D__PGLLVM__ -D__extension__= -DUPCXX_ASSERT_ENABLED=1 -DUPCXX_BACKEND=1 -DUPCXX_BACKEND_GASNET_SEQ=1 -DUPCXX_MPSC_QUEUE_ATOMIC=1 -D_GNU_SOURCE=1 -DGASNET_SEQ -D_REENTRANT --preinclude _cplus_preinclude.h --preinclude_macros _cplus_macros.h --gnu_version=80300 -D__pgnu_vsn=80300 --c++11 -g --dwarf2 -g --dwarf2 -q -o /tmp/pgc++6PNcUA7zlAI9.il future-crash.cpp

The problem appears specific to the construct upcxx::when_all(one_future).wait() or upcxx::when_all(one_future).result(). This luckily would be a silly thing to write directly (but might be less silly in a generic programming situation).

The problem also manifests if the when_all and wait are separated with an auto temporary:

 auto x = when_all(ans0); 
 int y = x.wait(); // crash

but NOT if the second line is omitted, nor if the temporary is given a specific type:

 future<int> x = when_all(ans0);
 int y = x.wait(); // ok

so I'm guessing the crash is happening in the type inference for the expression `upcxx::when_all(one_future) and its subsequent application to wait or result.

As an ICE this is very clearly an external bug, but it's our responsibility to at least report it and possibly seek a workaround.

Comments (6)