Chapel: internal error: MIS0558 chpl Version 1.16.0

[HouseK]

Summary of Problem

main.chpl:28: internal error: MIS0558 chpl Version 1.16.0 Note: This source location is a guess.

Internal errors indicate a bug in the Chapel compiler ("It's us, not you"), and we're sorry for the hassle. We would appreciate your reporting this bug -- please see http://chapel.cray.com/bugs.html for instructions. In the meantime, the filename + line number above may be useful in working around the issue.

Steps to Reproduce

I'm using the docker image chapel/chapel:latest
version is 1.16.0 at this moment.

Source Code:
https://github.com/HouseK/k_means/tree/MIS0558

Compile command:
chpl -o output/main main.chpl
(see run.sh)

Execution command:
output/main
(see run.sh)

Configuration Information

• Output of chpl --version:
  chpl Version 1.16.0
  Copyright (c) 2004-2017, Cray Inc. (See LICENSE file for more details)

• Output of $CHPL_HOME/util/printchplenv --anonymize:
  HPL_TARGET_PLATFORM: linux64
  CHPL_TARGET_COMPILER: gnu
  CHPL_TARGET_ARCH: native
  CHPL_LOCALE_MODEL: flat
  CHPL_COMM: none
  CHPL_TASKS: qthreads
  CHPL_LAUNCHER: none
  CHPL_TIMERS: generic
  CHPL_UNWIND: none
  CHPL_MEM: jemalloc
  CHPL_MAKE: make
  CHPL_ATOMICS: intrinsics
  CHPL_GMP: system *
  CHPL_HWLOC: hwloc
  CHPL_REGEXP: re2
  CHPL_WIDE_POINTERS: struct
  CHPL_AUX_FILESYS: none

• Back-end compiler and version, e.g. gcc --version or clang --version:
  gcc (Debian 6.3.0-18) 6.3.0 20170516
  Copyright (C) 2016 Free Software Foundation, Inc.
  This is free software; see the source for copying conditions. There is NO
  warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

[ben-albrecht]

@HouseK - thanks for reporting this.

I compiled with --devel and found that the internal error was caused by an unresolved call:

      res[i] = addAverageAccumulator(ma1[i], ma2[i]);

should actually be:

      res[i] = addAverageAccumulators(ma1[i], ma2[i]);

That should help you get past this for now.

[HouseK]

Thanks for the solution!

[bradcray]

@ben-albrecht: I think we should re-open this since it is a bug (we shouldn't be generating an internal error). Let me know if you disagree (I realize you didn't close it, but you didn't reopen it either). Thanks.

[ben-albrecht]

I think we should re-open this since it is a bug (we shouldn't be generating an internal error). Let me know if you disagree (I realize you didn't close it, but you didn't reopen it either). Thanks.

Agreed. I was hoping to distill the example down further and open a new issue with a more concrete description, but I haven't gotten as far as I'd like yet.

Here's how far I got on the distillation so far:

proc main() {
  param num_means = 5;
  var D : [0..99999] real;
  var M: [1..num_means] real;
  M = [1,2,3,4,5];
  forall a in (AddMeansAccumulators reduce toMeansAccumulator(M, D)) do a;
}

class AddMeansAccumulators : ReduceScanOp {
  type eltType;
  var ma : eltType;

  proc accumulate(val: [?] (real,int)){
    forall i in ma.domain do
      unresolvedCall();
  }
}

proc toMeansAccumulator(means: [?dom] real , elem: real) {
  var res: [dom] (real, int);
  return res;
}

The error seems to depend on being an unresolved call being inside a forall loop that is within a custom reduce intent class. Also, declaring an array with over [dom] (real, int) instead of relying on toMeansAccumulator avoids the internal error.

[HouseK]

I closed it because my user issue was fixed but I'm happy to reopen it :)

[cassella]

I can't reproduce this with the original code or the distillation.

Instead, I get (for the original code)

fortytwo@magrathea:~/src/chapel (master)$ chpl main.chpl
main.chpl:54: In function 'kmeans':
main.chpl:62: error: unresolved call 'unmanaged AddMeansAccumulators(_array(unmanaged DefaultRectangularArr(1,int(64),false,(real(64),int(64)),int(64)))).identity()'
main.chpl:62: note:   instantiated from kmeans
$CHPL_HOME/modules/internal/ChapelReduce.chpl:162: note: candidates are: SumReduceScanOp.identity [79584]
...

[bradcray]

I believe this reflects that the user-defined reduction interface has evolved since the test was written and now requires an identity() method. So the question is whether, upon supplying it, the original code works or breaks.

[cassella]

I modified test/reductions/userdefined.chpl, adding to each class
method (one at a time) a forall loop over the vals array calling an
undefined function, e.g.

  proc accumulate(value: eltType) {
    forall a in state.vals do unresolvedCall();
    accumulateOntoState(state, value);
  }

All of them failed with error: unresolved call 'unresolvedCall()'.

Except the one above, which compiled fine. I guess that accumulate
variant isn't used? (It compiles and runs without it entirely.)

[bradcray]

It may depend on which reduction / scan forms you use (e.g., standalone operator versus reduce intents). Or it may just not be used anymore. The UDR (User-Defined Reduction) interface is still heavily in flux, but we hope to stabilize in this coming release cycle.

[vasslitvinov]

It may well be that one of the accumulate() methods is not used.

We are switching to a new reduction interface soon, so hopefully the quirks of the current interface will not live for much longer. Except for users of the 1.19 release distro. :(

[ben-albrecht]

I added a clone and identity method (both required), and was not able to reproduce the original internal error from my previous reproducer:

proc main() {
  param num_means = 5;
  var D : [0..99999] real;
  var M: [1..num_means] real;
  M = [1,2,3,4,5];
  forall a in (AddMeansAccumulators reduce toMeansAccumulator(M, D)) do a;
}

class AddMeansAccumulators : ReduceScanOp {
  type eltType;
  var ma : eltType;

  proc accumulate(val: [?] (real,int)){
    forall i in ma.domain do
      unresolvedCall();
  }

  proc clone() { return new unmanaged AddMeansAccumulators(eltType=eltType); }
  proc identity() {
    var res: [1..0] (real, int);
    return res;
  }
}

proc toMeansAccumulator(means: [?dom] real , elem: real) {
  var res: [dom] (real, int);
  return res;
}

> chpl repro.chpl
repro.chpl:13: In function 'accumulate':
repro.chpl:15: error: unresolved call 'unresolvedCall()'

Maybe we should close this?

[bradcray]

Works for me, thanks Ben! (and Paul)