Tpetra: GCC 5.3.0 internal compiler error

[brian-kelley]

Bug Report

@trilinos/tpetra
@csiefer2

Description

GCC 5.3.0 (on trappist, rocketman and enigma) has an internal compiler error in
Tpetra_Details_residual.hpp, line 216, even though the line looks very innocuous.

/home/nightlyTesting/OPENMPI-1.10.1_RELEASE_PERFORMANCE/Trilinos/packages/tpetra/coreTpetra_Details_residual.hpp:216:48: internal compiler error: in tsubst_copy, at cp/pt.c:13151
     residual_launch_parameters<execution_space>(A_lcl.numRows (), A_lcl.nnz (), rows_per_thread, team_size, vector_length);
                                                ^
0x65f242 tsubst_copy
	../../.cp/pt.c:13149
0x65f878 tsubst_copy_and_build(tree_node*, tree_node*, int, tree_node*, bool, bool)
	../../.cp/pt.c:15747
...etc

I think this started with commit ed2c660. This affects most of the "Experimental" CDash builds on several different machines so hopefully it's easy to reproduce and find a workaround.

Steps to Reproduce

1. SHA1: 1f4c2d is the one showing in the failing dashboard builds, but probably ed2c660 or later will have it.
2. Build with GCC 5.3, it doesn't seem like the exact configuration or machine details matter.
3. Shake fist angrily at compiler.

[csiefer2]

@mhoemmen Any ideas as to how I tickled the compiler incorrectly?

[rppawlo]

Are you using nested lambdas? Gcc 5 and 6 have a nasty bug with nested lambdas.

[csiefer2]

I am using nested lambdas, but the compiler woofs long before that.

[mhoemmen]

@csiefer2 Sometimes it helps to make code simpler -- e.g., make a temporary variable for A_lcl.numRows() etc.

[csiefer2]

@mhoemmen Since I'm mucking with it anyway, do you mind if I yank the launch parameters function out of Ifpack2's scaledDampedResidual and just have it call this?

[csiefer2]

Buh?

Tpetra_Details_residual.hpp:219:48: error: no matching function for call to ‘residual_launch_parameters(const int64_t&, const int64_t&, const int64_t&, const int&, const int&)’
residual_launch_parameters<execution_space>(numLocalRows, myNnz, rows_per_thread, team_size, vector_length);

Tpetra_Details_residual.hpp:64:1: note: candidate: template int64_t Tpetra::Details::residual_launch_parameters(int64_t, int64_t, int64_t, int&, int&)

[csiefer2]

@dsunder helped me. Yet another compiler bug! Yay!

[csiefer2]

PR merged. @brian-kelley please verify the nightlies worked and close issue, if so.