search

gdtk-uq / gdtk

The Gas Dynamics Toolkit (GDTk) is a set of software tools for simulating high speed fluid flow, maintained at The University of Queensland and the University of Southern Queensland, Australia.
https://gdtk.uqcloud.net/
Other
59 stars 15 forks source link

LDC compiler gives garbage instead of stack trace when exceptions are thrown #29

Closed uqngibbo closed 1 year ago

uqngibbo commented 1 year ago

By default the LDC compiler doesn't include the symbols needed for a nice stack trace, even when compiling with -g

Example:

core.exception.RangeError@sfluidblock.d(2689): Range violation
----------------
??:? [0x5567cb998a25]
??:? [0x5567cb9c257e]
??:? [0x5567cb9a3ddd]
??:? [0x5567cb9980f8]
??:? [0x5567cb9987e7]
??:? [0x5567cb712f8e]
??:? [0x5567cb7169e0]
??:? [0x5567cb914532]
??:? [0x5567cb7bac66]
??:? [0x5567cb7967ae]
??:? [0x5567cb735cc5]
??:? [0x5567cb901fde]
??:? [0x5567cb92617f]
??:? [0x5567cb9a3a5f]
??:? [0x5567cb9a3952]
??:? [0x5567cb9a37ad]
??:? [0x5567cb926684]
??:? __libc_start_main [0x7f829e299082]
??:? [0x5567cb25c5ad]

This LDC issue (https://github.com/ldc-developers/ldc/issues/863) explains that adding -L-export-dynamic to the compile returns the expected output, which it does:

core.exception.RangeError@sfluidblock.d(2689): Range violation
----------------
??:? @nogc void sfluidblock.SFluidBlock.first_order_flux_calc(ulong) [0x5604b1957b53]
??:? @nogc void sfluidblock.SFluidBlock.convective_flux_phase0(bool, ulong) [0x5604b195bbe0]
??:? int steadystate_core.evalRHS(double, int).__foreachbody5(ref fluidblock.FluidBlock) [0x5604b1bce812]
??:? int std.parallelism.doSizeZeroCase!(fluidblock.FluidBlock[], int delegate(ref fluidblock.FluidBlock)).doSizeZeroCase(ref std.parallelism.ParallelForeach!(fluidblock.FluidBlock[]).ParallelForeach, int delegate(ref fluidblock.FluidBlock)) [0x5604b1a36ba6]
??:? int std.parallelism.ParallelForeach!(fluidblock.FluidBlock[]).ParallelForeach.opApply(scope int delegate(ref fluidblock.FluidBlock)) [0x5604b19ec69e]
??:? void steadystate_core.evalRHS(double, int) [0x5604b197b015]
??:? void steadystate_core.iterate_to_steady_state(int, int, int, bool, bool) [0x5604b1bbc15e]
??:? _Dmain [0x5604b1be2a4f]

The stated reason for this is that the option increases the size of the executable (it's off by default), but I have found that this only changes the size of e4-nk-shared from 18M to 25M. If there's no performance penalty for doing so, I propose including -L-export-dynamic in the debug flavour of the code.

uqngibbo commented 1 year ago

Test results with commit 8b5e6e24, using the example code in examples/eilmer/2D/flat-plate-turbulent-mabey/su2-steady-state-solver, compiled with debug mode and run with a single thread:

Execution speed (three trials):

 $ time e4-nk-shared --job=mabey --verbosity=1 --max-cpus=1

With -L-export-dynamic -link-defaultlib-debug
real    0m38.405s, 0m38.386s, 0m38.046s

Without (reference commit 8b5e6e24)
real    0m36.413s, 0m37.048s, 0m37.270s

File sizes:
                      WITH   WITHOUT
  e4mpi*              28M      17M
  e4-nk-dist*         30M      19M
  e4-nk-dist-real*    29M      17M
  e4-nk-shared*       30M      18M
  e4-nk-shared-real*  29M      17M
  e4shared*           28M      17M
  e4zmpi*             30M      18M
  e4zshared*          30M      18M
uqngibbo commented 1 year ago

Note that the above stack trace does not have line numbers/filenames in it. This is because of a bug in older ldc2 compilers:

compile v1.24 and 1.28:

object.Error@(0): RIP
----------------
??:? @nogc void ufluidblock.UFluidBlock.convective_flux_phase2(bool, ulong, fvcell.FVCell[], fvinterface.FVInterface[], fvvertex.FVVertex[]) [0x562ca66d6ad4]

Compile with v1.32:

object.Error@(0): RIP
----------------
ufluidblock.d:887 @nogc void ufluidblock.UFluidBlock.convective_flux_phase2(bool, ulong, fvcell.FVCell[], fvinterface.FVInterface[], fvvertex.FVVertex[]) [0x564391322942]

A workaround for older compilers is to use the full path:

$ /home/uqngibbo/programs/gdtk/bin/e4-nk-shared --job=mabey --verbosity=1 --max-cpus=1

object.Error@(0): RIP
----------------
/home/uqngibbo/source/gdtk.fresh/src/eilmer/ufluidblock.d:887 @nogc void ufluidblock.UFluidBlock.convective_flux_phase2(bool, ulong, fvcell.FVCell[], fvinterface.FVInterface[], fvvertex.FVVertex[]) [0x559d54b31ad4]
uqngibbo commented 1 year ago

I think that the above noted performance penalty and exe file sizes are acceptable in debug mode, so I have added the new flags in commit 34851ea4.