On Wed, Jan 19, 2022 at 11:29:18PM +0100, Thomas Schwinge wrote:
> (The pass is still disabled by default, by the way.)
>
> We've found that 'gcc/omp-oacc-kernels-decompose.cc' is currently not at
> all considerate of 'GIMPLE_DEBUG' statements -- and it's not always
> straight forward how to handle these (not rocket science either; but
> needs proper understanding and testing).
The general rule is that debug stmts shouldn't affect code generation
decisions, so when deciding what to optimize/how, they should be ignored,
and during actual transformation adjusted or worst case reset as needed.
> Actually fixing it is a separate task, but it seems prudent to at least
> catch it, and document via a few test cases. OK to push
> "Catch 'GIMPLE_DEBUG' misbehavior in OpenACC 'kernels' decomposition
> [PR100400, PR103836, PR104061]", see attached?
> --- a/gcc/omp-oacc-kernels-decompose.cc
> +++ b/gcc/omp-oacc-kernels-decompose.cc
> @@ -1255,6 +1255,16 @@ decompose_kernels_region_body (gimple *kernels_region,
> tree kernels_clauses)
> gsi_next (&gsi_n);
>
> gimple *stmt = gsi_stmt (gsi);
> + if (gimple_code (stmt) == GIMPLE_DEBUG)
> + {
> + if (flag_compare_debug_opt || flag_compare_debug)
> + /* Let the usual '-fcompare-debug' analysis bail out, as
> + necessary. */
> + ;
> + else
> + sorry_at (loc, "%qs not yet supported",
> + gimple_code_name[gimple_code (stmt)]);
> + }
This is wrong. It shouldn't be dependent on flag_compare_debug* options,
those are just debugging aids to verify that -g/-g0 don't affect code
generation. With the above you'd pretend they don't, but they actually
would (with -g you'd get sorry, without it it would compile fine).
If this code is analysing whether the kernels region body should be
decomposed or not, it should be if (is_gimple_debug (stmt)) continue;
or whatever else to just ignore them (in some opts already during analysis
phase we remember they are present and something about them, but not in
a way that would actually affect the code generation decisions).
And then when actually transforming it, it depends on what transformations
are done to the variables/values referenced in the debug stmts.
gimple_debug_bind_reset_value (stmt); update_stmt (stmt); is
what resets them and can be used as last resort, it will keep saying
that it describes some var, but will say that the var is optimized out.
Jakub
