This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [PR58315] reset inlined debug vars at return-to point


On Wed, Feb 25, 2015 at 10:40 AM, Alexandre Oliva <aoliva@redhat.com> wrote:
> This patch fixes a problem that has been with us for several years.
> Variable tracking has no idea about the end of the lifetime of inlined
> variables, so it keeps on computing locations for them over and over,
> even though the computed locations make no sense whatsoever because the
> variable can't even be accessed any more.
>
> With this patch, we unbind all inlined variables at the point the
> inlined function returns to, so that the locations for those variables
> will not be touched any further.
>
> In theory, we could do something similar to non-inlined auto variables,
> when they go out of scope, but their decls apply to the entire function
> and I'm told gdb sort-of expects the variables to be accessible
> throughout the function, so I'm not tackling that in this patch, for I'm
> happy enough with what this patch gets us:
>
> - almost 99% reduction in the output asm for the PR testcase
>
> - more than 90% reduction in the peak memory use compiling that testcase
>
> - 63% reduction in the compile time for that testcase
>
> What's scary is that the testcase is not particularly pathological.  Any
> function that calls a longish sequence of inlined functions, that in
> turn call other inline functions, and so on, something that's not
> particularly unusual in C++, will likely observe significant
> improvement, as we won't see growing sequences of var_location notes
> after each call or so, as var-tracking computes a new in-stack location
> for the implicit this argument of each previously-inlined function.
>
> Regstrapped on x86_64-linux-gnu and i686-linux-gnu.  Ok to install?

But code-motion could still move stmts from the inlined functions
across these resets?  That said - shouldn't this simply performed
by proper var-tracking u-ops produced by a backward scan over the
function for "live" scope-blocks?  That is, when you see a scope
block becoming live from exit then add u-ops resetting all
vars from that scope block?

Your patch as-is would add very many debug stmts to for example
tramp3d.  And as you say, the same reasoning applies to scopes
in general, not just for inlines.

Richard.

> Reset inlined debug variables at the end of the inlined function
>
> From: Alexandre Oliva <aoliva@redhat.com>
>
> for  gcc/ChangeLog
>
>         PR debug/58315
>         * tree-inline.c (reset_debug_binding): New.
>         (reset_debug_bindings): Likewise.
>         (expand_call_inline): Call it.
> ---
>  gcc/tree-inline.c |   56 +++++++++++++++++++++++++++++++++++++++++++++++++++++
>  1 file changed, 56 insertions(+)
>
> diff --git a/gcc/tree-inline.c b/gcc/tree-inline.c
> index d8abe03..5b58d8b 100644
> --- a/gcc/tree-inline.c
> +++ b/gcc/tree-inline.c
> @@ -4345,6 +4345,60 @@ add_local_variables (struct function *callee, struct function *caller,
>        }
>  }
>
> +/* Add to BINDINGS a debug stmt resetting SRCVAR if inlining might
> +   have brought in or introduced any debug stmts for SRCVAR.  */
> +
> +static inline void
> +reset_debug_binding (copy_body_data *id, tree srcvar, gimple_seq *bindings)
> +{
> +  tree *remappedvarp = id->decl_map->get (srcvar);
> +
> +  if (!remappedvarp)
> +    return;
> +
> +  if (TREE_CODE (*remappedvarp) != VAR_DECL)
> +    return;
> +
> +  if (*remappedvarp == id->retvar || *remappedvarp == id->retbnd)
> +    return;
> +
> +  tree tvar = target_for_debug_bind (*remappedvarp);
> +  if (!tvar)
> +    return;
> +
> +  gdebug *stmt = gimple_build_debug_bind (tvar, NULL_TREE,
> +                                         id->call_stmt);
> +  gimple_seq_add_stmt (bindings, stmt);
> +}
> +
> +/* For each inlined variable for which we may have debug bind stmts,
> +   add before GSI a final debug stmt resetting it, marking the end of
> +   its life, so that var-tracking knows it doesn't have to compute
> +   further locations for it.  */
> +
> +static inline void
> +reset_debug_bindings (copy_body_data *id, gimple_stmt_iterator gsi)
> +{
> +  tree var;
> +  unsigned ix;
> +  gimple_seq bindings = NULL;
> +
> +  if (!gimple_in_ssa_p (id->src_cfun))
> +    return;
> +
> +  if (!opt_for_fn (id->dst_fn, flag_var_tracking_assignments))
> +    return;
> +
> +  for (var = DECL_ARGUMENTS (id->src_fn);
> +       var; var = DECL_CHAIN (var))
> +    reset_debug_binding (id, var, &bindings);
> +
> +  FOR_EACH_LOCAL_DECL (id->src_cfun, ix, var)
> +    reset_debug_binding (id, var, &bindings);
> +
> +  gsi_insert_seq_before_without_update (&gsi, bindings, GSI_SAME_STMT);
> +}
> +
>  /* If STMT is a GIMPLE_CALL, replace it with its inline expansion.  */
>
>  static bool
> @@ -4650,6 +4704,8 @@ expand_call_inline (basic_block bb, gimple stmt, copy_body_data *id)
>              GCOV_COMPUTE_SCALE (cg_edge->frequency, CGRAPH_FREQ_BASE),
>              bb, return_block, NULL);
>
> +  reset_debug_bindings (id, stmt_gsi);
> +
>    /* Reset the escaped solution.  */
>    if (cfun->gimple_df)
>      pt_solution_reset (&cfun->gimple_df->escaped);
>
>
> --
> Alexandre Oliva, freedom fighter    http://FSFLA.org/~lxoliva/
> You must be the change you wish to see in the world. -- Gandhi
> Be Free! -- http://FSFLA.org/   FSF Latin America board member
> Free Software Evangelist|Red Hat Brasil GNU Toolchain Engineer


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]