On Wed, 2025-08-13 at 11:24 +0200, Richard Biener wrote:
> On Wed, Jul 30, 2025 at 4:42 PM Qing Zhao <qing.z...@oracle.com> wrote:
[...snip...]
>
> >
> >
> > diff --git a/gcc/diagnostic-context-rich-location.h
> > b/gcc/diagnostic-context-rich-location.h
> > new file mode 100644
> > index 00000000000..676b9d23b05
> > --- /dev/null
> > +++ b/gcc/diagnostic-context-rich-location.h
> > @@ -0,0 +1,73 @@
> > +/* A rich_location subclass that lazily populates a diagnostic_path
> > + with diagnostic context events, but only if the path is actually to be
> > + used.
[...snip...]
> > +
> > +class lazy_diagnostic_context_path : public diagnostics::paths::lazy_path
> > +{
> > +public:
> > + lazy_diagnostic_context_path (const tree_logical_location_manager
> > + &logical_loc_mgr,
> > + location_t location, gimple *stmt)
> > + : diagnostics::paths::lazy_path (logical_loc_mgr),
> > + m_logical_loc_mgr (logical_loc_mgr),
> > + m_location (location), m_stmt (stmt)
> > + {
> > + }
> > +
> > + std::unique_ptr<diagnostics::paths::path>
> > + make_inner_path () const final override;
> > + /* This method will be called on demand if a diagnostic is actually
> > + emitted for this rich_location. */
> > +
> > + const tree_logical_location_manager &m_logical_loc_mgr;
> > + location_t m_location;
> > + gimple *m_stmt;
> > +};
> > +
> > +class rich_location_with_details : public gcc_rich_location
> > +{
> > +public:
> > + rich_location_with_details (location_t location, gimple *stmt)
> > + : gcc_rich_location (location),
> > + m_lazy_diagnostic_context_path (m_logical_loc_mgr, location, stmt)
> > + {
> > + set_path (&m_lazy_diagnostic_context_path);
> > + }
> > +
> > + rich_location_with_details (location_t location, tree exp
> > ATTRIBUTE_UNUSED)
> > + : gcc_rich_location (location),
> > + m_lazy_diagnostic_context_path (m_logical_loc_mgr, location, nullptr)
> > + {
> > + }
To answer your question below: note from the ctor here that
constructing a rich_location_with_details instance on the stack is, by
design, merely some field initializations, (for it, and for the
lazy_diagnostic_context_path stored within it). There is no expensive
computation or heap allocation.
> > +
> > +private:
> > + const tree_logical_location_manager m_logical_loc_mgr;
> > + lazy_diagnostic_context_path m_lazy_diagnostic_context_path;
> > +};
> > +
> > +#endif // GCC_DIAGNOSTIC_CONTEXT_RICH_LOCATION_H
[...snip...]
> > diff --git a/gcc/gimple-ssa-warn-access.cc b/gcc/gimple-ssa-warn-access.cc
> > index 0f4aff6b59b..abfb8685ce7 100644
> > --- a/gcc/gimple-ssa-warn-access.cc
> > +++ b/gcc/gimple-ssa-warn-access.cc
> > @@ -57,6 +57,7 @@
> > #include "pointer-query.h"
> > #include "pretty-print-markup.h"
> > #include "gcc-urlifier.h"
> > +#include "diagnostic-context-rich-location.h"
> >
> > /* Return true if tree node X has an associated location. */
> >
> > @@ -169,17 +170,19 @@ warn_string_no_nul (location_t loc, GimpleOrTree
> > expr, const char *fname,
> > if (expr)
> > {
> > tree func = get_callee_fndecl (expr);
> > + rich_location_with_details richloc (loc, expr);
> > +
> > if (bndrng)
> > {
> > if (wi::ltu_p (maxsiz, bndrng[0]))
> > - warned = warning_at (loc, opt,
> > + warned = warning_at (&richloc, opt,
> > "%qD specified bound %s exceeds "
> > "maximum object size %E",
> > func, bndstr, maxobjsize);
> > else
> > {
> > bool maybe = wi::to_wide (size) == bndrng[0];
> > - warned = warning_at (loc, opt,
> > + warned = warning_at (&richloc, opt,
> > exact
> > ? G_("%qD specified bound %s exceeds "
> > "the size %E of unterminated array")
> > @@ -194,7 +197,7 @@ warn_string_no_nul (location_t loc, GimpleOrTree expr,
> > const char *fname,
> > }
> > }
> > else
> > - warned = warning_at (loc, opt,
> > + warned = warning_at (&richloc, opt,
> > "%qD argument missing terminating nul",
> > func);
> > }
> > @@ -486,14 +489,16 @@ maybe_warn_nonstring_arg (tree fndecl, GimpleOrTree
> > exp)
> > tree maxobjsize = max_object_size ();
> > if (tree_int_cst_lt (maxobjsize, bndrng[0]))
> > {
> > + rich_location_with_details richloc (loc, exp);
> > +
>
> At which point do we perform the CFG walk to identify the control chain?
> My original idea was that we can stuff almost arbitrary compile-time into
> improving diagnostic output when we know we are going to emit it? But
> I suppose this now happens unconditionally, before we know we are going
> to warn? Maybe not in this instance, but as a general issue. Or does the
> computation happen lazily when the rich location is "evaluated"?
Qing, please confirm, but I believe it happens lazily: constructing an
on-stack rich_location_with_details instance involves just initializing
some fields and pointers to vtables. In particular, the
rich_location_with_details has a diagnostics::paths::lazy_path: the
inner path is only populated (with the CFG walk) if the the path is
actually used (when the diagnostic is emitted).
[...snip...]
Dave
