https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83336
David Malcolm <dmalcolm at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |jakub at gcc dot gnu.org
--- Comment #1 from David Malcolm <dmalcolm at gcc dot gnu.org> ---
(In reply to David Malcolm from comment #0)
[...]
> There are various issues here:
> * non-display of inlining information
[...]
[CCing Jakub]
I tracked this down to what appears to me to be an overzealous removal of a
block in tree-ssa-live.c's remove_unused_scope_block_p, when optimization is on
and debuginfo generation is off (and various other conditions).
Specifically, when "foo" is inlined into "bar", various lexical BLOCKs are
created, including one with a nonzero BLOCK_SOURCE_LOCATION, representing "foo"
as a whole, where the location is that of the callsite.
inlined_function_outer_scope_p () is true for this block.
However, this block is later removed by tree-ssa-live.c's
remove_unused_scope_block_p.
remove_unused_scope_block_p has this code:
563 /* See if this block is important for representation of inlined
function.
564 Inlined functions are always represented by block with
565 block_ultimate_origin being set to FUNCTION_DECL and
DECL_SOURCE_LOCATION
566 set... */
567 else if (inlined_function_outer_scope_p (scope))
568 unused = false;
569 else
which I'd hoped would ensure that these blocks are preserved, but it isn't
reached.
Instead this clause is reached:
538 /* When not generating debug info we can eliminate info on unused
539 variables. */
540 else if (!flag_auto_profile && debug_info_level == DINFO_LEVEL_NONE)
541 {
542 /* Even for -g0 don't prune outer scopes from artificial
543 functions, otherwise diagnostics using
tree_nonartificial_location
544 will not be emitted properly. */
545 if (inlined_function_outer_scope_p (scope))
546 {
547 tree ao = scope;
548
549 while (ao
550 && TREE_CODE (ao) == BLOCK
551 && BLOCK_ABSTRACT_ORIGIN (ao) != ao)
552 ao = BLOCK_ABSTRACT_ORIGIN (ao);
553 if (ao
554 && TREE_CODE (ao) == FUNCTION_DECL
555 && DECL_DECLARED_INLINE_P (ao)
556 && lookup_attribute ("artificial", DECL_ATTRIBUTES
(ao)))
557 unused = false;
558 }
559 }
which was added by Jakub in r148212 (f63d3ecc34e8b969b3c940f14e725815c1ef3215)
to fix PR middle-end/40340.
The logic in this clause seems wrong to me: the comment reads:
542 /* Even for -g0 don't prune outer scopes from artificial
543 functions, otherwise diagnostics using
tree_nonartificial_location
544 will not be emitted properly. */
but it appears to only apply to artificial functions.
If an inlining of a non-artificial function is reached, then we enter the
clause, but fail the final condition (lookup_attribute), and hence unused is
left as true, and hence the block is removed, and the diagnostic fails to
mention the callsite in the inlining chain.
If I remove these lines, the test cases for pr40340 still pass:
make -k -j64 && make check-gcc RUNTESTFLAGS="-v -v dg.exp=pr40340*.c"
# of expected passes 13
(though this may be due to other patches I have locally)
As is, using gcc (GCC) 8.0.0 20171113:
$ gcc -c d.c -O2
d.c: In function ‘bar’:
d.c:3:19: warning: ‘__builtin_memcpy’ writing 8 bytes into a region of size 3
overflows the destination [-Wstringop-overflow=]
#define FOO(d, s) strcpy (d, s)
^~~~~~~~~~~~~
d.c:5:37: note: in expansion of macro ‘FOO’
void foo (char *d, const char *s) { FOO (d, s); }
^~~
...note the absence of a report on function 'foo' inlined from bar).
Whereas if I supply "-g", thus avoiding the clause above:
$ gcc -c d.c -O2 -g
In function ‘foo’,
inlined from ‘bar’ at d.c:13:3:
d.c:3:19: warning: ‘__builtin_memcpy’ writing 8 bytes into a region of size 3
overflows the destination [-Wstringop-overflow=]
#define FOO(d, s) strcpy (d, s)
^~~~~~~~~~~~~
d.c:5:37: note: in expansion of macro ‘FOO’
void foo (char *d, const char *s) { FOO (d, s); }
^~~
...and it correctly reports the inlining information.
