https://gcc.gnu.org/bugzilla/show_bug.cgi?id=62032
--- Comment #4 from rguenther at suse dot de <rguenther at suse dot de> --- On Thu, 7 Aug 2014, amker.cheng at gmail dot com wrote: > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=62032 > > --- Comment #3 from bin.cheng <amker.cheng at gmail dot com> --- > I did some investigation and think it's a latent bug of lto which reveaded by > r213585. > > Before the revision, pass_fold_builtins::execute calls gimple_fold_builtins > directly to fold __builtin___vsnprintf_chk (_26, _25, 0, 4294967295, "%d - > %c", > ap);. It works fine by simplifying the builtin into __builtin_vsnprintf (_26, > _25, "%d - %c", ap);. > > After the revision, it calls to > fold_stmt->gimple_fold_call->gimple_fold_builtin to do the job. But before > that, it calls gimple_call_builtin_p to verify that it IS a valid builtin > call, > which in turn calls gimple_builtin_call_types_compatible_p to check that types > between parameter and argument match each other. Well, the check fails on the > last parameter/argument, as dumped below: > > TREE_TYPE of ARGUMENT: > <record_type 0x7fc34cc9fc78 va_list SI > size <integer_cst 0x7fc34ce71dc8 type <integer_type 0x7fc34ce7c0a8 > bitsizetype> constant 32> > unit size <integer_cst 0x7fc34ce71de0 type <integer_type 0x7fc34ce7c000 > sizetype> constant 4> > align 32 symtab 0 alias set -1 canonical type 0x7fc34ce899d8 > fields <field_decl 0x7fc34ce87130 __ap > type <pointer_type 0x7fc34ce84000 type <void_type 0x7fc34ce7cf18 void> > public unsigned SI size <integer_cst 0x7fc34ce71dc8 32> unit size > <integer_cst 0x7fc34ce71de0 4> > align 32 symtab 0 alias set 3 canonical type 0x7fc34ce84000 > pointer_to_this <pointer_type 0x7fc34ce8a348>> > unsigned SI file <built-in> line 0 col 0 size <integer_cst > 0x7fc34ce71dc8 32> unit size <integer_cst 0x7fc34ce71de0 4> > align 32 offset_align 64 > offset <integer_cst 0x7fc34ce71df8 constant 0> > bit offset <integer_cst 0x7fc34ce71e58 constant 0> context > <record_type > 0x7fc34ce899d8 __va_list>> context <translation_unit_decl 0x7fc34df310f0 > D.4035> > pointer_to_this <pointer_type 0x7fc34ccaa2a0> chain <type_decl > 0x7fc34ce87da8 __va_list>> > > TREE_TYPE of PARAMETER of function BUILT_IN_VSNPRINTF_CHK: > <reference_type 0x7fc34ce89c78 > type <record_type 0x7fc34ce899d8 __va_list SI > size <integer_cst 0x7fc34ce71dc8 constant 32> > unit size <integer_cst 0x7fc34ce71de0 constant 4> > align 32 symtab 0 alias set 7 canonical type 0x7fc34ce899d8 > fields <field_decl 0x7fc34ce87130 __ap type <pointer_type > 0x7fc34ce84000> > unsigned SI file <built-in> line 0 col 0 size <integer_cst > 0x7fc34ce71dc8 32> unit size <integer_cst 0x7fc34ce71de0 4> > align 32 offset_align 64 > offset <integer_cst 0x7fc34ce71df8 constant 0> > bit offset <integer_cst 0x7fc34ce71e58 constant 0> context > <record_type 0x7fc34ce899d8 __va_list>> > pointer_to_this <pointer_type 0x7fc34ccc4888> reference_to_this > <reference_type 0x7fc34ccaa0a8> chain <type_decl 0x7fc34ce87098 __va_list>> > unsigned SI size <integer_cst 0x7fc34ce71dc8 32> unit size <integer_cst > 0x7fc34ce71de0 4> > align 32 symtab 0 alias set -1 canonical type 0x7fc34ce89c78> > > > The root cause lies in gcc/lto/lto-lang.c:lto_init, there is below code > snippet: > > if (TREE_CODE (va_list_type_node) == ARRAY_TYPE) > { > tree x = build_pointer_type (TREE_TYPE (va_list_type_node)); > lto_define_builtins (x, x); > } > else > { > lto_define_builtins (va_list_type_node, > build_reference_type (va_list_type_node)); > } > > While prototype of lto_define_builtins is like below. > static void > lto_define_builtins (tree va_list_ref_type_node ATTRIBUTE_UNUSED, > tree va_list_arg_type_node ATTRIBUTE_UNUSED); > > Apparently, the arguments passed to lto_define_builtins are mis-matched. In > fact, I think it should be consistent with function calls to c_define_builtins > in c front-end. > > I think above analysis also explains why it happens only with lto. > > I will send a patch fixing this soon. Thanks for the detective work! IMHO this fix is worth backporting to at least 4.9. Richard.
