https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90746
--- Comment #3 from Dmitry Vyukov <dvyukov at google dot com> ---
> It would probably useful to have __attribute__((notailcall)) which the
> sanitizers then could stick on the function types.
It should be on _declaration_, right? Sanitizers only provide a definition. The
declaration is provided by gcc itself:
sanitizer.def-DEF_SANITIZER_BUILTIN(BUILT_IN_SANITIZER_COV_TRACE_PC,
sanitizer.def: "__sanitizer_cov_trace_pc",
sanitizer.def- BT_FN_VOID, ATTR_NOTHROW_LEAF_LIST)
But I guess we could still plum that attribute there.
This also makes me wonder: find_tail_calls has fndecl_built_in_p check, the
callback is declared with DEF_SANITIZER_BUILTIN, yet it's not a builtin
somehow?
> We could also make noinline imply no tailcalls, noinline is already mentioned
> in the docs for __builtin_return_address. That wouldn't work for indirect
> calls though.
I would expect that C programmers want more control over this. C programmers
really like to keep things in control :)
That aside, this will introduce performance regressions for existing code,
right?
> Of course, for a builtin we can check that manually in tree-tailcall.c
This looks like the simplest solution for the problem at hand. What would it
be? I guess something more fancy that this: :)
diff --git a/gcc/tree-tailcall.c b/gcc/tree-tailcall.c
index afe8931b5f0..b83b8ba446e 100644
--- a/gcc/tree-tailcall.c
+++ b/gcc/tree-tailcall.c
@@ -477,6 +477,7 @@ find_tail_calls (basic_block bb, struct tailcall **ret)
func = gimple_call_fndecl (call);
if (func
&& !fndecl_built_in_p (func)
+ && func != builtin_decl_implicit (BUILT_IN_SANITIZER_COV_TRACE_PC)
&& recursive_call_p (current_function_decl, func))
{
tree arg;
