The just posted inliner patch can regress musttail calls if we perform
function splitting and then inline the outlined body back into the original
(or inline both the small function and outlined large body into something
else).
If there are any musttail calls, I think we need to call the outlined
body using a musttail call, so that the inliner will preserve musttail
attributes in the body.
2025-03-21 Jakub Jelinek <jakub@redhat.com>
PR ipa/119376
* ipa-split.cc (split_function): Call gimple_call_set_must_tail
on the call to outlined partition if has_musttail and
!add_tsan_func_exit.
* g++.dg/opt/musttail1.C: New test.
args_to_pass[i] = arg;
}
call = gimple_build_call_vec (node->decl, args_to_pass);
+ if (cur_node->get_fun ()->has_musttail && !add_tsan_func_exit)
+ gimple_call_set_must_tail (call, true);
gimple_set_block (call, DECL_INITIAL (current_function_decl));
args_to_pass.release ();
--- /dev/null
+// PR ipa/119376
+// { dg-do compile { target musttail } }
+// { dg-options "-O2 -fdump-tree-optimized" }
+// { dg-final { scan-tree-dump-times " \[^\n\r]* = foo \\\(\[^\n\r]*\\\); \\\[tail call\\\] \\\[must tail call\\\]" 2 "optimized" } }
+// { dg-final { scan-tree-dump-times " \[^\n\r]* = foo \\\(\[^\n\r]*\\\); \\\[tail call\\\]" 3 "optimized" } }
+
+int foo (int x);
+typedef int (*F) (int);
+int v;
+
+inline int
+bar (int x)
+{
+ if (__builtin_expect (x == 42, 1))
+ return 1;
+ [[gnu::musttail]] return foo (x + v * (x | v) * (x & v) * (x - v) * (x + v * v));
+}
+
+int
+baz (int x)
+{
+ [[gnu::musttail]] return bar (x);
+}
+
+int
+qux (int x)
+{
+ return bar (x + 1);
+}
+
+F
+corge ()
+{
+ return &bar;
+}