On Fri, 21 Mar 2025, Jakub Jelinek wrote:
> Hi!
>
> The following testcase shows we were ignoring musttail flags on calls
> when deciding if two functions are the same.
> That can result in problems in both directions, either we silently
> lose musttail attribute because there is a similar function without it
> earlier and then we e.g. don't diagnose if it can't be tail called
> or don't try harder to do a tail call, or we get it even in functions
> which didn't have it before.
>
> The following patch for now just punts if it differs. Perhaps we could
> just merge it and get musttail flag if any of the merged functions had
> one in such position, but it feels to me that it is now too late in GCC 15
> cycle to play with this.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
OK.
Richard.
> 2025-03-20 Jakub Jelinek <ja...@redhat.com>
>
> PR ipa/119376
> * ipa-icf-gimple.cc (func_checker::compare_gimple_call): Return false
> for gimple_call_must_tail_p mismatches.
>
> * c-c++-common/musttail27.c: New test.
>
> --- gcc/ipa-icf-gimple.cc.jj 2025-02-27 22:04:06.082435026 +0100
> +++ gcc/ipa-icf-gimple.cc 2025-03-20 15:33:38.996614656 +0100
> @@ -708,7 +708,8 @@ func_checker::compare_gimple_call (gcall
> || gimple_call_from_thunk_p (s1) != gimple_call_from_thunk_p (s2)
> || gimple_call_from_new_or_delete (s1) !=
> gimple_call_from_new_or_delete (s2)
> || gimple_call_va_arg_pack_p (s1) != gimple_call_va_arg_pack_p (s2)
> - || gimple_call_alloca_for_var_p (s1) != gimple_call_alloca_for_var_p
> (s2))
> + || gimple_call_alloca_for_var_p (s1) != gimple_call_alloca_for_var_p
> (s2)
> + || gimple_call_must_tail_p (s1) != gimple_call_must_tail_p (s2))
> return false;
>
> unsigned check_arg_types_from = 0;
> --- gcc/testsuite/c-c++-common/musttail27.c.jj 2025-03-20
> 15:42:30.200261715 +0100
> +++ gcc/testsuite/c-c++-common/musttail27.c 2025-03-20 15:44:24.990674296
> +0100
> @@ -0,0 +1,31 @@
> +/* 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\\\]" 4 "optimized" } } */
> +
> +int foo (int);
> +
> +int
> +bar (int x)
> +{
> + [[gnu::musttail]] return foo (x + 1);
> +}
> +
> +int
> +baz (int x)
> +{
> + return foo (x + 1);
> +}
> +
> +int
> +qux (int x)
> +{
> + return foo (x + 2);
> +}
> +
> +int
> +corge (int x)
> +{
> + [[gnu::musttail]] return foo (x + 2);
> +}
>
> Jakub
>
>
--
Richard Biener <rguent...@suse.de>
SUSE Software Solutions Germany GmbH,
Frankenstrasse 146, 90461 Nuernberg, Germany;
GF: Ivo Totev, Andrew McDonald, Werner Knoblich; (HRB 36809, AG Nuernberg)
