https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128341
--- Comment #3 from Yongqiang Tian <yqtian668 at gmail dot com> --- (In reply to Drea Pinski from comment #2) > mode-switching is used also for a few other things. x86_64 uses it in some > cases. So does sh. I wonder if you could figure out a testcase for those > targets too. Thanks for pointing this out. I checked both x86 and SH. They use the generic mode-switching pass, but neither target defines TARGET_MODE_BACKPROP. This particular anchor becomes observable when backprop_transfer assigns a mode to an otherwise requirement-free segment, which is what happens in the AArch64 SME case. Without that hook, an ordinary requirement-free x86 or SH segment retains no_mode and emits nothing at the saved anchor. I also tried focused cld, vzeroupper, and SH FPSCR probes. In addition, I swept the GCC torture tests and the relevant x86 and SH target tests with -fcompare-debug, but found no trigger on either target. I therefore do not think x86 or SH can provide a faithful regression test for this specific anchor change, although I will include them when doing the generic regression testing. BTW, our earlier testing found several debug-dependent failures in older GCC ports. After deduplication, there are three distinct C6X cases and one SH case without reports. They have reduced natural reproducers and candidate fixes, but I have held them because these ports are lightly maintained. I am wondering if you have any suggestions?
