https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106585
--- Comment #13 from Jeffrey A. Law <law at gcc dot gnu.org> ---
Yea, I definitely explored the ext-dce approach before for a different
testcase. But I think it's really the only viable path I see. A quick POC
results in:
bclr a0,a0,a1 # 12 [c=4 l=4] *bitmanip.md:751
sext.w a0,a0 # 18 [c=4 l=4] *extendsidi2_internal/0
ret # 27 [c=0 l=4] simple_return
My notes indicate that the transformation mucked up other bitmanip cases, so
I'd fully expect the POC to fail in fun and interesting ways. But as I said,
it's the only viable path I see. Making it work would help other cases, so
I'll poke at it a bit more in my copious spare time.
