https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97417
--- Comment #20 from Jim Wilson <wilson at gcc dot gnu.org> ---
Maybe convert_to_mode is recursively calling convert_to_mode until you run out
of stack space. You can use --save-temps to generate a .i preprocessed file
form your input testcasxe, then run cc1 under gdb. You need to give the
default arch/abi options or you will get an error
(gdb) run -march=rv64gc -mabi=lp64d -O2 tmp.i
when look at the backtrace when you hit the segfault.
There are other ways to get the zero extend we need here. We could just
generate the RTL for the insn directly instead of calling the
gen_zero_extendXiYi2 functions because we know that they exist and what form
they are in. This is inconvenient though. We could try querying the optabs
table to get the right insn code. There is a gen_extend_insn function that
looks like it would work and is more direct than convert_to_mode.
gen_extend_insn does require that the input is in a form that the convert will
accept, so it must be forced to a register if it isn't already a register.
Another solution would be to use one of the rtx simplier functions, e.g. you
can do
simplify_gen_unary (ZERO_EXTEND, word_mode, src, mode);
but the simplify functions are really for simplifying complex RTL to simpler
RTL, so I think not the right approach here.
I think gen_extend_insn is the best approach if we can't use convert_to_mode.
