https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126454
Luke Wren <wren6991 at gmail dot com> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |wren6991 at gmail dot com
--- Comment #3 from Luke Wren <wren6991 at gmail dot com> ---
I also hit this on 16.1.0 while debugging a regression in MicroPython from
15.2.0 to 16.1.0.
Reduced from the MicroPython source, compile with: riscv32-unknown-elf-gcc -c
repro.c -march=rv32i_zca_zcmp -Os
extern int mp_obj_get_type(int o);
extern int type_get_iternext(int type);
extern void mp_load_method_maybe(int o, int qstr, int *dest);
extern int mp_call_method_n_kw(int n_args, int n_kw, const int *args);
extern const char *mp_obj_get_type_str(int o);
extern void mp_raise_msg_varg(int type, const char *fmt, const char *arg)
__attribute__((noreturn));
int mp_iternext_allow_raise(int o_in) {
int dest[2];
int type = mp_obj_get_type(o_in);
if (type & 0x180) {
int (*it)(int) = (void *)(long)type_get_iternext;
return it(o_in);
}
mp_load_method_maybe(o_in, 25, dest);
if (dest[0]) {
return mp_call_method_n_kw(0, 0, dest);
}
const char *s = mp_obj_get_type_str(o_in);
mp_raise_msg_varg(0, "'%s' object isn't an iterator", s);
}
Disassembly of the call to mp_call_method_n_kw:
0000004e <.L6>:
4e: 0030 addi a2,sp,8
50: 4581 li a1,0
52: 00000097 auipc ra,0x0
52: R_RISCV_CALL_PLT mp_call_method_n_kw
52: R_RISCV_RELAX *ABS*
56: 000080e7 jalr ra # 52 <.L6+0x4>
5a: bc56 cm.popretz {ra,s0},32
There are two distinct bugs caused by the absorption of the `li a0, 0` into
`cm.popretz`:
* First arg of mp_call_method_n_kw() is not initialised to 0, gets garbage
instead
* Return value from mp_iternext_allow_raise() is replaced with 0 instead of the
result of mp_call_method_n_kw()
