https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101979
Bug ID: 101979
Summary: aarch64: wrong code with -O3
Product: gcc
Version: 12.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: cooper.qu at linux dot alibaba.com
Target Milestone: ---
The C code:
typedef int (*init_fn_t)(void);
extern int rti_start(void);
extern int rti_board_start(void);
extern int rti_board_end(void);
extern int rti_end(void);
__attribute__((used)) const init_fn_t __rt_init_rti_start
__attribute__((section(".rti_fn." "0"))) = rti_start;
__attribute__((used)) const init_fn_t __rt_init_rti_board_start
__attribute__((section(".rti_fn." "0.end"))) = rti_board_start;
__attribute__((used)) const init_fn_t __rt_init_rti_board_end
__attribute__((section(".rti_fn." "1.end"))) = rti_board_end;
__attribute__((used)) const init_fn_t __rt_init_rti_end
__attribute__((section(".rti_fn." "6.end"))) = rti_end;
void rt_components_init(void)
{ const init_fn_t *fn_ptr;
for (fn_ptr = &__rt_init_rti_board_end; fn_ptr < &__rt_init_rti_end; fn_ptr
++)
{
(*fn_ptr)();
}
}
The key loop assemble code generated by -O3 is :
.L3:
bl rti_board_end
cmp x19, x20
add x19, x19, 1
bne .L3
It only calls function 'rti_board_end' instead of moving the pointer 'fn_ptr'
forward.
The (*fn_ptr)() optimized to MEM[(int (*init_fn_t) (void)
*)&__rt_init_rti_board_end + ivtmp.5_11 * 8] in the ivopts pass,
it then fold to const in the fre5 pass as following dump information:
"RHS MEM[(int (*init_fn_t) (void) *)&__rt_init_rti_board_end + ivtmp.5_11 *
8] simplified to rti_board_end"
And I found a similar issue mentioned in an email a long time ago
(https://gcc.gnu.org/legacy-ml/gcc-patches/2012-06/msg01221.html), the
conclusion in the email is that the code written incorrectly. Is there a
problem with the code above, or it's a bug of tree optimization?
