On 04/04/2016 02:19 PM, Dan wrote:
Greetings!
GCC is usually so perfect, that I hate to write, but ... I think I'm
chasing down quite the bug in it and would appreciate some thought to
the following.
The code that causes the bug looks like:
if (ptr) {
*ptr = 1;
}
This code evaluates, in the instruction set I am working with, into the
following RTL:
(set (cc0) (compare (reg 3) (reg 1)))
(set (pc) (if-then-else (eq (cc0) (const_int 0)) (label_ref 258) (pc)))
(set (mem (reg 3)) (reg 1))
(all modes are SI)
This would be all fine and dandy, up until the if_convert modifications.
if_convert() rightly decides that the store instruction is prime for a
conditional execution conversion. Therefore, if_convert() transforms
this code into:
(cond_exec (ne (cc0) (const_int 0)) (set (mem (reg 3)) (reg 1)))
This, by itself, is a valid conversion.
The problem is that when cond_exec_process_if_block calls
merge_if_block(ce_info) to package up the changes and make them
permanent, merge_if_block then calls merge_block which deletes not only
the conditional jump statement (set (pc) ... etc.), but also the compare
that set the conditions prior to the jump statement (see
rtl_merge_blocks within cfgrtl.c). Hence instead of the working RTL,
(set (cc0) (compare (reg 3) (reg 1)))
(cond_exec (ne (cc0) (const_int 0)) (set (mem (reg 3)) (reg 1)))
I'm left with the broken RTL:
(cond_exec (ne (cc0) (const_int 0)) (set (mem (reg 3)) (reg 1)))
Can someone tell me if I am missing something, or whether this really is
a bug in GCC?
It's a bug in GCC. I don't think we currently have any targets that use
cc0 and conditional execution, thus other targets aren't stumbling over
this problem.
It sounds like you've got a custom ISA and thus a custom GCC backend. I
would strongly recommend against using cc0 in your backend. cc0 is an
old deprecated way to express condition code handling.
Jeff
