ICE: SEGFAULT in mem rtx - GCC-4.5.1

Jeff Law law@redhat.com
Mon Nov 23 17:30:00 GMT 2015


On 11/22/2015 11:54 PM, vikram1729 wrote:
> Hi
> I am building gcc-4.5.1 cross compiler. I am getting SEGFAULT ICE at
> simplify-rtx.c:167, i.e. at this function
>
> /* If X is a MEM referencing the constant pool, return the real value.
>      Otherwise return X.  */
> rtx
> avoid_constant_pool_reference (rtx x)
> {
>     rtx c, tmp, addr;
>     ....
>     ....
> -->   addr = targetm.delegitimize_address (addr); //SEGFAULT here
>     /* Split the address into a base and integer offset.  */
>     if (GET_CODE (addr) == CONST
>     ....
> }
>
> When debugging, call debug_rtx(x) gives this pattern:
> (mem/c/i:SI (plus:SI (reg/f:SI 129 virtual-stack-vars)
>          (const_int -8 [0xfffffff8])) [0 __ul+0 S4 A32])
>
> To solve it, I tried to implement my own delegitimize function:
> static rtx
> target_delegitimize_address (rtx orig_x)
> {
>     //rtx x = orig_x;
>     rtx x = delegitimize_mem_from_attrs (orig_x);
>
>     if (GET_CODE (x) != MEM)
>       return XEXP(x, 0);
>
>     x = XEXP (x, 0);
>     if (GET_CODE (x) == PLUS
>            && GET_CODE (XEXP (x, 1)) == CONST_INT
>            && GET_CODE (XEXP (x, 0)) == REG)
>       //DON'T KNOW WHAT TO WRITE HERE. Tried this but failed: return XVECEXP
> (XEXP(x, 1), 0, 0);
>
>    return orig_x;
> }
>
> Is it a backend problem?
This is most definitely a problem with your target files.

You certainly don't want to be returning XVECEXP (anything) since you 
don't have a vector type.

In fact, for (plus (virtual-stack-vars) (const_int)), I'd think you'd 
want to return the address as-is.

The delegitimize_address look, to the best of my knowledge, is primarily 
supposed to be dealing with embedded UNSPECs or other obfuscated RTL for 
PIC support and the like.  A simple base+offset address shouldn't need 
delegitimization.

jeff



More information about the Gcc-help mailing list