Crash when cross compiling for ARM with GCC-8-2-0 and -ftree-loop-distribute-patterns

[Richard Earnshaw (lists)]

On 18/10/2019 09:53, Josef Wolf wrote:
> Thanks for your help, Richard!
> 
> On Thu, Oct 17, 2019 at 03:55:31PM +0100, Richard Earnshaw (lists) wrote:
>> On 17/10/2019 15:04, Josef Wolf wrote:
>>> On Thu, Oct 17, 2019 at 02:37:11PM +0200, Matthias Pfaller wrote:
>>>> Why is the stack pointer so low at this point of execution? Using
>>>> 0x20018000-0x20017d20 == 0x2e0 bytes of stack seems a little excessive
>>>> for just one call.
>>>
>>> Ah!... Looks like you've spotted the problem! Actually, the SP is decremented
>>> on every cycle of the loop:
>>>
>>>    (gdb) disass
>>>    Dump of assembler code for function memset:
>>>       0x08001008 <+0>:    push {r4, lr}
>>>       0x0800100a <+2>:    mov  r4, r0
>>>       0x0800100c <+4>:    cbz  r2, 0x8001014 <memset+12>
>>>    => 0x0800100e <+6>:    uxtb r1, r1
>>>       0x08001010 <+8>:    bl   0x8001008 <memset>
>>>       0x08001014 <+12>:   mov  r0, r4
>>>       0x08001016 <+14>:   pop  {r4, pc}
>>>    End of assembler dump.
>>>
>>> This looks REALLY suspicous to me. Every cycle of the loop in memset() is
>>> pushing something onto the stack?!?
>>>
>>> Without the  -ftree-loop-distribute-patterns option, the memset() function
>>> looks entirely different:
>>>
>>>           cbz    r2, <memset+18>
>>>           add    r2, r0
>>>           subs   r2, #1
>>>           uxtb   r1, r1
>>>           subs   r3, r0, #1
>>>    <+10>: strb.w r1, [r3, #1]!
>>>           cmp    r3, r2
>>>           bne.n  <memset+10>
>>>    <+18>: bx     lr
>>
>> The compiler has spotted that you've written something that acts like memset
>> and optimized it into a function call to memset.  So now you're recursing to
>> oblivion.  Try adding -fno-builtin-memset to your compile options.
> 
> This sounds reasonable, and I was actually thinking it would solve the
> problem.
> 
> Unfortunately, -fno-built-memset doesn't have any effect. The same code is
> generated.
> 

Ah, yes.  Looking at some libc sources it puts an explicit optimization 
attribute onto the memset (and similar mem... functions) to disable 
-ftree-loop-distribute-patterns for such functions.  So something like

void *
__attribute__ ((__optimize__ ("-fno-tree-loop-distribute-patterns")))
memset (void *s, int c, size_t n)
{
   int i;
   for (i=0; i<n; i++)
     ((char *)s)[i] = c;

   return s;
}

Though, of course, it's wrapped up in a macro to make it look a bit 
prettier ;-)

This is just one of those gotchas that you have to be aware of when 
implementing the standard library.

R.