Re: GCC 4.0.1-beta20050507 miscompiles openssl-0.9.7g with -ftree-loop-linear

[Sebastian Pop <sebastian dot pop at cri dot ensmp dot fr>]

Daniel Berlin wrote:
> On Thu, 2005-05-12 at 10:01 -0600, Zan Lynx wrote:
> > I'm not subscribed to the list (please CC replies to me) and this isn't
> > a real bug report, just a sort of quick check to see if its a known
> > problem.
> > 
> > When I compiled openssl-0.9.7g using -O3 and -ftree-loop-linear as
> > CFLAGS, openssl failed its self-tests for the MD2 code.  Interestingly,
> > it succeeded at the tests for other algorithms.
> > 
> 
> -ftree-loop-linear uses the dependence analyzer which has a number of
> issues where it will claim things are legal that are not.
> 
> You shouldn't expect -ftree-loop-linear to work on random code that you
> don't know anything about :)
> At least, not yet.
> Hopefully these should be fixed in 4.1
> 
> 

The loops that get interchanged are in the following function:

static void md2_block(MD2_CTX *c, const unsigned char *d)
 {
 register unsigned int t,*sp1,*sp2;
 register int i,j;
 unsigned int state[48];

 sp1=c->state;
 sp2=c->cksm;
 j=sp2[16 -1];
 for (i=0; i<16; i++)
  {
  state[i]=sp1[i];
  state[i+16]=t=d[i];
  state[i+32]=(t^sp1[i]);
  j=sp2[i]^=S[t^j];
  }
 t=0;
 for (i=0; i<18; i++)
  {
  for (j=0; j<48; j+=8)
   {
   t= state[j+ 0]^=S[t];
   t= state[j+ 1]^=S[t];
   t= state[j+ 2]^=S[t];
   t= state[j+ 3]^=S[t];
   t= state[j+ 4]^=S[t];
   t= state[j+ 5]^=S[t];
   t= state[j+ 6]^=S[t];
   t= state[j+ 7]^=S[t];
   }
  t=(t+i)&0xff;
  }
 memcpy(sp1,state,16*sizeof(unsigned int));
 OPENSSL_cleanse(state,48*sizeof(unsigned int));
 }

It is failing on mainline with:

md2_dgst.c: In function 'md2_block':
md2_dgst.c:170: error: Definition in block 13 does not dominate use in block 9
for SSA_NAME: t_23 in statement:
t_143 = PHI <t_23(9), 0(3)>;
PHI argument
t_23
for PHI node
t_143 = PHI <t_23(9), 0(3)>;
md2_dgst.c:170: internal compiler error: verify_ssa failed.

after transformation in loop linear.