Re: [PATCH] Repair TYPE_REF_CAN_ALIAS_ALL machinery

[Eric Botcazou <ebotcazou at adacore dot com>]

> I don't like the tree-ssa-operands.c part of the patch as we should
> have a mechanism that deals with this problem already, as we have char*
> pointers. 

Unless I'm mistaken, we don't special-case 'char*' pointers, we special-case 
the type 'char'.  However the tree-ssa-operands.c part might indeed be 
replaced by a test in may_alias_p that would cause the function to return 
true for a ref-all pointer and a call-clobbered variable.

> Btw. I couldn't construct a C testcase that failed.

#include <stdlib.h>

void foo(void)
{
  float *pf;
  int *pi;

  pf = malloc (sizeof(float));
  pi = (int *)pf;

  *pf = 1.0;
  *pi = 0;

  if (*pf != 0.0)
    abort ();
}

int main(void)
{
  foo ();
  return 0;
}

Compile at -O2 and -O2 -fno-strict-aliasing on x86.

> So, as you noticed I confused the type 
> somewhat so one would
> need to amend the get_alias_set calls in tree-ssa-alias.c:get_tmt_for
> and friends to
> check for TYPE_REF_CAN_ALIAS_ALL on the pointer type and use alias set 0 in
> this case.

Thanks, interesting suggestion, I'm going to give it a whirl too.

-- 
Eric Botcazou