Re: Fix folding of EQ/NE_EXPR WRT symtab aliases

[Richard Biener <richard dot guenther at gmail dot com>]

On Wed, Dec 17, 2014 at 11:30 AM, Tejas Belagod <tejas.belagod@arm.com> wrote:
> On 07/12/14 07:38, Jan Hubicka wrote:
>>
>> Hi,
>> this is simplified patch that only adds the equal_address_to predicate
>> (and thus fixes issues
>> with inccorect folding of speculative calls). Hopefully it will mek it
>> easier to handle
>> the rest of fold-const incrementally.
>> Bootstrapped/regtested x86_64-linux, comitted
>>
>>          * symtab.c (symtab_node::equal_address_to): New function.
>>          * cgraph.h (symtab_node::equal_address_to): Declare.
>>          * fold-const.c (fold_comparison, fold_binary_loc): Use it.
>>
>>          * c-family/c-common.c: Refuse weaks for symbols that can not
>> change
>>          visibility.
>>
>>          * gcc.dg/addr_equal-1.c: New testcase.
>> Index: symtab.c
>> ===================================================================
>> --- symtab.c    (revision 218457)
>> +++ symtab.c    (working copy)
>> @@ -1860,3 +1860,90 @@ symtab_node::nonzero_address ()
>>       return true;
>>     return false;
>>   }
>> +
>> +/* Return 0 if symbol is known to have different address than S2,
>> +   Return 1 if symbol is known to have same address as S2,
>> +   return 2 otherwise.   */
>> +int
>> +symtab_node::equal_address_to (symtab_node *s2)
>> +{
>> +  enum availability avail1, avail2;
>> +
>> +  /* A Shortcut: equivalent symbols are always equivalent.  */
>> +  if (this == s2)
>> +    return 1;
>> +
>> +  /* For non-interposable aliases, lookup and compare their actual
>> definitions.
>> +     Also check if the symbol needs to bind to given definition.  */
>> +  symtab_node *rs1 = ultimate_alias_target (&avail1);
>> +  symtab_node *rs2 = s2->ultimate_alias_target (&avail2);
>> +  bool binds_local1 = rs1->analyzed && decl_binds_to_current_def_p
>> (this->decl);
>> +  bool binds_local2 = rs2->analyzed && decl_binds_to_current_def_p
>> (s2->decl);
>> +  bool really_binds_local1 = binds_local1;
>> +  bool really_binds_local2 = binds_local2;
>> +
>> +  /* Addresses of vtables and virtual functions can not be used by user
>> +     code and are used only within speculation.  In this case we may make
>> +     symbol equivalent to its alias even if interposition may break this
>> +     rule.  Doing so will allow us to turn speculative inlining into
>> +     non-speculative more agressively.  */
>> +  if (DECL_VIRTUAL_P (this->decl) && avail1 >= AVAIL_AVAILABLE)
>> +    binds_local1 = true;
>> +  if (DECL_VIRTUAL_P (s2->decl) && avail2 >= AVAIL_AVAILABLE)
>> +    binds_local2 = true;
>> +
>> +  /* If both definitions are available we know that even if they are
>> bound
>> +     to other unit they must be defined same way and therefore we can use
>> +     equivalence test.  */
>> +  if (rs1 != rs2 && avail1 >= AVAIL_AVAILABLE && avail2 >=
>> AVAIL_AVAILABLE)
>> +    binds_local1 = binds_local2 = true;
>> +
>> +  if ((binds_local1 ? rs1 : this)
>> +       == (binds_local2 ? rs2 : s2))
>> +    {
>> +      /* We made use of the fact that alias is not weak.  */
>> +      if (binds_local1 && rs1 != this)
>> +        refuse_visibility_changes = true;
>> +      if (binds_local2 && rs2 != s2)
>> +        s2->refuse_visibility_changes = true;
>> +      return 1;
>> +    }
>> +
>> +  /* If both symbols may resolve to NULL, we can not really prove them
>> different.  */
>> +  if (!nonzero_address () && !s2->nonzero_address ())
>> +    return 2;
>> +
>> +  /* Except for NULL, functions and variables never overlap.  */
>> +  if (TREE_CODE (decl) != TREE_CODE (s2->decl))
>> +    return 0;
>> +
>> +  /* If one of the symbols is unresolved alias, punt.  */
>> +  if (rs1->alias || rs2->alias)
>> +    return 2;
>> +
>> +  /* If we have a non-interposale definition of at least one of the
>> symbols
>> +     and the other symbol is different, we know other unit can not
>> interpose
>> +     it to the first symbol; all aliases of the definition needs to be
>> +     present in the current unit.  */
>> +  if (((really_binds_local1 || really_binds_local2)
>> +      /* If we have both definitions and they are different, we know they
>> +        will be different even in units they binds to.  */
>> +       || (binds_local1 && binds_local2))
>> +      && rs1 != rs2)
>> +    {
>> +      /* We make use of the fact that one symbol is not alias of the
>> other
>> +        and that the definition is non-interposable.  */
>> +      refuse_visibility_changes = true;
>> +      s2->refuse_visibility_changes = true;
>> +      rs1->refuse_visibility_changes = true;
>> +      rs2->refuse_visibility_changes = true;
>> +      return 0;
>> +    }
>> +
>> +  /* TODO: Alias oracle basically assume that addresses of global
>> variables
>> +     are different unless they are declared as alias of one to another.
>> +     We probably should be consistent and use this fact here, too, and
>> update
>> +     alias oracle to use this predicate.  */
>> +
>> +  return 2;
>> +}
>> Index: c-family/c-common.c
>> ===================================================================
>> --- c-family/c-common.c (revision 218457)
>> +++ c-family/c-common.c (working copy)
>> @@ -7781,7 +7781,12 @@ handle_weak_attribute (tree *node, tree
>>       }
>>     else if (TREE_CODE (*node) == FUNCTION_DECL
>>             || TREE_CODE (*node) == VAR_DECL)
>> -    declare_weak (*node);
>> +    {
>> +      struct symtab_node *n = symtab_node::get (*node);
>> +      if (n && n->refuse_visibility_changes)
>> +       error ("%+D declared weak after being used", *node);
>> +      declare_weak (*node);
>> +    }
>>     else
>>       warning (OPT_Wattributes, "%qE attribute ignored", name);
>>
>> Index: cgraph.h
>> ===================================================================
>> --- cgraph.h    (revision 218457)
>> +++ cgraph.h    (working copy)
>> @@ -332,6 +332,11 @@ public:
>>     /* Return true if symbol is known to be nonzero.  */
>>     bool nonzero_address ();
>>
>> +  /* Return 0 if symbol is known to have different address than S2,
>> +     Return 1 if symbol is known to have same address as S2,
>> +     return 2 otherwise.   */
>> +  int equal_address_to (symtab_node *s2);
>> +
>>     /* Return symbol table node associated with DECL, if any,
>>        and NULL otherwise.  */
>>     static inline symtab_node *get (const_tree decl)
>> Index: fold-const.c
>> ===================================================================
>> --- fold-const.c        (revision 218457)
>> +++ fold-const.c        (working copy)
>> @@ -8985,7 +8985,7 @@ fold_comparison (location_t loc, enum tr
>>              }
>>          }
>>         /* For non-equal bases we can simplify if they are addresses
>> -        of local binding decls or constants.  */
>> +        declarations with different addresses.  */
>>         else if (indirect_base0 && indirect_base1
>>                 /* We know that !operand_equal_p (base0, base1, 0)
>>                    because the if condition was false.  But make
>> @@ -8993,16 +8993,13 @@ fold_comparison (location_t loc, enum tr
>>                 && base0 != base1
>>                 && TREE_CODE (arg0) == ADDR_EXPR
>>                 && TREE_CODE (arg1) == ADDR_EXPR
>> -              && (((TREE_CODE (base0) == VAR_DECL
>> -                    || TREE_CODE (base0) == PARM_DECL)
>> -                   && (targetm.binds_local_p (base0)
>> -                       || CONSTANT_CLASS_P (base1)))
>> -                  || CONSTANT_CLASS_P (base0))
>> -              && (((TREE_CODE (base1) == VAR_DECL
>> -                    || TREE_CODE (base1) == PARM_DECL)
>> -                   && (targetm.binds_local_p (base1)
>> -                       || CONSTANT_CLASS_P (base0)))
>> -                  || CONSTANT_CLASS_P (base1)))
>> +              && DECL_P (base0)
>> +              && DECL_P (base1)
>> +              /* Watch for aliases.  */
>> +              && (!decl_in_symtab_p (base0)
>> +                  || !decl_in_symtab_p (base1)
>> +                  || !symtab_node::get_create (base0)->equal_address_to
>> +                        (symtab_node::get_create (base1))))
>>          {
>>            if (code == EQ_EXPR)
>>              return omit_two_operands_loc (loc, type, boolean_false_node,
>> @@ -12257,33 +12254,23 @@ fold_binary_loc (location_t loc,
>>           unaliased symbols neither of which are extern (since we do not
>>           have access to attributes for externs), then we know the result.
>> */
>>         if (TREE_CODE (arg0) == ADDR_EXPR
>> -         && VAR_OR_FUNCTION_DECL_P (TREE_OPERAND (arg0, 0))
>> -         && ! DECL_WEAK (TREE_OPERAND (arg0, 0))
>> -         && ! lookup_attribute ("alias",
>> -                                DECL_ATTRIBUTES (TREE_OPERAND (arg0, 0)))
>> -         && ! DECL_EXTERNAL (TREE_OPERAND (arg0, 0))
>> +         && DECL_P (TREE_OPERAND (arg0, 0))
>>            && TREE_CODE (arg1) == ADDR_EXPR
>> -         && VAR_OR_FUNCTION_DECL_P (TREE_OPERAND (arg1, 0))
>> -         && ! DECL_WEAK (TREE_OPERAND (arg1, 0))
>> -         && ! lookup_attribute ("alias",
>> -                                DECL_ATTRIBUTES (TREE_OPERAND (arg1, 0)))
>> -         && ! DECL_EXTERNAL (TREE_OPERAND (arg1, 0)))
>> -       {
>> -         /* We know that we're looking at the address of two
>> -            non-weak, unaliased, static _DECL nodes.
>> -
>> -            It is both wasteful and incorrect to call operand_equal_p
>> -            to compare the two ADDR_EXPR nodes.  It is wasteful in that
>> -            all we need to do is test pointer equality for the arguments
>> -            to the two ADDR_EXPR nodes.  It is incorrect to use
>> -            operand_equal_p as that function is NOT equivalent to a
>> -            C equality test.  It can in fact return false for two
>> -            objects which would test as equal using the C equality
>> -            operator.  */
>> -         bool equal = TREE_OPERAND (arg0, 0) == TREE_OPERAND (arg1, 0);
>> -         return constant_boolean_node (equal
>> -                                       ? code == EQ_EXPR : code !=
>> EQ_EXPR,
>> -                                       type);
>> +         && DECL_P (TREE_OPERAND (arg1, 0)))
>> +       {
>> +         int equal;
>> +
>> +         if (decl_in_symtab_p (TREE_OPERAND (arg0, 0))
>> +             && decl_in_symtab_p (TREE_OPERAND (arg1, 0)))
>> +           equal = symtab_node::get_create (TREE_OPERAND (arg0, 0))
>> +                   ->equal_address_to (symtab_node::get_create
>> +                                         (TREE_OPERAND (arg1, 0)));
>> +         else
>> +           equal = TREE_OPERAND (arg0, 0) == TREE_OPERAND (arg1, 0);
>> +         if (equal != 2)
>> +           return constant_boolean_node (equal
>> +                                         ? code == EQ_EXPR : code !=
>> EQ_EXPR,
>> +                                         type);
>>          }
>>
>>         /* Similarly for a NEGATE_EXPR.  */
>> Index: testsuite/gcc.dg/addr_equal-1.c
>> ===================================================================
>> --- testsuite/gcc.dg/addr_equal-1.c     (revision 0)
>> +++ testsuite/gcc.dg/addr_equal-1.c     (revision 0)
>> @@ -0,0 +1,101 @@
>> +/* { dg-do run } */
>> +/* { dg-require-weak "" } */
>> +/* { dg-require-alias "" } */
>> +/* { dg-options "-O2" } */
>> +void abort (void);
>> +extern int undef_var0, undef_var1;
>> +extern __attribute__ ((weak)) int weak_undef_var0;
>> +extern __attribute__ ((weak)) int weak_undef_var1;
>> +__attribute__ ((weak)) int weak_def_var0;
>> +int def_var0=0, def_var1=0;
>> +static int alias_var0 __attribute__ ((alias("def_var0")));
>> +extern int weak_alias_var0 __attribute__ ((alias("def_var0")))
>> __attribute__ ((weak));
>> +void undef_fn0(void);
>> +void undef_fn1(void);
>> +void def_fn0(void)
>> +{
>> +}
>> +void def_fn1(void)
>> +{
>> +}
>> +__attribute__ ((weak))
>> +void weak_def_fn0(void)
>> +{
>> +}
>> +__attribute__ ((weak))
>> +void weak_def_fn1(void)
>> +{
>> +}
>> +__attribute__ ((weak)) void weak_undef_fn0(void);
>> +
>> +inline
>> +void inline_fn0(void)
>> +{
>> +}
>> +inline
>> +void inline_fn1(void)
>> +{
>> +}
>> +
>> +int
>> +main(int argc, char **argv)
>> +{
>> +  /* Two definitions are always different unless they can be interposed.
>> */
>> +  if (!__builtin_constant_p (def_fn0 == def_fn1))
>> +    abort();
>> +  if (def_fn0 == def_fn1)
>> +    abort();
>> +
>> +  if (!__builtin_constant_p (&def_var0 == &def_var1))
>> +    abort();
>> +  if (&def_var0 == &def_var1)
>> +    abort();
>> +
>> +  /* Same symbol is the same no matter on interposition.  */
>> +  if (!__builtin_constant_p (undef_fn0 != undef_fn0))
>> +    abort ();
>> +  if (undef_fn0 != undef_fn0)
>> +    abort ();
>> +
>> +  /* Do not get confused by same offset.  */
>> +  if (!__builtin_constant_p (&undef_var0 + argc != &undef_var0 + argc))
>> +    abort ();
>> +  if (&undef_var0 + argc != &undef_var0 + argc)
>> +    abort ();
>> +
>> +  /* Alias and its target is equivalent unless one of them can be
>> interposed.  */
>> +  if (!__builtin_constant_p (&def_var0 != &alias_var0))
>> +    abort ();
>> +  if (&def_var0 != &alias_var0 )
>> +    abort ();
>> +
>> +  if (__builtin_constant_p (&def_var0 != &weak_alias_var0))
>> +    abort ();
>> +  if (&def_var0 != &weak_alias_var0)
>> +    abort ();
>> +
>> +  /* Weak definitions may be both NULL.  */
>> +  if (__builtin_constant_p ((void *)weak_undef_fn0 == (void
>> *)&weak_undef_var0))
>> +    abort ();
>> +  if ((void *)weak_undef_fn0 != (void *)&weak_undef_var0)
>> +    abort ();
>> +
>> +  /* Variables and functions do not share same memory locations
>> otherwise.  */
>> +  if (!__builtin_constant_p ((void *)undef_fn0 == (void *)&undef_var0))
>> +    abort ();
>> +  if ((void *)undef_fn0 == (void *)&undef_var0)
>> +    abort ();
>> +
>> +  /* This works for cases where one object is just weakly defined, too.
>> */
>> +  if (!__builtin_constant_p ((void *)weak_undef_fn0 == (void
>> *)&weak_def_var0))
>> +    abort ();
>> +  if ((void *)weak_undef_fn0 == (void *)&weak_def_var0)
>> +    abort ();
>> +
>> +  /* Inline functions are known to be different.  */
>> +  if (!__builtin_constant_p (inline_fn0 != inline_fn1))
>> +    abort ();
>> +  if (inline_fn0 == inline_fn1)
>> +    abort ();
>> +  return 0;
>> +}
>>
>
> addr_equal-1.c is a nopic test as agreed in this thread.
>
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64328
>
> Here is a patch. Tested on aarch64-none-elf/-fPIC
> FAIL -> UNSUPPORTED
>
> OK to apply?

Ok.

Thanks,
Richard.

> Thanks,
> Tejas.
>
> Changelog:
>
> 2014-12-17  Tejas Belagod  <tejas.belagod@arm.com>
>
>         PR testsuite/64328
>         * gcc.dg/addr_equal-1.c: Not supported for -fPIC.