This is the mail archive of the mailing list for the GCC project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [patch] Fix array type merging in LTO mode

On Sun, May 20, 2012 at 7:56 PM, Eric Botcazou <> wrote:
> Hi,
> since, canonical type
> merging for arrays takes hours instead of minutes for big Ada applications.
> The problem is that iterative_hash_canonical_type doesn't hash TYPE_MIN_VALUE
> and TYPE_MAX_VALUE for integer types anymore, so TYPE_DOMAIN is effectively
> not hashed anymore and the number of collisions goes to the roof in Ada.
> Fixed by the attached patch, which also removes a bogus comparison of the
> TYPE_SIZE of TYPE_DOMAIN in gimple_[canonical]types_compatible_p. ?LTO
> bootstrapped on x86_64-suse-linux, OK for mainline and 4.7 branch?

-  /* For integer types hash the types min/max values and the string flag.  */
+  /* For integer types hash the sizetype flag and the string flag.  */
   if (TREE_CODE (type) == INTEGER_TYPE)
     v = iterative_hash_hashval_t (TYPE_STRING_FLAG (type), v);

The sizetype flag is no longer present (the patch will differ for the branch).

+  /* For array types hash the domain and its bounds, and the string flag.  */
+  if (TREE_CODE (type) == ARRAY_TYPE && TYPE_DOMAIN (type))
       v = iterative_hash_hashval_t (TYPE_STRING_FLAG (type), v);
       v = iterative_hash_canonical_type (TYPE_DOMAIN (type), v);
+      /* OMP lowering can introduce error_mark_node in place of
+        random local decls in types.  */
+      if (TYPE_MIN_VALUE (TYPE_DOMAIN (type)) != error_mark_node)
+       v = iterative_hash_expr (TYPE_MIN_VALUE (TYPE_DOMAIN (type)), v);
+      if (TYPE_MAX_VALUE (TYPE_DOMAIN (type)) != error_mark_node)
+       v = iterative_hash_expr (TYPE_MAX_VALUE (TYPE_DOMAIN (type)), v);

I suppose hashing TYPE_DOMAIN itself is then useless?  Practically
it will be always sizetype anyway ... and if we start to see different
domain types arrays with the same actual domain [0, 1] should be
still merged for canonical merging.

Ok with the comment adjustment and not hashing TYPE_DOMAIN itself.


> 2012-05-20 ?Eric Botcazou ?<>
> ? ? ? ?* gimple.c (gimple_types_compatible_p_1) <ARRAY_TYPE>: Remove bogus
> ? ? ? ?size handling.
> ? ? ? ?(gimple_canonical_types_compatible_p) <ARRAY_TYPE>: Likewise.
> ? ? ? ?(iterative_hash_gimple_type): Adjust comment.
> ? ? ? ?(iterative_hash_canonical_type): Likewise. ?Hash the bounds of the
> ? ? ? ?domain for an array type.
> --
> Eric Botcazou

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]