[PATCH] Fix sra_explode_bitfield_assignment

[Jakub Jelinek]

Hi!

The PR middle-end/37248 patch I've just posted regressed bf64-1.c
test at -O[23s] on x86_64-linux, -fno-tree-sra fixed that.

The problem is that SRA is expecting that unsigned_type_for (and
lang_hooks.types.type_for_size) will return a type with
TYPE_PRECISION it was asking for, but it can return a larger type.
So for a signed 52 bit bitfield, it just casted the signed:52
value to (unsigned long) and so no masking of the topmost 12 bits
was performed.

Fixed thusly, bootstrapped/regtested on x86_64-linux, ok for trunk?

2008-12-05  Jakub Jelinek  <jakub@redhat.com>

	* tree-sra.c (sra_explode_bitfield_assignment): Always
	call unsigned_type_for, if the precision is higher than
	needed, call build_nonstandard_integer_type.

--- gcc/tree-sra.c.jj	2008-12-03 16:32:17.000000000 +0100
+++ gcc/tree-sra.c	2008-12-05 13:23:59.000000000 +0100
@@ -3013,11 +3013,9 @@ sra_explode_bitfield_assignment (tree va
 
 	  infld = fld->replacement;
 
-	  type = TREE_TYPE (infld);
+	  type = unsigned_type_for (TREE_TYPE (infld));
 	  if (TYPE_PRECISION (type) != TREE_INT_CST_LOW (flen))
-	    type = lang_hooks.types.type_for_size (TREE_INT_CST_LOW (flen), 1);
-	  else
-	    type = unsigned_type_for (type);
+	    type = build_nonstandard_integer_type (TREE_INT_CST_LOW (flen), 1);
 
 	  if (TREE_CODE (infld) == BIT_FIELD_REF)
 	    {

	Jakub