genmodes.c (struct mode_data): Add contained and next_cont fields.

	* genmodes.c (struct mode_data): Add contained and next_cont
	fields.
	(complete_mode): Maintain linked list of modes that have a
	given component.
	(emit_mode_unit_size): Delete.
	(emit_mode_nunits): New.
	(emit_insn_modes_c): Update to match.
	(emit_mode_adjustments): Propagate size and alignment
	adjustments from component modes to their containers.
	* machmode.h (mode_unit_size): Delete.
	(mode_nunits): New.
	(GET_MODE_NUNITS): Just return the value in the table.
	(GET_MODE_UNIT_SIZE): Compute using GET_MODE_INNER and
	GET_MODE_SIZE.
	* expmed.c (store_bit_field, extract_bit_field): Can use a
	plain move instruction if bitsize >= GET_MODE_BITSIZE of
	destination/source mode, respectively.
	* varasm.c (assemble_real): Write out the full size of the
	constant, not just its bitsize.
	(output_constant): Honor TYPE_MODE of TREE_REAL_CSTs.

	* config/ia64/ia64-modes.def: Define XFmode as well as TFmode.
	Use ADJUST_BYTESIZE and ADJUST_ALIGNMENT to set size and
	alignment of XF and TF modes in compliance with ia64 ABIs.
	Can now hardwire the format of both modes.
	* config/ia64/ia64.c: Change TFmode to XFmode wherever appropriate.
	(general_tfmode_operand, destination_tfmode_operand)
	(tfreg_or_fp01_operand, spill_tfmode_operand): Rename to
	general_xfmode_operand, destination_xfmode_operand,
	xfreg_or_fp01_operand, spill_xfmode_operand respectively.
	(ia64_init_builtins): Make TYPE_PRECISION of fpreg_type
	and float80_type be 96 so they get XFmode.  Use !TARGET_HPUX,
	not INTEL_EXTENDED_IEEE_FORMAT, to decide how to define
	__float128.
	* config/ia64/ia64.h: Default TARGET_HPUX to 0.
	Change TFmode to XFmode wherever appropriate.  Remove all
	references to INTEL_EXTENDED_IEEE_FORMAT.
	(LONG_DOUBLE_TYPE_SIZE): Varies with TARGET_HPUX.
	(LIBGCC2_LONG_DOUBLE_TYPE_SIZE): Define (always 96).
	(PREDICATE_CODES): Update to match function renames.
	* config/ia64/ia64.md: Change TF to XF throughout; rename all
	patterns to match.  Remove all references to
	INTEL_EXTENDED_IEEE_FORMAT.  Update predicate calls to match
	function renames.
	* config/ia64/ia64-protos.c: Update all prototypes to match
	renamed functions.
	* config/ia64/hpux.h: Redefine TARGET_HPUX to 1.
	Remove all references to INTEL_EXTENDED_IEEE_FORMAT.
	* config/ia64/lib1funcs.asm: Add __divxf3 as new name for
	__divtf3; keep old name for backward compatibility.
	(L__compat): New section providing forwarding stubs for
	__fixtfti, __fixunstfti, __floattitf.
	* config/ia64/t-ia64: Add __compat to LIB1ASMFUNCS.

From-SVN: r72916

1 files changed, 11 insertions, 8 deletions

diff --git a/gcc/varasm.c b/gcc/varasm.c
index 29bc605..791b8ad 100644
--- a/gcc/varasm.c
+++ b/gcc/varasm.c
@@ -1905,15 +1905,20 @@ assemble_real (REAL_VALUE_TYPE d, enum machine_mode mode, unsigned int align)
   int i;
   int bitsize, nelts, nunits, units_per;
 
-  /* This is hairy.  We have a quantity of known bitsize.  real_to_target
+  /* This is hairy.  We have a quantity of known size.  real_to_target
      will put it into an array of *host* longs, 32 bits per element
      (even if long is more than 32 bits).  We need to determine the
      number of array elements that are occupied (nelts) and the number
      of *target* min-addressable units that will be occupied in the
-     object file (nunits).  We can assume that BITS_PER_UNIT divides
-     the mode's bitsize evenly, but we can not assume that 32 does.  */
-  bitsize = GET_MODE_BITSIZE (mode);
-  nunits = bitsize / BITS_PER_UNIT;
+     object file (nunits).  We cannot assume that 32 divides the
+     mode's bitsize (size * BITS_PER_UNIT) evenly.
+
+     size * BITS_PER_UNIT is used here to make sure that padding bits
+     (which might appear at either end of the value; real_to_target
+     will include the padding bits in its output array) are included.  */
+
+  nunits = GET_MODE_SIZE (mode);
+  bitsize = nunits * BITS_PER_UNIT;
   nelts = CEIL (bitsize, 32);
   units_per = 32 / BITS_PER_UNIT;
 
@@ -3756,9 +3761,7 @@ output_constant (tree exp, unsigned HOST_WIDE_INT size, unsigned int align)
       if (TREE_CODE (exp) != REAL_CST)
 	error ("initializer for floating value is not a floating constant");
 
-      assemble_real (TREE_REAL_CST (exp),
-		     mode_for_size (size * BITS_PER_UNIT, MODE_FLOAT, 0),
-		     align);
+      assemble_real (TREE_REAL_CST (exp), TYPE_MODE (TREE_TYPE (exp)), align);
       break;
 
     case COMPLEX_TYPE: