The attached problem fixes a regression in handling DFmode aggregates
in pa_function_value.
Sometimes, an aggregrate with a single DFmode field could not be
handled by the existing code
which used a BLKmode parallel. We can just use a register when the
aggregate fits exactly in a
single or double word. This avoids an ICE in store_field.
Tested on hppa2.0w-hp-hpux11.11, hppa-unknown-linux-gnu and hppa64-hp-
hpux11.11. Committed
to trunk and 4.9 branch.
Dave
--
John David Anglin dave.ang...@bell.net
2014-09-16 John David Anglin <dang...@gcc.gnu.org>
PR target/61853
* config/pa/pa.c (pa_function_value): Directly handle aggregates
that fit exactly in a word or double word.
Index: config/pa/pa.c
===================================================================
--- config/pa/pa.c (revision 215242)
+++ config/pa/pa.c (working copy)
@@ -9298,6 +9298,12 @@
|| TREE_CODE (valtype) == COMPLEX_TYPE
|| TREE_CODE (valtype) == VECTOR_TYPE)
{
+ HOST_WIDE_INT valsize = int_size_in_bytes (valtype);
+
+ /* Handle aggregates that fit exactly in a word or double word. */
+ if ((valsize & (UNITS_PER_WORD - 1)) == 0)
+ return gen_rtx_REG (TYPE_MODE (valtype), 28);
+
if (TARGET_64BIT)
{
/* Aggregates with a size less than or equal to 128 bits are
@@ -9306,7 +9312,7 @@
memory. */
rtx loc[2];
int i, offset = 0;
- int ub = int_size_in_bytes (valtype) <= UNITS_PER_WORD ? 1 : 2;
+ int ub = valsize <= UNITS_PER_WORD ? 1 : 2;
for (i = 0; i < ub; i++)
{
@@ -9318,7 +9324,7 @@
return gen_rtx_PARALLEL (BLKmode, gen_rtvec_v (ub, loc));
}
- else if (int_size_in_bytes (valtype) > UNITS_PER_WORD)
+ else if (valsize > UNITS_PER_WORD)
{
/* Aggregates 5 to 8 bytes in size are returned in general
registers r28-r29 in the same manner as other non
