Hi,
In little endian mode, we managed to convert a load of the V4SI vector
{3, 3, 3, 7} into a vspltisw of 3, apparently taking offense at the
number 7. It turns out we only looked at the first N-1 elements of an
N-element vector in little endian mode, and verified the zeroth element
twice. Adjusting the loop boundaries fixes the problem.
Currently bootstrapping for powerpc64{,le}-unknown-linux-gnu. Ok to
commit to trunk if no regressions?
Thanks,
Bill
2013-10-18 Bill Schmidt <wschm...@linux.vnet.ibm.com>
* config/rs6000/rs6000.c (vspltis_constant): Make sure we check
all elements for both endian flavors.
Index: gcc/config/rs6000/rs6000.c
===================================================================
@@ -4932,6 +4932,8 @@ vspltis_constant (rtx op, unsigned step, unsigned
unsigned nunits;
unsigned bitsize;
unsigned mask;
+ unsigned start;
+ unsigned end;
HOST_WIDE_INT val;
HOST_WIDE_INT splat_val;
@@ -4981,7 +4983,10 @@ vspltis_constant (rtx op, unsigned step, unsigned
/* Check if VAL is present in every STEP-th element, and the
other elements are filled with its most significant bit. */
- for (i = 0; i < nunits - 1; ++i)
+ start = BYTES_BIG_ENDIAN ? 0 : 1;
+ end = BYTES_BIG_ENDIAN ? nunits - 1 : nunits;
+
+ for (i = start; i < end; ++i)
{
HOST_WIDE_INT desired_val;
if (((BYTES_BIG_ENDIAN ? i + 1 : i) & (step - 1)) == 0)
