In little-endian mode, the field selected for use in a vector splat is
numbered differently than in big-endian mode. This patch corrects the
code generated for little-endian. This addresses 45 test failures when
running the test-suite in little-endian mode.
Bootstrapped and tested on powerpc64-unknown-linux-gnu with no new
regressions. Ok for trunk?
Patch by Anton Blanchard.
Thanks,
Bill
2013-07-22 Bill Schmidt <wschm...@vnet.linux.ibm.com>
for Anton Blanchard <an...@au1.ibm.com>
* config/rs6000/rs6000.c (rs6000_expand_vector_init): Fix
endianness when selecting field to splat.
Index: gcc/config/rs6000/rs6000.c
===================================================================
--- gcc/config/rs6000/rs6000.c (revision 201131)
+++ gcc/config/rs6000/rs6000.c (working copy)
@@ -5177,6 +5177,7 @@ rs6000_expand_vector_init (rtx target, rtx vals)
of 64-bit items is not supported on Altivec. */
if (all_same && GET_MODE_SIZE (inner_mode) <= 4)
{
+ rtx field;
mem = assign_stack_temp (mode, GET_MODE_SIZE (inner_mode));
emit_move_insn (adjust_address_nv (mem, inner_mode, 0),
XVECEXP (vals, 0, 0));
@@ -5187,9 +5188,11 @@ rs6000_expand_vector_init (rtx target, rtx vals)
gen_rtx_SET (VOIDmode,
target, mem),
x)));
+ field = (BYTES_BIG_ENDIAN ? const0_rtx
+ : GEN_INT (GET_MODE_NUNITS (mode) - 1));
x = gen_rtx_VEC_SELECT (inner_mode, target,
gen_rtx_PARALLEL (VOIDmode,
- gen_rtvec (1, const0_rtx)));
+ gen_rtvec (1, field)));
emit_insn (gen_rtx_SET (VOIDmode, target,
gen_rtx_VEC_DUPLICATE (mode, x)));
return;
