This patch fixes another small little-endian problem with vectors on PowerPC. Element numbering is reversed when selecting a field for a vector splat operation. Fixing this removes a regression from the test suite when run in LE mode (gcc.dg/vect/slp-perm-3.c).
Patch by Anton Blanchard. Bootstrapped and tested on powerpc64-unknown-linux-gnu with no new regressions. Ok for trunk? Thanks, Bill 2013-07-23 Bill Schmidt <wschm...@linux.vnet.ibm.com> Anton Blanchard <an...@au1.ibm.com> * config/rs6000/rs6000.c (altivec_expand_vec_perm_const): Correct selection of field for vector splat in little endian mode. Index: gcc/config/rs6000/rs6000.c =================================================================== --- gcc/config/rs6000/rs6000.c (revision 201149) +++ gcc/config/rs6000/rs6000.c (working copy) @@ -28460,9 +28460,10 @@ altivec_expand_vec_perm_const (rtx operands[4]) break; if (i == 16) { + int field = BYTES_BIG_ENDIAN ? elt / 2 : 7 - elt / 2; x = gen_reg_rtx (V8HImode); emit_insn (gen_altivec_vsplth (x, gen_lowpart (V8HImode, op0), - GEN_INT (elt / 2))); + GEN_INT (field))); emit_move_insn (target, gen_lowpart (V16QImode, x)); return true; } @@ -28478,9 +28479,10 @@ altivec_expand_vec_perm_const (rtx operands[4]) break; if (i == 16) { + int field = BYTES_BIG_ENDIAN ? elt / 4 : 3 - elt / 4; x = gen_reg_rtx (V4SImode); emit_insn (gen_altivec_vspltw (x, gen_lowpart (V4SImode, op0), - GEN_INT (elt / 4))); + GEN_INT (field))); emit_move_insn (target, gen_lowpart (V16QImode, x)); return true; }