On Sun, Aug 30, 2015 at 7:19 PM, Bill Schmidt <wschm...@linux.vnet.ibm.com> wrote: > Hi, > > The VSX swap optimization currently misses opportunities to optimize > loops when expressions corresponding to xxpermdi instructions are > present (other than xxswapd instructions associated with loads and > stores). These occur commonly when interleaving vector double or vector > unsigned long operands, or when concatenating two doubles or unsigned > longs to make a V2DI or V2DF result. This patch adds logic to recognize > these insns and adjust them to account for swapped doublewords in the > computation. > > Both opportunities arise in a simple test case that performs a reduction > on complex multiplications, which I've added here. > > Bootstrapped and tested on powerpc64le-unknown-linux-gnu with no > regressions. Is this ok for trunk? > > Thanks, > Bill > > > [gcc] > > 2015-08-30 Bill Schmidt <wschm...@linux.vnet.ibm.com> > > * config/rs6000/rs6000.c (swap_web_entry): Enlarge > special_handling bitfield. > (special_handling_values): Add SH_XXPERMDI and SH_CONCAT. > (rtx_is_swappable_p): Add handling for vec_select/vec_concat > form > that represents a general xxpermdi. > (insn_is_swappable_p): Add handling for vec_concat of two > doublewords, which maps to a specific xxpermdi. > (adjust_xxpermdi): New function. > (adjust_concat): Likewise. > (handle_special_swappables): Call adjust_xxpermdi and > adjust_concat. > (dump_swap_insn_table): Handle SH_XXPERMDI and SH_CONCAT. > > [gcc/testsuite] > > 2015-08-30 Bill Schmidt <wschm...@linux.vnet.ibm.com> > > * gcc.target/powerpc/swaps-p8-19.c: New test.
Okay. Thanks, David
