On Sat, Nov 2, 2013 at 7:44 PM, Bill Schmidt
<wschm...@linux.vnet.ibm.com> wrote:
> Hi,
>
> When I created the patch to split VSX loads and stores to add permutes
> so the register image was correctly little endian, I forgot to implement
> a known requirement. When a VSX store is split after reload, we must
> reuse the source register for the permute, which leaves it in a swapped
> state after the store. If the register remains live, this is invalid.
> After the store, we need to permute the register back to its original
> value.
>
> For each of the little endian VSX store patterns, I've replaced the
> define_insn_and_split with a define_insn and two define_splits, one for
> prior to reload and one for after reload. The post-reload split has the
> extra behavior.
>
> I don't know of a way to set the insn's length attribute conditionally,
> so I'm optimistically setting this to 8, though it will be 12 in the
> post-reload case. Is there any concern with that? Is there a way to
> make it fully accurate?
>
> Bootstrapped and tested on powerpc64{,le}-unknown-linux-gnu with no
> regressions. This fixes two failing test cases. Is this ok for trunk?
>
> Thanks!
> Bill
>
>
> 2013-11-02 Bill Schmidt <wschm...@linux.vnet.ibm.com>
>
> * config/rs6000/vsx.md (*vsx_le_perm_store_<mode> for VSX_D):
> Replace the define_insn_and_split with a define_insn and two
> define_splits, with the split after reload re-permuting the source
> register to its original value.
> (*vsx_le_perm_store_<mode> for VSX_W): Likewise.
> (*vsx_le_perm_store_v8hi): Likewise.
> (*vsx_le_perm_store_v16qi): Likewise.
This is okay, but you must change the insn length to the conservative,
larger value. insn length is used to reverse branches when the
displacement is too large.
Thanks, David
