[PATCH, rs6000] Implement -maltivec=be for vec_splat builtins

[David Edelsohn]

On Wed, Jan 29, 2014 at 11:50 PM, Bill Schmidt
<wschmidt@linux.vnet.ibm.com> wrote:
> Hi,
>
> This continues the series of -maltivec=be patches, this one handling
> vec_splat.  Since vec_splat is no longer treated as a true intrinsic,
> its default behavior for little endian is changed to use right-to-left
> element indexing for selecting the element to splat.  With -maltivec=be
> for little endian, this reverts to left-to-right element indexing.
>
> The main changes are in altivec.md, and are quite similar to what was
> done for vector merge high/low.  Each of the vector splat builtins is
> split into a define_expand and an "_internal" define_insn.  The expand
> uses the natural element order for the target, unless we have
> -maltivec=be for an LE target, in which case the order is reversed.  The
> _internal insn modifies the element number on the hardware instruction
> for LE to account for the instruction's big-endian bias.
>
> For those of the vector splat instructions that are generated
> internally, rather than through a builtin, a "_direct" insn is supplied
> that generates precisely the instruction requested.
>
> vsx.md contains the same changes for V4SI and V4SF patterns when VSX
> instructions are enabled.  We don't need a separate define_expand for
> this one because the define_expand for altivec_vspltw is used to
> generate the pattern that is recognized by vsx_xxspltw_<mode>.  Either
> vsx_xxspltw_<mode> or *altivec_vspltw_internal handles the pattern,
> depending on whether or not VSX instructions are enabled.
>
> Most of the changes in rs6000.c are to use the new _direct forms.  There
> is one other change to rs6000_expand_vector_init, where the modification
> of the selector field for a generated splat pattern is removed.  That is
> properly handled instead by the code in altivec.md and vsx.md.
>
> As usual, there are four new test cases to cover the various vector
> types for VMX and VSX.  Two existing test cases require adjustment
> because of the change in the default semantics of vec_splat for little
> endian.
>
> Bootstrapped and tested on powerpc64{,le}-unknown-linux-gnu with no
> regressions.  Ok for trunk?
>
> Thanks,
> Bill
>
>
> gcc:
>
> 2014-01-29  Bill Schmidt  <wschmidt@linux.vnet.ibm.com>
>
>         * gcc/config/rs6000/rs6000.c (rs6000_expand_vector_init): Use
>         gen_vsx_xxspltw_v4sf_direct instead of gen_vsx_xxspltw_v4sf;
>         remove element index adjustment for endian (now handled in vsx.md
>         and altivec.md).
>         (altivec_expand_vec_perm_const): Use
>         gen_altivec_vsplt[bhw]_direct instead of gen_altivec_vsplt[bhw].
>         * gcc/config/rs6000/vsx.md (UNSPEC_VSX_XXSPLTW): New unspec.
>         (vsx_xxspltw_<mode>): Adjust element index for little endian.
>         * gcc/config/rs6000/altivec.md (altivec_vspltb): Divide into a
>         define_expand and a new define_insn *altivec_vspltb_internal;
>         adjust for -maltivec=be on a little endian target.
>         (altivec_vspltb_direct): New.
>         (altivec_vsplth): Divide into a define_expand and a new
>         define_insn *altivec_vsplth_internal; adjust for -maltivec=be on a
>         little endian target.
>         (altivec_vsplth_direct): New.
>         (altivec_vspltw): Divide into a define_expand and a new
>         define_insn *altivec_vspltw_internal; adjust for -maltivec=be on a
>         little endian target.
>         (altivec_vspltw_direct): New.
>         (altivec_vspltsf): Divide into a define_expand and a new
>         define_insn *altivec_vspltsf_internal; adjust for -maltivec=be on
>         a little endian target.
>
> gcc/testsuite:
>
> 2014-01-29  Bill Schmidt  <wschmidt@linux.vnet.ibm.com>
>
>         * gcc.dg/vmx/splat.c: New.
>         * gcc.dg/vmx/splat-vsx.c: New.
>         * gcc.dg/vmx/splat-be-order.c: New.
>         * gcc.dg/vmx/splat-vsx-be-order.c: New.
>         * gcc.dg/vmx/eg-5.c: Remove special casing for little endian.
>         * gcc.dg/vmx/sn7153.c: Add special casing for little endian.

Okay.

Thanks David