On Fri, 2017-10-13 at 11:36 +0200, Richard Biener wrote:
> On Thu, Oct 12, 2017 at 10:03 PM, Will Schmidt
> <will_schm...@vnet.ibm.com> wrote:
> > Hi,
> >
> > Add support for gimple folding of vec_cmp_{eq,ge,gt,le,ne} for
> > the integer data types.
> >
> > This adds a handful of entries to the switch statement in
> > builtin_function_type
> > for those builtins having unsigned arguments.
> >
> > Three entries are added to vsx.md to enable vcmpne[bhw] instruction, where
> > we
> > would otherwise generate a vcmpeq + vnor.
> >
> > This patch requires the previously posted "allow integer return type from
> > vector compares" patch.
> >
> > A handful of existing tests required updates to their specified optimization
> > levels to continue to generate the desired code. builtins-3-p9.c in
> > particular
> > has been updated to reflect improved code gen with the higher specified
> > optimization level. Testcase coverage is otherwise handled by the
> > already-in-tree
> > gcc.target/powerpc/fold-vec-cmp-*.c tests.
> >
> > Tested OK on P6 and newer. OK for trunk?
> >
> > Thanks,
> > -Will
> >
> > [gcc]
> >
> > 2017-10-12 Will Schmidt <will_schm...@vnet.ibm.com>
> >
> > * config/rs6000/rs6000.c: (rs6000_gimple_fold_builtin) Add support
> > for
> > folding of vector compares. (builtin_function_type) Add compare
> > builtins to the list of functions having unsigned arguments.
> > * config/rs6000/vsx.md: Add vcmpne{b,h,w} instructions.
> >
> > [testsuite]
> >
> > 2017-10-12 Will Schmidt <will_schm...@vnet.ibm.com>
> >
> > * gcc.target/powerpc/builtins-3-p9.c: Add -O1, update
> > expected codegen checks.
> > * gcc.target/powerpc/vec-cmp-sel.c: Mark vars as volatile.
> > * gcc.target/powerpc/vsu/vec-cmpne-0.c: Add -O1.
> > * gcc.target/powerpc/vsu/vec-cmpne-1.c: Add -O1.
> > * gcc.target/powerpc/vsu/vec-cmpne-2.c: Add -O1.
> > * gcc.target/powerpc/vsu/vec-cmpne-3.c: Add -O1.
> > * gcc.target/powerpc/vsu/vec-cmpne-4.c: Add -O1.
> > * gcc.target/powerpc/vsu/vec-cmpne-5.c: Add -O1.
> > * gcc.target/powerpc/vsu/vec-cmpne-6.c: Add -O1.
> >
> > diff --git a/gcc/config/rs6000/rs6000.c b/gcc/config/rs6000/rs6000.c
> > index 12ddd97..7e73239 100644
> > --- a/gcc/config/rs6000/rs6000.c
> > +++ b/gcc/config/rs6000/rs6000.c
> > @@ -16605,17 +16605,93 @@ rs6000_gimple_fold_builtin (gimple_stmt_iterator
> > *gsi)
> > build_int_cst (arg2_type, 0)),
> > arg0);
> > gimple_set_location (g, loc);
> > gsi_replace (gsi, g, true);
> > return true;
> > }
> > + /* Vector compares (integer); EQ, NE, GE, GT, LE. */
> > + case ALTIVEC_BUILTIN_VCMPEQUB:
> > + case ALTIVEC_BUILTIN_VCMPEQUH:
> > + case ALTIVEC_BUILTIN_VCMPEQUW:
> > + case P8V_BUILTIN_VCMPEQUD:
> > + {
> > + arg0 = gimple_call_arg (stmt, 0);
> > + arg1 = gimple_call_arg (stmt, 1);
> > + lhs = gimple_call_lhs (stmt);
> > + gimple *g = gimple_build_assign (lhs, EQ_EXPR, arg0, arg1);
>
> As said elsewhere this needs to become either
>
> tree ctype = build_same_sized_truth_vector_type (TREE_TYPE (lhs));
> gimple_build_assign (make_ssa_name (ctype), EQ_EXPR, arg0, arg1)
> gimple_build_assign (lhs, VIEW_CONVERT_EXPR, lhs above);
>
> (eventually the VCE can be elided - try) or
>
> gimple_build_assign (lhs, VEC_COND_EXPR,
> fold_build2 (EQ_EXPR, ctype, arg0, arg1),
> vector-with-trues, vector-with-falses);
>
> depending on what your target can expand.
Alright, i'll work with this some more and see what I come up with.
Thanks for the review and feedback. :-)
-Will
