| Message ID | 1507838577.26707.176.camel@brimstone.rchland.ibm.com |
|---|---|
| State | New |
| Headers | show |
| Series | (gimple) Allow integer return type from vector compares | expand |
On Thu, Oct 12, 2017 at 10:02 PM, Will Schmidt <will_schmidt@vnet.ibm.com> wrote: > Hi, > > Update the logic in verify_gimple_comparision to allow a vector integer result > from a vector comparison, where it previously was limited to only allowing > compares with boolean results. This allows powerpc intrinsics such as this > one to build (after gimple folding): > vector bool int vec_cmpeq (vector signed int, vector signed int); > > This has been tested in conjunction with the "rs6000 GIMPLE folding for vector > compares" patch (posted separately) on p6 and newer. > > OK for trunk? Well. It was this way before AVX512 introduced those vector boolean types. There we decided to have comparisons always return a boolean vector. Previously there was an additional restriction in that the vector elements of the boolean result had to be the same size as the elements of the comparison operands. You are now supposed to use the "canonical" _1 = _2 != _3 ? { your true value vector } : { your false value vector}; with a VEC_COND_EXPR. Richard. > Thanks, > -Will > > [gcc] > > 2017-10-12 Will Schmidt <will_schmidt@vnet.ibm.com> > > * gcc/tree-cfg.c: (@ verify_gimple_comparison ): allow boolean result > from vector compares. > > diff --git a/gcc/tree-cfg.c b/gcc/tree-cfg.c > index b5e0460..adf3607 100644 > --- a/gcc/tree-cfg.c > +++ b/gcc/tree-cfg.c > @@ -3624,14 +3624,16 @@ verify_gimple_comparison (tree type, tree op0, tree op1, enum tree_code code) > debug_generic_expr (op0_type); > debug_generic_expr (op1_type); > return true; > } > } > - /* Or a boolean vector type with the same element count > - as the comparison operand types. */ > + /* Or a vector type with the same element count > + as the comparison operand types. The vector type may > + be boolean or integer. */ > else if (TREE_CODE (type) == VECTOR_TYPE > - && TREE_CODE (TREE_TYPE (type)) == BOOLEAN_TYPE) > + && (( TREE_CODE (TREE_TYPE (type)) == BOOLEAN_TYPE) > + || ( TREE_CODE (TREE_TYPE (type)) == INTEGER_TYPE))) > { > if (TREE_CODE (op0_type) != VECTOR_TYPE > || TREE_CODE (op1_type) != VECTOR_TYPE) > { > error ("non-vector operands in vector comparison"); > >
diff --git a/gcc/tree-cfg.c b/gcc/tree-cfg.c index b5e0460..adf3607 100644 --- a/gcc/tree-cfg.c +++ b/gcc/tree-cfg.c @@ -3624,14 +3624,16 @@ verify_gimple_comparison (tree type, tree op0, tree op1, enum tree_code code) debug_generic_expr (op0_type); debug_generic_expr (op1_type); return true; } } - /* Or a boolean vector type with the same element count - as the comparison operand types. */ + /* Or a vector type with the same element count + as the comparison operand types. The vector type may + be boolean or integer. */ else if (TREE_CODE (type) == VECTOR_TYPE - && TREE_CODE (TREE_TYPE (type)) == BOOLEAN_TYPE) + && (( TREE_CODE (TREE_TYPE (type)) == BOOLEAN_TYPE) + || ( TREE_CODE (TREE_TYPE (type)) == INTEGER_TYPE))) { if (TREE_CODE (op0_type) != VECTOR_TYPE || TREE_CODE (op1_type) != VECTOR_TYPE) { error ("non-vector operands in vector comparison");
Hi, Update the logic in verify_gimple_comparision to allow a vector integer result from a vector comparison, where it previously was limited to only allowing compares with boolean results. This allows powerpc intrinsics such as this one to build (after gimple folding): vector bool int vec_cmpeq (vector signed int, vector signed int); This has been tested in conjunction with the "rs6000 GIMPLE folding for vector compares" patch (posted separately) on p6 and newer. OK for trunk? Thanks, -Will [gcc] 2017-10-12 Will Schmidt <will_schmidt@vnet.ibm.com> * gcc/tree-cfg.c: (@ verify_gimple_comparison ): allow boolean result from vector compares.