Handle vector boolean types when calculating the SLP unroll factor
authorRichard Sandiford <richard.sandiford@linaro.org>
Wed, 16 May 2018 11:50:44 +0000 (11:50 +0000)
committerRichard Sandiford <rsandifo@gcc.gnu.org>
Wed, 16 May 2018 11:50:44 +0000 (11:50 +0000)
commit1f3cb66326d724cca10ac6097ab32c53191c53ff
tree1694c0be101b9998ee5b4b8a8082de6d8c95c76b
parentc448fedea9abae497aaebccf55d1b1db2c3e5593
Handle vector boolean types when calculating the SLP unroll factor

The SLP unrolling factor is calculated by finding the smallest
scalar type for each SLP statement and taking the number of required
lanes from the vector versions of those scalar types.  E.g. for an
int32->int64 conversion, it's the vector of int32s rather than the
vector of int64s that determines the unroll factor.

We rely on tree-vect-patterns.c to replace boolean operations like:

   bool a, b, c;
   a = b & c;

with integer operations of whatever the best size is in context.
E.g. if b and c are fed by comparisons of ints, a, b and c will become
the appropriate size for an int comparison.  For most targets this means
that a, b and c will end up as int-sized themselves, but on targets like
SVE and AVX512 with packed vector booleans, they'll instead become a
small bitfield like :1, padded to a byte for memory purposes.
The SLP code would then take these scalar types and try to calculate
the vector type for them, causing the unroll factor to be much higher
than necessary.

This patch tries to make the SLP code use the same approach as the
loop vectorizer, by splitting out the code that calculates the
statement vector type and the vector type that should be used for
the number of units.

2018-05-16  Richard Sandiford  <richard.sandiford@linaro.org>

gcc/
* tree-vectorizer.h (vect_get_vector_types_for_stmt): Declare.
(vect_get_mask_type_for_stmt): Likewise.
* tree-vect-slp.c (vect_two_operations_perm_ok_p): New function,
split out from...
(vect_build_slp_tree_1): ...here.  Use vect_get_vector_types_for_stmt
to determine the statement's vector type and the vector type that
should be used for calculating nunits.  Deal with cases in which
the type has to be deferred.
(vect_slp_analyze_node_operations): Use vect_get_vector_types_for_stmt
and vect_get_mask_type_for_stmt to calculate STMT_VINFO_VECTYPE.
* tree-vect-loop.c (vect_determine_vf_for_stmt_1)
(vect_determine_vf_for_stmt): New functions, split out from...
(vect_determine_vectorization_factor): ...here.
* tree-vect-stmts.c (vect_get_vector_types_for_stmt)
(vect_get_mask_type_for_stmt): New functions, split out from
vect_determine_vectorization_factor.

gcc/testsuite/
* gcc.target/aarch64/sve/vcond_10.c: New test.
* gcc.target/aarch64/sve/vcond_10_run.c: Likewise.
* gcc.target/aarch64/sve/vcond_11.c: Likewise.
* gcc.target/aarch64/sve/vcond_11_run.c: Likewise.

From-SVN: r260287
gcc/ChangeLog
gcc/testsuite/ChangeLog
gcc/testsuite/gcc.target/aarch64/sve/vcond_10.c [new file with mode: 0644]
gcc/testsuite/gcc.target/aarch64/sve/vcond_10_run.c [new file with mode: 0644]
gcc/testsuite/gcc.target/aarch64/sve/vcond_11.c [new file with mode: 0644]
gcc/testsuite/gcc.target/aarch64/sve/vcond_11_run.c [new file with mode: 0644]
gcc/tree-vect-loop.c
gcc/tree-vect-slp.c
gcc/tree-vect-stmts.c
gcc/tree-vectorizer.h