This patch changes the vec_perm_indices element type from HOST_WIDE_INT to poly_int64, so that it can represent indices into a variable-length vector.
Tested on aarch64-linux-gnu, x86_64-linux-gnu and powerpc64le-linux-gnu. Also tested by comparing the before and after assembly output for at least one target per CPU directory. Committed as pre-approved by Jeff here: https://gcc.gnu.org/ml/gcc-patches/2017-11/msg01409.html . Thanks, Richard 2018-01-03 Richard Sandiford <richard.sandif...@linaro.org> gcc/ * vec-perm-indices.h (vec_perm_builder): Change element type from HOST_WIDE_INT to poly_int64. (vec_perm_indices::element_type): Update accordingly. (vec_perm_indices::clamp): Handle polynomial element_types. * vec-perm-indices.c (vec_perm_indices::series_p): Likewise. (vec_perm_indices::all_in_range_p): Likewise. (tree_to_vec_perm_builder): Check for poly_int64 trees rather than shwi trees. * vector-builder.h (vector_builder::stepped_sequence_p): Handle polynomial vec_perm_indices element types. * int-vector-builder.h (int_vector_builder::equal_p): Likewise. * fold-const.c (fold_vec_perm): Likewise. * optabs.c (shift_amt_for_vec_perm_mask): Likewise. * tree-vect-generic.c (lower_vec_perm): Likewise. * tree-vect-slp.c (vect_transform_slp_perm_load): Likewise. * config/aarch64/aarch64.c (aarch64_evpc_tbl): Cast d->perm element type to HOST_WIDE_INT. Index: gcc/vec-perm-indices.h =================================================================== --- gcc/vec-perm-indices.h 2018-01-02 18:27:07.346639775 +0000 +++ gcc/vec-perm-indices.h 2018-01-03 08:46:30.347663443 +0000 @@ -25,7 +25,7 @@ #define GCC_VEC_PERN_INDICES_H 1 /* A vector_builder for building constant permutation vectors. The elements do not need to be clamped to a particular range of input elements. */ -typedef int_vector_builder<HOST_WIDE_INT> vec_perm_builder; +typedef int_vector_builder<poly_int64> vec_perm_builder; /* This class represents a constant permutation vector, such as that used as the final operand to a VEC_PERM_EXPR. @@ -49,7 +49,7 @@ typedef int_vector_builder<HOST_WIDE_INT different numbers of elements. */ class vec_perm_indices { - typedef HOST_WIDE_INT element_type; + typedef poly_int64 element_type; public: vec_perm_indices (); @@ -118,13 +118,17 @@ vec_perm_indices::vec_perm_indices (cons inline vec_perm_indices::element_type vec_perm_indices::clamp (element_type elt) const { - element_type limit = input_nelts (); - elt %= limit; + element_type limit = input_nelts (), elem_within_input; + int input; + if (!can_div_trunc_p (elt, limit, &input, &elem_within_input)) + return elt; + /* Treat negative elements as counting from the end. This only matters if the vector size is not a power of 2. */ - if (elt < 0) - elt += limit; - return elt; + if (known_lt (elem_within_input, 0)) + return elem_within_input + limit; + + return elem_within_input; } /* Return the value of vector element I, which might or might not be Index: gcc/vec-perm-indices.c =================================================================== --- gcc/vec-perm-indices.c 2018-01-02 18:27:52.375765255 +0000 +++ gcc/vec-perm-indices.c 2018-01-03 08:46:30.347663443 +0000 @@ -95,7 +95,7 @@ vec_perm_indices::series_p (unsigned int element_type in_base, element_type in_step) const { /* Check the base value. */ - if (clamp (m_encoding.elt (out_base)) != clamp (in_base)) + if (maybe_ne (clamp (m_encoding.elt (out_base)), clamp (in_base))) return false; unsigned int full_nelts = m_encoding.full_nelts (); @@ -127,7 +127,7 @@ vec_perm_indices::series_p (unsigned int element_type v0 = m_encoding.elt (out_base - out_step); element_type v1 = m_encoding.elt (out_base); - if (clamp (v1 - v0) != in_step) + if (maybe_ne (clamp (v1 - v0), in_step)) return false; out_base += out_step; @@ -146,7 +146,7 @@ vec_perm_indices::all_in_range_p (elemen unsigned int nelts_per_pattern = m_encoding.nelts_per_pattern (); unsigned int base_nelts = npatterns * MIN (nelts_per_pattern, 2); for (unsigned int i = 0; i < base_nelts; ++i) - if (m_encoding[i] < start || (m_encoding[i] - start) >= size) + if (!known_in_range_p (m_encoding[i], start, size)) return false; /* For stepped encodings, check the full range of the series. */ @@ -174,8 +174,11 @@ vec_perm_indices::all_in_range_p (elemen wide enough for overflow not to be a problem. */ element_type headroom_down = base1 - start; element_type headroom_up = size - headroom_down - 1; - if (headroom_up < step * step_nelts - && headroom_down < (limit - step) * step_nelts) + HOST_WIDE_INT diff; + if ((!step.is_constant (&diff) + || maybe_lt (headroom_up, diff * step_nelts)) + && (!(limit - step).is_constant (&diff) + || maybe_lt (headroom_down, diff * step_nelts))) return false; } } @@ -191,14 +194,14 @@ tree_to_vec_perm_builder (vec_perm_build { unsigned int encoded_nelts = vector_cst_encoded_nelts (cst); for (unsigned int i = 0; i < encoded_nelts; ++i) - if (!tree_fits_shwi_p (VECTOR_CST_ENCODED_ELT (cst, i))) + if (!tree_fits_poly_int64_p (VECTOR_CST_ENCODED_ELT (cst, i))) return false; builder->new_vector (TYPE_VECTOR_SUBPARTS (TREE_TYPE (cst)), VECTOR_CST_NPATTERNS (cst), VECTOR_CST_NELTS_PER_PATTERN (cst)); for (unsigned int i = 0; i < encoded_nelts; ++i) - builder->quick_push (tree_to_shwi (VECTOR_CST_ENCODED_ELT (cst, i))); + builder->quick_push (tree_to_poly_int64 (VECTOR_CST_ENCODED_ELT (cst, i))); return true; } Index: gcc/vector-builder.h =================================================================== --- gcc/vector-builder.h 2018-01-02 18:27:07.346639775 +0000 +++ gcc/vector-builder.h 2018-01-03 08:46:30.347663443 +0000 @@ -284,7 +284,8 @@ vector_builder<T, Derived>::stepped_sequ || !derived ()->integral_p (elt3)) return false; - if (derived ()->step (elt1, elt2) != derived ()->step (elt2, elt3)) + if (maybe_ne (derived ()->step (elt1, elt2), + derived ()->step (elt2, elt3))) return false; if (!derived ()->can_elide_p (elt3)) Index: gcc/int-vector-builder.h =================================================================== --- gcc/int-vector-builder.h 2018-01-02 18:26:37.619823164 +0000 +++ gcc/int-vector-builder.h 2018-01-03 08:46:30.345663760 +0000 @@ -66,7 +66,7 @@ int_vector_builder<T>::int_vector_builde inline bool int_vector_builder<T>::equal_p (T elt1, T elt2) const { - return elt1 == elt2; + return known_eq (elt1, elt2); } /* Return the value of element ELT2 minus the value of element ELT1. */ Index: gcc/fold-const.c =================================================================== --- gcc/fold-const.c 2018-01-03 07:17:07.401983352 +0000 +++ gcc/fold-const.c 2018-01-03 08:46:30.345663760 +0000 @@ -8945,9 +8945,12 @@ fold_vec_perm (tree type, tree arg0, tre tree_vector_builder out_elts (type, nelts, 1); for (i = 0; i < nelts; i++) { - if (!CONSTANT_CLASS_P (in_elts[sel[i]])) + HOST_WIDE_INT index; + if (!sel[i].is_constant (&index)) + return NULL_TREE; + if (!CONSTANT_CLASS_P (in_elts[index])) need_ctor = true; - out_elts.quick_push (unshare_expr (in_elts[sel[i]])); + out_elts.quick_push (unshare_expr (in_elts[index])); } if (need_ctor) Index: gcc/optabs.c =================================================================== --- gcc/optabs.c 2018-01-02 18:27:52.375765255 +0000 +++ gcc/optabs.c 2018-01-03 08:46:30.346663602 +0000 @@ -5398,16 +5398,18 @@ shift_amt_for_vec_perm_mask (machine_mod { unsigned int nelt = GET_MODE_NUNITS (mode); unsigned int bitsize = GET_MODE_UNIT_BITSIZE (mode); - unsigned int first = sel[0]; - if (first >= nelt) + poly_int64 first = sel[0]; + if (maybe_ge (sel[0], nelt)) return NULL_RTX; if (!sel.series_p (0, 1, first, 1)) for (unsigned int i = 1; i < nelt; i++) { - unsigned int expected = i + first; + poly_int64 expected = i + first; /* Indices into the second vector are all equivalent. */ - if (MIN (nelt, sel[i]) != MIN (nelt, expected)) + if (maybe_lt (sel[i], nelt) + ? maybe_ne (sel[i], expected) + : maybe_lt (expected, nelt)) return NULL_RTX; } Index: gcc/tree-vect-generic.c =================================================================== --- gcc/tree-vect-generic.c 2018-01-03 07:16:49.037017897 +0000 +++ gcc/tree-vect-generic.c 2018-01-03 08:46:30.346663602 +0000 @@ -1337,18 +1337,19 @@ lower_vec_perm (gimple_stmt_iterator *gs != CODE_FOR_nothing && TREE_CODE (vec1) == VECTOR_CST && initializer_zerop (vec1) - && indices[0] - && indices[0] < elements) + && maybe_ne (indices[0], 0) + && known_lt (indices[0], elements)) { bool ok_p = indices.series_p (0, 1, indices[0], 1); if (!ok_p) { for (i = 1; i < elements; ++i) { - unsigned int expected = i + indices[0]; + poly_int64 expected = i + indices[0]; /* Indices into the second vector are all equivalent. */ - if (MIN (elements, (unsigned) indices[i]) - != MIN (elements, expected)) + if (maybe_lt (indices[i], elements) + ? maybe_ne (indices[i], expected) + : maybe_lt (expected, elements)) break; } ok_p = i == elements; Index: gcc/tree-vect-slp.c =================================================================== --- gcc/tree-vect-slp.c 2018-01-03 07:16:36.566760459 +0000 +++ gcc/tree-vect-slp.c 2018-01-03 08:46:30.347663443 +0000 @@ -3727,8 +3727,10 @@ vect_transform_slp_perm_load (slp_tree n vect_location, "unsupported vect permute { "); for (i = 0; i < nunits; ++i) - dump_printf (MSG_MISSED_OPTIMIZATION, - HOST_WIDE_INT_PRINT_DEC " ", mask[i]); + { + dump_dec (MSG_MISSED_OPTIMIZATION, mask[i]); + dump_printf (MSG_MISSED_OPTIMIZATION, " "); + } dump_printf (MSG_MISSED_OPTIMIZATION, "}\n"); } gcc_assert (analyze_only); Index: gcc/config/aarch64/aarch64.c =================================================================== --- gcc/config/aarch64/aarch64.c 2018-01-03 07:14:58.947066762 +0000 +++ gcc/config/aarch64/aarch64.c 2018-01-03 08:46:30.343664078 +0000 @@ -13625,7 +13625,7 @@ aarch64_evpc_tbl (struct expand_vec_perm mode on NEON. Reverse the index within each word but not the word itself. */ rperm[i] = GEN_INT (BYTES_BIG_ENDIAN ? d->perm[i] ^ (nunits - 1) - : d->perm[i]); + : (HOST_WIDE_INT) d->perm[i]); } sel = gen_rtx_CONST_VECTOR (vmode, gen_rtvec_v (nelt, rperm)); sel = force_reg (vmode, sel);
