This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [14/n] PR85694: Rework overwidening detection
- From: Christophe Lyon <christophe dot lyon at linaro dot org>
- To: Richard Biener <richard dot guenther at gmail dot com>, gcc Patches <gcc-patches at gcc dot gnu dot org>, Richard Sandiford <richard dot sandiford at arm dot com>
- Date: Tue, 3 Jul 2018 22:07:42 +0200
- Subject: Re: [14/n] PR85694: Rework overwidening detection
- References: <878t79eob8.fsf@arm.com> <87a7rdbzam.fsf@arm.com> <CAFiYyc3yeFp-kXh-Mrg-e1pc7k210rJXEmq9U-TpncBPZ6SVsQ@mail.gmail.com> <87wouc7i2w.fsf@arm.com>
On Tue, 3 Jul 2018 at 12:02, Richard Sandiford
<richard.sandiford@arm.com> wrote:
>
> Richard Biener <richard.guenther@gmail.com> writes:
> > On Fri, Jun 29, 2018 at 1:36 PM Richard Sandiford
> > <richard.sandiford@arm.com> wrote:
> >>
> >> Richard Sandiford <richard.sandiford@arm.com> writes:
> >> > This patch is the main part of PR85694. The aim is to recognise at least:
> >> >
> >> > signed char *a, *b, *c;
> >> > ...
> >> > for (int i = 0; i < 2048; i++)
> >> > c[i] = (a[i] + b[i]) >> 1;
> >> >
> >> > as an over-widening pattern, since the addition and shift can be done
> >> > on shorts rather than ints. However, it ended up being a lot more
> >> > general than that.
> >> >
> >> > The current over-widening pattern detection is limited to a few simple
> >> > cases: logical ops with immediate second operands, and shifts by a
> >> > constant. These cases are enough for common pixel-format conversion
> >> > and can be detected in a peephole way.
> >> >
> >> > The loop above requires two generalisations of the current code: support
> >> > for addition as well as logical ops, and support for non-constant second
> >> > operands. These are harder to detect in the same peephole way, so the
> >> > patch tries to take a more global approach.
> >> >
> >> > The idea is to get information about the minimum operation width
> >> > in two ways:
> >> >
> >> > (1) by using the range information attached to the SSA_NAMEs
> >> > (effectively a forward walk, since the range info is
> >> > context-independent).
> >> >
> >> > (2) by back-propagating the number of output bits required by
> >> > users of the result.
> >> >
> >> > As explained in the comments, there's a balance to be struck between
> >> > narrowing an individual operation and fitting in with the surrounding
> >> > code. The approach is pretty conservative: if we could narrow an
> >> > operation to N bits without changing its semantics, it's OK to do that if:
> >> >
> >> > - no operations later in the chain require more than N bits; or
> >> >
> >> > - all internally-defined inputs are extended from N bits or fewer,
> >> > and at least one of them is single-use.
> >> >
> >> > See the comments for the rationale.
> >> >
> >> > I didn't bother adding STMT_VINFO_* wrappers for the new fields
> >> > since the code seemed more readable without.
> >> >
> >> > Tested on aarch64-linux-gnu and x86_64-linux-gnu. OK to install?
> >>
> >> Here's a version rebased on top of current trunk. Changes from last time:
> >>
> >> - reintroduce dump_generic_expr_loc, with the obvious change to the
> >> prototype
> >>
> >> - fix a typo in a comment
> >>
> >> - use vect_element_precision from the new version of 12/n.
> >>
> >> Tested as before. OK to install?
> >
> > OK.
>
> Thanks. For the record, here's what I installed (updated on top of
> Dave's recent patch, and with an obvious fix to vect-widen-mult-u8-u32.c).
>
> Richard
>
Hi,
It seems the new bb-slp-over-widen tests lack a -fdump option:
gcc.dg/vect/bb-slp-over-widen-2.c -flto -ffat-lto-objects : dump file
does not exist
UNRESOLVED: gcc.dg/vect/bb-slp-over-widen-2.c -flto -ffat-lto-objects
scan-tree-dump-times vect "basic block vectorized" 2
Christophe
>
> 2018-07-03 Richard Sandiford <richard.sandiford@arm.com>
>
> gcc/
> * poly-int.h (print_hex): New function.
> * dumpfile.h (dump_dec, dump_hex): Declare.
> * dumpfile.c (dump_dec, dump_hex): New poly_wide_int functions.
> * tree-vectorizer.h (_stmt_vec_info): Add min_output_precision,
> min_input_precision, operation_precision and operation_sign.
> * tree-vect-patterns.c (vect_get_range_info): New function.
> (vect_same_loop_or_bb_p, vect_single_imm_use)
> (vect_operation_fits_smaller_type): Delete.
> (vect_look_through_possible_promotion): Add an optional
> single_use_p parameter.
> (vect_recog_over_widening_pattern): Rewrite to use new
> stmt_vec_info infomration. Handle one operation at a time.
> (vect_recog_cast_forwprop_pattern, vect_narrowable_type_p)
> (vect_truncatable_operation_p, vect_set_operation_type)
> (vect_set_min_input_precision): New functions.
> (vect_determine_min_output_precision_1): Likewise.
> (vect_determine_min_output_precision): Likewise.
> (vect_determine_precisions_from_range): Likewise.
> (vect_determine_precisions_from_users): Likewise.
> (vect_determine_stmt_precisions, vect_determine_precisions): Likewise.
> (vect_vect_recog_func_ptrs): Put over_widening first.
> Add cast_forwprop.
> (vect_pattern_recog): Call vect_determine_precisions.
>
> gcc/testsuite/
> * gcc.dg/vect/vect-widen-mult-u8-u32.c: Check specifically for a
> widen_mult pattern.
> * gcc.dg/vect/vect-over-widen-1.c: Update the scan tests for new
> over-widening messages.
> * gcc.dg/vect/vect-over-widen-1-big-array.c: Likewise.
> * gcc.dg/vect/vect-over-widen-2.c: Likewise.
> * gcc.dg/vect/vect-over-widen-2-big-array.c: Likewise.
> * gcc.dg/vect/vect-over-widen-3.c: Likewise.
> * gcc.dg/vect/vect-over-widen-3-big-array.c: Likewise.
> * gcc.dg/vect/vect-over-widen-4.c: Likewise.
> * gcc.dg/vect/vect-over-widen-4-big-array.c: Likewise.
> * gcc.dg/vect/bb-slp-over-widen-1.c: New test.
> * gcc.dg/vect/bb-slp-over-widen-2.c: Likewise.
> * gcc.dg/vect/vect-over-widen-5.c: Likewise.
> * gcc.dg/vect/vect-over-widen-6.c: Likewise.
> * gcc.dg/vect/vect-over-widen-7.c: Likewise.
> * gcc.dg/vect/vect-over-widen-8.c: Likewise.
> * gcc.dg/vect/vect-over-widen-9.c: Likewise.
> * gcc.dg/vect/vect-over-widen-10.c: Likewise.
> * gcc.dg/vect/vect-over-widen-11.c: Likewise.
> * gcc.dg/vect/vect-over-widen-12.c: Likewise.
> * gcc.dg/vect/vect-over-widen-13.c: Likewise.
> * gcc.dg/vect/vect-over-widen-14.c: Likewise.
> * gcc.dg/vect/vect-over-widen-15.c: Likewise.
> * gcc.dg/vect/vect-over-widen-16.c: Likewise.
> * gcc.dg/vect/vect-over-widen-17.c: Likewise.
> * gcc.dg/vect/vect-over-widen-18.c: Likewise.
> * gcc.dg/vect/vect-over-widen-19.c: Likewise.
> * gcc.dg/vect/vect-over-widen-20.c: Likewise.
> * gcc.dg/vect/vect-over-widen-21.c: Likewise.
> ------------------------------------------------------------------------------
>
> Index: gcc/poly-int.h
> ===================================================================
> --- gcc/poly-int.h 2018-07-03 09:01:31.075962445 +0100
> +++ gcc/poly-int.h 2018-07-03 09:02:36.563413564 +0100
> @@ -2420,6 +2420,25 @@ print_dec (const poly_int_pod<N, C> &val
> poly_coeff_traits<C>::signedness ? SIGNED : UNSIGNED);
> }
>
> +/* Use print_hex to print VALUE to FILE. */
> +
> +template<unsigned int N, typename C>
> +void
> +print_hex (const poly_int_pod<N, C> &value, FILE *file)
> +{
> + if (value.is_constant ())
> + print_hex (value.coeffs[0], file);
> + else
> + {
> + fprintf (file, "[");
> + for (unsigned int i = 0; i < N; ++i)
> + {
> + print_hex (value.coeffs[i], file);
> + fputc (i == N - 1 ? ']' : ',', file);
> + }
> + }
> +}
> +
> /* Helper for calculating the distance between two points P1 and P2,
> in cases where known_le (P1, P2). T1 and T2 are the types of the
> two positions, in either order. The coefficients of P2 - P1 have
> Index: gcc/dumpfile.h
> ===================================================================
> --- gcc/dumpfile.h 2018-07-02 14:30:09.280175397 +0100
> +++ gcc/dumpfile.h 2018-07-03 09:02:36.563413564 +0100
> @@ -436,6 +436,8 @@ extern bool enable_rtl_dump_file (void);
>
> template<unsigned int N, typename C>
> void dump_dec (dump_flags_t, const poly_int<N, C> &);
> +extern void dump_dec (dump_flags_t, const poly_wide_int &, signop);
> +extern void dump_hex (dump_flags_t, const poly_wide_int &);
>
> /* In tree-dump.c */
> extern void dump_node (const_tree, dump_flags_t, FILE *);
> Index: gcc/dumpfile.c
> ===================================================================
> --- gcc/dumpfile.c 2018-07-03 09:01:31.071962478 +0100
> +++ gcc/dumpfile.c 2018-07-03 09:02:36.563413564 +0100
> @@ -597,6 +597,28 @@ template void dump_dec (dump_flags_t, co
> template void dump_dec (dump_flags_t, const poly_offset_int &);
> template void dump_dec (dump_flags_t, const poly_widest_int &);
>
> +void
> +dump_dec (dump_flags_t dump_kind, const poly_wide_int &value, signop sgn)
> +{
> + if (dump_file && (dump_kind & pflags))
> + print_dec (value, dump_file, sgn);
> +
> + if (alt_dump_file && (dump_kind & alt_flags))
> + print_dec (value, alt_dump_file, sgn);
> +}
> +
> +/* Output VALUE in hexadecimal to appropriate dump streams. */
> +
> +void
> +dump_hex (dump_flags_t dump_kind, const poly_wide_int &value)
> +{
> + if (dump_file && (dump_kind & pflags))
> + print_hex (value, dump_file);
> +
> + if (alt_dump_file && (dump_kind & alt_flags))
> + print_hex (value, alt_dump_file);
> +}
> +
> /* The current dump scope-nesting depth. */
>
> static int dump_scope_depth;
> Index: gcc/tree-vectorizer.h
> ===================================================================
> --- gcc/tree-vectorizer.h 2018-07-03 09:01:31.079962411 +0100
> +++ gcc/tree-vectorizer.h 2018-07-03 09:02:36.567413531 +0100
> @@ -899,6 +899,21 @@ typedef struct _stmt_vec_info {
>
> /* The number of scalar stmt references from active SLP instances. */
> unsigned int num_slp_uses;
> +
> + /* If nonzero, the lhs of the statement could be truncated to this
> + many bits without affecting any users of the result. */
> + unsigned int min_output_precision;
> +
> + /* If nonzero, all non-boolean input operands have the same precision,
> + and they could each be truncated to this many bits without changing
> + the result. */
> + unsigned int min_input_precision;
> +
> + /* If OPERATION_BITS is nonzero, the statement could be performed on
> + an integer with the sign and number of bits given by OPERATION_SIGN
> + and OPERATION_BITS without changing the result. */
> + unsigned int operation_precision;
> + signop operation_sign;
> } *stmt_vec_info;
>
> /* Information about a gather/scatter call. */
> Index: gcc/tree-vect-patterns.c
> ===================================================================
> --- gcc/tree-vect-patterns.c 2018-07-03 09:01:31.035962780 +0100
> +++ gcc/tree-vect-patterns.c 2018-07-03 09:02:36.567413531 +0100
> @@ -47,6 +47,40 @@ Software Foundation; either version 3, o
> #include "omp-simd-clone.h"
> #include "predict.h"
>
> +/* Return true if we have a useful VR_RANGE range for VAR, storing it
> + in *MIN_VALUE and *MAX_VALUE if so. Note the range in the dump files. */
> +
> +static bool
> +vect_get_range_info (tree var, wide_int *min_value, wide_int *max_value)
> +{
> + value_range_type vr_type = get_range_info (var, min_value, max_value);
> + wide_int nonzero = get_nonzero_bits (var);
> + signop sgn = TYPE_SIGN (TREE_TYPE (var));
> + if (intersect_range_with_nonzero_bits (vr_type, min_value, max_value,
> + nonzero, sgn) == VR_RANGE)
> + {
> + if (dump_enabled_p ())
> + {
> + dump_generic_expr_loc (MSG_NOTE, vect_location, TDF_SLIM, var);
> + dump_printf (MSG_NOTE, " has range [");
> + dump_hex (MSG_NOTE, *min_value);
> + dump_printf (MSG_NOTE, ", ");
> + dump_hex (MSG_NOTE, *max_value);
> + dump_printf (MSG_NOTE, "]\n");
> + }
> + return true;
> + }
> + else
> + {
> + if (dump_enabled_p ())
> + {
> + dump_generic_expr_loc (MSG_NOTE, vect_location, TDF_SLIM, var);
> + dump_printf (MSG_NOTE, " has no range info\n");
> + }
> + return false;
> + }
> +}
> +
> /* Report that we've found an instance of pattern PATTERN in
> statement STMT. */
>
> @@ -190,40 +224,6 @@ vect_supportable_direct_optab_p (tree ot
> return true;
> }
>
> -/* Check whether STMT2 is in the same loop or basic block as STMT1.
> - Which of the two applies depends on whether we're currently doing
> - loop-based or basic-block-based vectorization, as determined by
> - the vinfo_for_stmt for STMT1 (which must be defined).
> -
> - If this returns true, vinfo_for_stmt for STMT2 is guaranteed
> - to be defined as well. */
> -
> -static bool
> -vect_same_loop_or_bb_p (gimple *stmt1, gimple *stmt2)
> -{
> - stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt1);
> - return vect_stmt_in_region_p (stmt_vinfo->vinfo, stmt2);
> -}
> -
> -/* If the LHS of DEF_STMT has a single use, and that statement is
> - in the same loop or basic block, return it. */
> -
> -static gimple *
> -vect_single_imm_use (gimple *def_stmt)
> -{
> - tree lhs = gimple_assign_lhs (def_stmt);
> - use_operand_p use_p;
> - gimple *use_stmt;
> -
> - if (!single_imm_use (lhs, &use_p, &use_stmt))
> - return NULL;
> -
> - if (!vect_same_loop_or_bb_p (def_stmt, use_stmt))
> - return NULL;
> -
> - return use_stmt;
> -}
> -
> /* Round bit precision PRECISION up to a full element. */
>
> static unsigned int
> @@ -347,7 +347,9 @@ vect_unpromoted_value::set_op (tree op_i
> is possible to convert OP' back to OP using a possible sign change
> followed by a possible promotion P. Return this OP', or null if OP is
> not a vectorizable SSA name. If there is a promotion P, describe its
> - input in UNPROM, otherwise describe OP' in UNPROM.
> + input in UNPROM, otherwise describe OP' in UNPROM. If SINGLE_USE_P
> + is nonnull, set *SINGLE_USE_P to false if any of the SSA names involved
> + have more than one user.
>
> A successful return means that it is possible to go from OP' to OP
> via UNPROM. The cast from OP' to UNPROM is at most a sign change,
> @@ -374,7 +376,8 @@ vect_unpromoted_value::set_op (tree op_i
>
> static tree
> vect_look_through_possible_promotion (vec_info *vinfo, tree op,
> - vect_unpromoted_value *unprom)
> + vect_unpromoted_value *unprom,
> + bool *single_use_p = NULL)
> {
> tree res = NULL_TREE;
> tree op_type = TREE_TYPE (op);
> @@ -420,7 +423,14 @@ vect_look_through_possible_promotion (ve
> if (!def_stmt)
> break;
> if (dt == vect_internal_def)
> - caster = vinfo_for_stmt (def_stmt);
> + {
> + caster = vinfo_for_stmt (def_stmt);
> + /* Ignore pattern statements, since we don't link uses for them. */
> + if (single_use_p
> + && !STMT_VINFO_RELATED_STMT (caster)
> + && !has_single_use (res))
> + *single_use_p = false;
> + }
> else
> caster = NULL;
> gassign *assign = dyn_cast <gassign *> (def_stmt);
> @@ -1371,363 +1381,318 @@ vect_recog_widen_sum_pattern (vec<gimple
> return pattern_stmt;
> }
>
> +/* Recognize cases in which an operation is performed in one type WTYPE
> + but could be done more efficiently in a narrower type NTYPE. For example,
> + if we have:
> +
> + ATYPE a; // narrower than NTYPE
> + BTYPE b; // narrower than NTYPE
> + WTYPE aw = (WTYPE) a;
> + WTYPE bw = (WTYPE) b;
> + WTYPE res = aw + bw; // only uses of aw and bw
> +
> + then it would be more efficient to do:
> +
> + NTYPE an = (NTYPE) a;
> + NTYPE bn = (NTYPE) b;
> + NTYPE resn = an + bn;
> + WTYPE res = (WTYPE) resn;
> +
> + Other situations include things like:
> +
> + ATYPE a; // NTYPE or narrower
> + WTYPE aw = (WTYPE) a;
> + WTYPE res = aw + b;
> +
> + when only "(NTYPE) res" is significant. In that case it's more efficient
> + to truncate "b" and do the operation on NTYPE instead:
> +
> + NTYPE an = (NTYPE) a;
> + NTYPE bn = (NTYPE) b; // truncation
> + NTYPE resn = an + bn;
> + WTYPE res = (WTYPE) resn;
> +
> + All users of "res" should then use "resn" instead, making the final
> + statement dead (not marked as relevant). The final statement is still
> + needed to maintain the type correctness of the IR.
> +
> + vect_determine_precisions has already determined the minimum
> + precison of the operation and the minimum precision required
> + by users of the result. */
>
> -/* Return TRUE if the operation in STMT can be performed on a smaller type.
> +static gimple *
> +vect_recog_over_widening_pattern (vec<gimple *> *stmts, tree *type_out)
> +{
> + gassign *last_stmt = dyn_cast <gassign *> (stmts->pop ());
> + if (!last_stmt)
> + return NULL;
>
> - Input:
> - STMT - a statement to check.
> - DEF - we support operations with two operands, one of which is constant.
> - The other operand can be defined by a demotion operation, or by a
> - previous statement in a sequence of over-promoted operations. In the
> - later case DEF is used to replace that operand. (It is defined by a
> - pattern statement we created for the previous statement in the
> - sequence).
> -
> - Input/output:
> - NEW_TYPE - Output: a smaller type that we are trying to use. Input: if not
> - NULL, it's the type of DEF.
> - STMTS - additional pattern statements. If a pattern statement (type
> - conversion) is created in this function, its original statement is
> - added to STMTS.
> + /* See whether we have found that this operation can be done on a
> + narrower type without changing its semantics. */
> + stmt_vec_info last_stmt_info = vinfo_for_stmt (last_stmt);
> + unsigned int new_precision = last_stmt_info->operation_precision;
> + if (!new_precision)
> + return NULL;
>
> - Output:
> - OP0, OP1 - if the operation fits a smaller type, OP0 and OP1 are the new
> - operands to use in the new pattern statement for STMT (will be created
> - in vect_recog_over_widening_pattern ()).
> - NEW_DEF_STMT - in case DEF has to be promoted, we create two pattern
> - statements for STMT: the first one is a type promotion and the second
> - one is the operation itself. We return the type promotion statement
> - in NEW_DEF_STMT and further store it in STMT_VINFO_PATTERN_DEF_SEQ of
> - the second pattern statement. */
> + vec_info *vinfo = last_stmt_info->vinfo;
> + tree lhs = gimple_assign_lhs (last_stmt);
> + tree type = TREE_TYPE (lhs);
> + tree_code code = gimple_assign_rhs_code (last_stmt);
> +
> + /* Keep the first operand of a COND_EXPR as-is: only the other two
> + operands are interesting. */
> + unsigned int first_op = (code == COND_EXPR ? 2 : 1);
>
> -static bool
> -vect_operation_fits_smaller_type (gimple *stmt, tree def, tree *new_type,
> - tree *op0, tree *op1, gimple **new_def_stmt,
> - vec<gimple *> *stmts)
> -{
> - enum tree_code code;
> - tree const_oprnd, oprnd;
> - tree interm_type = NULL_TREE, half_type, new_oprnd, type;
> - gimple *def_stmt, *new_stmt;
> - bool first = false;
> - bool promotion;
> + /* Check the operands. */
> + unsigned int nops = gimple_num_ops (last_stmt) - first_op;
> + auto_vec <vect_unpromoted_value, 3> unprom (nops);
> + unprom.quick_grow (nops);
> + unsigned int min_precision = 0;
> + bool single_use_p = false;
> + for (unsigned int i = 0; i < nops; ++i)
> + {
> + tree op = gimple_op (last_stmt, first_op + i);
> + if (TREE_CODE (op) == INTEGER_CST)
> + unprom[i].set_op (op, vect_constant_def);
> + else if (TREE_CODE (op) == SSA_NAME)
> + {
> + bool op_single_use_p = true;
> + if (!vect_look_through_possible_promotion (vinfo, op, &unprom[i],
> + &op_single_use_p))
> + return NULL;
> + /* If:
>
> - *op0 = NULL_TREE;
> - *op1 = NULL_TREE;
> - *new_def_stmt = NULL;
> + (1) N bits of the result are needed;
> + (2) all inputs are widened from M<N bits; and
> + (3) one operand OP is a single-use SSA name
> +
> + we can shift the M->N widening from OP to the output
> + without changing the number or type of extensions involved.
> + This then reduces the number of copies of STMT_INFO.
> +
> + If instead of (3) more than one operand is a single-use SSA name,
> + shifting the extension to the output is even more of a win.
> +
> + If instead:
> +
> + (1) N bits of the result are needed;
> + (2) one operand OP2 is widened from M2<N bits;
> + (3) another operand OP1 is widened from M1<M2 bits; and
> + (4) both OP1 and OP2 are single-use
> +
> + the choice is between:
> +
> + (a) truncating OP2 to M1, doing the operation on M1,
> + and then widening the result to N
> +
> + (b) widening OP1 to M2, doing the operation on M2, and then
> + widening the result to N
> +
> + Both shift the M2->N widening of the inputs to the output.
> + (a) additionally shifts the M1->M2 widening to the output;
> + it requires fewer copies of STMT_INFO but requires an extra
> + M2->M1 truncation.
> +
> + Which is better will depend on the complexity and cost of
> + STMT_INFO, which is hard to predict at this stage. However,
> + a clear tie-breaker in favor of (b) is the fact that the
> + truncation in (a) increases the length of the operation chain.
> +
> + If instead of (4) only one of OP1 or OP2 is single-use,
> + (b) is still a win over doing the operation in N bits:
> + it still shifts the M2->N widening on the single-use operand
> + to the output and reduces the number of STMT_INFO copies.
> +
> + If neither operand is single-use then operating on fewer than
> + N bits might lead to more extensions overall. Whether it does
> + or not depends on global information about the vectorization
> + region, and whether that's a good trade-off would again
> + depend on the complexity and cost of the statements involved,
> + as well as things like register pressure that are not normally
> + modelled at this stage. We therefore ignore these cases
> + and just optimize the clear single-use wins above.
> +
> + Thus we take the maximum precision of the unpromoted operands
> + and record whether any operand is single-use. */
> + if (unprom[i].dt == vect_internal_def)
> + {
> + min_precision = MAX (min_precision,
> + TYPE_PRECISION (unprom[i].type));
> + single_use_p |= op_single_use_p;
> + }
> + }
> + }
>
> - if (!is_gimple_assign (stmt))
> - return false;
> + /* Although the operation could be done in operation_precision, we have
> + to balance that against introducing extra truncations or extensions.
> + Calculate the minimum precision that can be handled efficiently.
> +
> + The loop above determined that the operation could be handled
> + efficiently in MIN_PRECISION if SINGLE_USE_P; this would shift an
> + extension from the inputs to the output without introducing more
> + instructions, and would reduce the number of instructions required
> + for STMT_INFO itself.
> +
> + vect_determine_precisions has also determined that the result only
> + needs min_output_precision bits. Truncating by a factor of N times
> + requires a tree of N - 1 instructions, so if TYPE is N times wider
> + than min_output_precision, doing the operation in TYPE and truncating
> + the result requires N + (N - 1) = 2N - 1 instructions per output vector.
> + In contrast:
> +
> + - truncating the input to a unary operation and doing the operation
> + in the new type requires at most N - 1 + 1 = N instructions per
> + output vector
> +
> + - doing the same for a binary operation requires at most
> + (N - 1) * 2 + 1 = 2N - 1 instructions per output vector
> +
> + Both unary and binary operations require fewer instructions than
> + this if the operands were extended from a suitable truncated form.
> + Thus there is usually nothing to lose by doing operations in
> + min_output_precision bits, but there can be something to gain. */
> + if (!single_use_p)
> + min_precision = last_stmt_info->min_output_precision;
> + else
> + min_precision = MIN (min_precision, last_stmt_info->min_output_precision);
>
> - code = gimple_assign_rhs_code (stmt);
> - if (code != LSHIFT_EXPR && code != RSHIFT_EXPR
> - && code != BIT_IOR_EXPR && code != BIT_XOR_EXPR && code != BIT_AND_EXPR)
> - return false;
> + /* Apply the minimum efficient precision we just calculated. */
> + if (new_precision < min_precision)
> + new_precision = min_precision;
> + if (new_precision >= TYPE_PRECISION (type))
> + return NULL;
>
> - oprnd = gimple_assign_rhs1 (stmt);
> - const_oprnd = gimple_assign_rhs2 (stmt);
> - type = gimple_expr_type (stmt);
> + vect_pattern_detected ("vect_recog_over_widening_pattern", last_stmt);
>
> - if (TREE_CODE (oprnd) != SSA_NAME
> - || TREE_CODE (const_oprnd) != INTEGER_CST)
> - return false;
> + *type_out = get_vectype_for_scalar_type (type);
> + if (!*type_out)
> + return NULL;
>
> - /* If oprnd has other uses besides that in stmt we cannot mark it
> - as being part of a pattern only. */
> - if (!has_single_use (oprnd))
> - return false;
> + /* We've found a viable pattern. Get the new type of the operation. */
> + bool unsigned_p = (last_stmt_info->operation_sign == UNSIGNED);
> + tree new_type = build_nonstandard_integer_type (new_precision, unsigned_p);
> +
> + /* We specifically don't check here whether the target supports the
> + new operation, since it might be something that a later pattern
> + wants to rewrite anyway. If targets have a minimum element size
> + for some optabs, we should pattern-match smaller ops to larger ops
> + where beneficial. */
> + tree new_vectype = get_vectype_for_scalar_type (new_type);
> + if (!new_vectype)
> + return NULL;
>
> - /* If we are in the middle of a sequence, we use DEF from a previous
> - statement. Otherwise, OPRND has to be a result of type promotion. */
> - if (*new_type)
> - {
> - half_type = *new_type;
> - oprnd = def;
> - }
> - else
> + if (dump_enabled_p ())
> {
> - first = true;
> - if (!type_conversion_p (oprnd, stmt, false, &half_type, &def_stmt,
> - &promotion)
> - || !promotion
> - || !vect_same_loop_or_bb_p (stmt, def_stmt))
> - return false;
> + dump_printf_loc (MSG_NOTE, vect_location, "demoting ");
> + dump_generic_expr (MSG_NOTE, TDF_SLIM, type);
> + dump_printf (MSG_NOTE, " to ");
> + dump_generic_expr (MSG_NOTE, TDF_SLIM, new_type);
> + dump_printf (MSG_NOTE, "\n");
> }
>
> - /* Can we perform the operation on a smaller type? */
> - switch (code)
> - {
> - case BIT_IOR_EXPR:
> - case BIT_XOR_EXPR:
> - case BIT_AND_EXPR:
> - if (!int_fits_type_p (const_oprnd, half_type))
> - {
> - /* HALF_TYPE is not enough. Try a bigger type if possible. */
> - if (TYPE_PRECISION (type) < (TYPE_PRECISION (half_type) * 4))
> - return false;
> -
> - interm_type = build_nonstandard_integer_type (
> - TYPE_PRECISION (half_type) * 2, TYPE_UNSIGNED (type));
> - if (!int_fits_type_p (const_oprnd, interm_type))
> - return false;
> - }
> -
> - break;
> -
> - case LSHIFT_EXPR:
> - /* Try intermediate type - HALF_TYPE is not enough for sure. */
> - if (TYPE_PRECISION (type) < (TYPE_PRECISION (half_type) * 4))
> - return false;
> -
> - /* Check that HALF_TYPE size + shift amount <= INTERM_TYPE size.
> - (e.g., if the original value was char, the shift amount is at most 8
> - if we want to use short). */
> - if (compare_tree_int (const_oprnd, TYPE_PRECISION (half_type)) == 1)
> - return false;
> -
> - interm_type = build_nonstandard_integer_type (
> - TYPE_PRECISION (half_type) * 2, TYPE_UNSIGNED (type));
> -
> - if (!vect_supportable_shift (code, interm_type))
> - return false;
> -
> - break;
> -
> - case RSHIFT_EXPR:
> - if (vect_supportable_shift (code, half_type))
> - break;
> -
> - /* Try intermediate type - HALF_TYPE is not supported. */
> - if (TYPE_PRECISION (type) < (TYPE_PRECISION (half_type) * 4))
> - return false;
> -
> - interm_type = build_nonstandard_integer_type (
> - TYPE_PRECISION (half_type) * 2, TYPE_UNSIGNED (type));
> -
> - if (!vect_supportable_shift (code, interm_type))
> - return false;
> -
> - break;
> -
> - default:
> - gcc_unreachable ();
> - }
> -
> - /* There are four possible cases:
> - 1. OPRND is defined by a type promotion (in that case FIRST is TRUE, it's
> - the first statement in the sequence)
> - a. The original, HALF_TYPE, is not enough - we replace the promotion
> - from HALF_TYPE to TYPE with a promotion to INTERM_TYPE.
> - b. HALF_TYPE is sufficient, OPRND is set as the RHS of the original
> - promotion.
> - 2. OPRND is defined by a pattern statement we created.
> - a. Its type is not sufficient for the operation, we create a new stmt:
> - a type conversion for OPRND from HALF_TYPE to INTERM_TYPE. We store
> - this statement in NEW_DEF_STMT, and it is later put in
> - STMT_VINFO_PATTERN_DEF_SEQ of the pattern statement for STMT.
> - b. OPRND is good to use in the new statement. */
> - if (first)
> - {
> - if (interm_type)
> - {
> - /* Replace the original type conversion HALF_TYPE->TYPE with
> - HALF_TYPE->INTERM_TYPE. */
> - if (STMT_VINFO_RELATED_STMT (vinfo_for_stmt (def_stmt)))
> - {
> - new_stmt = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (def_stmt));
> - /* Check if the already created pattern stmt is what we need. */
> - if (!is_gimple_assign (new_stmt)
> - || !CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (new_stmt))
> - || TREE_TYPE (gimple_assign_lhs (new_stmt)) != interm_type)
> - return false;
> -
> - stmts->safe_push (def_stmt);
> - oprnd = gimple_assign_lhs (new_stmt);
> - }
> - else
> - {
> - /* Create NEW_OPRND = (INTERM_TYPE) OPRND. */
> - oprnd = gimple_assign_rhs1 (def_stmt);
> - new_oprnd = make_ssa_name (interm_type);
> - new_stmt = gimple_build_assign (new_oprnd, NOP_EXPR, oprnd);
> - STMT_VINFO_RELATED_STMT (vinfo_for_stmt (def_stmt)) = new_stmt;
> - stmts->safe_push (def_stmt);
> - oprnd = new_oprnd;
> - }
> - }
> - else
> - {
> - /* Retrieve the operand before the type promotion. */
> - oprnd = gimple_assign_rhs1 (def_stmt);
> - }
> - }
> - else
> - {
> - if (interm_type)
> - {
> - /* Create a type conversion HALF_TYPE->INTERM_TYPE. */
> - new_oprnd = make_ssa_name (interm_type);
> - new_stmt = gimple_build_assign (new_oprnd, NOP_EXPR, oprnd);
> - oprnd = new_oprnd;
> - *new_def_stmt = new_stmt;
> - }
> + /* Calculate the rhs operands for an operation on NEW_TYPE. */
> + STMT_VINFO_PATTERN_DEF_SEQ (last_stmt_info) = NULL;
> + tree ops[3] = {};
> + for (unsigned int i = 1; i < first_op; ++i)
> + ops[i - 1] = gimple_op (last_stmt, i);
> + vect_convert_inputs (last_stmt_info, nops, &ops[first_op - 1],
> + new_type, &unprom[0], new_vectype);
> +
> + /* Use the operation to produce a result of type NEW_TYPE. */
> + tree new_var = vect_recog_temp_ssa_var (new_type, NULL);
> + gimple *pattern_stmt = gimple_build_assign (new_var, code,
> + ops[0], ops[1], ops[2]);
> + gimple_set_location (pattern_stmt, gimple_location (last_stmt));
>
> - /* Otherwise, OPRND is already set. */
> + if (dump_enabled_p ())
> + {
> + dump_printf_loc (MSG_NOTE, vect_location,
> + "created pattern stmt: ");
> + dump_gimple_stmt (MSG_NOTE, TDF_SLIM, pattern_stmt, 0);
> }
>
> - if (interm_type)
> - *new_type = interm_type;
> - else
> - *new_type = half_type;
> + pattern_stmt = vect_convert_output (last_stmt_info, type,
> + pattern_stmt, new_vectype);
>
> - *op0 = oprnd;
> - *op1 = fold_convert (*new_type, const_oprnd);
> -
> - return true;
> + stmts->safe_push (last_stmt);
> + return pattern_stmt;
> }
>
> +/* Recognize cases in which the input to a cast is wider than its
> + output, and the input is fed by a widening operation. Fold this
> + by removing the unnecessary intermediate widening. E.g.:
>
> -/* Try to find a statement or a sequence of statements that can be performed
> - on a smaller type:
> + unsigned char a;
> + unsigned int b = (unsigned int) a;
> + unsigned short c = (unsigned short) b;
>
> - type x_t;
> - TYPE x_T, res0_T, res1_T;
> - loop:
> - S1 x_t = *p;
> - S2 x_T = (TYPE) x_t;
> - S3 res0_T = op (x_T, C0);
> - S4 res1_T = op (res0_T, C1);
> - S5 ... = () res1_T; - type demotion
> -
> - where type 'TYPE' is at least double the size of type 'type', C0 and C1 are
> - constants.
> - Check if S3 and S4 can be done on a smaller type than 'TYPE', it can either
> - be 'type' or some intermediate type. For now, we expect S5 to be a type
> - demotion operation. We also check that S3 and S4 have only one use. */
> + -->
>
> -static gimple *
> -vect_recog_over_widening_pattern (vec<gimple *> *stmts, tree *type_out)
> -{
> - gimple *stmt = stmts->pop ();
> - gimple *pattern_stmt = NULL, *new_def_stmt, *prev_stmt = NULL,
> - *use_stmt = NULL;
> - tree op0, op1, vectype = NULL_TREE, use_lhs, use_type;
> - tree var = NULL_TREE, new_type = NULL_TREE, new_oprnd;
> - bool first;
> - tree type = NULL;
> -
> - first = true;
> - while (1)
> - {
> - if (!vinfo_for_stmt (stmt)
> - || STMT_VINFO_IN_PATTERN_P (vinfo_for_stmt (stmt)))
> - return NULL;
> -
> - new_def_stmt = NULL;
> - if (!vect_operation_fits_smaller_type (stmt, var, &new_type,
> - &op0, &op1, &new_def_stmt,
> - stmts))
> - {
> - if (first)
> - return NULL;
> - else
> - break;
> - }
> + unsigned short c = (unsigned short) a;
>
> - /* STMT can be performed on a smaller type. Check its uses. */
> - use_stmt = vect_single_imm_use (stmt);
> - if (!use_stmt || !is_gimple_assign (use_stmt))
> - return NULL;
> -
> - /* Create pattern statement for STMT. */
> - vectype = get_vectype_for_scalar_type (new_type);
> - if (!vectype)
> - return NULL;
> -
> - /* We want to collect all the statements for which we create pattern
> - statetments, except for the case when the last statement in the
> - sequence doesn't have a corresponding pattern statement. In such
> - case we associate the last pattern statement with the last statement
> - in the sequence. Therefore, we only add the original statement to
> - the list if we know that it is not the last. */
> - if (prev_stmt)
> - stmts->safe_push (prev_stmt);
> + Although this is rare in input IR, it is an expected side-effect
> + of the over-widening pattern above.
>
> - var = vect_recog_temp_ssa_var (new_type, NULL);
> - pattern_stmt
> - = gimple_build_assign (var, gimple_assign_rhs_code (stmt), op0, op1);
> - STMT_VINFO_RELATED_STMT (vinfo_for_stmt (stmt)) = pattern_stmt;
> - new_pattern_def_seq (vinfo_for_stmt (stmt), new_def_stmt);
> + This is beneficial also for integer-to-float conversions, if the
> + widened integer has more bits than the float, and if the unwidened
> + input doesn't. */
>
> - if (dump_enabled_p ())
> - {
> - dump_printf_loc (MSG_NOTE, vect_location,
> - "created pattern stmt: ");
> - dump_gimple_stmt (MSG_NOTE, TDF_SLIM, pattern_stmt, 0);
> - }
> +static gimple *
> +vect_recog_cast_forwprop_pattern (vec<gimple *> *stmts, tree *type_out)
> +{
> + /* Check for a cast, including an integer-to-float conversion. */
> + gassign *last_stmt = dyn_cast <gassign *> (stmts->pop ());
> + if (!last_stmt)
> + return NULL;
> + tree_code code = gimple_assign_rhs_code (last_stmt);
> + if (!CONVERT_EXPR_CODE_P (code) && code != FLOAT_EXPR)
> + return NULL;
>
> - type = gimple_expr_type (stmt);
> - prev_stmt = stmt;
> - stmt = use_stmt;
> -
> - first = false;
> - }
> -
> - /* We got a sequence. We expect it to end with a type demotion operation.
> - Otherwise, we quit (for now). There are three possible cases: the
> - conversion is to NEW_TYPE (we don't do anything), the conversion is to
> - a type bigger than NEW_TYPE and/or the signedness of USE_TYPE and
> - NEW_TYPE differs (we create a new conversion statement). */
> - if (CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (use_stmt)))
> - {
> - use_lhs = gimple_assign_lhs (use_stmt);
> - use_type = TREE_TYPE (use_lhs);
> - /* Support only type demotion or signedess change. */
> - if (!INTEGRAL_TYPE_P (use_type)
> - || TYPE_PRECISION (type) <= TYPE_PRECISION (use_type))
> - return NULL;
> + /* Make sure that the rhs is a scalar with a natural bitsize. */
> + tree lhs = gimple_assign_lhs (last_stmt);
> + if (!lhs)
> + return NULL;
> + tree lhs_type = TREE_TYPE (lhs);
> + scalar_mode lhs_mode;
> + if (VECT_SCALAR_BOOLEAN_TYPE_P (lhs_type)
> + || !is_a <scalar_mode> (TYPE_MODE (lhs_type), &lhs_mode))
> + return NULL;
>
> - /* Check that NEW_TYPE is not bigger than the conversion result. */
> - if (TYPE_PRECISION (new_type) > TYPE_PRECISION (use_type))
> - return NULL;
> + /* Check for a narrowing operation (from a vector point of view). */
> + tree rhs = gimple_assign_rhs1 (last_stmt);
> + tree rhs_type = TREE_TYPE (rhs);
> + if (!INTEGRAL_TYPE_P (rhs_type)
> + || VECT_SCALAR_BOOLEAN_TYPE_P (rhs_type)
> + || TYPE_PRECISION (rhs_type) <= GET_MODE_BITSIZE (lhs_mode))
> + return NULL;
>
> - if (TYPE_UNSIGNED (new_type) != TYPE_UNSIGNED (use_type)
> - || TYPE_PRECISION (new_type) != TYPE_PRECISION (use_type))
> - {
> - *type_out = get_vectype_for_scalar_type (use_type);
> - if (!*type_out)
> - return NULL;
> + /* Try to find an unpromoted input. */
> + stmt_vec_info last_stmt_info = vinfo_for_stmt (last_stmt);
> + vec_info *vinfo = last_stmt_info->vinfo;
> + vect_unpromoted_value unprom;
> + if (!vect_look_through_possible_promotion (vinfo, rhs, &unprom)
> + || TYPE_PRECISION (unprom.type) >= TYPE_PRECISION (rhs_type))
> + return NULL;
>
> - /* Create NEW_TYPE->USE_TYPE conversion. */
> - new_oprnd = make_ssa_name (use_type);
> - pattern_stmt = gimple_build_assign (new_oprnd, NOP_EXPR, var);
> - STMT_VINFO_RELATED_STMT (vinfo_for_stmt (use_stmt)) = pattern_stmt;
> -
> - /* We created a pattern statement for the last statement in the
> - sequence, so we don't need to associate it with the pattern
> - statement created for PREV_STMT. Therefore, we add PREV_STMT
> - to the list in order to mark it later in vect_pattern_recog_1. */
> - if (prev_stmt)
> - stmts->safe_push (prev_stmt);
> - }
> - else
> - {
> - if (prev_stmt)
> - STMT_VINFO_PATTERN_DEF_SEQ (vinfo_for_stmt (use_stmt))
> - = STMT_VINFO_PATTERN_DEF_SEQ (vinfo_for_stmt (prev_stmt));
> + /* If the bits above RHS_TYPE matter, make sure that they're the
> + same when extending from UNPROM as they are when extending from RHS. */
> + if (!INTEGRAL_TYPE_P (lhs_type)
> + && TYPE_SIGN (rhs_type) != TYPE_SIGN (unprom.type))
> + return NULL;
>
> - *type_out = vectype;
> - }
> + /* We can get the same result by casting UNPROM directly, to avoid
> + the unnecessary widening and narrowing. */
> + vect_pattern_detected ("vect_recog_cast_forwprop_pattern", last_stmt);
>
> - stmts->safe_push (use_stmt);
> - }
> - else
> - /* TODO: support general case, create a conversion to the correct type. */
> + *type_out = get_vectype_for_scalar_type (lhs_type);
> + if (!*type_out)
> return NULL;
>
> - /* Pattern detected. */
> - vect_pattern_detected ("vect_recog_over_widening_pattern", stmts->last ());
> + tree new_var = vect_recog_temp_ssa_var (lhs_type, NULL);
> + gimple *pattern_stmt = gimple_build_assign (new_var, code, unprom.op);
> + gimple_set_location (pattern_stmt, gimple_location (last_stmt));
>
> + stmts->safe_push (last_stmt);
> return pattern_stmt;
> }
>
> @@ -4205,6 +4170,390 @@ vect_recog_gather_scatter_pattern (vec<g
> return pattern_stmt;
> }
>
> +/* Return true if TYPE is a non-boolean integer type. These are the types
> + that we want to consider for narrowing. */
> +
> +static bool
> +vect_narrowable_type_p (tree type)
> +{
> + return INTEGRAL_TYPE_P (type) && !VECT_SCALAR_BOOLEAN_TYPE_P (type);
> +}
> +
> +/* Return true if the operation given by CODE can be truncated to N bits
> + when only N bits of the output are needed. This is only true if bit N+1
> + of the inputs has no effect on the low N bits of the result. */
> +
> +static bool
> +vect_truncatable_operation_p (tree_code code)
> +{
> + switch (code)
> + {
> + case PLUS_EXPR:
> + case MINUS_EXPR:
> + case MULT_EXPR:
> + case BIT_AND_EXPR:
> + case BIT_IOR_EXPR:
> + case BIT_XOR_EXPR:
> + case COND_EXPR:
> + return true;
> +
> + default:
> + return false;
> + }
> +}
> +
> +/* Record that STMT_INFO could be changed from operating on TYPE to
> + operating on a type with the precision and sign given by PRECISION
> + and SIGN respectively. PRECISION is an arbitrary bit precision;
> + it might not be a whole number of bytes. */
> +
> +static void
> +vect_set_operation_type (stmt_vec_info stmt_info, tree type,
> + unsigned int precision, signop sign)
> +{
> + /* Round the precision up to a whole number of bytes. */
> + precision = vect_element_precision (precision);
> + if (precision < TYPE_PRECISION (type)
> + && (!stmt_info->operation_precision
> + || stmt_info->operation_precision > precision))
> + {
> + stmt_info->operation_precision = precision;
> + stmt_info->operation_sign = sign;
> + }
> +}
> +
> +/* Record that STMT_INFO only requires MIN_INPUT_PRECISION from its
> + non-boolean inputs, all of which have type TYPE. MIN_INPUT_PRECISION
> + is an arbitrary bit precision; it might not be a whole number of bytes. */
> +
> +static void
> +vect_set_min_input_precision (stmt_vec_info stmt_info, tree type,
> + unsigned int min_input_precision)
> +{
> + /* This operation in isolation only requires the inputs to have
> + MIN_INPUT_PRECISION of precision, However, that doesn't mean
> + that MIN_INPUT_PRECISION is a natural precision for the chain
> + as a whole. E.g. consider something like:
> +
> + unsigned short *x, *y;
> + *y = ((*x & 0xf0) >> 4) | (*y << 4);
> +
> + The right shift can be done on unsigned chars, and only requires the
> + result of "*x & 0xf0" to be done on unsigned chars. But taking that
> + approach would mean turning a natural chain of single-vector unsigned
> + short operations into one that truncates "*x" and then extends
> + "(*x & 0xf0) >> 4", with two vectors for each unsigned short
> + operation and one vector for each unsigned char operation.
> + This would be a significant pessimization.
> +
> + Instead only propagate the maximum of this precision and the precision
> + required by the users of the result. This means that we don't pessimize
> + the case above but continue to optimize things like:
> +
> + unsigned char *y;
> + unsigned short *x;
> + *y = ((*x & 0xf0) >> 4) | (*y << 4);
> +
> + Here we would truncate two vectors of *x to a single vector of
> + unsigned chars and use single-vector unsigned char operations for
> + everything else, rather than doing two unsigned short copies of
> + "(*x & 0xf0) >> 4" and then truncating the result. */
> + min_input_precision = MAX (min_input_precision,
> + stmt_info->min_output_precision);
> +
> + if (min_input_precision < TYPE_PRECISION (type)
> + && (!stmt_info->min_input_precision
> + || stmt_info->min_input_precision > min_input_precision))
> + stmt_info->min_input_precision = min_input_precision;
> +}
> +
> +/* Subroutine of vect_determine_min_output_precision. Return true if
> + we can calculate a reduced number of output bits for STMT_INFO,
> + whose result is LHS. */
> +
> +static bool
> +vect_determine_min_output_precision_1 (stmt_vec_info stmt_info, tree lhs)
> +{
> + /* Take the maximum precision required by users of the result. */
> + unsigned int precision = 0;
> + imm_use_iterator iter;
> + use_operand_p use;
> + FOR_EACH_IMM_USE_FAST (use, iter, lhs)
> + {
> + gimple *use_stmt = USE_STMT (use);
> + if (is_gimple_debug (use_stmt))
> + continue;
> + if (!vect_stmt_in_region_p (stmt_info->vinfo, use_stmt))
> + return false;
> + stmt_vec_info use_stmt_info = vinfo_for_stmt (use_stmt);
> + if (!use_stmt_info->min_input_precision)
> + return false;
> + precision = MAX (precision, use_stmt_info->min_input_precision);
> + }
> +
> + if (dump_enabled_p ())
> + {
> + dump_printf_loc (MSG_NOTE, vect_location, "only the low %d bits of ",
> + precision);
> + dump_generic_expr (MSG_NOTE, TDF_SLIM, lhs);
> + dump_printf (MSG_NOTE, " are significant\n");
> + }
> + stmt_info->min_output_precision = precision;
> + return true;
> +}
> +
> +/* Calculate min_output_precision for STMT_INFO. */
> +
> +static void
> +vect_determine_min_output_precision (stmt_vec_info stmt_info)
> +{
> + /* We're only interested in statements with a narrowable result. */
> + tree lhs = gimple_get_lhs (stmt_info->stmt);
> + if (!lhs
> + || TREE_CODE (lhs) != SSA_NAME
> + || !vect_narrowable_type_p (TREE_TYPE (lhs)))
> + return;
> +
> + if (!vect_determine_min_output_precision_1 (stmt_info, lhs))
> + stmt_info->min_output_precision = TYPE_PRECISION (TREE_TYPE (lhs));
> +}
> +
> +/* Use range information to decide whether STMT (described by STMT_INFO)
> + could be done in a narrower type. This is effectively a forward
> + propagation, since it uses context-independent information that applies
> + to all users of an SSA name. */
> +
> +static void
> +vect_determine_precisions_from_range (stmt_vec_info stmt_info, gassign *stmt)
> +{
> + tree lhs = gimple_assign_lhs (stmt);
> + if (!lhs || TREE_CODE (lhs) != SSA_NAME)
> + return;
> +
> + tree type = TREE_TYPE (lhs);
> + if (!vect_narrowable_type_p (type))
> + return;
> +
> + /* First see whether we have any useful range information for the result. */
> + unsigned int precision = TYPE_PRECISION (type);
> + signop sign = TYPE_SIGN (type);
> + wide_int min_value, max_value;
> + if (!vect_get_range_info (lhs, &min_value, &max_value))
> + return;
> +
> + tree_code code = gimple_assign_rhs_code (stmt);
> + unsigned int nops = gimple_num_ops (stmt);
> +
> + if (!vect_truncatable_operation_p (code))
> + /* Check that all relevant input operands are compatible, and update
> + [MIN_VALUE, MAX_VALUE] to include their ranges. */
> + for (unsigned int i = 1; i < nops; ++i)
> + {
> + tree op = gimple_op (stmt, i);
> + if (TREE_CODE (op) == INTEGER_CST)
> + {
> + /* Don't require the integer to have RHS_TYPE (which it might
> + not for things like shift amounts, etc.), but do require it
> + to fit the type. */
> + if (!int_fits_type_p (op, type))
> + return;
> +
> + min_value = wi::min (min_value, wi::to_wide (op, precision), sign);
> + max_value = wi::max (max_value, wi::to_wide (op, precision), sign);
> + }
> + else if (TREE_CODE (op) == SSA_NAME)
> + {
> + /* Ignore codes that don't take uniform arguments. */
> + if (!types_compatible_p (TREE_TYPE (op), type))
> + return;
> +
> + wide_int op_min_value, op_max_value;
> + if (!vect_get_range_info (op, &op_min_value, &op_max_value))
> + return;
> +
> + min_value = wi::min (min_value, op_min_value, sign);
> + max_value = wi::max (max_value, op_max_value, sign);
> + }
> + else
> + return;
> + }
> +
> + /* Try to switch signed types for unsigned types if we can.
> + This is better for two reasons. First, unsigned ops tend
> + to be cheaper than signed ops. Second, it means that we can
> + handle things like:
> +
> + signed char c;
> + int res = (int) c & 0xff00; // range [0x0000, 0xff00]
> +
> + as:
> +
> + signed char c;
> + unsigned short res_1 = (unsigned short) c & 0xff00;
> + int res = (int) res_1;
> +
> + where the intermediate result res_1 has unsigned rather than
> + signed type. */
> + if (sign == SIGNED && !wi::neg_p (min_value))
> + sign = UNSIGNED;
> +
> + /* See what precision is required for MIN_VALUE and MAX_VALUE. */
> + unsigned int precision1 = wi::min_precision (min_value, sign);
> + unsigned int precision2 = wi::min_precision (max_value, sign);
> + unsigned int value_precision = MAX (precision1, precision2);
> + if (value_precision >= precision)
> + return;
> +
> + if (dump_enabled_p ())
> + {
> + dump_printf_loc (MSG_NOTE, vect_location, "can narrow to %s:%d"
> + " without loss of precision: ",
> + sign == SIGNED ? "signed" : "unsigned",
> + value_precision);
> + dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0);
> + }
> +
> + vect_set_operation_type (stmt_info, type, value_precision, sign);
> + vect_set_min_input_precision (stmt_info, type, value_precision);
> +}
> +
> +/* Use information about the users of STMT's result to decide whether
> + STMT (described by STMT_INFO) could be done in a narrower type.
> + This is effectively a backward propagation. */
> +
> +static void
> +vect_determine_precisions_from_users (stmt_vec_info stmt_info, gassign *stmt)
> +{
> + tree_code code = gimple_assign_rhs_code (stmt);
> + unsigned int opno = (code == COND_EXPR ? 2 : 1);
> + tree type = TREE_TYPE (gimple_op (stmt, opno));
> + if (!vect_narrowable_type_p (type))
> + return;
> +
> + unsigned int precision = TYPE_PRECISION (type);
> + unsigned int operation_precision, min_input_precision;
> + switch (code)
> + {
> + CASE_CONVERT:
> + /* Only the bits that contribute to the output matter. Don't change
> + the precision of the operation itself. */
> + operation_precision = precision;
> + min_input_precision = stmt_info->min_output_precision;
> + break;
> +
> + case LSHIFT_EXPR:
> + case RSHIFT_EXPR:
> + {
> + tree shift = gimple_assign_rhs2 (stmt);
> + if (TREE_CODE (shift) != INTEGER_CST
> + || !wi::ltu_p (wi::to_widest (shift), precision))
> + return;
> + unsigned int const_shift = TREE_INT_CST_LOW (shift);
> + if (code == LSHIFT_EXPR)
> + {
> + /* We need CONST_SHIFT fewer bits of the input. */
> + operation_precision = stmt_info->min_output_precision;
> + min_input_precision = (MAX (operation_precision, const_shift)
> + - const_shift);
> + }
> + else
> + {
> + /* We need CONST_SHIFT extra bits to do the operation. */
> + operation_precision = (stmt_info->min_output_precision
> + + const_shift);
> + min_input_precision = operation_precision;
> + }
> + break;
> + }
> +
> + default:
> + if (vect_truncatable_operation_p (code))
> + {
> + /* Input bit N has no effect on output bits N-1 and lower. */
> + operation_precision = stmt_info->min_output_precision;
> + min_input_precision = operation_precision;
> + break;
> + }
> + return;
> + }
> +
> + if (operation_precision < precision)
> + {
> + if (dump_enabled_p ())
> + {
> + dump_printf_loc (MSG_NOTE, vect_location, "can narrow to %s:%d"
> + " without affecting users: ",
> + TYPE_UNSIGNED (type) ? "unsigned" : "signed",
> + operation_precision);
> + dump_gimple_stmt (MSG_NOTE, TDF_SLIM, stmt, 0);
> + }
> + vect_set_operation_type (stmt_info, type, operation_precision,
> + TYPE_SIGN (type));
> + }
> + vect_set_min_input_precision (stmt_info, type, min_input_precision);
> +}
> +
> +/* Handle vect_determine_precisions for STMT_INFO, given that we
> + have already done so for the users of its result. */
> +
> +void
> +vect_determine_stmt_precisions (stmt_vec_info stmt_info)
> +{
> + vect_determine_min_output_precision (stmt_info);
> + if (gassign *stmt = dyn_cast <gassign *> (stmt_info->stmt))
> + {
> + vect_determine_precisions_from_range (stmt_info, stmt);
> + vect_determine_precisions_from_users (stmt_info, stmt);
> + }
> +}
> +
> +/* Walk backwards through the vectorizable region to determine the
> + values of these fields:
> +
> + - min_output_precision
> + - min_input_precision
> + - operation_precision
> + - operation_sign. */
> +
> +void
> +vect_determine_precisions (vec_info *vinfo)
> +{
> + DUMP_VECT_SCOPE ("vect_determine_precisions");
> +
> + if (loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo))
> + {
> + struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
> + basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo);
> + unsigned int nbbs = loop->num_nodes;
> +
> + for (unsigned int i = 0; i < nbbs; i++)
> + {
> + basic_block bb = bbs[nbbs - i - 1];
> + for (gimple_stmt_iterator si = gsi_last_bb (bb);
> + !gsi_end_p (si); gsi_prev (&si))
> + vect_determine_stmt_precisions (vinfo_for_stmt (gsi_stmt (si)));
> + }
> + }
> + else
> + {
> + bb_vec_info bb_vinfo = as_a <bb_vec_info> (vinfo);
> + gimple_stmt_iterator si = bb_vinfo->region_end;
> + gimple *stmt;
> + do
> + {
> + if (!gsi_stmt (si))
> + si = gsi_last_bb (bb_vinfo->bb);
> + else
> + gsi_prev (&si);
> + stmt = gsi_stmt (si);
> + stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
> + if (stmt_info && STMT_VINFO_VECTORIZABLE (stmt_info))
> + vect_determine_stmt_precisions (stmt_info);
> + }
> + while (stmt != gsi_stmt (bb_vinfo->region_begin));
> + }
> +}
> +
> typedef gimple *(*vect_recog_func_ptr) (vec<gimple *> *, tree *);
>
> struct vect_recog_func
> @@ -4217,13 +4566,14 @@ struct vect_recog_func
> taken which means usually the more complex one needs to preceed the
> less comples onex (widen_sum only after dot_prod or sad for example). */
> static vect_recog_func vect_vect_recog_func_ptrs[] = {
> + { vect_recog_over_widening_pattern, "over_widening" },
> + { vect_recog_cast_forwprop_pattern, "cast_forwprop" },
> { vect_recog_widen_mult_pattern, "widen_mult" },
> { vect_recog_dot_prod_pattern, "dot_prod" },
> { vect_recog_sad_pattern, "sad" },
> { vect_recog_widen_sum_pattern, "widen_sum" },
> { vect_recog_pow_pattern, "pow" },
> { vect_recog_widen_shift_pattern, "widen_shift" },
> - { vect_recog_over_widening_pattern, "over_widening" },
> { vect_recog_rotate_pattern, "rotate" },
> { vect_recog_vector_vector_shift_pattern, "vector_vector_shift" },
> { vect_recog_divmod_pattern, "divmod" },
> @@ -4502,6 +4852,8 @@ vect_pattern_recog (vec_info *vinfo)
> unsigned int i, j;
> auto_vec<gimple *, 1> stmts_to_replace;
>
> + vect_determine_precisions (vinfo);
> +
> DUMP_VECT_SCOPE ("vect_pattern_recog");
>
> if (loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo))
> Index: gcc/testsuite/gcc.dg/vect/vect-widen-mult-u8-u32.c
> ===================================================================
> --- gcc/testsuite/gcc.dg/vect/vect-widen-mult-u8-u32.c 2016-11-11 17:07:36.776796115 +0000
> +++ gcc/testsuite/gcc.dg/vect/vect-widen-mult-u8-u32.c 2018-07-03 09:02:36.567413531 +0100
> @@ -43,5 +43,5 @@ int main (void)
>
> /* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" { target { vect_widen_mult_qi_to_hi || vect_unpack } } } } */
> /* { dg-final { scan-tree-dump-times "vect_recog_widen_mult_pattern: detected" 1 "vect" { target vect_widen_mult_qi_to_hi_pattern } } } */
> -/* { dg-final { scan-tree-dump-times "pattern recognized" 1 "vect" { target vect_widen_mult_qi_to_hi_pattern } } } */
> +/* { dg-final { scan-tree-dump-times "widen_mult pattern recognized" 1 "vect" { target vect_widen_mult_qi_to_hi_pattern } } } */
>
> Index: gcc/testsuite/gcc.dg/vect/vect-over-widen-1.c
> ===================================================================
> --- gcc/testsuite/gcc.dg/vect/vect-over-widen-1.c 2018-07-03 09:01:31.075962445 +0100
> +++ gcc/testsuite/gcc.dg/vect/vect-over-widen-1.c 2018-07-03 09:02:36.563413564 +0100
> @@ -62,8 +62,9 @@ int main (void)
> }
>
> /* { dg-final { scan-tree-dump-times "vect_recog_widen_shift_pattern: detected" 2 "vect" { target vect_widen_shift } } } */
> -/* { dg-final { scan-tree-dump-times "vect_recog_over_widening_pattern: detected" 2 "vect" { target vect_widen_shift } } } */
> -/* { dg-final { scan-tree-dump-times "vect_recog_over_widening_pattern: detected" 4 "vect" { target { { ! vect_sizes_32B_16B } && { ! vect_widen_shift } } } } } */
> -/* { dg-final { scan-tree-dump-times "vect_recog_over_widening_pattern: detected" 8 "vect" { target vect_sizes_32B_16B } } } */
> +/* { dg-final { scan-tree-dump {vect_recog_over_widening_pattern: detected:[^\n]* << 3} "vect" } } */
> +/* { dg-final { scan-tree-dump {vect_recog_over_widening_pattern: detected:[^\n]* >> 3} "vect" } } */
> +/* { dg-final { scan-tree-dump {vect_recog_over_widening_pattern: detected:[^\n]* << 8} "vect" } } */
> +/* { dg-final { scan-tree-dump {vect_recog_over_widening_pattern: detected:[^\n]* >> 5} "vect" } } */
> /* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" } } */
>
> Index: gcc/testsuite/gcc.dg/vect/vect-over-widen-1-big-array.c
> ===================================================================
> --- gcc/testsuite/gcc.dg/vect/vect-over-widen-1-big-array.c 2018-07-03 09:01:31.075962445 +0100
> +++ gcc/testsuite/gcc.dg/vect/vect-over-widen-1-big-array.c 2018-07-03 09:02:36.563413564 +0100
> @@ -58,7 +58,9 @@ int main (void)
> }
>
> /* { dg-final { scan-tree-dump-times "vect_recog_widen_shift_pattern: detected" 2 "vect" { target vect_widen_shift } } } */
> -/* { dg-final { scan-tree-dump-times "vect_recog_over_widening_pattern: detected" 2 "vect" { target vect_widen_shift } } } */
> -/* { dg-final { scan-tree-dump-times "vect_recog_over_widening_pattern: detected" 4 "vect" { target { ! vect_widen_shift } } } } */
> +/* { dg-final { scan-tree-dump {vect_recog_over_widening_pattern: detected:[^\n]* << 3} "vect" } } */
> +/* { dg-final { scan-tree-dump {vect_recog_over_widening_pattern: detected:[^\n]* >> 3} "vect" } } */
> +/* { dg-final { scan-tree-dump {vect_recog_over_widening_pattern: detected:[^\n]* << 8} "vect" } } */
> +/* { dg-final { scan-tree-dump {vect_recog_over_widening_pattern: detected:[^\n]* >> 5} "vect" } } */
> /* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" } } */
>
> Index: gcc/testsuite/gcc.dg/vect/vect-over-widen-2.c
> ===================================================================
> --- gcc/testsuite/gcc.dg/vect/vect-over-widen-2.c 2018-07-03 09:01:31.075962445 +0100
> +++ gcc/testsuite/gcc.dg/vect/vect-over-widen-2.c 2018-07-03 09:02:36.563413564 +0100
> @@ -57,7 +57,12 @@ int main (void)
> return 0;
> }
>
> -/* Final value stays in int, so no over-widening is detected at the moment. */
> -/* { dg-final { scan-tree-dump-times "vect_recog_over_widening_pattern: detected" 0 "vect" } } */
> +/* This is an over-widening even though the final result is still an int.
> + It's better to do one vector of ops on chars and then widen than to
> + widen and then do 4 vectors of ops on ints. */
> +/* { dg-final { scan-tree-dump {vect_recog_over_widening_pattern: detected:[^\n]* << 3} "vect" } } */
> +/* { dg-final { scan-tree-dump {vect_recog_over_widening_pattern: detected:[^\n]* >> 3} "vect" } } */
> +/* { dg-final { scan-tree-dump {vect_recog_over_widening_pattern: detected:[^\n]* << 8} "vect" } } */
> +/* { dg-final { scan-tree-dump {vect_recog_over_widening_pattern: detected:[^\n]* >> 5} "vect" } } */
> /* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" } } */
>
> Index: gcc/testsuite/gcc.dg/vect/vect-over-widen-2-big-array.c
> ===================================================================
> --- gcc/testsuite/gcc.dg/vect/vect-over-widen-2-big-array.c 2018-07-03 09:01:31.075962445 +0100
> +++ gcc/testsuite/gcc.dg/vect/vect-over-widen-2-big-array.c 2018-07-03 09:02:36.563413564 +0100
> @@ -57,7 +57,12 @@ int main (void)
> return 0;
> }
>
> -/* Final value stays in int, so no over-widening is detected at the moment. */
> -/* { dg-final { scan-tree-dump-times "vect_recog_over_widening_pattern: detected" 0 "vect" } } */
> +/* This is an over-widening even though the final result is still an int.
> + It's better to do one vector of ops on chars and then widen than to
> + widen and then do 4 vectors of ops on ints. */
> +/* { dg-final { scan-tree-dump {vect_recog_over_widening_pattern: detected:[^\n]* << 3} "vect" } } */
> +/* { dg-final { scan-tree-dump {vect_recog_over_widening_pattern: detected:[^\n]* >> 3} "vect" } } */
> +/* { dg-final { scan-tree-dump {vect_recog_over_widening_pattern: detected:[^\n]* << 8} "vect" } } */
> +/* { dg-final { scan-tree-dump {vect_recog_over_widening_pattern: detected:[^\n]* >> 5} "vect" } } */
> /* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" } } */
>
> Index: gcc/testsuite/gcc.dg/vect/vect-over-widen-3.c
> ===================================================================
> --- gcc/testsuite/gcc.dg/vect/vect-over-widen-3.c 2018-07-03 09:01:31.075962445 +0100
> +++ gcc/testsuite/gcc.dg/vect/vect-over-widen-3.c 2018-07-03 09:02:36.563413564 +0100
> @@ -57,6 +57,9 @@ int main (void)
> return 0;
> }
>
> -/* { dg-final { scan-tree-dump "vect_recog_over_widening_pattern: detected" "vect" } } */
> +/* { dg-final { scan-tree-dump {vect_recog_over_widening_pattern: detected:[^\n]* << 3} "vect" } } */
> +/* { dg-final { scan-tree-dump {vect_recog_over_widening_pattern: detected:[^\n]* >> 3} "vect" } } */
> +/* { dg-final { scan-tree-dump {vect_recog_over_widening_pattern: detected:[^\n]* >> 8} "vect" } } */
> +/* { dg-final { scan-tree-dump {vect_recog_over_widening_pattern: detected:[^\n]* << 9} "vect" } } */
> /* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" } } */
>
> Index: gcc/testsuite/gcc.dg/vect/vect-over-widen-3-big-array.c
> ===================================================================
> --- gcc/testsuite/gcc.dg/vect/vect-over-widen-3-big-array.c 2018-07-03 09:01:31.075962445 +0100
> +++ gcc/testsuite/gcc.dg/vect/vect-over-widen-3-big-array.c 2018-07-03 09:02:36.563413564 +0100
> @@ -59,7 +59,9 @@ int main (void)
> return 0;
> }
>
> -/* { dg-final { scan-tree-dump-times "vect_recog_over_widening_pattern: detected" 2 "vect" { target { ! vect_widen_shift } } } } */
> -/* { dg-final { scan-tree-dump-times "vect_recog_over_widening_pattern: detected" 1 "vect" { target vect_widen_shift } } } */
> +/* { dg-final { scan-tree-dump {vect_recog_over_widening_pattern: detected:[^\n]* << 3} "vect" } } */
> +/* { dg-final { scan-tree-dump {vect_recog_over_widening_pattern: detected:[^\n]* >> 3} "vect" } } */
> +/* { dg-final { scan-tree-dump {vect_recog_over_widening_pattern: detected:[^\n]* >> 8} "vect" } } */
> +/* { dg-final { scan-tree-dump {vect_recog_over_widening_pattern: detected:[^\n]* << 9} "vect" } } */
> /* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" } } */
>
> Index: gcc/testsuite/gcc.dg/vect/vect-over-widen-4.c
> ===================================================================
> --- gcc/testsuite/gcc.dg/vect/vect-over-widen-4.c 2018-07-03 09:01:31.075962445 +0100
> +++ gcc/testsuite/gcc.dg/vect/vect-over-widen-4.c 2018-07-03 09:02:36.563413564 +0100
> @@ -66,8 +66,9 @@ int main (void)
> }
>
> /* { dg-final { scan-tree-dump-times "vect_recog_widen_shift_pattern: detected" 2 "vect" { target vect_widen_shift } } } */
> -/* { dg-final { scan-tree-dump-times "vect_recog_over_widening_pattern: detected" 2 "vect" { target vect_widen_shift } } } */
> -/* { dg-final { scan-tree-dump-times "vect_recog_over_widening_pattern: detected" 4 "vect" { target { { ! vect_sizes_32B_16B } && { ! vect_widen_shift } } } } } */
> -/* { dg-final { scan-tree-dump-times "vect_recog_over_widening_pattern: detected" 8 "vect" { target vect_sizes_32B_16B } } } */
> +/* { dg-final { scan-tree-dump {vect_recog_over_widening_pattern: detected:[^\n]* << 3} "vect" } } */
> +/* { dg-final { scan-tree-dump {vect_recog_over_widening_pattern: detected:[^\n]* >> 3} "vect" } } */
> +/* { dg-final { scan-tree-dump {vect_recog_over_widening_pattern: detected:[^\n]* << 8} "vect" } } */
> +/* { dg-final { scan-tree-dump {vect_recog_over_widening_pattern: detected:[^\n]* >> 5} "vect" } } */
> /* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" } } */
>
> Index: gcc/testsuite/gcc.dg/vect/vect-over-widen-4-big-array.c
> ===================================================================
> --- gcc/testsuite/gcc.dg/vect/vect-over-widen-4-big-array.c 2018-07-03 09:01:31.075962445 +0100
> +++ gcc/testsuite/gcc.dg/vect/vect-over-widen-4-big-array.c 2018-07-03 09:02:36.563413564 +0100
> @@ -62,7 +62,9 @@ int main (void)
> }
>
> /* { dg-final { scan-tree-dump-times "vect_recog_widen_shift_pattern: detected" 2 "vect" { target vect_widen_shift } } } */
> -/* { dg-final { scan-tree-dump-times "vect_recog_over_widening_pattern: detected" 2 "vect" { target vect_widen_shift } } } */
> -/* { dg-final { scan-tree-dump-times "vect_recog_over_widening_pattern: detected" 4 "vect" { target { ! vect_widen_shift } } } } */
> +/* { dg-final { scan-tree-dump {vect_recog_over_widening_pattern: detected:[^\n]* << 3} "vect" } } */
> +/* { dg-final { scan-tree-dump {vect_recog_over_widening_pattern: detected:[^\n]* >> 3} "vect" } } */
> +/* { dg-final { scan-tree-dump {vect_recog_over_widening_pattern: detected:[^\n]* << 8} "vect" } } */
> +/* { dg-final { scan-tree-dump {vect_recog_over_widening_pattern: detected:[^\n]* >> 5} "vect" } } */
> /* { dg-final { scan-tree-dump-times "vectorized 1 loops" 1 "vect" } } */
>
> Index: gcc/testsuite/gcc.dg/vect/bb-slp-over-widen-1.c
> ===================================================================
> --- /dev/null 2018-06-13 14:36:57.192460992 +0100
> +++ gcc/testsuite/gcc.dg/vect/bb-slp-over-widen-1.c 2018-07-03 09:02:36.563413564 +0100
> @@ -0,0 +1,66 @@
> +/* { dg-require-effective-target vect_int } */
> +/* { dg-require-effective-target vect_shift } */
> +/* { dg-require-effective-target vect_pack_trunc } */
> +/* { dg-require-effective-target vect_unpack } */
> +
> +#include "tree-vect.h"
> +
> +/* Deliberate use of signed >>. */
> +#define DEF_LOOP(SIGNEDNESS) \
> + void __attribute__ ((noipa)) \
> + f_##SIGNEDNESS (SIGNEDNESS char *restrict a, \
> + SIGNEDNESS char *restrict b, \
> + SIGNEDNESS char *restrict c) \
> + { \
> + a[0] = (b[0] + c[0]) >> 1; \
> + a[1] = (b[1] + c[1]) >> 1; \
> + a[2] = (b[2] + c[2]) >> 1; \
> + a[3] = (b[3] + c[3]) >> 1; \
> + a[4] = (b[4] + c[4]) >> 1; \
> + a[5] = (b[5] + c[5]) >> 1; \
> + a[6] = (b[6] + c[6]) >> 1; \
> + a[7] = (b[7] + c[7]) >> 1; \
> + a[8] = (b[8] + c[8]) >> 1; \
> + a[9] = (b[9] + c[9]) >> 1; \
> + a[10] = (b[10] + c[10]) >> 1; \
> + a[11] = (b[11] + c[11]) >> 1; \
> + a[12] = (b[12] + c[12]) >> 1; \
> + a[13] = (b[13] + c[13]) >> 1; \
> + a[14] = (b[14] + c[14]) >> 1; \
> + a[15] = (b[15] + c[15]) >> 1; \
> + }
> +
> +DEF_LOOP (signed)
> +DEF_LOOP (unsigned)
> +
> +#define N 16
> +
> +#define TEST_LOOP(SIGNEDNESS, BASE_B, BASE_C) \
> + { \
> + SIGNEDNESS char a[N], b[N], c[N]; \
> + for (int i = 0; i < N; ++i)