[PATCH 4/5] vect: Make reduction code handle calls
Richard Sandiford
richard.sandiford@arm.com
Tue Nov 16 16:24:40 GMT 2021
Richard Biener via Gcc-patches <gcc-patches@gcc.gnu.org> writes:
> On Wed, Nov 10, 2021 at 1:48 PM Richard Sandiford via Gcc-patches
> <gcc-patches@gcc.gnu.org> wrote:
>>
>> This patch extends the reduction code to handle calls. So far
>> it's a structural change only; a later patch adds support for
>> specific function reductions.
>>
>> Most of the patch consists of using code_helper and gimple_match_op
>> to describe the reduction operations. The other main change is that
>> vectorizable_call now needs to handle fully-predicated reductions.
>>
>> Tested on aarch64-linux-gnu and x86_64-linux-gnu. OK to install?
>>
>> Richard
>>
>>
>> gcc/
>> * builtins.h (associated_internal_fn): Declare overload that
>> takes a (combined_cfn, return type) pair.
>> * builtins.c (associated_internal_fn): Split new overload out
>> of original fndecl version. Also provide an overload that takes
>> a (combined_cfn, return type) pair.
>> * internal-fn.h (commutative_binary_fn_p): Declare.
>> (associative_binary_fn_p): Likewise.
>> * internal-fn.c (commutative_binary_fn_p): New function,
>> split out from...
>> (first_commutative_argument): ...here.
>> (associative_binary_fn_p): New function.
>> * gimple-match.h (code_helper): Add a constructor that takes
>> internal functions.
>> (commutative_binary_op_p): Declare.
>> (associative_binary_op_p): Likewise.
>> (canonicalize_code): Likewise.
>> (directly_supported_p): Likewise.
>> (get_conditional_internal_fn): Likewise.
>> (gimple_build): New overload that takes a code_helper.
>> * gimple-fold.c (gimple_build): Likewise.
>> * gimple-match-head.c (commutative_binary_op_p): New function.
>> (associative_binary_op_p): Likewise.
>> (canonicalize_code): Likewise.
>> (directly_supported_p): Likewise.
>> (get_conditional_internal_fn): Likewise.
>> * tree-vectorizer.h: Include gimple-match.h.
>> (neutral_op_for_reduction): Take a code_helper instead of a tree_code.
>> (needs_fold_left_reduction_p): Likewise.
>> (reduction_fn_for_scalar_code): Likewise.
>> (vect_can_vectorize_without_simd_p): Declare a nNew overload that takes
>> a code_helper.
>> * tree-vect-loop.c: Include case-cfn-macros.h.
>> (fold_left_reduction_fn): Take a code_helper instead of a tree_code.
>> (reduction_fn_for_scalar_code): Likewise.
>> (neutral_op_for_reduction): Likewise.
>> (needs_fold_left_reduction_p): Likewise.
>> (use_mask_by_cond_expr_p): Likewise.
>> (build_vect_cond_expr): Likewise.
>> (vect_create_partial_epilog): Likewise. Use gimple_build rather
>> than gimple_build_assign.
>> (check_reduction_path): Handle calls and operate on code_helpers
>> rather than tree_codes.
>> (vect_is_simple_reduction): Likewise.
>> (vect_model_reduction_cost): Likewise.
>> (vect_find_reusable_accumulator): Likewise.
>> (vect_create_epilog_for_reduction): Likewise.
>> (vect_transform_cycle_phi): Likewise.
>> (vectorizable_reduction): Likewise. Make more use of
>> lane_reduc_code_p.
>> (vect_transform_reduction): Use gimple_extract_op but expect
>> a tree_code for now.
>> (vect_can_vectorize_without_simd_p): New overload that takes
>> a code_helper.
>> * tree-vect-stmts.c (vectorizable_call): Handle reductions in
>> fully-masked loops.
>> * tree-vect-patterns.c (vect_mark_pattern_stmts): Use
>> gimple_extract_op when updating STMT_VINFO_REDUC_IDX.
>> ---
>> gcc/builtins.c | 46 ++++-
>> gcc/builtins.h | 1 +
>> gcc/gimple-fold.c | 9 +
>> gcc/gimple-match-head.c | 70 +++++++
>> gcc/gimple-match.h | 20 ++
>> gcc/internal-fn.c | 46 ++++-
>> gcc/internal-fn.h | 2 +
>> gcc/tree-vect-loop.c | 420 +++++++++++++++++++--------------------
>> gcc/tree-vect-patterns.c | 23 ++-
>> gcc/tree-vect-stmts.c | 66 ++++--
>> gcc/tree-vectorizer.h | 10 +-
>> 11 files changed, 455 insertions(+), 258 deletions(-)
>>
>> diff --git a/gcc/builtins.c b/gcc/builtins.c
>> index 384864bfb3a..03829c03a5a 100644
>> --- a/gcc/builtins.c
>> +++ b/gcc/builtins.c
>> @@ -2139,17 +2139,17 @@ mathfn_built_in_type (combined_fn fn)
>> #undef SEQ_OF_CASE_MATHFN
>> }
>>
>> -/* If BUILT_IN_NORMAL function FNDECL has an associated internal function,
>> - return its code, otherwise return IFN_LAST. Note that this function
>> - only tests whether the function is defined in internals.def, not whether
>> - it is actually available on the target. */
>> +/* Check whether there is an internal function associated with function FN
>> + and return type RETURN_TYPE. Return the function if so, otherwise return
>> + IFN_LAST.
>>
>> -internal_fn
>> -associated_internal_fn (tree fndecl)
>> + Note that this function only tests whether the function is defined in
>> + internals.def, not whether it is actually available on the target. */
>> +
>> +static internal_fn
>> +associated_internal_fn (built_in_function fn, tree return_type)
>> {
>> - gcc_checking_assert (DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL);
>> - tree return_type = TREE_TYPE (TREE_TYPE (fndecl));
>> - switch (DECL_FUNCTION_CODE (fndecl))
>> + switch (fn)
>> {
>> #define DEF_INTERNAL_FLT_FN(NAME, FLAGS, OPTAB, TYPE) \
>> CASE_FLT_FN (BUILT_IN_##NAME): return IFN_##NAME;
>> @@ -2177,6 +2177,34 @@ associated_internal_fn (tree fndecl)
>> }
>> }
>>
>> +/* If BUILT_IN_NORMAL function FNDECL has an associated internal function,
>> + return its code, otherwise return IFN_LAST. Note that this function
>> + only tests whether the function is defined in internals.def, not whether
>> + it is actually available on the target. */
>> +
>> +internal_fn
>> +associated_internal_fn (tree fndecl)
>> +{
>> + gcc_checking_assert (DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL);
>> + return associated_internal_fn (DECL_FUNCTION_CODE (fndecl),
>> + TREE_TYPE (TREE_TYPE (fndecl)));
>> +}
>> +
>> +/* Check whether there is an internal function associated with function CFN
>> + and return type RETURN_TYPE. Return the function if so, otherwise return
>> + IFN_LAST.
>> +
>> + Note that this function only tests whether the function is defined in
>> + internals.def, not whether it is actually available on the target. */
>> +
>> +internal_fn
>> +associated_internal_fn (combined_fn cfn, tree return_type)
>> +{
>> + if (internal_fn_p (cfn))
>> + return as_internal_fn (cfn);
>> + return associated_internal_fn (as_builtin_fn (cfn), return_type);
>> +}
>> +
>> /* If CALL is a call to a BUILT_IN_NORMAL function that could be replaced
>> on the current target by a call to an internal function, return the
>> code of that internal function, otherwise return IFN_LAST. The caller
>> diff --git a/gcc/builtins.h b/gcc/builtins.h
>> index 5e4d86e9c37..c99670b12f1 100644
>> --- a/gcc/builtins.h
>> +++ b/gcc/builtins.h
>> @@ -148,6 +148,7 @@ extern char target_percent_s_newline[4];
>> extern bool target_char_cst_p (tree t, char *p);
>> extern rtx get_memory_rtx (tree exp, tree len);
>>
>> +extern internal_fn associated_internal_fn (combined_fn, tree);
>> extern internal_fn associated_internal_fn (tree);
>> extern internal_fn replacement_internal_fn (gcall *);
>>
>> diff --git a/gcc/gimple-fold.c b/gcc/gimple-fold.c
>> index 9daf2cc590c..a937f130815 100644
>> --- a/gcc/gimple-fold.c
>> +++ b/gcc/gimple-fold.c
>> @@ -8808,6 +8808,15 @@ gimple_build (gimple_seq *seq, location_t loc, combined_fn fn,
>> return res;
>> }
>
> Toplevel comment missing. You add this for two operands, please
> also add it for one and three (even if unused).
OK. On the comment side: I was hoping to piggy-back on the comment
for the previous overload :-) I'll add a new one though.
>> +tree
>> +gimple_build (gimple_seq *seq, location_t loc, code_helper code,
>> + tree type, tree op0, tree op1)
>> +{
>> + if (code.is_tree_code ())
>> + return gimple_build (seq, loc, tree_code (code), type, op0, op1);
>> + return gimple_build (seq, loc, combined_fn (code), type, op0, op1);
>> +}
>> +
>> /* Build the conversion (TYPE) OP with a result of type TYPE
>> with location LOC if such conversion is neccesary in GIMPLE,
>> simplifying it first.
>> diff --git a/gcc/gimple-match-head.c b/gcc/gimple-match-head.c
>> index d4d7d767075..4558a3db5fc 100644
>> --- a/gcc/gimple-match-head.c
>> +++ b/gcc/gimple-match-head.c
>> @@ -1304,3 +1304,73 @@ optimize_successive_divisions_p (tree divisor, tree inner_div)
>> }
>> return true;
>> }
>> +
>> +/* If CODE, operating on TYPE, represents a built-in function that has an
>> + associated internal function, return the associated internal function,
>> + otherwise return CODE. This function does not check whether the
>> + internal function is supported, only that it exists. */
>
> Hmm, why not name the function associated_internal_fn then, or
> have it contain internal_fn?
I didn't want to call it associated_internal_fn because the existing
forms of that function return an internal_fn. Here the idea is to
avoid having multiple representations of the same operation, so that
code_helpers can be compared for equality. I guess the fact that that
currently means mapping built-in functions to internal functions
(and nothing else) is more of an implementation detail.
So I guess the emphasis in the comment is wrong. How about if I change
it to:
/* Return a canonical form for CODE when operating on TYPE. The idea
is to remove redundant ways of representing the same operation so
that code_helpers can be hashed and compared for equality.
The only current canonicalization is to replace built-in functions
with internal functions, in cases where internal-fn.def defines
such an internal function.
Note that the new code_helper cannot necessarily be used in place of
the original code_helper. For example, the new code_helper might be
an internal function that the target does not support. */
> I also wonder why all the functions below are not member functions
> of code_helper?
TBH I don't really like that style very much. :-) E.g. it's not obvious
that directly_supported_p should be a member of code_helper when it's
really querying information in the optab array. If code_helper was
available more widely then the code would probably be in optab*.c
instead.
The queries are also about (code_helper, type) pairs rather than about
code_helpers in isolation.
IMO the current code_helper member functions seem like the right set,
in that they're providing the abstraction “tree_code or combined_fn”.
That abstraction can then be used in all sorts of places.
>> +code_helper
>> +canonicalize_code (code_helper code, tree type)
>> +{
>> + if (code.is_fn_code ())
>> + return associated_internal_fn (combined_fn (code), type);
>> + return code;
>> +}
>> +
>> +/* Return true if CODE is a binary operation that is commutative when
>> + operating on type TYPE. */
>> +
>> +bool
>> +commutative_binary_op_p (code_helper code, tree type)
>> +{
>> + if (code.is_tree_code ())
>> + return commutative_tree_code (tree_code (code));
>> + auto cfn = combined_fn (code);
>> + return commutative_binary_fn_p (associated_internal_fn (cfn, type));
>> +}
>
> Do we need commutative_ternary_op_p? Can we do a more generic
> commutative_p instead?
How about using the first_commutative_argument interface from internal-fn.c,
which returns the first argument in a commutative pair or -1 if none?
>> +
>> +/* Return true if CODE is a binary operation that is associative when
>> + operating on type TYPE. */
>> +
>> +bool
>> +associative_binary_op_p (code_helper code, tree type)
>
> We only have associative_tree_code, is _binary relevant here?
But we do have commutative_ternary_tree_code, like you say :-)
I guess it didn't seem worth going back and renaming all the
uses of commutative_tree_code to commutative_binary_tree_code
to account for that.
So this was partly future-proofing. It was also partly to emphasise
that the caller doesn't need to check that the operator is a binary
operator first (although I'm not sure the name actually achieves
that, oh well).
>> +{
>> + if (code.is_tree_code ())
>> + return associative_tree_code (tree_code (code));
>> + auto cfn = combined_fn (code);
>> + return associative_binary_fn_p (associated_internal_fn (cfn, type));
>> +}
>> +
>> +/* Return true if the target directly supports operation CODE on type TYPE.
>> + QUERY_TYPE acts as for optab_for_tree_code. */
>> +
>> +bool
>> +directly_supported_p (code_helper code, tree type, optab_subtype query_type)
>> +{
>> + if (code.is_tree_code ())
>> + {
>> + direct_optab optab = optab_for_tree_code (tree_code (code), type,
>> + query_type);
>> + return (optab != unknown_optab
>> + && optab_handler (optab, TYPE_MODE (type)) != CODE_FOR_nothing);
>> + }
>> + gcc_assert (query_type == optab_default
>> + || (query_type == optab_vector && VECTOR_TYPE_P (type))
>> + || (query_type == optab_scalar && !VECTOR_TYPE_P (type)));
>> + internal_fn ifn = associated_internal_fn (combined_fn (code), type);
>> + return (direct_internal_fn_p (ifn)
>> + && direct_internal_fn_supported_p (ifn, type, OPTIMIZE_FOR_SPEED));
>> +}
>> +
>> +/* A wrapper around the internal-fn.c versions of get_conditional_internal_fn
>> + for a code_helper CODE operating on type TYPE. */
>> +
>> +internal_fn
>> +get_conditional_internal_fn (code_helper code, tree type)
>> +{
>> + if (code.is_tree_code ())
>> + return get_conditional_internal_fn (tree_code (code));
>> + auto cfn = combined_fn (code);
>> + return get_conditional_internal_fn (associated_internal_fn (cfn, type));
>> +}
>> diff --git a/gcc/gimple-match.h b/gcc/gimple-match.h
>> index 1b9dc3851c2..6d24a8a2378 100644
>> --- a/gcc/gimple-match.h
>> +++ b/gcc/gimple-match.h
>> @@ -31,6 +31,7 @@ public:
>> code_helper () {}
>> code_helper (tree_code code) : rep ((int) code) {}
>> code_helper (combined_fn fn) : rep (-(int) fn) {}
>> + code_helper (internal_fn fn) : rep (-(int) as_combined_fn (fn)) {}
>> explicit operator tree_code () const { return (tree_code) rep; }
>> explicit operator combined_fn () const { return (combined_fn) -rep; }
>
> Do we want a
>
> explicit operator internal_fn () const { ... }
>
> for completeness?
Yeah, guess that would simplify some things. Maybe a built_in_function
one too.
>
>> bool is_tree_code () const { return rep > 0; }
>> @@ -346,4 +347,23 @@ tree maybe_push_res_to_seq (gimple_match_op *, gimple_seq *,
>> void maybe_build_generic_op (gimple_match_op *);
>>
>>
>> +bool commutative_binary_op_p (code_helper, tree);
>> +bool associative_binary_op_p (code_helper, tree);
>> +code_helper canonicalize_code (code_helper, tree);
>> +
>> +#ifdef GCC_OPTABS_TREE_H
>> +bool directly_supported_p (code_helper, tree, optab_subtype = optab_default);
>> +#endif
>> +
>> +internal_fn get_conditional_internal_fn (code_helper, tree);
>> +
>> +extern tree gimple_build (gimple_seq *, location_t,
>> + code_helper, tree, tree, tree);
>> +inline tree
>> +gimple_build (gimple_seq *seq, code_helper code, tree type, tree op0,
>> + tree op1)
>> +{
>> + return gimple_build (seq, UNKNOWN_LOCATION, code, type, op0, op1);
>> +}
>
> That looks a bit misplaced and should be in gimple-fold.h, no?
Files don't need to include gimple-match.h before gimple-fold.h.
So I saw this as being a bit like the optab stuff: it's generalising
interfaces provided elsewhere for the "tree_code or combined_fn" union.
One alternative would be to put the functions in gimple-fold.h but
protect them with #ifdef GCC_GIMPLE_MATCH_H. Another would be to
move code_helper somewhere else, such as gimple.h.
Thanks,
Richard
>> +
>> #endif /* GCC_GIMPLE_MATCH_H */
>> diff --git a/gcc/internal-fn.c b/gcc/internal-fn.c
>> index da7d8355214..7b13db6dfe3 100644
>> --- a/gcc/internal-fn.c
>> +++ b/gcc/internal-fn.c
>> @@ -3815,6 +3815,43 @@ direct_internal_fn_supported_p (gcall *stmt, optimization_type opt_type)
>> return direct_internal_fn_supported_p (fn, types, opt_type);
>> }
>>
>> +/* Return true if FN is a commutative binary operation. */
>> +
>> +bool
>> +commutative_binary_fn_p (internal_fn fn)
>> +{
>> + switch (fn)
>> + {
>> + case IFN_AVG_FLOOR:
>> + case IFN_AVG_CEIL:
>> + case IFN_MULH:
>> + case IFN_MULHS:
>> + case IFN_MULHRS:
>> + case IFN_FMIN:
>> + case IFN_FMAX:
>> + return true;
>> +
>> + default:
>> + return false;
>> + }
>> +}
>> +
>> +/* Return true if FN is an associative binary operation. */
>> +
>> +bool
>> +associative_binary_fn_p (internal_fn fn)
>
> See above - without _binary?
>
>> +{
>> + switch (fn)
>> + {
>> + case IFN_FMIN:
>> + case IFN_FMAX:
>> + return true;
>> +
>> + default:
>> + return false;
>> + }
>> +}
>> +
>> /* If FN is commutative in two consecutive arguments, return the
>> index of the first, otherwise return -1. */
>>
>> @@ -3827,13 +3864,6 @@ first_commutative_argument (internal_fn fn)
>> case IFN_FMS:
>> case IFN_FNMA:
>> case IFN_FNMS:
>> - case IFN_AVG_FLOOR:
>> - case IFN_AVG_CEIL:
>> - case IFN_MULH:
>> - case IFN_MULHS:
>> - case IFN_MULHRS:
>> - case IFN_FMIN:
>> - case IFN_FMAX:
>> return 0;
>>
>> case IFN_COND_ADD:
>> @@ -3852,7 +3882,7 @@ first_commutative_argument (internal_fn fn)
>> return 1;
>>
>> default:
>> - return -1;
>> + return commutative_binary_fn_p (fn) ? 0 : -1;
>> }
>> }
>>
>> diff --git a/gcc/internal-fn.h b/gcc/internal-fn.h
>> index 19d0f849a5a..82ef4b0d792 100644
>> --- a/gcc/internal-fn.h
>> +++ b/gcc/internal-fn.h
>> @@ -206,6 +206,8 @@ direct_internal_fn_supported_p (internal_fn fn, tree type0, tree type1,
>> opt_type);
>> }
>>
>> +extern bool commutative_binary_fn_p (internal_fn);
>
> I'm somewhat missing commutative_ternary_fn_p which would work
> on FMAs?
>
> So that was all API comments, the real changes below look good to me.
>
> Thanks,
> Richard.
>
>> +extern bool associative_binary_fn_p (internal_fn);
>> extern int first_commutative_argument (internal_fn);
>>
>> extern bool set_edom_supported_p (void);
>> diff --git a/gcc/tree-vect-loop.c b/gcc/tree-vect-loop.c
>> index 1cd5dbcb6f7..cae895a88f2 100644
>> --- a/gcc/tree-vect-loop.c
>> +++ b/gcc/tree-vect-loop.c
>> @@ -54,6 +54,7 @@ along with GCC; see the file COPYING3. If not see
>> #include "tree-vector-builder.h"
>> #include "vec-perm-indices.h"
>> #include "tree-eh.h"
>> +#include "case-cfn-macros.h"
>>
>> /* Loop Vectorization Pass.
>>
>> @@ -3125,17 +3126,14 @@ vect_analyze_loop (class loop *loop, vec_info_shared *shared)
>> it in *REDUC_FN if so. */
>>
>> static bool
>> -fold_left_reduction_fn (tree_code code, internal_fn *reduc_fn)
>> +fold_left_reduction_fn (code_helper code, internal_fn *reduc_fn)
>> {
>> - switch (code)
>> + if (code == PLUS_EXPR)
>> {
>> - case PLUS_EXPR:
>> *reduc_fn = IFN_FOLD_LEFT_PLUS;
>> return true;
>> -
>> - default:
>> - return false;
>> }
>> + return false;
>> }
>>
>> /* Function reduction_fn_for_scalar_code
>> @@ -3152,21 +3150,22 @@ fold_left_reduction_fn (tree_code code, internal_fn *reduc_fn)
>> Return FALSE if CODE currently cannot be vectorized as reduction. */
>>
>> bool
>> -reduction_fn_for_scalar_code (enum tree_code code, internal_fn *reduc_fn)
>> +reduction_fn_for_scalar_code (code_helper code, internal_fn *reduc_fn)
>> {
>> - switch (code)
>> - {
>> + if (code.is_tree_code ())
>> + switch (tree_code (code))
>> + {
>> case MAX_EXPR:
>> - *reduc_fn = IFN_REDUC_MAX;
>> - return true;
>> + *reduc_fn = IFN_REDUC_MAX;
>> + return true;
>>
>> case MIN_EXPR:
>> - *reduc_fn = IFN_REDUC_MIN;
>> - return true;
>> + *reduc_fn = IFN_REDUC_MIN;
>> + return true;
>>
>> case PLUS_EXPR:
>> - *reduc_fn = IFN_REDUC_PLUS;
>> - return true;
>> + *reduc_fn = IFN_REDUC_PLUS;
>> + return true;
>>
>> case BIT_AND_EXPR:
>> *reduc_fn = IFN_REDUC_AND;
>> @@ -3182,12 +3181,13 @@ reduction_fn_for_scalar_code (enum tree_code code, internal_fn *reduc_fn)
>>
>> case MULT_EXPR:
>> case MINUS_EXPR:
>> - *reduc_fn = IFN_LAST;
>> - return true;
>> + *reduc_fn = IFN_LAST;
>> + return true;
>>
>> default:
>> - return false;
>> + break;
>> }
>> + return false;
>> }
>>
>> /* If there is a neutral value X such that a reduction would not be affected
>> @@ -3197,32 +3197,35 @@ reduction_fn_for_scalar_code (enum tree_code code, internal_fn *reduc_fn)
>> then INITIAL_VALUE is that value, otherwise it is null. */
>>
>> tree
>> -neutral_op_for_reduction (tree scalar_type, tree_code code, tree initial_value)
>> +neutral_op_for_reduction (tree scalar_type, code_helper code,
>> + tree initial_value)
>> {
>> - switch (code)
>> - {
>> - case WIDEN_SUM_EXPR:
>> - case DOT_PROD_EXPR:
>> - case SAD_EXPR:
>> - case PLUS_EXPR:
>> - case MINUS_EXPR:
>> - case BIT_IOR_EXPR:
>> - case BIT_XOR_EXPR:
>> - return build_zero_cst (scalar_type);
>> + if (code.is_tree_code ())
>> + switch (tree_code (code))
>> + {
>> + case WIDEN_SUM_EXPR:
>> + case DOT_PROD_EXPR:
>> + case SAD_EXPR:
>> + case PLUS_EXPR:
>> + case MINUS_EXPR:
>> + case BIT_IOR_EXPR:
>> + case BIT_XOR_EXPR:
>> + return build_zero_cst (scalar_type);
>>
>> - case MULT_EXPR:
>> - return build_one_cst (scalar_type);
>> + case MULT_EXPR:
>> + return build_one_cst (scalar_type);
>>
>> - case BIT_AND_EXPR:
>> - return build_all_ones_cst (scalar_type);
>> + case BIT_AND_EXPR:
>> + return build_all_ones_cst (scalar_type);
>>
>> - case MAX_EXPR:
>> - case MIN_EXPR:
>> - return initial_value;
>> + case MAX_EXPR:
>> + case MIN_EXPR:
>> + return initial_value;
>>
>> - default:
>> - return NULL_TREE;
>> - }
>> + default:
>> + break;
>> + }
>> + return NULL_TREE;
>> }
>>
>> /* Error reporting helper for vect_is_simple_reduction below. GIMPLE statement
>> @@ -3239,26 +3242,27 @@ report_vect_op (dump_flags_t msg_type, gimple *stmt, const char *msg)
>> overflow must wrap. */
>>
>> bool
>> -needs_fold_left_reduction_p (tree type, tree_code code)
>> +needs_fold_left_reduction_p (tree type, code_helper code)
>> {
>> /* CHECKME: check for !flag_finite_math_only too? */
>> if (SCALAR_FLOAT_TYPE_P (type))
>> - switch (code)
>> - {
>> - case MIN_EXPR:
>> - case MAX_EXPR:
>> - return false;
>> + {
>> + if (code.is_tree_code ())
>> + switch (tree_code (code))
>> + {
>> + case MIN_EXPR:
>> + case MAX_EXPR:
>> + return false;
>>
>> - default:
>> - return !flag_associative_math;
>> - }
>> + default:
>> + break;
>> + }
>> + return !flag_associative_math;
>> + }
>>
>> if (INTEGRAL_TYPE_P (type))
>> - {
>> - if (!operation_no_trapping_overflow (type, code))
>> - return true;
>> - return false;
>> - }
>> + return (!code.is_tree_code ()
>> + || !operation_no_trapping_overflow (type, tree_code (code)));
>>
>> if (SAT_FIXED_POINT_TYPE_P (type))
>> return true;
>> @@ -3272,7 +3276,7 @@ needs_fold_left_reduction_p (tree type, tree_code code)
>>
>> static bool
>> check_reduction_path (dump_user_location_t loc, loop_p loop, gphi *phi,
>> - tree loop_arg, enum tree_code *code,
>> + tree loop_arg, code_helper *code,
>> vec<std::pair<ssa_op_iter, use_operand_p> > &path)
>> {
>> auto_bitmap visited;
>> @@ -3347,45 +3351,57 @@ pop:
>> for (unsigned i = 1; i < path.length (); ++i)
>> {
>> gimple *use_stmt = USE_STMT (path[i].second);
>> - tree op = USE_FROM_PTR (path[i].second);
>> - if (! is_gimple_assign (use_stmt)
>> + gimple_match_op op;
>> + if (!gimple_extract_op (use_stmt, &op))
>> + {
>> + fail = true;
>> + break;
>> + }
>> + unsigned int opi = op.num_ops;
>> + if (gassign *assign = dyn_cast<gassign *> (use_stmt))
>> + {
>> /* The following make sure we can compute the operand index
>> easily plus it mostly disallows chaining via COND_EXPR condition
>> operands. */
>> - || (gimple_assign_rhs1_ptr (use_stmt) != path[i].second->use
>> - && (gimple_num_ops (use_stmt) <= 2
>> - || gimple_assign_rhs2_ptr (use_stmt) != path[i].second->use)
>> - && (gimple_num_ops (use_stmt) <= 3
>> - || gimple_assign_rhs3_ptr (use_stmt) != path[i].second->use)))
>> + for (opi = 0; opi < op.num_ops; ++opi)
>> + if (gimple_assign_rhs1_ptr (assign) + opi == path[i].second->use)
>> + break;
>> + }
>> + else if (gcall *call = dyn_cast<gcall *> (use_stmt))
>> + {
>> + for (opi = 0; opi < op.num_ops; ++opi)
>> + if (gimple_call_arg_ptr (call, opi) == path[i].second->use)
>> + break;
>> + }
>> + if (opi == op.num_ops)
>> {
>> fail = true;
>> break;
>> }
>> - tree_code use_code = gimple_assign_rhs_code (use_stmt);
>> - if (use_code == MINUS_EXPR)
>> + op.code = canonicalize_code (op.code, op.type);
>> + if (op.code == MINUS_EXPR)
>> {
>> - use_code = PLUS_EXPR;
>> + op.code = PLUS_EXPR;
>> /* Track whether we negate the reduction value each iteration. */
>> - if (gimple_assign_rhs2 (use_stmt) == op)
>> + if (op.ops[1] == op.ops[opi])
>> neg = ! neg;
>> }
>> - if (CONVERT_EXPR_CODE_P (use_code)
>> - && tree_nop_conversion_p (TREE_TYPE (gimple_assign_lhs (use_stmt)),
>> - TREE_TYPE (gimple_assign_rhs1 (use_stmt))))
>> + if (CONVERT_EXPR_CODE_P (op.code)
>> + && tree_nop_conversion_p (op.type, TREE_TYPE (op.ops[0])))
>> ;
>> else if (*code == ERROR_MARK)
>> {
>> - *code = use_code;
>> - sign = TYPE_SIGN (TREE_TYPE (gimple_assign_lhs (use_stmt)));
>> + *code = op.code;
>> + sign = TYPE_SIGN (op.type);
>> }
>> - else if (use_code != *code)
>> + else if (op.code != *code)
>> {
>> fail = true;
>> break;
>> }
>> - else if ((use_code == MIN_EXPR
>> - || use_code == MAX_EXPR)
>> - && sign != TYPE_SIGN (TREE_TYPE (gimple_assign_lhs (use_stmt))))
>> + else if ((op.code == MIN_EXPR
>> + || op.code == MAX_EXPR)
>> + && sign != TYPE_SIGN (op.type))
>> {
>> fail = true;
>> break;
>> @@ -3397,7 +3413,7 @@ pop:
>> imm_use_iterator imm_iter;
>> gimple *op_use_stmt;
>> unsigned cnt = 0;
>> - FOR_EACH_IMM_USE_STMT (op_use_stmt, imm_iter, op)
>> + FOR_EACH_IMM_USE_STMT (op_use_stmt, imm_iter, op.ops[opi])
>> if (!is_gimple_debug (op_use_stmt)
>> && (*code != ERROR_MARK
>> || flow_bb_inside_loop_p (loop, gimple_bb (op_use_stmt))))
>> @@ -3427,7 +3443,7 @@ check_reduction_path (dump_user_location_t loc, loop_p loop, gphi *phi,
>> tree loop_arg, enum tree_code code)
>> {
>> auto_vec<std::pair<ssa_op_iter, use_operand_p> > path;
>> - enum tree_code code_;
>> + code_helper code_;
>> return (check_reduction_path (loc, loop, phi, loop_arg, &code_, path)
>> && code_ == code);
>> }
>> @@ -3596,9 +3612,9 @@ vect_is_simple_reduction (loop_vec_info loop_info, stmt_vec_info phi_info,
>> gimple *def1 = SSA_NAME_DEF_STMT (op1);
>> if (gimple_bb (def1)
>> && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt))
>> - && loop->inner
>> - && flow_bb_inside_loop_p (loop->inner, gimple_bb (def1))
>> - && is_gimple_assign (def1)
>> + && loop->inner
>> + && flow_bb_inside_loop_p (loop->inner, gimple_bb (def1))
>> + && (is_gimple_assign (def1) || is_gimple_call (def1))
>> && is_a <gphi *> (phi_use_stmt)
>> && flow_bb_inside_loop_p (loop->inner, gimple_bb (phi_use_stmt)))
>> {
>> @@ -3615,7 +3631,7 @@ vect_is_simple_reduction (loop_vec_info loop_info, stmt_vec_info phi_info,
>>
>> /* Look for the expression computing latch_def from then loop PHI result. */
>> auto_vec<std::pair<ssa_op_iter, use_operand_p> > path;
>> - enum tree_code code;
>> + code_helper code;
>> if (check_reduction_path (vect_location, loop, phi, latch_def, &code,
>> path))
>> {
>> @@ -3633,15 +3649,24 @@ vect_is_simple_reduction (loop_vec_info loop_info, stmt_vec_info phi_info,
>> {
>> gimple *stmt = USE_STMT (path[i].second);
>> stmt_vec_info stmt_info = loop_info->lookup_stmt (stmt);
>> - STMT_VINFO_REDUC_IDX (stmt_info)
>> - = path[i].second->use - gimple_assign_rhs1_ptr (stmt);
>> - enum tree_code stmt_code = gimple_assign_rhs_code (stmt);
>> - bool leading_conversion = (CONVERT_EXPR_CODE_P (stmt_code)
>> + gimple_match_op op;
>> + if (!gimple_extract_op (stmt, &op))
>> + gcc_unreachable ();
>> + if (gassign *assign = dyn_cast<gassign *> (stmt))
>> + STMT_VINFO_REDUC_IDX (stmt_info)
>> + = path[i].second->use - gimple_assign_rhs1_ptr (assign);
>> + else
>> + {
>> + gcall *call = as_a<gcall *> (stmt);
>> + STMT_VINFO_REDUC_IDX (stmt_info)
>> + = path[i].second->use - gimple_call_arg_ptr (call, 0);
>> + }
>> + bool leading_conversion = (CONVERT_EXPR_CODE_P (op.code)
>> && (i == 1 || i == path.length () - 1));
>> - if ((stmt_code != code && !leading_conversion)
>> + if ((op.code != code && !leading_conversion)
>> /* We can only handle the final value in epilogue
>> generation for reduction chains. */
>> - || (i != 1 && !has_single_use (gimple_assign_lhs (stmt))))
>> + || (i != 1 && !has_single_use (gimple_get_lhs (stmt))))
>> is_slp_reduc = false;
>> /* For reduction chains we support a trailing/leading
>> conversions. We do not store those in the actual chain. */
>> @@ -4390,8 +4415,6 @@ vect_model_reduction_cost (loop_vec_info loop_vinfo,
>> int ncopies, stmt_vector_for_cost *cost_vec)
>> {
>> int prologue_cost = 0, epilogue_cost = 0, inside_cost = 0;
>> - enum tree_code code;
>> - optab optab;
>> tree vectype;
>> machine_mode mode;
>> class loop *loop = NULL;
>> @@ -4407,7 +4430,9 @@ vect_model_reduction_cost (loop_vec_info loop_vinfo,
>> mode = TYPE_MODE (vectype);
>> stmt_vec_info orig_stmt_info = vect_orig_stmt (stmt_info);
>>
>> - code = gimple_assign_rhs_code (orig_stmt_info->stmt);
>> + gimple_match_op op;
>> + if (!gimple_extract_op (orig_stmt_info->stmt, &op))
>> + gcc_unreachable ();
>>
>> if (reduction_type == EXTRACT_LAST_REDUCTION)
>> /* No extra instructions are needed in the prologue. The loop body
>> @@ -4501,20 +4526,16 @@ vect_model_reduction_cost (loop_vec_info loop_vinfo,
>> else
>> {
>> int vec_size_in_bits = tree_to_uhwi (TYPE_SIZE (vectype));
>> - tree bitsize =
>> - TYPE_SIZE (TREE_TYPE (gimple_assign_lhs (orig_stmt_info->stmt)));
>> + tree bitsize = TYPE_SIZE (op.type);
>> int element_bitsize = tree_to_uhwi (bitsize);
>> int nelements = vec_size_in_bits / element_bitsize;
>>
>> - if (code == COND_EXPR)
>> - code = MAX_EXPR;
>> -
>> - optab = optab_for_tree_code (code, vectype, optab_default);
>> + if (op.code == COND_EXPR)
>> + op.code = MAX_EXPR;
>>
>> /* We have a whole vector shift available. */
>> - if (optab != unknown_optab
>> - && VECTOR_MODE_P (mode)
>> - && optab_handler (optab, mode) != CODE_FOR_nothing
>> + if (VECTOR_MODE_P (mode)
>> + && directly_supported_p (op.code, vectype)
>> && have_whole_vector_shift (mode))
>> {
>> /* Final reduction via vector shifts and the reduction operator.
>> @@ -4855,7 +4876,7 @@ vect_find_reusable_accumulator (loop_vec_info loop_vinfo,
>> initialize the accumulator with a neutral value instead. */
>> if (!operand_equal_p (initial_value, main_adjustment))
>> return false;
>> - tree_code code = STMT_VINFO_REDUC_CODE (reduc_info);
>> + code_helper code = STMT_VINFO_REDUC_CODE (reduc_info);
>> initial_values[0] = neutral_op_for_reduction (TREE_TYPE (initial_value),
>> code, initial_value);
>> }
>> @@ -4870,7 +4891,7 @@ vect_find_reusable_accumulator (loop_vec_info loop_vinfo,
>> CODE emitting stmts before GSI. Returns a vector def of VECTYPE. */
>>
>> static tree
>> -vect_create_partial_epilog (tree vec_def, tree vectype, enum tree_code code,
>> +vect_create_partial_epilog (tree vec_def, tree vectype, code_helper code,
>> gimple_seq *seq)
>> {
>> unsigned nunits = TYPE_VECTOR_SUBPARTS (TREE_TYPE (vec_def)).to_constant ();
>> @@ -4953,9 +4974,7 @@ vect_create_partial_epilog (tree vec_def, tree vectype, enum tree_code code,
>> gimple_seq_add_stmt_without_update (seq, epilog_stmt);
>> }
>>
>> - new_temp = make_ssa_name (vectype1);
>> - epilog_stmt = gimple_build_assign (new_temp, code, dst1, dst2);
>> - gimple_seq_add_stmt_without_update (seq, epilog_stmt);
>> + new_temp = gimple_build (seq, code, vectype1, dst1, dst2);
>> }
>>
>> return new_temp;
>> @@ -5032,7 +5051,7 @@ vect_create_epilog_for_reduction (loop_vec_info loop_vinfo,
>> }
>> gphi *reduc_def_stmt
>> = as_a <gphi *> (STMT_VINFO_REDUC_DEF (vect_orig_stmt (stmt_info))->stmt);
>> - enum tree_code code = STMT_VINFO_REDUC_CODE (reduc_info);
>> + code_helper code = STMT_VINFO_REDUC_CODE (reduc_info);
>> internal_fn reduc_fn = STMT_VINFO_REDUC_FN (reduc_info);
>> tree vectype;
>> machine_mode mode;
>> @@ -5699,14 +5718,9 @@ vect_create_epilog_for_reduction (loop_vec_info loop_vinfo,
>> tree vectype1 = get_related_vectype_for_scalar_type (TYPE_MODE (vectype),
>> stype, nunits1);
>> reduce_with_shift = have_whole_vector_shift (mode1);
>> - if (!VECTOR_MODE_P (mode1))
>> + if (!VECTOR_MODE_P (mode1)
>> + || !directly_supported_p (code, vectype1))
>> reduce_with_shift = false;
>> - else
>> - {
>> - optab optab = optab_for_tree_code (code, vectype1, optab_default);
>> - if (optab_handler (optab, mode1) == CODE_FOR_nothing)
>> - reduce_with_shift = false;
>> - }
>>
>> /* First reduce the vector to the desired vector size we should
>> do shift reduction on by combining upper and lower halves. */
>> @@ -5944,7 +5958,7 @@ vect_create_epilog_for_reduction (loop_vec_info loop_vinfo,
>> for (k = 0; k < live_out_stmts.size (); k++)
>> {
>> stmt_vec_info scalar_stmt_info = vect_orig_stmt (live_out_stmts[k]);
>> - scalar_dest = gimple_assign_lhs (scalar_stmt_info->stmt);
>> + scalar_dest = gimple_get_lhs (scalar_stmt_info->stmt);
>>
>> phis.create (3);
>> /* Find the loop-closed-use at the loop exit of the original scalar
>> @@ -6277,7 +6291,7 @@ is_nonwrapping_integer_induction (stmt_vec_info stmt_vinfo, class loop *loop)
>> CODE is the code for the operation. COND_FN is the conditional internal
>> function, if it exists. VECTYPE_IN is the type of the vector input. */
>> static bool
>> -use_mask_by_cond_expr_p (enum tree_code code, internal_fn cond_fn,
>> +use_mask_by_cond_expr_p (code_helper code, internal_fn cond_fn,
>> tree vectype_in)
>> {
>> if (cond_fn != IFN_LAST
>> @@ -6285,15 +6299,17 @@ use_mask_by_cond_expr_p (enum tree_code code, internal_fn cond_fn,
>> OPTIMIZE_FOR_SPEED))
>> return false;
>>
>> - switch (code)
>> - {
>> - case DOT_PROD_EXPR:
>> - case SAD_EXPR:
>> - return true;
>> + if (code.is_tree_code ())
>> + switch (tree_code (code))
>> + {
>> + case DOT_PROD_EXPR:
>> + case SAD_EXPR:
>> + return true;
>>
>> - default:
>> - return false;
>> - }
>> + default:
>> + break;
>> + }
>> + return false;
>> }
>>
>> /* Insert a conditional expression to enable masked vectorization. CODE is the
>> @@ -6301,10 +6317,10 @@ use_mask_by_cond_expr_p (enum tree_code code, internal_fn cond_fn,
>> mask. GSI is a statement iterator used to place the new conditional
>> expression. */
>> static void
>> -build_vect_cond_expr (enum tree_code code, tree vop[3], tree mask,
>> +build_vect_cond_expr (code_helper code, tree vop[3], tree mask,
>> gimple_stmt_iterator *gsi)
>> {
>> - switch (code)
>> + switch (tree_code (code))
>> {
>> case DOT_PROD_EXPR:
>> {
>> @@ -6390,12 +6406,10 @@ vectorizable_reduction (loop_vec_info loop_vinfo,
>> slp_instance slp_node_instance,
>> stmt_vector_for_cost *cost_vec)
>> {
>> - tree scalar_dest;
>> tree vectype_in = NULL_TREE;
>> class loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
>> enum vect_def_type cond_reduc_dt = vect_unknown_def_type;
>> stmt_vec_info cond_stmt_vinfo = NULL;
>> - tree scalar_type;
>> int i;
>> int ncopies;
>> bool single_defuse_cycle = false;
>> @@ -6508,18 +6522,18 @@ vectorizable_reduction (loop_vec_info loop_vinfo,
>> info_for_reduction to work. */
>> if (STMT_VINFO_LIVE_P (vdef))
>> STMT_VINFO_REDUC_DEF (def) = phi_info;
>> - gassign *assign = dyn_cast <gassign *> (vdef->stmt);
>> - if (!assign)
>> + gimple_match_op op;
>> + if (!gimple_extract_op (vdef->stmt, &op))
>> {
>> if (dump_enabled_p ())
>> dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
>> - "reduction chain includes calls.\n");
>> + "reduction chain includes unsupported"
>> + " statement type.\n");
>> return false;
>> }
>> - if (CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (assign)))
>> + if (CONVERT_EXPR_CODE_P (op.code))
>> {
>> - if (!tree_nop_conversion_p (TREE_TYPE (gimple_assign_lhs (assign)),
>> - TREE_TYPE (gimple_assign_rhs1 (assign))))
>> + if (!tree_nop_conversion_p (op.type, TREE_TYPE (op.ops[0])))
>> {
>> if (dump_enabled_p ())
>> dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
>> @@ -6530,7 +6544,7 @@ vectorizable_reduction (loop_vec_info loop_vinfo,
>> else if (!stmt_info)
>> /* First non-conversion stmt. */
>> stmt_info = vdef;
>> - reduc_def = gimple_op (vdef->stmt, 1 + STMT_VINFO_REDUC_IDX (vdef));
>> + reduc_def = op.ops[STMT_VINFO_REDUC_IDX (vdef)];
>> reduc_chain_length++;
>> if (!stmt_info && slp_node)
>> slp_for_stmt_info = SLP_TREE_CHILDREN (slp_for_stmt_info)[0];
>> @@ -6588,26 +6602,24 @@ vectorizable_reduction (loop_vec_info loop_vinfo,
>>
>> tree vectype_out = STMT_VINFO_VECTYPE (stmt_info);
>> STMT_VINFO_REDUC_VECTYPE (reduc_info) = vectype_out;
>> - gassign *stmt = as_a <gassign *> (stmt_info->stmt);
>> - enum tree_code code = gimple_assign_rhs_code (stmt);
>> - bool lane_reduc_code_p
>> - = (code == DOT_PROD_EXPR || code == WIDEN_SUM_EXPR || code == SAD_EXPR);
>> - int op_type = TREE_CODE_LENGTH (code);
>> + gimple_match_op op;
>> + if (!gimple_extract_op (stmt_info->stmt, &op))
>> + gcc_unreachable ();
>> + bool lane_reduc_code_p = (op.code == DOT_PROD_EXPR
>> + || op.code == WIDEN_SUM_EXPR
>> + || op.code == SAD_EXPR);
>> enum optab_subtype optab_query_kind = optab_vector;
>> - if (code == DOT_PROD_EXPR
>> - && TYPE_SIGN (TREE_TYPE (gimple_assign_rhs1 (stmt)))
>> - != TYPE_SIGN (TREE_TYPE (gimple_assign_rhs2 (stmt))))
>> + if (op.code == DOT_PROD_EXPR
>> + && (TYPE_SIGN (TREE_TYPE (op.ops[0]))
>> + != TYPE_SIGN (TREE_TYPE (op.ops[1]))))
>> optab_query_kind = optab_vector_mixed_sign;
>>
>> -
>> - scalar_dest = gimple_assign_lhs (stmt);
>> - scalar_type = TREE_TYPE (scalar_dest);
>> - if (!POINTER_TYPE_P (scalar_type) && !INTEGRAL_TYPE_P (scalar_type)
>> - && !SCALAR_FLOAT_TYPE_P (scalar_type))
>> + if (!POINTER_TYPE_P (op.type) && !INTEGRAL_TYPE_P (op.type)
>> + && !SCALAR_FLOAT_TYPE_P (op.type))
>> return false;
>>
>> /* Do not try to vectorize bit-precision reductions. */
>> - if (!type_has_mode_precision_p (scalar_type))
>> + if (!type_has_mode_precision_p (op.type))
>> return false;
>>
>> /* For lane-reducing ops we're reducing the number of reduction PHIs
>> @@ -6626,25 +6638,23 @@ vectorizable_reduction (loop_vec_info loop_vinfo,
>> The last use is the reduction variable. In case of nested cycle this
>> assumption is not true: we use reduc_index to record the index of the
>> reduction variable. */
>> - slp_tree *slp_op = XALLOCAVEC (slp_tree, op_type);
>> + slp_tree *slp_op = XALLOCAVEC (slp_tree, op.num_ops);
>> /* We need to skip an extra operand for COND_EXPRs with embedded
>> comparison. */
>> unsigned opno_adjust = 0;
>> - if (code == COND_EXPR
>> - && COMPARISON_CLASS_P (gimple_assign_rhs1 (stmt)))
>> + if (op.code == COND_EXPR && COMPARISON_CLASS_P (op.ops[0]))
>> opno_adjust = 1;
>> - for (i = 0; i < op_type; i++)
>> + for (i = 0; i < (int) op.num_ops; i++)
>> {
>> /* The condition of COND_EXPR is checked in vectorizable_condition(). */
>> - if (i == 0 && code == COND_EXPR)
>> + if (i == 0 && op.code == COND_EXPR)
>> continue;
>>
>> stmt_vec_info def_stmt_info;
>> enum vect_def_type dt;
>> - tree op;
>> if (!vect_is_simple_use (loop_vinfo, stmt_info, slp_for_stmt_info,
>> - i + opno_adjust, &op, &slp_op[i], &dt, &tem,
>> - &def_stmt_info))
>> + i + opno_adjust, &op.ops[i], &slp_op[i], &dt,
>> + &tem, &def_stmt_info))
>> {
>> if (dump_enabled_p ())
>> dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
>> @@ -6669,13 +6679,13 @@ vectorizable_reduction (loop_vec_info loop_vinfo,
>> < GET_MODE_SIZE (SCALAR_TYPE_MODE (TREE_TYPE (tem))))))
>> vectype_in = tem;
>>
>> - if (code == COND_EXPR)
>> + if (op.code == COND_EXPR)
>> {
>> /* Record how the non-reduction-def value of COND_EXPR is defined. */
>> if (dt == vect_constant_def)
>> {
>> cond_reduc_dt = dt;
>> - cond_reduc_val = op;
>> + cond_reduc_val = op.ops[i];
>> }
>> if (dt == vect_induction_def
>> && def_stmt_info
>> @@ -6845,7 +6855,7 @@ vectorizable_reduction (loop_vec_info loop_vinfo,
>> (and also the same tree-code) when generating the epilog code and
>> when generating the code inside the loop. */
>>
>> - enum tree_code orig_code = STMT_VINFO_REDUC_CODE (phi_info);
>> + code_helper orig_code = STMT_VINFO_REDUC_CODE (phi_info);
>> STMT_VINFO_REDUC_CODE (reduc_info) = orig_code;
>>
>> vect_reduction_type reduction_type = STMT_VINFO_REDUC_TYPE (reduc_info);
>> @@ -6864,7 +6874,7 @@ vectorizable_reduction (loop_vec_info loop_vinfo,
>> && !REDUC_GROUP_FIRST_ELEMENT (stmt_info)
>> && known_eq (LOOP_VINFO_VECT_FACTOR (loop_vinfo), 1u))
>> ;
>> - else if (needs_fold_left_reduction_p (scalar_type, orig_code))
>> + else if (needs_fold_left_reduction_p (op.type, orig_code))
>> {
>> /* When vectorizing a reduction chain w/o SLP the reduction PHI
>> is not directy used in stmt. */
>> @@ -6879,8 +6889,8 @@ vectorizable_reduction (loop_vec_info loop_vinfo,
>> STMT_VINFO_REDUC_TYPE (reduc_info)
>> = reduction_type = FOLD_LEFT_REDUCTION;
>> }
>> - else if (!commutative_tree_code (orig_code)
>> - || !associative_tree_code (orig_code))
>> + else if (!commutative_binary_op_p (orig_code, op.type)
>> + || !associative_binary_op_p (orig_code, op.type))
>> {
>> if (dump_enabled_p ())
>> dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
>> @@ -6935,7 +6945,7 @@ vectorizable_reduction (loop_vec_info loop_vinfo,
>> else if (reduction_type == COND_REDUCTION)
>> {
>> int scalar_precision
>> - = GET_MODE_PRECISION (SCALAR_TYPE_MODE (scalar_type));
>> + = GET_MODE_PRECISION (SCALAR_TYPE_MODE (op.type));
>> cr_index_scalar_type = make_unsigned_type (scalar_precision);
>> cr_index_vector_type = get_same_sized_vectype (cr_index_scalar_type,
>> vectype_out);
>> @@ -7121,28 +7131,19 @@ vectorizable_reduction (loop_vec_info loop_vinfo,
>>
>> if (single_defuse_cycle || lane_reduc_code_p)
>> {
>> - gcc_assert (code != COND_EXPR);
>> + gcc_assert (op.code != COND_EXPR);
>>
>> /* 4. Supportable by target? */
>> bool ok = true;
>>
>> /* 4.1. check support for the operation in the loop */
>> - optab optab = optab_for_tree_code (code, vectype_in, optab_query_kind);
>> - if (!optab)
>> - {
>> - if (dump_enabled_p ())
>> - dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
>> - "no optab.\n");
>> - ok = false;
>> - }
>> -
>> machine_mode vec_mode = TYPE_MODE (vectype_in);
>> - if (ok && optab_handler (optab, vec_mode) == CODE_FOR_nothing)
>> + if (!directly_supported_p (op.code, vectype_in, optab_query_kind))
>> {
>> if (dump_enabled_p ())
>> dump_printf (MSG_NOTE, "op not supported by target.\n");
>> if (maybe_ne (GET_MODE_SIZE (vec_mode), UNITS_PER_WORD)
>> - || !vect_can_vectorize_without_simd_p (code))
>> + || !vect_can_vectorize_without_simd_p (op.code))
>> ok = false;
>> else
>> if (dump_enabled_p ())
>> @@ -7150,7 +7151,7 @@ vectorizable_reduction (loop_vec_info loop_vinfo,
>> }
>>
>> if (vect_emulated_vector_p (vectype_in)
>> - && !vect_can_vectorize_without_simd_p (code))
>> + && !vect_can_vectorize_without_simd_p (op.code))
>> {
>> if (dump_enabled_p ())
>> dump_printf (MSG_NOTE, "using word mode not possible.\n");
>> @@ -7183,11 +7184,9 @@ vectorizable_reduction (loop_vec_info loop_vinfo,
>>
>> if (slp_node
>> && !(!single_defuse_cycle
>> - && code != DOT_PROD_EXPR
>> - && code != WIDEN_SUM_EXPR
>> - && code != SAD_EXPR
>> + && !lane_reduc_code_p
>> && reduction_type != FOLD_LEFT_REDUCTION))
>> - for (i = 0; i < op_type; i++)
>> + for (i = 0; i < (int) op.num_ops; i++)
>> if (!vect_maybe_update_slp_op_vectype (slp_op[i], vectype_in))
>> {
>> if (dump_enabled_p ())
>> @@ -7206,10 +7205,7 @@ vectorizable_reduction (loop_vec_info loop_vinfo,
>> /* Cost the reduction op inside the loop if transformed via
>> vect_transform_reduction. Otherwise this is costed by the
>> separate vectorizable_* routines. */
>> - if (single_defuse_cycle
>> - || code == DOT_PROD_EXPR
>> - || code == WIDEN_SUM_EXPR
>> - || code == SAD_EXPR)
>> + if (single_defuse_cycle || lane_reduc_code_p)
>> record_stmt_cost (cost_vec, ncopies, vector_stmt, stmt_info, 0, vect_body);
>>
>> if (dump_enabled_p ()
>> @@ -7220,9 +7216,7 @@ vectorizable_reduction (loop_vec_info loop_vinfo,
>> /* All but single defuse-cycle optimized, lane-reducing and fold-left
>> reductions go through their own vectorizable_* routines. */
>> if (!single_defuse_cycle
>> - && code != DOT_PROD_EXPR
>> - && code != WIDEN_SUM_EXPR
>> - && code != SAD_EXPR
>> + && !lane_reduc_code_p
>> && reduction_type != FOLD_LEFT_REDUCTION)
>> {
>> stmt_vec_info tem
>> @@ -7238,10 +7232,10 @@ vectorizable_reduction (loop_vec_info loop_vinfo,
>> else if (loop_vinfo && LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo))
>> {
>> vec_loop_masks *masks = &LOOP_VINFO_MASKS (loop_vinfo);
>> - internal_fn cond_fn = get_conditional_internal_fn (code);
>> + internal_fn cond_fn = get_conditional_internal_fn (op.code, op.type);
>>
>> if (reduction_type != FOLD_LEFT_REDUCTION
>> - && !use_mask_by_cond_expr_p (code, cond_fn, vectype_in)
>> + && !use_mask_by_cond_expr_p (op.code, cond_fn, vectype_in)
>> && (cond_fn == IFN_LAST
>> || !direct_internal_fn_supported_p (cond_fn, vectype_in,
>> OPTIMIZE_FOR_SPEED)))
>> @@ -7294,24 +7288,11 @@ vect_transform_reduction (loop_vec_info loop_vinfo,
>> gcc_assert (STMT_VINFO_DEF_TYPE (reduc_info) == vect_double_reduction_def);
>> }
>>
>> - gassign *stmt = as_a <gassign *> (stmt_info->stmt);
>> - enum tree_code code = gimple_assign_rhs_code (stmt);
>> - int op_type = TREE_CODE_LENGTH (code);
>> -
>> - /* Flatten RHS. */
>> - tree ops[3];
>> - switch (get_gimple_rhs_class (code))
>> - {
>> - case GIMPLE_TERNARY_RHS:
>> - ops[2] = gimple_assign_rhs3 (stmt);
>> - /* Fall thru. */
>> - case GIMPLE_BINARY_RHS:
>> - ops[0] = gimple_assign_rhs1 (stmt);
>> - ops[1] = gimple_assign_rhs2 (stmt);
>> - break;
>> - default:
>> - gcc_unreachable ();
>> - }
>> + gimple_match_op op;
>> + if (!gimple_extract_op (stmt_info->stmt, &op))
>> + gcc_unreachable ();
>> + gcc_assert (op.code.is_tree_code ());
>> + auto code = tree_code (op.code);
>>
>> /* All uses but the last are expected to be defined in the loop.
>> The last use is the reduction variable. In case of nested cycle this
>> @@ -7359,7 +7340,7 @@ vect_transform_reduction (loop_vec_info loop_vinfo,
>> internal_fn reduc_fn = STMT_VINFO_REDUC_FN (reduc_info);
>> return vectorize_fold_left_reduction
>> (loop_vinfo, stmt_info, gsi, vec_stmt, slp_node, reduc_def_phi, code,
>> - reduc_fn, ops, vectype_in, reduc_index, masks);
>> + reduc_fn, op.ops, vectype_in, reduc_index, masks);
>> }
>>
>> bool single_defuse_cycle = STMT_VINFO_FORCE_SINGLE_CYCLE (reduc_info);
>> @@ -7369,22 +7350,22 @@ vect_transform_reduction (loop_vec_info loop_vinfo,
>> || code == SAD_EXPR);
>>
>> /* Create the destination vector */
>> - tree scalar_dest = gimple_assign_lhs (stmt);
>> + tree scalar_dest = gimple_assign_lhs (stmt_info->stmt);
>> tree vec_dest = vect_create_destination_var (scalar_dest, vectype_out);
>>
>> vect_get_vec_defs (loop_vinfo, stmt_info, slp_node, ncopies,
>> single_defuse_cycle && reduc_index == 0
>> - ? NULL_TREE : ops[0], &vec_oprnds0,
>> + ? NULL_TREE : op.ops[0], &vec_oprnds0,
>> single_defuse_cycle && reduc_index == 1
>> - ? NULL_TREE : ops[1], &vec_oprnds1,
>> - op_type == ternary_op
>> + ? NULL_TREE : op.ops[1], &vec_oprnds1,
>> + op.num_ops == 3
>> && !(single_defuse_cycle && reduc_index == 2)
>> - ? ops[2] : NULL_TREE, &vec_oprnds2);
>> + ? op.ops[2] : NULL_TREE, &vec_oprnds2);
>> if (single_defuse_cycle)
>> {
>> gcc_assert (!slp_node);
>> vect_get_vec_defs_for_operand (loop_vinfo, stmt_info, 1,
>> - ops[reduc_index],
>> + op.ops[reduc_index],
>> reduc_index == 0 ? &vec_oprnds0
>> : (reduc_index == 1 ? &vec_oprnds1
>> : &vec_oprnds2));
>> @@ -7414,7 +7395,7 @@ vect_transform_reduction (loop_vec_info loop_vinfo,
>> }
>> else
>> {
>> - if (op_type == ternary_op)
>> + if (op.num_ops == 3)
>> vop[2] = vec_oprnds2[i];
>>
>> if (masked_loop_p && mask_by_cond_expr)
>> @@ -7546,7 +7527,7 @@ vect_transform_cycle_phi (loop_vec_info loop_vinfo,
>> {
>> tree initial_value
>> = (num_phis == 1 ? initial_values[0] : NULL_TREE);
>> - tree_code code = STMT_VINFO_REDUC_CODE (reduc_info);
>> + code_helper code = STMT_VINFO_REDUC_CODE (reduc_info);
>> tree neutral_op
>> = neutral_op_for_reduction (TREE_TYPE (vectype_out),
>> code, initial_value);
>> @@ -7603,7 +7584,7 @@ vect_transform_cycle_phi (loop_vec_info loop_vinfo,
>> if (!reduc_info->reduc_initial_values.is_empty ())
>> {
>> initial_def = reduc_info->reduc_initial_values[0];
>> - enum tree_code code = STMT_VINFO_REDUC_CODE (reduc_info);
>> + code_helper code = STMT_VINFO_REDUC_CODE (reduc_info);
>> tree neutral_op
>> = neutral_op_for_reduction (TREE_TYPE (initial_def),
>> code, initial_def);
>> @@ -7901,6 +7882,15 @@ vect_can_vectorize_without_simd_p (tree_code code)
>> }
>> }
>>
>> +/* Likewise, but taking a code_helper. */
>> +
>> +bool
>> +vect_can_vectorize_without_simd_p (code_helper code)
>> +{
>> + return (code.is_tree_code ()
>> + && vect_can_vectorize_without_simd_p (tree_code (code)));
>> +}
>> +
>> /* Function vectorizable_induction
>>
>> Check if STMT_INFO performs an induction computation that can be vectorized.
>> diff --git a/gcc/tree-vect-patterns.c b/gcc/tree-vect-patterns.c
>> index 854cbcff390..26421ee5511 100644
>> --- a/gcc/tree-vect-patterns.c
>> +++ b/gcc/tree-vect-patterns.c
>> @@ -5594,8 +5594,10 @@ vect_mark_pattern_stmts (vec_info *vinfo,
>> /* Transfer reduction path info to the pattern. */
>> if (STMT_VINFO_REDUC_IDX (orig_stmt_info_saved) != -1)
>> {
>> - tree lookfor = gimple_op (orig_stmt_info_saved->stmt,
>> - 1 + STMT_VINFO_REDUC_IDX (orig_stmt_info));
>> + gimple_match_op op;
>> + if (!gimple_extract_op (orig_stmt_info_saved->stmt, &op))
>> + gcc_unreachable ();
>> + tree lookfor = op.ops[STMT_VINFO_REDUC_IDX (orig_stmt_info)];
>> /* Search the pattern def sequence and the main pattern stmt. Note
>> we may have inserted all into a containing pattern def sequence
>> so the following is a bit awkward. */
>> @@ -5615,14 +5617,15 @@ vect_mark_pattern_stmts (vec_info *vinfo,
>> do
>> {
>> bool found = false;
>> - for (unsigned i = 1; i < gimple_num_ops (s); ++i)
>> - if (gimple_op (s, i) == lookfor)
>> - {
>> - STMT_VINFO_REDUC_IDX (vinfo->lookup_stmt (s)) = i - 1;
>> - lookfor = gimple_get_lhs (s);
>> - found = true;
>> - break;
>> - }
>> + if (gimple_extract_op (s, &op))
>> + for (unsigned i = 0; i < op.num_ops; ++i)
>> + if (op.ops[i] == lookfor)
>> + {
>> + STMT_VINFO_REDUC_IDX (vinfo->lookup_stmt (s)) = i;
>> + lookfor = gimple_get_lhs (s);
>> + found = true;
>> + break;
>> + }
>> if (s == pattern_stmt)
>> {
>> if (!found && dump_enabled_p ())
>> diff --git a/gcc/tree-vect-stmts.c b/gcc/tree-vect-stmts.c
>> index 03cc7267cf8..1e197023b98 100644
>> --- a/gcc/tree-vect-stmts.c
>> +++ b/gcc/tree-vect-stmts.c
>> @@ -3202,7 +3202,6 @@ vectorizable_call (vec_info *vinfo,
>> int ndts = ARRAY_SIZE (dt);
>> int ncopies, j;
>> auto_vec<tree, 8> vargs;
>> - auto_vec<tree, 8> orig_vargs;
>> enum { NARROW, NONE, WIDEN } modifier;
>> size_t i, nargs;
>> tree lhs;
>> @@ -3426,6 +3425,8 @@ vectorizable_call (vec_info *vinfo,
>> needs to be generated. */
>> gcc_assert (ncopies >= 1);
>>
>> + int reduc_idx = STMT_VINFO_REDUC_IDX (stmt_info);
>> + internal_fn cond_fn = get_conditional_internal_fn (ifn);
>> vec_loop_masks *masks = (loop_vinfo ? &LOOP_VINFO_MASKS (loop_vinfo) : NULL);
>> if (!vec_stmt) /* transformation not required. */
>> {
>> @@ -3446,14 +3447,33 @@ vectorizable_call (vec_info *vinfo,
>> record_stmt_cost (cost_vec, ncopies / 2,
>> vec_promote_demote, stmt_info, 0, vect_body);
>>
>> - if (loop_vinfo && mask_opno >= 0)
>> + if (loop_vinfo
>> + && LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo)
>> + && (reduc_idx >= 0 || mask_opno >= 0))
>> {
>> - unsigned int nvectors = (slp_node
>> - ? SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node)
>> - : ncopies);
>> - tree scalar_mask = gimple_call_arg (stmt_info->stmt, mask_opno);
>> - vect_record_loop_mask (loop_vinfo, masks, nvectors,
>> - vectype_out, scalar_mask);
>> + if (reduc_idx >= 0
>> + && (cond_fn == IFN_LAST
>> + || !direct_internal_fn_supported_p (cond_fn, vectype_out,
>> + OPTIMIZE_FOR_SPEED)))
>> + {
>> + if (dump_enabled_p ())
>> + dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
>> + "can't use a fully-masked loop because no"
>> + " conditional operation is available.\n");
>> + LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P (loop_vinfo) = false;
>> + }
>> + else
>> + {
>> + unsigned int nvectors
>> + = (slp_node
>> + ? SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node)
>> + : ncopies);
>> + tree scalar_mask = NULL_TREE;
>> + if (mask_opno >= 0)
>> + scalar_mask = gimple_call_arg (stmt_info->stmt, mask_opno);
>> + vect_record_loop_mask (loop_vinfo, masks, nvectors,
>> + vectype_out, scalar_mask);
>> + }
>> }
>> return true;
>> }
>> @@ -3468,12 +3488,17 @@ vectorizable_call (vec_info *vinfo,
>> vec_dest = vect_create_destination_var (scalar_dest, vectype_out);
>>
>> bool masked_loop_p = loop_vinfo && LOOP_VINFO_FULLY_MASKED_P (loop_vinfo);
>> + unsigned int vect_nargs = nargs;
>> + if (masked_loop_p && reduc_idx >= 0)
>> + {
>> + ifn = cond_fn;
>> + vect_nargs += 2;
>> + }
>>
>> if (modifier == NONE || ifn != IFN_LAST)
>> {
>> tree prev_res = NULL_TREE;
>> - vargs.safe_grow (nargs, true);
>> - orig_vargs.safe_grow (nargs, true);
>> + vargs.safe_grow (vect_nargs, true);
>> auto_vec<vec<tree> > vec_defs (nargs);
>> for (j = 0; j < ncopies; ++j)
>> {
>> @@ -3488,12 +3513,23 @@ vectorizable_call (vec_info *vinfo,
>> /* Arguments are ready. Create the new vector stmt. */
>> FOR_EACH_VEC_ELT (vec_oprnds0, i, vec_oprnd0)
>> {
>> + int varg = 0;
>> + if (masked_loop_p && reduc_idx >= 0)
>> + {
>> + unsigned int vec_num = vec_oprnds0.length ();
>> + /* Always true for SLP. */
>> + gcc_assert (ncopies == 1);
>> + vargs[varg++] = vect_get_loop_mask (gsi, masks, vec_num,
>> + vectype_out, i);
>> + }
>> size_t k;
>> for (k = 0; k < nargs; k++)
>> {
>> vec<tree> vec_oprndsk = vec_defs[k];
>> - vargs[k] = vec_oprndsk[i];
>> + vargs[varg++] = vec_oprndsk[i];
>> }
>> + if (masked_loop_p && reduc_idx >= 0)
>> + vargs[varg++] = vargs[reduc_idx + 1];
>> gimple *new_stmt;
>> if (modifier == NARROW)
>> {
>> @@ -3546,6 +3582,10 @@ vectorizable_call (vec_info *vinfo,
>> continue;
>> }
>>
>> + int varg = 0;
>> + if (masked_loop_p && reduc_idx >= 0)
>> + vargs[varg++] = vect_get_loop_mask (gsi, masks, ncopies,
>> + vectype_out, j);
>> for (i = 0; i < nargs; i++)
>> {
>> op = gimple_call_arg (stmt, i);
>> @@ -3556,8 +3596,10 @@ vectorizable_call (vec_info *vinfo,
>> op, &vec_defs[i],
>> vectypes[i]);
>> }
>> - orig_vargs[i] = vargs[i] = vec_defs[i][j];
>> + vargs[varg++] = vec_defs[i][j];
>> }
>> + if (masked_loop_p && reduc_idx >= 0)
>> + vargs[varg++] = vargs[reduc_idx + 1];
>>
>> if (mask_opno >= 0 && masked_loop_p)
>> {
>> diff --git a/gcc/tree-vectorizer.h b/gcc/tree-vectorizer.h
>> index f8f30641512..8330cd897b8 100644
>> --- a/gcc/tree-vectorizer.h
>> +++ b/gcc/tree-vectorizer.h
>> @@ -28,6 +28,7 @@ typedef class _stmt_vec_info *stmt_vec_info;
>> #include "target.h"
>> #include "internal-fn.h"
>> #include "tree-ssa-operands.h"
>> +#include "gimple-match.h"
>>
>> /* Used for naming of new temporaries. */
>> enum vect_var_kind {
>> @@ -1192,7 +1193,7 @@ public:
>> enum vect_reduction_type reduc_type;
>>
>> /* The original reduction code, to be used in the epilogue. */
>> - enum tree_code reduc_code;
>> + code_helper reduc_code;
>> /* An internal function we should use in the epilogue. */
>> internal_fn reduc_fn;
>>
>> @@ -2151,7 +2152,7 @@ extern tree vect_create_addr_base_for_vector_ref (vec_info *,
>> tree);
>>
>> /* In tree-vect-loop.c. */
>> -extern tree neutral_op_for_reduction (tree, tree_code, tree);
>> +extern tree neutral_op_for_reduction (tree, code_helper, tree);
>> extern widest_int vect_iv_limit_for_partial_vectors (loop_vec_info loop_vinfo);
>> bool vect_rgroup_iv_might_wrap_p (loop_vec_info, rgroup_controls *);
>> /* Used in tree-vect-loop-manip.c */
>> @@ -2160,7 +2161,7 @@ extern opt_result vect_determine_partial_vectors_and_peeling (loop_vec_info,
>> /* Used in gimple-loop-interchange.c and tree-parloops.c. */
>> extern bool check_reduction_path (dump_user_location_t, loop_p, gphi *, tree,
>> enum tree_code);
>> -extern bool needs_fold_left_reduction_p (tree, tree_code);
>> +extern bool needs_fold_left_reduction_p (tree, code_helper);
>> /* Drive for loop analysis stage. */
>> extern opt_loop_vec_info vect_analyze_loop (class loop *, vec_info_shared *);
>> extern tree vect_build_loop_niters (loop_vec_info, bool * = NULL);
>> @@ -2178,7 +2179,7 @@ extern tree vect_get_loop_len (loop_vec_info, vec_loop_lens *, unsigned int,
>> unsigned int);
>> extern gimple_seq vect_gen_len (tree, tree, tree, tree);
>> extern stmt_vec_info info_for_reduction (vec_info *, stmt_vec_info);
>> -extern bool reduction_fn_for_scalar_code (enum tree_code, internal_fn *);
>> +extern bool reduction_fn_for_scalar_code (code_helper, internal_fn *);
>>
>> /* Drive for loop transformation stage. */
>> extern class loop *vect_transform_loop (loop_vec_info, gimple *);
>> @@ -2216,6 +2217,7 @@ extern bool vectorizable_phi (vec_info *, stmt_vec_info, gimple **, slp_tree,
>> stmt_vector_for_cost *);
>> extern bool vect_emulated_vector_p (tree);
>> extern bool vect_can_vectorize_without_simd_p (tree_code);
>> +extern bool vect_can_vectorize_without_simd_p (code_helper);
>> extern int vect_get_known_peeling_cost (loop_vec_info, int, int *,
>> stmt_vector_for_cost *,
>> stmt_vector_for_cost *,
>> --
>> 2.25.1
>>
More information about the Gcc-patches
mailing list