@c Copyright (C) 2014-2020 Free Software Foundation, Inc.
@c Free Software Foundation, Inc.
@c This is part of the GCC manual.
@c For copying conditions, see the file gcc.texi.
@node Match and Simplify
@chapter Match and Simplify
@cindex Match and Simplify
The GIMPLE and GENERIC pattern matching project match-and-simplify
tries to address several issues.
@enumerate
@item unify expression simplifications currently spread and duplicated
over separate files like fold-const.c, gimple-fold.c and builtins.c
@item allow for a cheap way to implement building and simplifying
non-trivial GIMPLE expressions, avoiding the need to go through
building and simplifying GENERIC via fold_buildN and then
gimplifying via force_gimple_operand
@end enumerate
To address these the project introduces a simple domain specific language
to write expression simplifications from which code targeting GIMPLE
and GENERIC is auto-generated. The GENERIC variant follows the
fold_buildN API while for the GIMPLE variant and to address 2) new
APIs are introduced.
@menu
* GIMPLE API::
* The Language::
@end menu
@node GIMPLE API
@section GIMPLE API
@cindex GIMPLE API
@deftypefn {GIMPLE function} tree gimple_simplify (enum tree_code, tree, tree, gimple_seq *, tree (*)(tree))
@deftypefnx {GIMPLE function} tree gimple_simplify (enum tree_code, tree, tree, tree, gimple_seq *, tree (*)(tree))
@deftypefnx {GIMPLE function} tree gimple_simplify (enum tree_code, tree, tree, tree, tree, gimple_seq *, tree (*)(tree))
@deftypefnx {GIMPLE function} tree gimple_simplify (enum built_in_function, tree, tree, gimple_seq *, tree (*)(tree))
@deftypefnx {GIMPLE function} tree gimple_simplify (enum built_in_function, tree, tree, tree, gimple_seq *, tree (*)(tree))
@deftypefnx {GIMPLE function} tree gimple_simplify (enum built_in_function, tree, tree, tree, tree, gimple_seq *, tree (*)(tree))
The main GIMPLE API entry to the expression simplifications mimicing
that of the GENERIC fold_@{unary,binary,ternary@} functions.
@end deftypefn
thus providing n-ary overloads for operation or function. The
additional arguments are a gimple_seq where built statements are
inserted on (if @code{NULL} then simplifications requiring new statements
are not performed) and a valueization hook that can be used to
tie simplifications to a SSA lattice.
In addition to those APIs @code{fold_stmt} is overloaded with
a valueization hook:
@deftypefn bool fold_stmt (gimple_stmt_iterator *, tree (*)(tree));
@end deftypefn
Ontop of these a @code{fold_buildN}-like API for GIMPLE is introduced:
@deftypefn {GIMPLE function} tree gimple_build (gimple_seq *, location_t, enum tree_code, tree, tree, tree (*valueize) (tree) = NULL);
@deftypefnx {GIMPLE function} tree gimple_build (gimple_seq *, location_t, enum tree_code, tree, tree, tree, tree (*valueize) (tree) = NULL);
@deftypefnx {GIMPLE function} tree gimple_build (gimple_seq *, location_t, enum tree_code, tree, tree, tree, tree, tree (*valueize) (tree) = NULL);
@deftypefnx {GIMPLE function} tree gimple_build (gimple_seq *, location_t, enum built_in_function, tree, tree, tree (*valueize) (tree) = NULL);
@deftypefnx {GIMPLE function} tree gimple_build (gimple_seq *, location_t, enum built_in_function, tree, tree, tree, tree (*valueize) (tree) = NULL);
@deftypefnx {GIMPLE function} tree gimple_build (gimple_seq *, location_t, enum built_in_function, tree, tree, tree, tree, tree (*valueize) (tree) = NULL);
@deftypefnx {GIMPLE function} tree gimple_convert (gimple_seq *, location_t, tree, tree);
@end deftypefn
which is supposed to replace @code{force_gimple_operand (fold_buildN (...), ...)}
and calls to @code{fold_convert}. Overloads without the @code{location_t}
argument exist. Built statements are inserted on the provided sequence
and simplification is performed using the optional valueization hook.
@node The Language
@section The Language
@cindex The Language
The language to write expression simplifications in resembles other
domain-specific languages GCC uses. Thus it is lispy. Lets start
with an example from the match.pd file:
@smallexample
(simplify
(bit_and @@0 integer_all_onesp)
@@0)
@end smallexample
This example contains all required parts of an expression simplification.
A simplification is wrapped inside a @code{(simplify ...)} expression.
That contains at least two operands - an expression that is matched
with the GIMPLE or GENERIC IL and a replacement expression that is
returned if the match was successful.
Expressions have an operator ID, @code{bit_and} in this case. Expressions can
be lower-case tree codes with @code{_expr} stripped off or builtin
function code names in all-caps, like @code{BUILT_IN_SQRT}.
@code{@@n} denotes a so-called capture. It captures the operand and lets
you refer to it in other places of the match-and-simplify. In the
above example it is refered to in the replacement expression. Captures
are @code{@@} followed by a number or an identifier.
@smallexample
(simplify
(bit_xor @@0 @@0)
@{ build_zero_cst (type); @})
@end smallexample
In this example @code{@@0} is mentioned twice which constrains the matched
expression to have two equal operands. Usually matches are constraint
to equal types. If operands may be constants and conversions are involved
matching by value might be preferred in which case use @code{@@@@0} to
denote a by value match and the specific operand you want to refer to
in the result part. This example also introduces
operands written in C code. These can be used in the expression
replacements and are supposed to evaluate to a tree node which has to
be a valid GIMPLE operand (so you cannot generate expressions in C code).
@smallexample
(simplify
(trunc_mod integer_zerop@@0 @@1)
(if (!integer_zerop (@@1))
@@0))
@end smallexample
Here @code{@@0} captures the first operand of the trunc_mod expression
which is also predicated with @code{integer_zerop}. Expression operands
may be either expressions, predicates or captures. Captures
can be unconstrained or capture expresions or predicates.
This example introduces an optional operand of simplify,
the if-expression. This condition is evaluated after the
expression matched in the IL and is required to evaluate to true
to enable the replacement expression in the second operand
position. The expression operand of the @code{if} is a standard C
expression which may contain references to captures. The @code{if}
has an optional third operand which may contain the replacement
expression that is enabled when the condition evaluates to false.
A @code{if} expression can be used to specify a common condition
for multiple simplify patterns, avoiding the need
to repeat that multiple times:
@smallexample
(if (!TYPE_SATURATING (type)
&& !FLOAT_TYPE_P (type) && !FIXED_POINT_TYPE_P (type))
(simplify
(minus (plus @@0 @@1) @@0)
@@1)
(simplify
(minus (minus @@0 @@1) @@0)
(negate @@1)))
@end smallexample
Note that @code{if}s in outer position do not have the optional
else clause but instead have multiple then clauses.
Ifs can be nested.
There exists a @code{switch} expression which can be used to
chain conditions avoiding nesting @code{if}s too much:
@smallexample
(simplify
(simple_comparison @@0 REAL_CST@@1)
(switch
/* a CMP (-0) -> a CMP 0 */
(if (REAL_VALUE_MINUS_ZERO (TREE_REAL_CST (@@1)))
(cmp @@0 @{ build_real (TREE_TYPE (@@1), dconst0); @}))
/* x != NaN is always true, other ops are always false. */
(if (REAL_VALUE_ISNAN (TREE_REAL_CST (@@1))
&& ! HONOR_SNANS (@@1))
@{ constant_boolean_node (cmp == NE_EXPR, type); @})))
@end smallexample
Is equal to
@smallexample
(simplify
(simple_comparison @@0 REAL_CST@@1)
(switch
/* a CMP (-0) -> a CMP 0 */
(if (REAL_VALUE_MINUS_ZERO (TREE_REAL_CST (@@1)))
(cmp @@0 @{ build_real (TREE_TYPE (@@1), dconst0); @})
/* x != NaN is always true, other ops are always false. */
(if (REAL_VALUE_ISNAN (TREE_REAL_CST (@@1))
&& ! HONOR_SNANS (@@1))
@{ constant_boolean_node (cmp == NE_EXPR, type); @}))))
@end smallexample
which has the second @code{if} in the else operand of the first.
The @code{switch} expression takes @code{if} expressions as
operands (which may not have else clauses) and as a last operand
a replacement expression which should be enabled by default if
no other condition evaluated to true.
Captures can also be used for capturing results of sub-expressions.
@smallexample
#if GIMPLE
(simplify
(pointer_plus (addr@@2 @@0) INTEGER_CST_P@@1)
(if (is_gimple_min_invariant (@@2)))
@{
poly_int64 off;
tree base = get_addr_base_and_unit_offset (@@0, &off);
off += tree_to_uhwi (@@1);
/* Now with that we should be able to simply write
(addr (mem_ref (addr @@base) (plus @@off @@1))) */
build1 (ADDR_EXPR, type,
build2 (MEM_REF, TREE_TYPE (TREE_TYPE (@@2)),
build_fold_addr_expr (base),
build_int_cst (ptr_type_node, off)));
@})
#endif
@end smallexample
In the above example, @code{@@2} captures the result of the expression
@code{(addr @@0)}. For outermost expression only its type can be captured,
and the keyword @code{type} is reserved for this purpose. The above
example also gives a way to conditionalize patterns to only apply
to @code{GIMPLE} or @code{GENERIC} by means of using the pre-defined
preprocessor macros @code{GIMPLE} and @code{GENERIC} and using
preprocessor directives.
@smallexample
(simplify
(bit_and:c integral_op_p@@0 (bit_ior:c (bit_not @@0) @@1))
(bit_and @@1 @@0))
@end smallexample
Here we introduce flags on match expressions. The flag used
above, @code{c}, denotes that the expression should
be also matched commutated. Thus the above match expression
is really the following four match expressions:
@smallexample
(bit_and integral_op_p@@0 (bit_ior (bit_not @@0) @@1))
(bit_and (bit_ior (bit_not @@0) @@1) integral_op_p@@0)
(bit_and integral_op_p@@0 (bit_ior @@1 (bit_not @@0)))
(bit_and (bit_ior @@1 (bit_not @@0)) integral_op_p@@0)
@end smallexample
Usual canonicalizations you know from GENERIC expressions are
applied before matching, so for example constant operands always
come second in commutative expressions.
The second supported flag is @code{s} which tells the code
generator to fail the pattern if the expression marked with
@code{s} does have more than one use and the simplification
results in an expression with more than one operator.
For example in
@smallexample
(simplify
(pointer_plus (pointer_plus:s @@0 @@1) @@3)
(pointer_plus @@0 (plus @@1 @@3)))
@end smallexample
this avoids the association if @code{(pointer_plus @@0 @@1)} is
used outside of the matched expression and thus it would stay
live and not trivially removed by dead code elimination.
Now consider @code{((x + 3) + -3)} with the temporary
holding @code{(x + 3)} used elsewhere. This simplifies down
to @code{x} which is desirable and thus flagging with @code{s}
does not prevent the transform. Now consider @code{((x + 3) + 1)}
which simplifies to @code{(x + 4)}. Despite being flagged with
@code{s} the simplification will be performed. The
simplification of @code{((x + a) + 1)} to @code{(x + (a + 1))} will
not performed in this case though.
More features exist to avoid too much repetition.
@smallexample
(for op (plus pointer_plus minus bit_ior bit_xor)
(simplify
(op @@0 integer_zerop)
@@0))
@end smallexample
A @code{for} expression can be used to repeat a pattern for each
operator specified, substituting @code{op}. @code{for} can be
nested and a @code{for} can have multiple operators to iterate.
@smallexample
(for opa (plus minus)
opb (minus plus)
(for opc (plus minus)
(simplify...
@end smallexample
In this example the pattern will be repeated four times with
@code{opa, opb, opc} being @code{plus, minus, plus};
@code{plus, minus, minus}; @code{minus, plus, plus};
@code{minus, plus, minus}.
To avoid repeating operator lists in @code{for} you can name
them via
@smallexample
(define_operator_list pmm plus minus mult)
@end smallexample
and use them in @code{for} operator lists where they get expanded.
@smallexample
(for opa (pmm trunc_div)
(simplify...
@end smallexample
So this example iterates over @code{plus}, @code{minus}, @code{mult}
and @code{trunc_div}.
Using operator lists can also remove the need to explicitely write
a @code{for}. All operator list uses that appear in a @code{simplify}
or @code{match} pattern in operator positions will implicitely
be added to a new @code{for}. For example
@smallexample
(define_operator_list SQRT BUILT_IN_SQRTF BUILT_IN_SQRT BUILT_IN_SQRTL)
(define_operator_list POW BUILT_IN_POWF BUILT_IN_POW BUILT_IN_POWL)
(simplify
(SQRT (POW @@0 @@1))
(POW (abs @@0) (mult @@1 @{ built_real (TREE_TYPE (@@1), dconsthalf); @})))
@end smallexample
is the same as
@smallexample
(for SQRT (BUILT_IN_SQRTF BUILT_IN_SQRT BUILT_IN_SQRTL)
POW (BUILT_IN_POWF BUILT_IN_POW BUILT_IN_POWL)
(simplify
(SQRT (POW @@0 @@1))
(POW (abs @@0) (mult @@1 @{ built_real (TREE_TYPE (@@1), dconsthalf); @}))))
@end smallexample
@code{for}s and operator lists can include the special identifier
@code{null} that matches nothing and can never be generated. This can
be used to pad an operator list so that it has a standard form,
even if there isn't a suitable operator for every form.
Another building block are @code{with} expressions in the
result expression which nest the generated code in a new C block
followed by its argument:
@smallexample
(simplify
(convert (mult @@0 @@1))
(with @{ tree utype = unsigned_type_for (type); @}
(convert (mult (convert:utype @@0) (convert:utype @@1)))))
@end smallexample
This allows code nested in the @code{with} to refer to the declared
variables. In the above case we use the feature to specify the
type of a generated expression with the @code{:type} syntax where
@code{type} needs to be an identifier that refers to the desired type.
Usually the types of the generated result expressions are
determined from the context, but sometimes like in the above case
it is required that you specify them explicitely.
As intermediate conversions are often optional there is a way to
avoid the need to repeat patterns both with and without such
conversions. Namely you can mark a conversion as being optional
with a @code{?}:
@smallexample
(simplify
(eq (convert@@0 @@1) (convert@? @@2))
(eq @@1 (convert @@2)))
@end smallexample
which will match both @code{(eq (convert @@1) (convert @@2))} and
@code{(eq (convert @@1) @@2)}. The optional converts are supposed
to be all either present or not, thus
@code{(eq (convert@? @@1) (convert@? @@2))} will result in two
patterns only. If you want to match all four combinations you
have access to two additional conditional converts as in
@code{(eq (convert1@? @@1) (convert2@? @@2))}.
The support for @code{?} marking extends to all unary operations
including predicates you declare yourself with @code{match}.
Predicates available from the GCC middle-end need to be made
available explicitely via @code{define_predicates}:
@smallexample
(define_predicates
integer_onep integer_zerop integer_all_onesp)
@end smallexample
You can also define predicates using the pattern matching language
and the @code{match} form:
@smallexample
(match negate_expr_p
INTEGER_CST
(if (TYPE_OVERFLOW_WRAPS (type)
|| may_negate_without_overflow_p (t))))
(match negate_expr_p
(negate @@0))
@end smallexample
This shows that for @code{match} expressions there is @code{t}
available which captures the outermost expression (something
not possible in the @code{simplify} context). As you can see
@code{match} has an identifier as first operand which is how
you refer to the predicate in patterns. Multiple @code{match}
for the same identifier add additional cases where the predicate
matches.
Predicates can also match an expression in which case you need
to provide a template specifying the identifier and where to
get its operands from:
@smallexample
(match (logical_inverted_value @@0)
(eq @@0 integer_zerop))
(match (logical_inverted_value @@0)
(bit_not truth_valued_p@@0))
@end smallexample
You can use the above predicate like
@smallexample
(simplify
(bit_and @@0 (logical_inverted_value @@0))
@{ build_zero_cst (type); @})
@end smallexample
Which will match a bitwise and of an operand with its logical
inverted value.