#include "config.h"#include "system.h"#include "coretypes.h"#include "tm.h"#include "flags.h"#include "tree.h"#include "realmpfr.h"#include "rtl.h"#include "expr.h"#include "tm_p.h"#include "target.h"#include "diagnostic-core.h"#include "intl.h"#include "ggc.h"#include "hash-table.h"#include "langhooks.h"#include "md5.h"#include "gimple.h"#include "tree-dfa.h"
Macros | |
| #define | OP_SAME(N) |
| #define | OP_SAME_WITH_NULL(N) |
| #define | LOGICAL_OP_NON_SHORT_CIRCUIT |
| #define | START_FOLD_INIT |
| #define | END_FOLD_INIT |
Enumerations | |
| enum | comparison_code { COMPCODE_FALSE = 0, COMPCODE_LT = 1, COMPCODE_EQ = 2, COMPCODE_LE = 3, COMPCODE_GT = 4, COMPCODE_LTGT = 5, COMPCODE_GE = 6, COMPCODE_ORD = 7, COMPCODE_UNORD = 8, COMPCODE_UNLT = 9, COMPCODE_UNEQ = 10, COMPCODE_UNLE = 11, COMPCODE_UNGT = 12, COMPCODE_NE = 13, COMPCODE_UNGE = 14, COMPCODE_TRUE = 15 } |
Variables | |
| int | folding_initializer = 0 |
| static int | fold_deferring_overflow_warnings |
| static const char * | fold_deferred_overflow_warning |
| static enum warn_strict_overflow_code | fold_deferred_overflow_code |
| int | pedantic_lvalues |
Macro Definition Documentation
| #define END_FOLD_INIT |
| #define LOGICAL_OP_NON_SHORT_CIRCUIT |
| #define OP_SAME | ( | N | ) |
| #define OP_SAME_WITH_NULL | ( | N | ) |
| #define START_FOLD_INIT |
Perform constant folding and related simplification of initializer expression EXPR. These behave identically to "fold_buildN" but ignore potential run-time traps and exceptions that fold must preserve.
Enumeration Type Documentation
| enum comparison_code |
The following constants represent a bit based encoding of GCC's comparison operators. This encoding simplifies transformations on relational comparison operators, such as AND and OR.
Function Documentation
|
static |
|
static |
Return nonzero if MASK represents a mask of SIZE ones in the low-order bit positions.
|
static |
|
static |
Re-associate trees split by the above function. T1 and T2 are either expressions to associate or null. Return the new expression, if any. LOC is the location of the new expression. If we build an operation, do it in TYPE and with CODE.
If either input is CODE, a PLUS_EXPR, or a MINUS_EXPR, don't try to fold this since we will have infinite recursion. But do deal with any NEGATE_EXPRs.
| tree build_fold_addr_expr_loc | ( | ) |
Build an expression for the address of T.
References fold_build2_loc, fold_convert_loc(), negate_expr(), TREE_OPERAND, and TREE_TYPE.
Referenced by initialize_argument_information(), lower_reduction_clauses(), and native_interpret_real().
| tree build_fold_addr_expr_with_type_loc | ( | ) |
Build an expression for the address of T. Folds away INDIRECT_REF to avoid confusing the gimplify process.
The size of the object is not relevant when talking about its address.
| tree build_fold_indirect_ref_loc | ( | ) |
Builds an expression for an indirection through T, simplifying some cases.
Given a range, LOW, HIGH, and IN_P, an expression, EXP, and a result type, TYPE, return an expression to test if EXP is in (or out of, depending on IN_P) the range. Return 0 if the test couldn't be created.
Optimize (c>=1) && (c<=127) into (signed char)c > 0.
Optimize (c>=low) && (c<=high) into (c-low>=0) && (c-low<=high-low). This requires wrap-around arithmetics for the type of the expression. First make sure that arithmetics in this type is valid, then make sure that it wraps around.
Check if (unsigned) INT_MAX + 1 == (unsigned) INT_MIN
for the type in question, as we rely on this here.
|
static |
Construct a vector of zero elements of vector type TYPE.
|
static |
Returns true if we can interpret the contents of a native encoding as TYPE.
| tree combine_comparisons | ( | location_t | loc, |
| enum tree_code | code, | ||
| enum tree_code | lcode, | ||
| enum tree_code | rcode, | ||
| tree | truth_type, | ||
| tree | ll_arg, | ||
| tree | lr_arg | ||
| ) |
Return a tree for the comparison which is the combination of doing the AND or OR (depending on CODE) of the two operations LCODE and RCODE on the identical operands LL_ARG and LR_ARG. Take into account the possibility of trapping if the mode has NaNs, and return NULL_TREE if this makes the transformation invalid.
Eliminate unordered comparisons, as well as LTGT and ORD
which are not used unless the mode has NaNs.
Check that the original operation and the optimized ones will trap
under the same condition. In a short-circuited boolean expression the LHS might be
such that the RHS, if evaluated, will never trap. For
example, in ORD (x, y) && (x < y), we evaluate the RHS only
if neither x nor y is NaN. (This is a mixed blessing: for
example, the expression above will never trap, hence
optimizing it to x < y would be invalid). If the comparison was short-circuited, and only the RHS
trapped, we may now generate a spurious trap. If we changed the conditions that cause a trap, we lose.
Referenced by fold_range_test().
|
static |
Referenced by swap_tree_comparison().
|
static |
Convert a comparison tree code from an enum tree_code representation into a compcode bit-based encoding. This function is the inverse of compcode_to_comparison.
|
static |
|
static |
Convert a compcode bit-based encoding of a comparison operator back to GCC's enum tree_code representation. This function is the inverse of comparison_to_compcode.
References error_mark_node, TREE_CODE, tree_int_cst_equal(), and TREE_TYPE.
Referenced by distribute_real_division(), and extract_muldiv().
|
static |
Combine two constants ARG1 and ARG2 under operation CODE to produce a new constant. We assume ARG1 and ARG2 have the same data type, or at least are the same kind of constant and the same machine mode. Return zero if combining the constants is not allowed in the current operating mode.
Sanity check for the recursive cases.
The following codes are handled by real_arithmetic.
Don't perform operation if we honor signaling NaNs and
either operand is a NaN. Don't perform operation if it would raise a division
by zero exception. If either operand is a NaN, just return it. Otherwise, set up
for floating-point trap; we return an overflow. Don't constant fold this floating point operation if
the result has overflowed and flag_trapping_math. Don't constant fold this floating point operation if the
result may dependent upon the run-time rounding mode and
flag_rounding_math is set, or if GCC's software emulation
is unable to accurately represent the result. The following codes are handled by fixed_arithmetic.
Propagate overflow flags.
do_nonfinite=
do_nonfinite=
Fallthru ...
Keep this algorithm in sync with
tree-complex.c:expand_complex_div_straight().
Expand complex division to scalars, straightforward algorithm.
a / b = ((ar*br + ai*bi)/t) + i((ai*br - ar*bi)/t)
t = br*br + bi*bi Keep this algorithm in sync with
tree-complex.c:expand_complex_div_wide().
Expand complex division to scalars, modified algorithm to minimize
overflow with wide input ranges. In the TRUE branch, we compute
ratio = br/bi;
div = (br * ratio) + bi;
tr = (ar * ratio) + ai;
ti = (ai * ratio) - ar;
tr = tr / div;
ti = ti / div; In the FALSE branch, we compute
ratio = d/c;
divisor = (d * ratio) + c;
tr = (b * ratio) + a;
ti = b - (a * ratio);
tr = tr / div;
ti = ti / div; It is possible that const_binop cannot handle the given
code and return NULL_TREE Shifts allow a scalar offset for a vector.
The direction of VEC_[LR]SHIFT_EXPR is endian dependent.
For reductions, compiler emits VEC_RSHIFT_EXPR always,
for !BYTES_BIG_ENDIAN picks first vector element, but
for BYTES_BIG_ENDIAN last element from the vector. It is possible that const_binop cannot handle the given
code and return NULL_TREE
| tree constant_boolean_node | ( | ) |
Return a node which has the indicated constant VALUE (either 0 or 1 for scalars or {-1,-1,..} or {0,0,...} for vectors), and is of the indicated TYPE.
Referenced by distribute_real_division(), and fold_comparison().
|
static |
Callback for walk_tree, looking for LABEL_EXPR. Return *TP if it is a LABEL_EXPR; otherwise return NULL_TREE. Do not check the subtrees of GOTO_EXPR.
... fall through ...
|
static |
Return whether the sub-tree ST contains a label which is accessible from outside the sub-tree.
|
static |
Subroutine for fold_truth_andor_1: decode a field reference.
If EXP is a comparison reference, we return the innermost reference.
*PBITSIZE is set to the number of bits in the reference, *PBITPOS is set to the starting bit number.
If the innermost field can be completely contained in a mode-sized unit, *PMODE is set to that mode. Otherwise, it is set to VOIDmode.
*PVOLATILEP is set to 1 if the any expression encountered is volatile; otherwise it is not changed.
*PUNSIGNEDP is set to the signedness of the field.
*PMASK is set to the mask used. This is either contained in a BIT_AND_EXPR or derived from the width of the field.
*PAND_MASK is set to the mask found in a BIT_AND_EXPR, if any.
Return 0 if this is not a component reference or is one that we can't do anything with.
All the optimizations using this function assume integer fields. There are problems with FP fields since the type_for_size call below can fail for, e.g., XFmode.
We are interested in the bare arrangement of bits, so strip everything that doesn't affect the machine mode. However, record the type of the outermost expression if it may matter below.
If the number of bits in the reference is the same as the bitsize of the outer type, then the outer type gives the signedness. Otherwise (in case of a small bitfield) the signedness is unchanged.
Compute the mask to access the bitfield.
Merge it with the mask we found in the BIT_AND_EXPR, if any.
References fold_build2, fold_convert, STRIP_NOPS, tcc_comparison, TREE_CODE, TREE_CODE_CLASS, and TREE_TYPE.
|
static |
Given a bit-wise operation CODE applied to ARG0 and ARG1, see if both operands are another bit-wise operation with a common input. If so, distribute the bit operations to save an operation and possibly two if constants are involved. For example, convert (A | B) & (A | C) into A | (B & C) Further simplification will occur if B and C are constants.
If this optimization cannot be done, 0 will be returned.
|
static |
Knowing that ARG0 and ARG1 are both RDIV_EXPRs, simplify a binary operation with code CODE. This optimization is unsafe.
(A / C) +- (B / C) -> (A +- B) / C.
(A / C1) +- (A / C2) -> A * (1 / C1 +- 1 / C2).
References const_binop(), constant_boolean_node(), fold_convert_loc(), integer_zerop(), size_int, and warning().
| tree div_if_zero_remainder | ( | ) |
If ARG2 divides ARG1 with zero remainder, carries out the division of type CODE and returns the quotient. Otherwise returns NULL_TREE.
The sign of the division is according to operand two, that does the correct thing for POINTER_PLUS_EXPR where we want a signed division.
|
static |
ARG is a tree that is known to contain just arithmetic operations and comparisons. Evaluate the operations in the tree substituting NEW0 for any occurrence of OLD0 as an operand of a comparison and likewise for NEW1 and OLD1.
We can handle some of the tcc_expression cases here.
Fall through - ???
We need to check both for exact equality and tree equality. The
former will be true if the operand has a side-effect. In that
case, we know the operand occurred exactly once.
Referenced by operand_equal_for_comparison_p().
|
static |
If the real or vector real constant CST of type TYPE has an exact inverse, return it, else return NULL.
References build2_loc, fold_convert_loc(), and TREE_OPERAND.
|
static |
Return EXPR_LOCATION of T if it is not UNKNOWN_LOCATION. Otherwise, return LOC.
References CAN_HAVE_LOCATION_P, EXPR_LOCATION, and TREE_CODE.
|
static |
T is an integer expression that is being multiplied, divided, or taken a modulus (CODE says which and what kind of divide or modulus) by a constant C. See if we can eliminate that operation by folding it with other operations already in T. WIDE_TYPE, if non-null, is a type that should be used for the computation if wider than our type.
For example, if we are dividing (X * 8) + (Y * 16) by 4, we can return (X * 2) + (Y * 4). We must, however, be assured that either the original expression would not overflow or that overflow is undefined for the type in the language in question.
If we return a non-null expression, it is an equivalent form of the original computation, but need not be in the original type.
We set *STRICT_OVERFLOW_P to true if the return values depends on signed overflow being undefined. Otherwise we do not change *STRICT_OVERFLOW_P.
To avoid exponential search depth, refuse to allow recursion past three levels. Beyond that (1) it's highly unlikely that we'll find something interesting and (2) we've probably processed it before when we built the inner expression.
References const_binop(), fold_build2, fold_convert, integer_zerop(), and TYPE_OVERFLOW_UNDEFINED.
|
static |
Don't deal with constants of zero here; they confuse the code below.
Note that we need not handle conditional operations here since fold already handles those cases. So just do arithmetic here.
For a constant, we can always simplify if we are a multiply
or (for divide and modulus) if it is a multiple of our constant. If op0 is an expression ...
... and has wrapping overflow, and its type is smaller
than ctype, then we cannot pass through as widening. ... or this is a truncation (t is narrower than op0),
then we cannot pass through this narrowing. ... or signedness changes for division or modulus,
then we cannot pass through this conversion. ... or has undefined overflow while the converted to
type has not, we cannot do the operation in the inner type
as that would introduce undefined overflow. Pass the constant down and see if we can make a simplification. If
we can, replace this expression with the inner simplification for
possible later conversion to our or some other type. If widening the type changes it from signed to unsigned, then we
must avoid building ABS_EXPR itself as unsigned. If the constant is negative, we cannot simplify this.
FALLTHROUGH
For division and modulus, type can't be unsigned, as e.g.
(-(x / 2U)) / 2U isn't equal to -((x / 2U) / 2U) for x >= 2.
For signed types, even with wrapping overflow, this is fine. If widening the type changes the signedness, then we can't perform
this optimization as that changes the result. MIN (a, b) / 5 -> MIN (a / 5, b / 5)
If the second operand is constant, this is a multiplication
or floor division, by a power of two, so we can treat it that
way unless the multiplier or divisor overflows. Signed
left-shift overflow is implementation-defined rather than
undefined in C90, so do not convert signed left shift into
multiplication. const_binop may not detect overflow correctly,
so check for it explicitly here. See if we can eliminate the operation on both sides. If we can, we
can return a new PLUS or MINUS. If we can't, the only remaining
cases where we can do anything are if the second operand is a
constant. If not multiplication, we can only do this if both operands
are divisible by c. If this was a subtraction, negate OP1 and set it to be an addition.
This simplifies the logic below. If OP1 was not easily negatable, the constant may be OP0.
If either OP1 or C are negative, this optimization is not safe for
some of the division and remainder types while for others we need
to change the code. If it's a multiply or a division/modulus operation of a multiple
of our constant, do the operation and verify it doesn't overflow. We allow the constant to overflow with wrapping semantics.
If we have an unsigned type, we cannot widen the operation since it
will change the result if the original computation overflowed. If we were able to eliminate our operation from the first side,
apply our operation to the second side and reform the PLUS. The last case is if we are a multiply. In that case, we can
apply the distributive law to commute the multiply and addition
if the multiplication of the constants doesn't overflow
and overflow is defined. With undefined overflow
op0 * c might overflow, while (op0 + orig_op1) * c doesn't. We have a special case here if we are doing something like
(C * 8) % 4 since we know that's zero. If the multiplication can overflow we cannot optimize this.
... fall through ...
If we can extract our operation from the LHS, do so and return a
new operation. Likewise for the RHS from a MULT_EXPR. Otherwise,
do something only if the second operand is a constant. If these are the same operation types, we can associate them
assuming no overflow. If these operations "cancel" each other, we have the main
optimizations of this pass, which occur when either constant is a
multiple of the other, in which case we replace this with either an
operation or CODE or TCODE.
If we have an unsigned type, we cannot do this since it will change
the result if the original computation overflowed.
| tree fold | ( | ) |
Perform constant folding and related simplification of EXPR. The related simplifications include x*1 => x, x*0 => 0, etc., and application of the associative law. NOP_EXPR conversions may be removed freely (as long as we are careful not to change the type of the overall expression). We cannot simplify through a CONVERT_EXPR, FIX_EXPR or FLOAT_EXPR, but we can constant-fold them if they have constant operands.
Return right away if a constant.
CALL_EXPR-like objects with variable numbers of operands are treated specially.
Find a matching index by means of a binary search.
Return a VECTOR_CST if possible.
Referenced by convert_to_real(), expand_expr_real_1(), and process_assert_insertions_for().
| tree fold_abs_const | ( | ) |
Return the tree for abs (ARG0) when ARG0 is known to be either an integer constant or real constant.
TYPE is the type of the result.
If the value is unsigned or non-negative, then the absolute value is the same as the ordinary value.
If the value is negative, then the absolute value is
its negation.
|
static |
Try to fold a pointer difference of type TYPE two address expressions of array references AREF0 and AREF1 using location LOC. Return a simplified expression for the difference or NULL_TREE.
If the bases are array references as well, recurse. If the bases are pointer indirections compute the difference of the pointers. If the bases are equal, we are set.
Fold a binary expression of code CODE and type TYPE with operands OP0 and OP1. LOC is the location of the resulting expression. Return the folded expression if folding is successful. Otherwise, return NULL_TREE.
Strip any conversions that don't change the mode. This is safe for every expression, except for a comparison expression because its signedness is derived from its operands. So, in the latter case, only strip conversions that don't change the signedness. MIN_EXPR/MAX_EXPR also need signedness of arguments preserved. Note that this is done as an internal manipulation within the constant folder, in order to find the simplest representation of the arguments so that their form can be studied. In any cases, the appropriate type conversions should be put back in the tree that will get out of the constant folder.
Note that TREE_CONSTANT isn't enough: static var addresses are constant but we can't do arithmetic on them.
Make sure type and arg0 have the same saturating flag.
If this is a commutative operation, and ARG0 is a constant, move it to ARG1 to reduce the number of tests below.
ARG0 is the first operand of EXPR, and ARG1 is the second operand.
First check for cases where an arithmetic operation is applied to a
compound, conditional, or comparison operation. Push the arithmetic
operation inside the compound or conditional to see if any folding
can then be done. Convert comparison to conditional for this purpose.
The also optimizes non-constant cases that used to be done in
expand_expr.
Before we do that, see if this is a BIT_AND_EXPR or a BIT_IOR_EXPR,
one of the operands is a comparison and the other is a comparison, a
BIT_AND_EXPR with the constant 1, or a truth value. In that case, the
code below would make the expression more complex. Change it to a
TRUTH_{AND,OR}_EXPR. Likewise, convert a similar NE_EXPR to
TRUTH_XOR_EXPR and an EQ_EXPR to the inversion of a TRUTH_XOR_EXPR. MEM[&MEM[p, CST1], CST2] -> MEM[p, CST1 + CST2].
MEM[&a.b, CST2] -> MEM[&a, offsetof (a, b) + CST2].
0 +p index -> (type)index
PTR +p 0 -> PTR
INT +p INT -> (PTR)(INT + INT). Stripping types allows for this.
(PTR +p B) +p A -> PTR +p (B + A)
PTR_CST +p CST -> CST1
Try replacing &a[i1] +p c * i2 with &a[i1 + i2], if c is step
of the array. Loop optimizer sometimes produce this type of
expressions. A + (-B) -> A - B
(-A) + B -> B - A
Convert ~A + 1 to -A.
~X + X is -1.
X + ~X is -1.
X + (X / CST) * -CST is X % CST.
Handle (A1 * C1) + (A2 * C2) with A1, A2 or C1, C2 being the same or
one. Make sure the type is not saturating and has the signedness of
the stripped operands, as fold_plusminus_mult_expr will re-associate.
??? The latter condition should use TYPE_OVERFLOW_* flags instead. If we are adding two BIT_AND_EXPR's, both of which are and'ing
with a constant, and the two constants have no bits in common,
we should treat this as a BIT_IOR_EXPR since this may produce more
simplifications. Reassociate (plus (plus (mult) (foo)) (mult)) as
(plus (plus (mult) (mult)) (foo)) so that we can
take advantage of the factoring cases below. See if ARG1 is zero and X + ARG1 reduces to X.
Likewise if the operands are reversed.
Convert X + -C into X - C.
Fold __complex__ ( x, 0 ) + __complex__ ( 0, y )
to __complex__ ( x, y ). This is not the same for SNaNs or
if signed zeros are involved. Convert x+x into x*2.0.
Convert a + (b*c + d*e) into (a + b*c) + d*e.
We associate floats only if the user has specified
-fassociative-math. Convert (b*c + d*e) + a into b*c + (d*e +a).
We associate floats only if the user has specified
-fassociative-math. (A << C1) + (A >> C2) if A is unsigned and C1+C2 is the size of A
is a rotate of A by C1 bits. (A << B) + (A >> (Z - B)) if A is unsigned and Z is the size of A
is a rotate of A by B bits. Only create rotates in complete modes. Other cases are not
expanded properly. In most languages, can't associate operations on floats through
parentheses. Rather than remember where the parentheses were, we
don't associate floats at all, unless the user has specified
-fassociative-math.
And, we need to make sure type is not saturating. Split both trees into variables, constants, and literals. Then
associate each group together, the constants with literals,
then the result with variables. This increases the chances of
literals being recombined later and of generating relocatable
expressions for the sum of a constant and literal. Recombine MINUS_EXPR operands by using PLUS_EXPR.
With undefined overflow prefer doing association in a type
which wraps on overflow, if that is one of the operand types. With undefined overflow we can only associate constants with one
variable, and constants whose association doesn't overflow. The only case we can still associate with two variables
is if they are the same, modulo negation and bit-pattern
preserving conversions. Only do something if we found more than two objects. Otherwise,
nothing has changed and we risk infinite recursion. Preserve the MINUS_EXPR if the negative part of the literal is
greater than the positive part. Otherwise, the multiplicative
folding code (i.e extract_muldiv) may be fooled in case
unsigned constants are subtracted, like in the following
example: ((X*2 + 4) - 8U)/2. Don't introduce overflows through reassociation.
Pointer simplifications for subtraction, simple reassociations.
(PTR0 p+ A) - (PTR1 p+ B) -> (PTR0 - PTR1) + (A - B)
(PTR0 p+ A) - PTR1 -> (PTR0 - PTR1) + A, assuming PTR0 - PTR1 simplifies.
A - (-B) -> A + B
(-A) - B -> (-B) - A where B is easily negated and we can swap.
Convert -A - 1 to ~A.
Convert -1 - A to ~A.
X - (X / Y) * Y is X % Y.
Fold A - (A & B) into ~B & A.
Fold (A & ~B) - (A & B) into (A ^ B) - B, where B is
any power of 2 minus 1. See if ARG1 is zero and X - ARG1 reduces to X.
(ARG0 - ARG1) is the same as (-ARG1 + ARG0). So check whether
ARG0 is zero and X + ARG0 reduces to X, since that would mean
(-ARG1 + ARG0) reduces to -ARG1. Fold __complex__ ( x, 0 ) - __complex__ ( 0, y ) to
__complex__ ( x, -y ). This is not the same for SNaNs or if
signed zeros are involved. Fold &x - &x. This can happen from &x.foo - &x.
This is unsafe for certain floats even in non-IEEE formats.
In IEEE, it is unsafe because it does wrong for NaNs.
Also note that operand_equal_p is always false if an operand
is volatile. A - B -> A + (-B) if B is easily negatable.
Avoid this transformation if B is a positive REAL_CST.
Try folding difference of addresses.
Fold &a[i] - &a[j] to i-j.
Handle (A1 * C1) - (A2 * C2) with A1, A2 or C1, C2 being the same or
one. Make sure the type is not saturating and has the signedness of
the stripped operands, as fold_plusminus_mult_expr will re-associate.
??? The latter condition should use TYPE_OVERFLOW_* flags instead. (-A) * (-B) -> A * B
Transform x * -1 into -x. Make sure to do the negation
on the original operand with conversions not stripped
because we can only strip non-sign-changing conversions. Transform x * -C into -x * C if x is easily negatable.
(a * (1 << b)) is (a << b)
(A + A) * C -> A * 2 * C
((T) (X /[ex] C)) * C cancels out if the conversion is
sign-changing only. Optimize z * conj(z) for integer complex numbers.
Maybe fold x * 0 to 0. The expressions aren't the same
when x is NaN, since x * 0 is also NaN. Nor are they the
same in modes with signed zeros, since multiplying a
negative value by 0 gives -0, not +0. In IEEE floating point, x*1 is not equivalent to x for snans.
Likewise for complex arithmetic with signed zeros. Transform x * -1.0 into -x.
Convert (C1/X)*C2 into (C1*C2)/X. This transformation may change
the result for floating point types due to rounding so it is applied
only if -fassociative-math was specify. Strip sign operations from X in X*X, i.e. -Y*-Y -> Y*Y.
Fold z * +-I to __complex__ (-+__imag z, +-__real z).
This is not the same for NaNs or if signed zeros are
involved. Optimize z * conj(z) for floating point complex numbers.
Guarded by flag_unsafe_math_optimizations as non-finite
imaginary components don't produce scalar results. Optimizations of root(...)*root(...).
Optimize sqrt(x)*sqrt(x) as x.
Optimize root(x)*root(y) as root(x*y).
Optimize expN(x)*expN(y) as expN(x+y).
Optimizations of pow(...)*pow(...).
Optimize pow(x,y)*pow(z,y) as pow(x*z,y).
Optimize pow(x,y)*pow(x,z) as pow(x,y+z).
Optimize tan(x)*cos(x) as sin(x).
Optimize x*pow(x,c) as pow(x,c+1).
Optimize pow(x,c)*x as pow(x,c+1).
Canonicalize x*x as pow(x,2.0), which is expanded as x*x.
~X | X is -1.
X | ~X is -1.
Canonicalize (X & C1) | C2.
If (C1&C2) == C1, then (X&C1)|C2 becomes (X,C2).
If (C1|C2) == ~0 then (X&C1)|C2 becomes X|C2.
Minimize the number of bits set in C1, i.e. C1 := C1 & ~C2,
unless (C1 & ~C2) | (C2 & C3) for some C3 is a mask of some
mode which allows further optimizations. If X is a tree of the form (Y * K1) & K2, this might conflict
with that optimization from the BIT_AND_EXPR optimizations.
This could end up in an infinite recursion. (X & Y) | Y is (X, Y).
(X & Y) | X is (Y, X).
X | (X & Y) is (Y, X).
X | (Y & X) is (Y, X).
(X & ~Y) | (~X & Y) is X ^ Y
Convert (or (not arg0) (not arg1)) to (not (and (arg0) (arg1))).
This results in more efficient code for machines without a NAND
instruction. Combine will canonicalize to the first form
which will allow use of NAND instructions provided by the
backend if they exist. See if this can be simplified into a rotate first. If that
is unsuccessful continue in the association code. ~X ^ X is -1.
X ^ ~X is -1.
If we are XORing two BIT_AND_EXPR's, both of which are and'ing
with a constant, and the two constants have no bits in common,
we should treat this as a BIT_IOR_EXPR since this may produce more
simplifications. (X | Y) ^ X -> Y & ~ X
(Y | X) ^ X -> Y & ~ X
X ^ (X | Y) -> Y & ~ X
X ^ (Y | X) -> Y & ~ X
Convert ~X ^ ~Y to X ^ Y.
Convert ~X ^ C to X ^ ~C.
Fold (X & 1) ^ 1 as (X & 1) == 0.
Fold (X & Y) ^ Y as ~X & Y.
Fold (X & Y) ^ X as ~Y & X.
Fold X ^ (X & Y) as X & ~Y.
Fold X ^ (Y & X) as ~Y & X.
See if this can be simplified into a rotate first. If that
is unsuccessful continue in the association code. ~X & X, (X == 0) & X, and !X & X are always zero.
X & ~X , X & (X == 0), and X & !X are always zero.
Canonicalize (X | C1) & C2 as (X & C2) | (C1 & C2).
(X | Y) & Y is (X, Y).
(X | Y) & X is (Y, X).
X & (X | Y) is (Y, X).
X & (Y | X) is (Y, X).
Fold (X ^ 1) & 1 as (X & 1) == 0.
Fold ~X & 1 as (X & 1) == 0.
Fold !X & 1 as X == 0.
Fold (X ^ Y) & Y as ~X & Y.
Fold (X ^ Y) & X as ~Y & X.
Fold X & (X ^ Y) as X & ~Y.
Fold X & (Y ^ X) as ~Y & X.
Fold (X * Y) & -(1 << CST) to X * Y if Y is a constant
multiple of 1 << CST. Fold (X * CST1) & CST2 to zero if we can, or drop known zero
bits from CST2. For constants M and N, if M == (1LL << cst) - 1 && (N & M) == M,
((A & N) + B) & M -> (A + B) & M
Similarly if (N & M) == 0,
((A | N) + B) & M -> (A + B) & M
and for - instead of + (or unary - instead of +)
and/or ^ instead of |.
If B is constant and (B & M) == 0, fold into A & M. Now we know that arg0 is (C + D) or (C - D) or
-C and arg1 (M) is == (1LL << cst) - 1.
Store C into PMOP[0] and D into PMOP[1]. tree_low_cst not used, because we don't care about
the upper bits. If C or D is of the form (A & N) where
(N & M) == M, or of the form (A | N) or
(A ^ N) where (N & M) == 0, replace it with A. If C or D is a N where (N & M) == 0, it can be
omitted (assumed 0). Only build anything new if we optimized one or both arguments
above. Perform the operations in a type that has defined
overflow behavior. TEM is now the new binary +, - or unary - replacement.
Simplify ((int)c & 0377) into (int)c, if c is unsigned char.
Convert (and (not arg0) (not arg1)) to (not (or (arg0) (arg1))).
This results in more efficient code for machines without a NOR
instruction. Combine will canonicalize to the first form
which will allow use of NOR instructions provided by the
backend if they exist. If arg0 is derived from the address of an object or function, we may
be able to fold this expression using the object or function's
alignment. This works because modulus is a power of 2. If this weren't the
case, we'd have to replace it by its greatest power-of-2
divisor: modulus & -modulus. Fold (X << C1) & C2 into (X << C1) & (C2 | ((1 << C1) - 1))
(X >> C1) & C2 into (X >> C1) & (C2 | ~((type) -1 >> C1))
if the new mask might be further optimized. See if more bits can be proven as zero because of
zero extension. See if we can shorten the right shift.
For arithmetic shift if sign bit could be set, zerobits
can contain actually sign bits, so no transformation is
possible, unless MASK masks them all away. In that
case the shift needs to be converted into logical shift. ((X << 16) & 0xff00) is (X, 0).
Only do the transformation if NEWMASK is some integer
mode's mask. Don't touch a floating-point divide by zero unless the mode
of the constant can represent infinity. Optimize A / A to 1.0 if we don't care about
NaNs or Infinities. Skip the transformation
for non-real operands. The complex version of the above A / A optimization.
omit_two_operands will call fold_convert for us.
(-A) / (-B) -> A / B
In IEEE floating point, x/1 is not equivalent to x for snans.
In IEEE floating point, x/-1 is not equivalent to -x for snans.
If ARG1 is a constant, we can convert this to a multiply by the
reciprocal. This does not have the same rounding properties,
so only do this if -freciprocal-math. We can actually
always safely do it if ARG1 is a power of two, but it's hard to
tell if it is or not in a portable manner. Find the reciprocal if optimizing and the result is exact.
TODO: Complex reciprocal not implemented. Convert A/B/C to A/(B*C).
Convert A/(B/C) to (A/B)*C.
Convert C1/(X*C2) into (C1/C2)/X.
Optimize sin(x)/cos(x) as tan(x).
Optimize cos(x)/sin(x) as 1.0/tan(x).
Optimize sin(x)/tan(x) as cos(x) if we don't care about
NaNs or Infinities. Optimize tan(x)/sin(x) as 1.0/cos(x) if we don't care about
NaNs or Infinities. Optimize pow(x,c)/x as pow(x,c-1).
Optimize a/root(b/c) into a*root(c/b).
Optimize x/expN(y) into x*expN(-y).
Optimize x/pow(y,z) into x*pow(y,-z).
Optimize (X & (-A)) / A where A is a power of 2,
to X >> log2(A) Fall through
Simplify A / (B << N) where A and B are positive and B is
a power of 2, to A >> (N + log2(B)). For unsigned integral types, FLOOR_DIV_EXPR is the same as
TRUNC_DIV_EXPR. Rewrite into the latter in this case. Fall through
X / -1 is -X.
Convert -A / -B to A / B when the type is signed and overflow is
undefined. If arg0 is a multiple of arg1, then rewrite to the fastest div
operation, EXACT_DIV_EXPR.
Note that only CEIL_DIV_EXPR and FLOOR_DIV_EXPR are rewritten now.
At one time others generated faster code, it's not clear if they do
after the last round to changes to the DIV code in expmed.c. X % 1 is always zero, but be sure to preserve any side
effects in X. X % 0, return X % 0 unchanged so that we can get the
proper warnings and errors. 0 % X is always zero, but be sure to preserve any side
effects in X. Place this after checking for X == 0. X % -1 is zero.
X % -C is the same as X % C.
Avoid this transformation if C is INT_MIN, i.e. C == -C.
X % -Y is the same as X % Y.
Optimize TRUNC_MOD_EXPR by a power of two into a BIT_AND_EXPR,
i.e. "X % C" into "X & (C - 1)", if X and C are positive. Also optimize A % (C << N) where C is a power of 2,
to A & ((C << N) - 1). Optimize -1 >> x for arithmetic right shifts.
... fall through ...
Prefer vector1 << scalar to vector1 << vector2
if vector2 is uniform. Since negative shift count is not well-defined,
don't try to compute it in the compiler. Turn (a OP c1) OP c2 into a OP (c1+c2).
Deal with a OP (c1 + c2) being undefined but (a OP c1) OP c2
being well defined. Transform (x >> c) << c into x & (-1<<c), or transform (x << c) >> c
into x & ((unsigned)-1 >> c) for unsigned types. Rewrite an LROTATE_EXPR by a constant into an
RROTATE_EXPR by a new constant. If we have a rotate of a bit operation with the rotate count and
the second operand of the bit operation both constant,
permute the two operations. Two consecutive rotates adding up to the precision of the
type can be ignored. Fold (X & C2) << C1 into (X << C1) & (C2 << C1)
(X & C2) >> C1 into (X >> C1) & (C2 >> C1)
if the latter can be further optimized. Note that the operands of this must be ints
and their values must be 0 or 1.
("true" is a fixed value perhaps depending on the language.) If first arg is constant zero, return it.
If either arg is constant true, drop it.
Preserve sequence points.
If second arg is constant zero, result is zero, but first arg
must be evaluated. Likewise for first arg, but note that only the TRUTH_AND_EXPR
case will be handled here. !X && X is always false.
X && !X is always false.
A < X && A + 1 > Y ==> A < X && A >= Y. Normally A + 1 > Y
means A >= Y && A != MAX, but in this case we know that
A < X <= MAX. Note that the operands of this must be ints
and their values must be 0 or true.
("true" is a fixed value perhaps depending on the language.) If first arg is constant true, return it.
If either arg is constant zero, drop it.
Preserve sequence points.
If second arg is constant true, result is true, but we must
evaluate first arg. Likewise for first arg, but note this only occurs here for
TRUTH_OR_EXPR. !X || X is always true.
X || !X is always true.
(X && !Y) || (!X && Y) is X ^ Y
If the second arg is constant zero, drop it.
If the second arg is constant true, this is a logical inversion.
Identical arguments cancel to zero.
!X ^ X is always true.
X ^ !X is always true.
bool_var != 0 becomes bool_var.
bool_var == 1 becomes bool_var.
bool_var != 1 becomes !bool_var.
bool_var == 0 becomes !bool_var.
!exp != 0 becomes !exp
If this is an equality comparison of the address of two non-weak,
unaliased symbols neither of which are extern (since we do not
have access to attributes for externs), then we know the result. We know that we're looking at the address of two
non-weak, unaliased, static _DECL nodes.
It is both wasteful and incorrect to call operand_equal_p
to compare the two ADDR_EXPR nodes. It is wasteful in that
all we need to do is test pointer equality for the arguments
to the two ADDR_EXPR nodes. It is incorrect to use
operand_equal_p as that function is NOT equivalent to a
C equality test. It can in fact return false for two
objects which would test as equal using the C equality
operator. If this is an EQ or NE comparison of a constant with a PLUS_EXPR or
a MINUS_EXPR of a constant, we can convert it into a comparison with
a revised constant as long as no overflow occurs. Similarly for a NEGATE_EXPR.
Similarly for a BIT_XOR_EXPR; X ^ C1 == C2 is X == (C1 ^ C2).
Transform comparisons of the form X +- Y CMP X to Y CMP 0.
Transform comparisons of the form C - X CMP X if C % 2 == 1.
If we have X - Y == 0, we can convert that to X == Y and similarly
for !=. Don't do this for ordered comparisons due to overflow. Convert ABS_EXPR<x> == 0 or ABS_EXPR<x> != 0 to x == 0 or x != 0.
If this is an EQ or NE comparison with zero and ARG0 is
(1 << foo) & bar, convert it to (bar >> foo) & 1. Both require
two operations, but the latter can be done in one less insn
on machines that have only two-operand insns or on which a
constant cannot be the first operand. If this is an NE or EQ comparison of zero against the result of a
signed MOD operation whose second operand is a power of 2, make
the MOD operation unsigned since it is simpler and equivalent. Fold ((X >> C1) & C2) == 0 and ((X >> C1) & C2) != 0 where
C1 is a valid shift constant, and C2 is a power of two, i.e.
a single bit. Check for a valid shift count.
If (C2 << C1) doesn't overflow, then ((X >> C1) & C2) != 0
can be rewritten as (X & (C2 << C1)) != 0. Otherwise, for signed (arithmetic) shifts,
((X >> C1) & C2) != 0 is rewritten as X < 0, and
((X >> C1) & C2) == 0 is rewritten as X >= 0. Otherwise, of unsigned (logical) shifts,
((X >> C1) & C2) != 0 is rewritten as (X,false), and
((X >> C1) & C2) == 0 is rewritten as (X,true). If we have (A & C) == C where C is a power of 2, convert this into
(A & C) != 0. Similarly for NE_EXPR. If we have (A & C) != 0 or (A & C) == 0 and C is the sign
bit, then fold the expression into A < 0 or A >= 0. If we have (A & C) == D where D & ~C != 0, convert this into 0.
Similarly for NE_EXPR. If we have (A | C) == D where C & ~D != 0, convert this into 0.
Similarly for NE_EXPR. If this is a comparison of a field, we may be able to simplify it.
Handle the constant case even without -O
to make sure the warnings are given. Optimize comparisons of strlen vs zero to a compare of the
first character of the string vs zero. To wit,
strlen(ptr) == 0 => *ptr == 0
strlen(ptr) != 0 => *ptr != 0
Other cases should reduce to one of these two (or a constant)
due to the return value of strlen being unsigned. Fold (X >> C) != 0 into X < 0 if C is one less than the width
of X. Similarly fold (X >> C) == 0 into X >= 0. (X ^ Y) == 0 becomes X == Y, and (X ^ Y) != 0 becomes X != Y.
(X ^ Y) == Y becomes X == 0. We know that Y has no side-effects.
Likewise (X ^ Y) == X becomes Y == 0. X has no side-effects.
(X ^ C1) op C2 can be rewritten as X op (C1 ^ C2).
Fold (~X & C) == 0 into (X & C) != 0 and (~X & C) != 0 into
(X & C) == 0 when C is a single bit. Fold ((X & C) ^ C) eq/ne 0 into (X & C) ne/eq 0, when the
constant C is a power of two, i.e. a single bit. Likewise, fold ((X ^ C) & C) eq/ne 0 into (X & C) ne/eq 0,
when is C is a power of two, i.e. a single bit. Fold -X op -Y as X op Y, where op is eq/ne.
Fold (X & C) op (Y & C) as (X ^ Y) & C op 0", and symmetries.
Optimize (X ^ Z) op (Y ^ Z) as X op Y, and symmetries.
operand_equal_p guarantees no side-effects so we don't need
to use omit_one_operand on Z. Optimize (X ^ C1) op (Y ^ C2) as (X ^ (C1 ^ C2)) op Y.
Attempt to simplify equality/inequality comparisons of complex
values. Only lower the comparison if the result is known or
can be simplified to a single scalar comparison. Transform comparisons of the form X +- C CMP X.
(X - c) > X becomes false.
Likewise (X + c) < X becomes false.
Convert (X - c) <= X to true.
Convert (X + c) >= X to true.
Convert X + c > X and X - c < X to true for integers.
Convert X + c <= X and X - c >= X to false for integers.
Comparisons with the highest or lowest possible integer of
the specified precision will have known values. The GE_EXPR and LT_EXPR cases above are not normally
reached because of previous transformations. We will flip the signedness of the comparison operator
associated with the mode of arg1, so the sign bit is
specified by this mode. Check that arg1 is the signed
max associated with this sign bit. signed_type does not work on pointer types.
The following case also applies to X < signed_max+1
and X >= signed_max+1 because previous transformations. If we are comparing an ABS_EXPR with a constant, we can
convert all the cases into explicit comparisons, but they may
well not be faster than doing the ABS and one comparison.
But ABS (X) <= C is a range comparison, which becomes a subtraction
and a comparison, and is probably faster. Convert ABS_EXPR<x> >= 0 to true.
Convert ABS_EXPR<x> < 0 to false.
If X is unsigned, convert X < (1 << Y) into X >> Y == 0
and similarly for >= into !=. Similarly for X < (cast) (1 << Y). But cast can't be narrowing,
otherwise Y might be >= # of bits in X's type and thus e.g.
(unsigned char) (1 << Y) for Y 15 might be 0.
If the cast is widening, then 1 << Y should have unsigned type,
otherwise if Y is number of bits in the signed shift type minus 1,
we can't optimize this. E.g. (unsigned long long) (1 << Y) for Y
31 might be 0xffffffff80000000. If the first operand is NaN, the result is constant.
If the second operand is NaN, the result is constant.
Simplify unordered comparison of something with itself.
Fold (double)float1 CMP (double)float2 into float1 CMP float2.
When pedantic, a compound expression can be neither an lvalue
nor an integer constant expression. Don't let (0, 0) be null pointer constant.
An ASSERT_EXPR should never be passed to fold_binary.
Referenced by create_bb(), fold_mult_zconjz(), optimize_stmt(), and rhs_to_tree().
|
static |
Transform `a + (b ? x : y)' into `b ? (a + x) : (a + y)'. Transform, `a + (x < y)' into `(x < y) ? (a + 1) : (a + 0)'. Here CODE corresponds to the `+', COND to the `(b ? x : y)' or `(x < y)' expression, and ARG to `a'. If COND_FIRST_P is nonzero, then the COND is the first argument to CODE; otherwise (as in the example given here), it is the second argument. TYPE is the type of the original expression. Return NULL_TREE if no simplification is possible.
If this operand throws an expression, then it does not make sense to try to perform a logical or arithmetic operation involving it.
This transformation is only worthwhile if we don't have to wrap ARG in a SAVE_EXPR and the operation can be simplified without recursing on at least one of the branches once its pushed inside the COND_EXPR. Check that we have simplified at least one of the branches.
| tree fold_binary_to_constant | ( | ) |
Given the components of a binary expression CODE, TYPE, OP0 and OP1, attempt to fold the expression to a constant without modifying TYPE, OP0 or OP1.
If the expression could be simplified to a constant, then return the constant. If the expression would not be simplified to a constant, then return NULL_TREE.
Referenced by maybe_fold_tmr(), and number_of_iterations_cond().
| tree fold_build1_initializer_loc | ( | location_t | loc, |
| enum tree_code | code, | ||
| tree | type, | ||
| tree | op | ||
| ) |
References INTEGRAL_TYPE_P, POINTER_TYPE_P, and TREE_TYPE.
| tree fold_build1_stat_loc | ( | location_t | loc, |
| enum tree_code | code, | ||
| tree | type, | ||
| tree op0 | MEM_STAT_DECL | ||
| ) |
Fold a unary tree expression with code CODE of type TYPE with an operand OP0. LOC is the location of the resulting expression. Return a folded expression if successful. Otherwise, return a tree expression with code CODE of type TYPE with an operand OP0.
| tree fold_build2_initializer_loc | ( | location_t | loc, |
| enum tree_code | code, | ||
| tree | type, | ||
| tree | op0, | ||
| tree | op1 | ||
| ) |
| tree fold_build2_stat_loc | ( | location_t | loc, |
| enum tree_code | code, | ||
| tree | type, | ||
| tree | op0, | ||
| tree op1 | MEM_STAT_DECL | ||
| ) |
Fold a binary tree expression with code CODE of type TYPE with operands OP0 and OP1. LOC is the location of the resulting expression. Return a folded expression if successful. Otherwise, return a tree expression with code CODE of type TYPE with operands OP0 and OP1.
References tree_expr_nonzero_warnv_p(), TREE_TYPE, type(), and TYPE_PRECISION.
| tree fold_build3_initializer_loc | ( | location_t | loc, |
| enum tree_code | code, | ||
| tree | type, | ||
| tree | op0, | ||
| tree | op1, | ||
| tree | op2 | ||
| ) |
| tree fold_build3_stat_loc | ( | location_t | loc, |
| enum tree_code | code, | ||
| tree | type, | ||
| tree | op0, | ||
| tree | op1, | ||
| tree op2 | MEM_STAT_DECL | ||
| ) |
Fold a ternary tree expression with code CODE of type TYPE with operands OP0, OP1, and OP2. Return a folded expression if successful. Otherwise, return a tree expression with code CODE of type TYPE with operands OP0, OP1, and OP2.
| tree fold_build_call_array_initializer_loc | ( | location_t | loc, |
| tree | type, | ||
| tree | fn, | ||
| int | nargs, | ||
| tree * | argarray | ||
| ) |
| tree fold_build_call_array_loc | ( | location_t | loc, |
| tree | type, | ||
| tree | fn, | ||
| int | nargs, | ||
| tree * | argarray | ||
| ) |
Fold a CALL_EXPR expression of type TYPE with operands FN and NARGS arguments in ARGARRAY, and a null static chain. Return a folded expression if successful. Otherwise, return a CALL_EXPR of type TYPE from the given operands as constructed by build_call_array.
| tree fold_build_cleanup_point_expr | ( | ) |
If necessary, return a CLEANUP_POINT_EXPR for EXPR with the indicated TYPE. If no CLEANUP_POINT_EXPR is necessary, return EXPR itself.
If the expression does not have side effects then we don't have to wrap it with a cleanup point expression.
If the expression is a return, check to see if the expression inside the return has no side effects or the right hand side of the modify expression inside the return. If either don't have side effects set we don't need to wrap the expression in a cleanup point expression. Note we don't check the left hand side of the modify because it should always be a return decl.
|
static |
Subroutine of fold_binary. This routine performs all of the transformations that are common to the equality/inequality operators (EQ_EXPR and NE_EXPR) and the ordering operators (LT_EXPR, LE_EXPR, GE_EXPR and GT_EXPR). Callers other than fold_binary should call fold_binary. Fold a comparison with tree code CODE and type TYPE with operands OP0 and OP1. Return the folded comparison or NULL_TREE.
If one arg is a real or integer constant, put it last.
Transform comparisons of the form X +- C1 CMP C2 to X CMP C2 +- C1.
If the constant operation overflowed this can be
simplified as a comparison against INT_MAX/INT_MIN. Get the sign of the constant on the lhs if the
operation were VARIABLE + CONST1. The sign of the constant determines if we overflowed
INT_MAX (const1_sgn == -1) or INT_MIN (const1_sgn == 1).
Canonicalize to the INT_MIN overflow by swapping the comparison
if necessary. We now can look at the canonicalized case
VARIABLE + 1 CODE2 INT_MIN
and decide on the result. For comparisons of pointers we can decompose it to a compile time comparison of the base objects and the offsets into the object. This requires at least one operand being an ADDR_EXPR or a POINTER_PLUS_EXPR to do more than the operand_equal_p test below.
Get base and offset for the access. Strip ADDR_EXPR for
get_inner_reference, but put it back by stripping INDIRECT_REF
off the base object if possible. indirect_baseN will be true
if baseN is not an address but refers to the object itself. A local variable can never be pointed to by
the default SSA name of an incoming parameter. If we have equivalent bases we might be able to simplify.
We can fold this expression to a constant if the non-constant
offset parts are equal. We can simplify the comparison to a comparison of the variable
offset parts if the constant offset parts are equal.
Be careful to use signed sizetype here because otherwise we
mess with array offsets in the wrong way. This is possible
because pointer arithmetic is restricted to retain within an
object and overflow on pointer differences is undefined as of
6.5.6/8 and /9 with respect to the signed ptrdiff_t. By converting to signed sizetype we cover middle-end pointer
arithmetic which operates on unsigned pointer types of size
type size and ARRAY_REF offsets which are properly sign or
zero extended from their type in case it is narrower than
sizetype. For non-equal bases we can simplify if they are addresses
of local binding decls or constants. We know that !operand_equal_p (base0, base1, 0)
because the if condition was false. But make
sure two decls are not the same. For equal offsets we can simplify to a comparison of the
base addresses. Transform comparisons of the form X +- C1 CMP Y +- C2 to X CMP Y +- C2 +- C1 for signed X, Y. This is valid if the resulting offset is smaller in absolute value than the original one.
Put the constant on the side where it doesn't overflow and is
of lower absolute value than before. Transform comparisons of the form X * C1 CMP 0 to X CMP 0 in the signed arithmetic case. That form is created by the compiler often enough for folding it to be of value. One example is in computing loop trip counts after Operator Strength Reduction.
Handle unfolded multiplication by zero.
If const1 is negative we swap the sense of the comparison.
Fold (double)float1 CMP (double)float2 into float1 CMP float2.
(-a) CMP (-b) -> b CMP a
(-a) CMP CST -> a swap(CMP) (-CST)
IEEE doesn't distinguish +0 and -0 in comparisons.
a CMP (-0) -> a CMP 0
x != NaN is always true, other ops are always false.
Fold comparisons against infinity.
If this is a comparison of a real constant with a PLUS_EXPR
or a MINUS_EXPR of a real constant, we can convert it into a
comparison with a revised real constant as long as no overflow
occurs when unsafe_math_optimizations are enabled. Likewise, we can simplify a comparison of a real constant with
a MINUS_EXPR whose first operand is also a real constant, i.e.
(c1 - x) < c2 becomes x > c1-c2. Reordering is allowed on
floating-point types only if -fassociative-math is set. Fold comparisons against built-in math functions.
If we are widening one operand of an integer comparison,
see if the other operand is similarly being widened. Perhaps we
can do the comparison in the narrower type. Or if we are changing signedness.
If this is comparing a constant with a MIN_EXPR or a MAX_EXPR of a constant, we can simplify it.
Simplify comparison of something with itself. (For IEEE floating-point, we can only do some of these simplifications.)
For NE, we can only do this simplification if integer
or we don't honor IEEE floating point NaNs. ... fall through ...
If we are comparing an expression that just has comparisons of two integer values, arithmetic expressions of those comparisons, and constants, we can simplify it. There are only three cases to check: the two values can either be equal, the first can be greater, or the second can be greater. Fold the expression for those three values. Since each value must be 0 or 1, we have eight possibilities, each of which corresponds to the constant 0 or 1 or one of the six possible comparisons. This handles common cases like (a > b) == 0 but also handles expressions like ((x > y) - (y > x)) > 0, which supposedly occur in macroized code.
Don't handle degenerate cases here; they should already
have been handled anyway. We can't just pass T to eval_subst in case cval1 or cval2
was the same as ARG1. All three of these results should be 0 or 1. Confirm they are.
Then use those values to select the proper code to use. Make a 3-bit mask with the high-order bit being the
value for `>', the next for '=', and the low for '<'. Always false.
Always true.
We can fold X/C1 op C2 where C1 and C2 are integer constants into a single range test.
Fold ~X op ~Y as Y op X.
Fold ~X op C as X op' ~C, where op' is the swapped comparison.
References constant_boolean_node(), FLOAT_TYPE_P, fold_build2_loc, gcc_unreachable, HONOR_NANS, and TYPE_MODE.
|
static |
Subroutine of fold, looking inside expressions of the form A op B ? A : C, where ARG0, ARG1 and ARG2 are the three operands of the COND_EXPR. This function is being used also to optimize A op B ? C : A, by reversing the comparison first.
Return a folded expression whose code is not a COND_EXPR anymore, or NULL_TREE if no folding opportunity is found.
If we have A op 0 ? A : -A, consider applying the following transformations: A == 0? A : -A same as -A A != 0? A : -A same as A A >= 0? A : -A same as abs (A) A > 0? A : -A same as abs (A) A <= 0? A : -A same as -abs (A) A < 0? A : -A same as -abs (A) None of these transformations work for modes with signed zeros. If A is +/-0, the first two transformations will change the sign of the result (from +0 to -0, or vice versa). The last four will fix the sign of the result, even though the original expressions could be positive or negative, depending on the sign of A. Note that all these transformations are correct if A is NaN, since the two alternatives (A and -A) are also NaNs.
In the case that A is of the form X-Y, '-A' (arg2) may
have already been folded to Y-X, check for that. Fall through.
A != 0 ? A : 0 is simply A, unless A is -0. Likewise A == 0 ? A : 0 is always 0 unless A is -0. Note that both transformations are correct when A is NaN: A != 0 is then true, and A == 0 is false.
Try some transformations of A op B ? A : B. A == B? A : B same as B A != B? A : B same as A A >= B? A : B same as max (A, B) A > B? A : B same as max (B, A) A <= B? A : B same as min (A, B) A < B? A : B same as min (B, A) As above, these transformations don't work in the presence of signed zeros. For example, if A and B are zeros of opposite sign, the first two transformations will change the sign of the result. In the last four, the original expressions give different results for (A=+0, B=-0) and (A=-0, B=+0), but the transformed expressions do not. The first two transformations are correct if either A or B is a NaN. In the first transformation, the condition will be false, and B will indeed be chosen. In the case of the second transformation, the condition A != B will be true, and A will be chosen. The conversions to max() and min() are not correct if B is a number and A is not. The conditions in the original expressions will be false, so all four give B. The min() and max() versions would give a NaN instead.
Avoid these transformations if the COND_EXPR may be used
as an lvalue in the C++ front-end. PR c++/19199. Avoid adding NOP_EXPRs in case this is an lvalue.
In C++ a ?: expression can be an lvalue, so put the
operand which will be used if they are equal first
so that we can convert this back to the
corresponding COND_EXPR. If this is A op C1 ? A : C2 with C1 and C2 constant integers, we might still be able to simplify this. For example, if C1 is one less or one more than C2, this might have started out as a MIN or MAX and been transformed by this function. Only good for INTEGER_TYPEs, because we need TYPE_MAX_VALUE.
We can replace A with C1 in this case.
If C1 is C2 + 1, this is min(A, C2), but use ARG00's type for
MIN_EXPR, to preserve the signedness of the comparison. If C1 is C2 - 1, this is min(A, C2), with the same care
as above. If C1 is C2 - 1, this is max(A, C2), but use ARG00's type for
MAX_EXPR, to preserve the signedness of the comparison. If C1 is C2 + 1, this is max(A, C2), with the same care as above.
|
static |
Attempt to fold type conversion operation CODE of expression ARG1 to type TYPE. If no simplification can be done return NULL_TREE.
|
static |
A subroutine of fold_convert_const handling conversions a FIXED_CST to another fixed-point type.
Propagate overflow flags.
|
static |
A subroutine of fold_convert_const handling conversions an INTEGER_CST to a fixed-point type.
Propagate overflow flags.
References INTEGRAL_TYPE_P, POINTER_TYPE_P, TREE_CODE, tree_int_cst_equal(), and TYPE_SIZE.
|
static |
A subroutine of fold_convert_const handling conversions a REAL_CST to a fixed-point type.
Propagate overflow flags.
References error_mark_node, POINTER_TYPE_P, TREE_CODE, TREE_TYPE, and TYPE_ADDR_SPACE.
|
static |
A subroutine of fold_convert_const handling conversions of a FIXED_CST to an integer type.
Right shift FIXED_CST to temp by fbit.
Left shift temp to temp_trunc by fbit.
If FIXED_CST is negative, we need to round the value toward 0. By checking if the fractional bits are not zero to add 1 to temp.
Given a fixed-point constant, make new constant with new type, appropriately sign-extended or truncated.
References build_fixed(), fixed_convert(), FIXED_VALUE_TYPE, TREE_FIXED_CST, TREE_OVERFLOW, TYPE_MODE, and TYPE_SATURATING.
|
static |
A subroutine of fold_convert_const handling conversions of an INTEGER_CST to another integer type.
Given an integer constant, make new constant with new type, appropriately sign-extended or truncated.
References GET_MODE_FBIT, HOST_BITS_PER_DOUBLE_INT, double_int::lshift(), double_int::rshift(), SIGNED_FIXED_POINT_MODE_P, and TREE_FIXED_CST.
|
static |
A subroutine of fold_convert_const handling conversions a REAL_CST to an integer type.
The following code implements the floating point to integer conversion rules required by the Java Language Specification, that IEEE NaNs are mapped to zero and values that overflow the target precision saturate, i.e. values greater than INT_MAX are mapped to INT_MAX, and values less than INT_MIN are mapped to INT_MIN. These semantics are allowed by the C and C++ standards that simply state that the behavior of FP-to-integer conversion is unspecified upon overflow.
If R is NaN, return zero and show we have an overflow.
See if R is less than the lower bound or greater than the upper bound.
References double_int_zero.
|
static |
A subroutine of fold_convert_const handling conversions a FIXED_CST to a floating point type.
|
static |
A subroutine of fold_convert_const handling conversions a REAL_CST to another floating point type.
If converting an infinity or NAN to a representation that doesn't have one, set the overflow bit so that we can produce some kind of error message at the appropriate point if necessary. It's not the most user-friendly message, but it's better than nothing.
Regular overflow, conversion produced an infinity in a mode that can't represent them.
| tree fold_convert_loc | ( | ) |
Convert expression ARG to type TYPE. Used by the middle-end for simple conversions in preference to calling the front-end's convert.
Handle conversions between pointers to different address spaces.
fall through
Referenced by build_fold_addr_expr_loc(), build_one_array(), cprop_into_successor_phis(), distribute_real_division(), exact_inverse(), expand_expr_real_1(), fold_builtin_4(), fold_builtin_bitop(), fold_builtin_exponent(), fold_builtin_fputs(), fold_builtin_n(), fold_builtin_strncpy(), fold_builtin_tan(), fold_mathfn_compare(), fold_mult_zconjz(), get_ws_args_for(), gimplify_modify_expr_complex_part(), ipa_modify_call_arguments(), lower_gimple_return(), lower_send_clauses(), merge_ranges(), native_interpret_complex(), native_interpret_expr(), native_interpret_real(), native_interpret_vector(), and workshare_safe_to_combine_p().
| bool fold_convertible_p | ( | ) |
Returns true, if ARG is convertible to TYPE using a NOP_EXPR.
Referenced by ipa_modify_call_arguments(), and read_replacements_section().
| void fold_defer_overflow_warnings | ( | void | ) |
Start deferring overflow warnings. We could use a stack here to permit nested calls, but at present it is not necessary.
References fold_deferring_overflow_warnings, gcc_assert, and NULL.
Referenced by bit_value_binop(), create_bb(), and rhs_to_tree().
| bool fold_deferring_overflow_warnings_p | ( | void | ) |
Whether we are deferring overflow warnings.
|
static |
Subroutine of fold() that optimizes comparisons of a division by a nonzero integer constant against an integer constant, i.e. X/C1 op C2.
CODE is the comparison operator: EQ_EXPR, NE_EXPR, GT_EXPR, LT_EXPR, GE_EXPR or LE_EXPR. TYPE is the type of the result and ARG0 and ARG1 are the operands of the comparison. ARG1 must be a TREE_REAL_CST.
The function returns the constant folded tree if a simplification can be made, and NULL_TREE otherwise.
We have to do this the hard way to detect unsigned overflow. prod = int_const_binop (MULT_EXPR, arg01, arg1);
Likewise hi = int_const_binop (PLUS_EXPR, prod, tmp).
A negative divisor reverses the relational operators.
| tree fold_ignored_result | ( | ) |
Strip non-trapping, non-side-effecting tree nodes from an expression whose result is ignored. The type of the returned tree need not be the same as the original expression.
| tree fold_indirect_ref_1 | ( | ) |
Given a pointer value OP0 and a type TYPE, return a simplified version of an indirection through OP0, or NULL_TREE if no simplification is possible.
&CONST_DECL -> to the value of the const decl.
&p => p; make sure to handle *&"str"[cst] here.
(foo *)&fooarray => fooarray[0]
(foo *)&complexfoo => __real__ complexfoo
(foo *)&vectorfoo => BIT_FIELD_REF<vectorfoo,...>
((foo*)&vectorfoo)[1] => BIT_FIELD_REF<vectorfoo,...>
((foo*)&complexfoo)[1] => __imag__ complexfoo
((foo *)&fooarray)[1] => fooarray[1]
(foo *)fooarrptr => (*fooarrptr)[0]
| tree fold_indirect_ref_loc | ( | ) |
Given an INDIRECT_REF T, return either T or a simplified version.
|
static |
Subroutine of fold() that optimizes comparisons against Infinities, either +Inf or -Inf.
CODE is the comparison operator: EQ_EXPR, NE_EXPR, GT_EXPR, LT_EXPR, GE_EXPR or LE_EXPR. TYPE is the type of the result and ARG0 and ARG1 are the operands of the comparison. ARG1 must be a TREE_REAL_CST.
The function returns the constant folded tree if a simplification can be made, and NULL_TREE otherwise.
For negative infinity swap the sense of the comparison.
x > +Inf is always false, if with ignore sNANs.
x <= +Inf is always true, if we don't case about NaNs.
x <= +Inf is the same as x == x, i.e. isfinite(x).
x == +Inf and x >= +Inf are always equal to x > DBL_MAX.
x < +Inf is always equal to x <= DBL_MAX.
x != +Inf is always equal to !(x > DBL_MAX).
|
static |
Fold the truth-negation of ARG. This never alters ARG itself. We assume that ARG is an operation that returns a truth value (0 or 1 for scalars, 0 or -1 for vectors). Return the folded expression if folding is successful. Otherwise, return NULL_TREE.
|
static |
Subroutine of fold() that checks comparisons of built-in math functions against real constants.
FCODE is the DECL_FUNCTION_CODE of the built-in, CODE is the comparison operator: EQ_EXPR, NE_EXPR, GT_EXPR, LT_EXPR, GE_EXPR or LE_EXPR. TYPE is the type of the result and ARG0 and ARG1 are the operands of the comparison. ARG1 must be a TREE_REAL_CST.
The function returns the constant folded tree if a simplification can be made, and NULL_TREE otherwise.
sqrt(x) < y is always false, if y is negative.
sqrt(x) > y is always true, if y is negative and we
don't care about NaNs, i.e. negative values of x. sqrt(x) > y is the same as x >= 0, if y is negative.
sqrt(x) > y is x == +Inf, when y is very large.
sqrt(x) > y is always false, when y is very large
and we don't care about infinities. sqrt(x) > c is the same as x > c*c.
sqrt(x) < y is always true, when y is a very large
value and we don't care about NaNs or Infinities. sqrt(x) < y is x != +Inf when y is very large and we
don't care about NaNs. sqrt(x) < y is x >= 0 when y is very large and we
don't care about Infinities. sqrt(x) < y is x >= 0 && x != +Inf, when y is large.
sqrt(x) < c is the same as x < c*c, if we ignore NaNs.
sqrt(x) < c is the same as x >= 0 && x < c*c.
References build2, build_int_cst(), fold_build2_loc, fold_convert_loc(), fold_single_bit_test_into_sign_test(), host_integerp(), integer_pow2p(), integer_zerop(), NULL_TREE, size_int, TREE_CODE, TREE_INT_CST_LOW, tree_log2(), TREE_OPERAND, TREE_TYPE, lang_hooks_for_types::type_for_mode, TYPE_MODE, TYPE_PRECISION, and lang_hooks::types.
|
static |
Fold a binary expression of code CODE and type TYPE with operands OP0 and OP1, containing either a MIN-MAX or a MAX-MIN combination. Return the folded expression if folding is successful. Otherwise, return NULL_TREE.
MIN (MAX (a, b), b) == b.
MIN (MAX (b, a), b) == b.
MIN (a, MAX (a, b)) == a.
MIN (a, MAX (b, a)) == a.
|
static |
Subroutine of fold_binary. Optimize complex multiplications of the form z * conj(z), as pow(realpart(z),2) + pow(imagpart(z),2). The argument EXPR represents the expression "z" of type TYPE.
References fold_binary_loc(), fold_build2_loc, fold_convert_loc(), integer_zerop(), and TREE_OPERAND.
|
static |
Return the tree for neg (ARG0) when ARG0 is known to be either an integer constant, real, or fixed-point constant.
TYPE is the type of the result.
Propagate overflow flags.
|
static |
Given T, an expression, return a folded tree for -T or NULL_TREE, if no simplification is possible. If negate_expr_p would return true for T, NULL_TREE will never be returned.
Convert - (~A) to A + 1.
Two's complement FP formats, such as c4x, may overflow.
-(A + B) -> (-B) - A.
-(A + B) -> (-A) - B.
- (A - B) -> B - A
Fall through.
In general we can't negate A / B, because if A is INT_MIN and
B is 1, we may turn this into INT_MIN / -1 which is undefined
and actually traps on some architectures. But if overflow is
undefined, we can negate, because - (INT_MIN / 1) is an
overflow. If overflow is undefined then we have to be careful because
we ask whether it's ok to associate the negate with the
division which is not ok for example for
-((a - b) / c) where (-(a - b)) / c may invoke undefined
overflow because of negating INT_MIN. So do not use
negate_expr_p here but open-code the two important cases. Convert -((double)float) into (double)(-float).
Negate -f(x) as f(-x).
Optimize -((int)x >> 31) into (unsigned)x >> 31.
References build_complex(), negate_expr(), TREE_CODE, TREE_IMAGPART, and TREE_REALPART.
|
static |
|
static |
Return the tree for not (ARG0) when ARG0 is known to be an integer constant. TYPE is the type of the result.
|
static |
This is called when we fold something based on the fact that signed overflow is undefined.
References CASE_FLT_FN.
|
static |
Fold a sum or difference of at least one multiplication. Returns the folded tree or NULL if no simplification could be made.
(A * C) +- (B * C) -> (A+-B) * C. (A * C) +- A -> A * (C+-1). We are most concerned about the case where C is a constant, but other combinations show up during loop reduction. Since it is not difficult, try all four possibilities.
We cannot generate constant 1 for fract.
As we canonicalize A - 2 to A + -2 get rid of that sign for
the purpose of this canonicalization. We cannot generate constant 1 for fract.
No identical multiplicands; see if we can find a common power-of-two factor in non-power-of-two multiplies. This can help in multi-dimensional array access.
Move min of absolute values to int11.
The remainder should not be a constant, otherwise we
end up folding i * 4 + 2 to (i * 2 + 1) * 2 which has
increased the number of multiplications necessary.
|
static |
EXP is some logical combination of boolean tests. See if we can merge it into some range test. Return the new tree if so.
If this is an OR operation, invert both sides; we will invert again at the end.
If both expressions are the same, if we can merge the ranges, and we can build the range test, return it or it inverted. If one of the ranges is always true or always false, consider it to be the same expression as the other.
On machines where the branch cost is expensive, if this is a short-circuited branch and the underlying object on both sides is the same, make a non-short-circuit operation.
If simple enough, just rewrite. Otherwise, make a SAVE_EXPR
unless we are at top level or LHS contains a PLACEHOLDER_EXPR, in
which cases we can't do this.
References combine_comparisons(), operand_equal_p(), and swap_tree_comparison().
| tree fold_read_from_constant_string | ( | ) |
If EXP represents referencing an element in a constant string (either via pointer arithmetic or array indexing), return the tree representing the value accessed, otherwise return NULL.
Optimize the special-case of a zero lower bound.
We convert the low_bound to sizetype to avoid some problems with constant folding. (E.g. suppose the lower bound is 1, and its mode is QI. Without the conversion,l (ARRAY +(INDEX-(unsigned char)1)) becomes ((ARRAY+(-(unsigned char)1)) +INDEX), which becomes (ARRAY+255+INDEX). Oops!)
| bool fold_real_zero_addition_p | ( | ) |
Subroutine of fold() that checks for the addition of +/- 0.0.
If !NEGATE, return true if ADDEND is +/-0.0 and, for all X of type TYPE, X + ADDEND is the same as X. If NEGATE, return true if X - ADDEND is the same as X.
X + 0 and X - 0 both give X when X is NaN, infinite, or nonzero and finite. The problematic cases are when X is zero, and its mode has signed zeros. In the case of rounding towards -infinity, X - 0 is not the same as X because 0 - 0 is -0. In other rounding modes, X + 0 is not the same as X because -0 + 0 is 0.
Don't allow the fold with -fsignaling-nans.
Allow the fold if zeros aren't signed, or their sign isn't important.
In a vector or complex, we would need to check the sign of all zeros.
Treat x + -0 as x - 0 and x - -0 as x + 0.
The mode has signed zeros, and we have to honor their sign. In this situation, there is only one case we can return true for. X - 0 is the same as X unless rounding towards -infinity is supported.
|
static |
Given CODE, a relational operator, the target type, TYPE and two constant operands OP0 and OP1, return the result of the relational operation. If the result is not a compile time constant, then return NULL_TREE.
From here on, the only cases we handle are when the result is known to be a constant.
Handle the cases where either operand is a NaN.
Handle equality/inequality of complex constants.
From here on we only handle LT, LE, GT, GE, EQ and NE. To compute GT, swap the arguments and do LT. To compute GE, do LT and invert the result. To compute LE, swap the arguments, do LT and invert the result. To compute NE, do EQ and invert the result. Therefore, the code below must handle only EQ and LT.
Note that it is safe to invert for real values here because we have already handled the one case that it matters.
Compute a result for LT or EQ if args permit; Otherwise return T.
|
static |
Fold comparison ARG0 CODE ARG1 (with result in TYPE), where for ARG0 just the signedness is changed.
| tree fold_single_bit_test | ( | location_t | loc, |
| enum tree_code | code, | ||
| tree | arg0, | ||
| tree | arg1, | ||
| tree | result_type | ||
| ) |
If CODE with arguments ARG0 and ARG1 represents a single bit equality/inequality test, then return a simplified form of the test using shifts and logical operations. Otherwise return NULL. TYPE is the desired result type.
If this is testing a single bit, we can optimize the test.
First, see if we can fold the single bit test into a sign-bit
test. Otherwise we have (A & C) != 0 where C is a single bit,
convert that into ((A >> C2) & 1). Where C2 = log2(C).
Similarly for (A & C) == 0. If INNER is a right shift of a constant and it plus BITNUM does
not overflow, adjust BITNUM and INNER. If we are going to be able to omit the AND below, we must do our
operations as unsigned. If we must use the AND, we have a choice.
Normally unsigned is faster, but for some machines signed is. Put the AND last so it can combine with more things.
Make sure to return the proper type.
|
static |
If CODE with arguments ARG0 and ARG1 represents a single bit equality/inequality test, then return a simplified form of the test using a sign testing. Otherwise return NULL. TYPE is the desired result type.
If this is testing a single bit, we can optimize the test.
If we have (A & C) != 0 where C is the sign bit of A, convert
this into A < 0. Similarly for (A & C) == 0 into A >= 0. This is only a win if casting to a signed type is cheap,
i.e. when arg00's type is not a partial mode.
Referenced by fold_mathfn_compare().
| tree fold_strip_sign_ops | ( | ) |
Simplify the floating point expression EXP when the sign of the result is not significant. Return NULL_TREE if no simplification is possible.
Strip copysign function call, return the 1st argument.
Strip sign ops from the argument of "odd" math functions.
Referenced by fold_trunc_transparent_mathfn().
| tree fold_ternary_loc | ( | location_t | loc, |
| enum tree_code | code, | ||
| tree | type, | ||
| tree | op0, | ||
| tree | op1, | ||
| tree | op2 | ||
| ) |
Fold a ternary expression of code CODE and type TYPE with operands OP0, OP1, and OP2. Return the folded expression if folding is successful. Otherwise, return NULL_TREE.
Strip any conversions that don't change the mode. This is safe for every expression, except for a comparison expression because its signedness is derived from its operands. So, in the latter case, only strip conversions that don't change the signedness. Note that this is done as an internal manipulation within the constant folder, in order to find the simplest representation of the arguments so that their form can be studied. In any cases, the appropriate type conversions should be put back in the tree that will get out of the constant folder.
Pedantic ANSI C says that a conditional expression is never an lvalue,
so all simple results must be passed through pedantic_non_lvalue. Only optimize constant conditions when the selected branch
has the same type as the COND_EXPR. This avoids optimizing
away "c ? x : throw", where the throw has a void type.
Avoid throwing away that operand which contains label. If we have A op B ? A : C, we may be able to convert this to a
simpler expression, depending on the operation and the values
of B and C. Signed zeros prevent all of these transformations,
for reasons given above each one.
Also try swapping the arguments and inverting the conditional. If the second operand is simpler than the third, swap them
since that produces better jump optimization results. See if this can be inverted. If it can't, possibly because
it was a floating-point inequality comparison, don't do
anything. Convert A ? 1 : 0 to simply A.
If we try to convert OP0 to our type, the
call to fold will try to move the conversion inside
a COND, which will recurse. In that case, the COND_EXPR
is probably the best choice, so leave it alone. Convert A ? 0 : 1 to !A. This prefers the use of NOT_EXPR
over COND_EXPR in cases such as floating point comparisons. A < 0 ? <sign bit of A> : 0 is simply (A & <sign bit of A>).
sign_bit_p looks through both zero and sign extensions,
but for this optimization only sign extensions are
usable. sign_bit_p only checks ARG1 bits within A's precision.
If <sign bit of A> has wider type than A, bits outside
of A's precision in <sign bit of A> need to be checked.
If they are all 0, this optimization needs to be done
in unsigned A's type, if they are all 1 in signed A's type,
otherwise this can't be done. (A >> N) & 1 ? (1 << N) : 0 is simply A & (1 << N). A & 1 was
already handled above. A & N ? N : 0 is simply A & N if N is a power of two. This
is probably obsolete because the first operand should be a
truth value (that's why we have the two cases above), but let's
leave it in until we can confirm this for all front-ends. Disable the transformations below for vectors, since
fold_binary_op_with_conditional_arg may undo them immediately,
yielding an infinite loop. Convert A ? B : 0 into A && B if A and B are truth values.
Convert A ? B : 1 into !A || B if A and B are truth values.
Only perform transformation if ARG0 is easily inverted.
Convert A ? 0 : B into !A && B if A and B are truth values.
Only perform transformation if ARG0 is easily inverted.
Convert A ? 1 : B into A || B if A and B are truth values.
CALL_EXPRs used to be ternary exprs. Catch any mistaken uses
of fold_ternary on them. Constructor elements can be subvectors.
We keep an exact subset of the constructor elements.
The bitfield references a single constructor element.
A bit-field-ref that referenced the full argument can be stripped.
On constants we can use native encode/interpret to constant
fold (nearly) all BIT_FIELD_REFs. This limitation should not be necessary, we just need to
round this up to mode size. Need bit-shifting of the buffer to relax the following.
??? We cannot tell native_encode_expr to start at
some random byte only. So limit us to a reasonable amount
of work. For integers we can decompose the FMA if possible.
|
static |
Fold A < X && A + 1 > Y to A < X && A >= Y. Normally A + 1 > Y means A >= Y && A != MAX, but in this case we know that A < X <= MAX. INEQ is A + 1 > Y, BOUND is A < X.
Convert the pointer types into integer before taking the difference.
|
static |
Fold a binary bitwise/truth expression of code CODE and type TYPE with operands OP0 and OP1. LOC is the location of the resulting expression. ARG0 and ARG1 are the NOP_STRIPed results of OP0 and OP1. Return the folded expression if folding is successful. Otherwise, return NULL_TREE.
We only do these simplifications if we are optimizing.
Check for things like (A || B) && (A || C). We can convert this to A || (B && C). Note that either operator can be any of the four truth and/or operations and the transformation will still be valid. Also note that we only care about order for the ANDIF and ORIF operators. If B contains side effects, this might change the truth-value of A.
This case if tricky because we must either have commutative
operators or else A10 must not have side-effects. See if we can build a range comparison.
Check for the possibility of merging component references. If our lhs is another similar operation, try to merge its rhs with our rhs. Then try to merge our lhs and rhs.
Transform ((A AND-IF B) AND[-IF] C) into (A AND-IF (B AND C)),
or ((A OR-IF B) OR[-IF] C) into (A OR-IF (B OR C))
We don't want to pack more than two leafs to a non-IF AND/OR
expression.
If tree-code of left-hand operand isn't an AND/OR-IF code and not
equal to IF-CODE, then we don't want to add right-hand operand.
If the inner right-hand side of left-hand operand has
side-effects, or isn't simple, then we can't add to it,
as otherwise we might destroy if-sequence. Needed for sequence points to handle trappings, and
side-effects. Same as abouve but for (A AND[-IF] (B AND-IF C)) -> ((A AND B) AND-IF C),
or (A OR[-IF] (B OR-IF C) -> ((A OR B) OR-IF C). Needed for sequence points to handle trappings, and
side-effects. Transform (A AND-IF B) into (A AND B), or (A OR-IF B)
into (A OR B).
For sequence point consistancy, we need to check for trapping,
and side-effects.
|
static |
Find ways of folding logical expressions of LHS and RHS: Try to merge two comparisons to the same innermost item. Look for range tests like "ch >= '0' && ch <= '9'". Look for combinations of simple terms on machines with expensive branches and evaluate the RHS unconditionally.
For example, if we have p->a == 2 && p->b == 4 and we can make an object large enough to span both A and B, we can do this with a comparison against the object ANDed with the a mask.
If we have p->a == q->a && p->b == q->b, we may be able to use bit masking operations to do this with one comparison.
We check for both normal comparisons and the BIT_AND_EXPRs made this by function and the one above.
CODE is the logical operation being done. It can be TRUTH_ANDIF_EXPR, TRUTH_AND_EXPR, TRUTH_ORIF_EXPR, or TRUTH_OR_EXPR.
TRUTH_TYPE is the type of the logical operand and LHS and RHS are its two operands.
We return the simplified tree or 0 if no optimization is possible.
If this is the "or" of two comparisons, we can do something if
the comparisons are NE_EXPR. If this is the "and", we can do something
if the comparisons are EQ_EXPR. I.e.,
(a->b == 2 && a->c == 4) can become (a->new == NEW).
WANTED_CODE is this operation code. For single bit fields, we can
convert EQ_EXPR to NE_EXPR so we need not reject the "wrong"
comparison for one-bit fields.
Start by getting the comparison codes. Fail if anything is volatile. If one operand is a BIT_AND_EXPR with the constant one, treat it as if it were surrounded with a NE_EXPR.
Simplify (x<y) && (x==y) into (x<=y) and related optimizations.
If the RHS can be evaluated unconditionally and its operands are simple, it wins to evaluate the RHS unconditionally on machines with expensive branches. In this case, this isn't a comparison that can be merged.
Convert (a != 0) || (b != 0) into (a | b) != 0.
Convert (a == 0) && (b == 0) into (a | b) == 0.
See if the comparisons can be merged. Then get all the parameters for each side.
It must be true that the inner operation on the lhs of each comparison must be the same if we are to be able to do anything. Then see if we have constants. If not, the same must be true for the rhs's.
If either comparison code is not correct for our logical operation, fail. However, we can convert a one-bit comparison against zero into the opposite comparison against that bit being set in the field.
Make the left operand unsigned, since we are only interested
in the value of one bit. Otherwise we are doing the wrong
thing below. This is analogous to the code for l_const above.
See if we can find a mode that contains both fields being compared on the left. If we can't, fail. Otherwise, update all constants and masks to be relative to a field of that size.
If the right sides are not constant, do the same for it. Also, disallow this optimization if a size or signedness mismatch occurs between the left and right sides.
Make sure the two fields on the right
correspond to the left without being swapped. Make a mask that corresponds to both fields being compared.
Do this for both items being compared. If the operands are the
same size and the bits being compared are in the same position
then we can do this by masking both and comparing the masked
results. There is still another way we can do something: If both pairs of
fields being compared are adjacent, we may be able to make a wider
field containing them both.
Note that we still must mask the lhs/rhs expressions. Furthermore,
the mask must be shifted to account for the shift done by
make_bit_field_ref. Convert to the smaller type before masking out unwanted bits.
Handle the case of comparisons with constants. If there is something in common between the masks, those bits of the constants must be the same. If not, the condition is always false. Test for this to avoid generating incorrect code below.
Construct the expression we will return. First get the component reference we will make. Unless the mask is all ones the width of that field, perform the mask operation. Then compare with the merged constant.
|
static |
Return a simplified tree node for the truth-negation of ARG. This never alters ARG itself. We assume that ARG is an operation that returns a truth value (0 or 1).
FIXME: one would think we would fold the result, but it causes problems with the dominator optimizer.
If this is a comparison, we can simply invert it, except for floating-point non-equality comparisons, in which case we just enclose a TRUTH_NOT_EXPR around what we have.
Here we can invert either operand. We invert the first operand
unless the second operand is a TRUTH_NOT_EXPR in which case our
result is the XOR of the first operand with the inside of the
negation of the second operand. A COND_EXPR may have a throw as one operand, which
then has void type. Just leave void operands
as they are. ... fall through ...
References TREE_OPERAND.
| tree fold_unary_ignore_overflow_loc | ( | location_t | loc, |
| enum tree_code | code, | ||
| tree | type, | ||
| tree | op0 | ||
| ) |
If the operation was a conversion do not mark a resulting constant with TREE_OVERFLOW if the original constant was not. These conversions have implementation defined behavior and retaining the TREE_OVERFLOW flag here would confuse later passes such as VRP.
References get_inner_reference(), NULL_TREE, TREE_CODE, and TREE_OPERAND.
| tree fold_unary_loc | ( | ) |
Fold a unary expression of code CODE and type TYPE with operand OP0. Return the folded expression if folding is successful. Otherwise, return NULL_TREE.
Don't use STRIP_NOPS, because signedness of argument type
matters.
Strip any conversions that don't change the mode. This
is safe for every expression, except for a comparison
expression because its signedness is derived from its
operands.
Note that this is done as an internal manipulation within
the constant folder, in order to find the simplest
representation of the arguments so that their form can be
studied. In any cases, the appropriate type conversions
should be put back in the tree that will get out of the
constant folder. If this was a conversion, and all we did was to move into
inside the COND_EXPR, bring it back out. But leave it if
it is a conversion from integer to integer and the
result precision is no wider than a word since such a
conversion is cheap and may be optimized away by combine,
while it couldn't if it were outside the COND_EXPR. Then return
so we don't get into an infinite recursion loop taking the
conversion out and then back in. Re-association barriers around constants and other re-association
barriers can be removed. If we have (type) (a CMP b) and type is an integral type, return
new expression involving the new type. Canonicalize
(type) (a CMP b) to (a CMP b) ? (type) true : (type) false for
non-integral type.
Do not fold the result as that would not simplify further, also
folding again results in recursions. Handle cases of two conversions in a row.
In addition to the cases of two conversions in a row
handled below, if we are converting something to its own
type via an object of identical or wider precision, neither
conversion is needed. Likewise, if the intermediate and initial types are either both
float or both integer, we don't need the middle conversion if the
former is wider than the latter and doesn't change the signedness
(for integers). Avoid this if the final type is a pointer since
then we sometimes need the middle conversion. Likewise if the
final type has a precision not equal to the size of its mode. If we have a sign-extension of a zero-extended value, we can
replace that by a single zero-extension. Likewise if the
final conversion does not change precision we can drop the
intermediate conversion. Two conversions in a row are not needed unless:
- some conversion is floating-point (overstrict for now), or
- some conversion is a vector (overstrict for now), or
- the intermediate type is narrower than both initial and
final, or
- the intermediate type and innermost type differ in signedness,
and the outermost type is wider than the intermediate, or
- the initial type is a pointer type and the precisions of the
intermediate and final types differ, or
- the final type is a pointer type and the precisions of the
initial and intermediate types differ. Handle (T *)&A.B.C for A being of type T and B and C
living at offset zero. This occurs frequently in
C++ upcasting and then accessing the base. If the reference was to a (constant) zero offset, we can use
the address of the base if it has the same base type
as the result type and the pointer type is unqualified. Detect assigning a bitfield.
Don't leave an assignment inside a conversion
unless assigning a bitfield. First do the assignment, then return converted constant.
Convert (T)(x & c) into (T)x & (T)c, if c is an integer
constants (if x has signed type, the sign bit cannot be set
in c). This folds extension into the BIT_AND_EXPR.
??? We don't do it for BOOLEAN_TYPE or ENUMERAL_TYPE because they
very likely don't have maximal range for their precision and this
transformation effectively doesn't preserve non-maximal ranges. Convert (T1)(X p+ Y) into ((T1)X p+ Y), for pointer type,
when one of the new casts will fold away. Conservatively we assume
that this happens when X or Y is NOP_EXPR or Y is INTEGER_CST. Convert (T1)(~(T2)X) into ~(T1)X if T1 and T2 are integral types
of the same precision, and X is an integer type not narrower than
types T1 or T2, i.e. the cast (T2)X isn't an extension. Convert (T1)(X * Y) into (T1)X * (T1)Y if T1 is narrower than the
type of X and Y (integer types only). Be careful not to introduce new overflows.
For integral conversions with the same precision or pointer
conversions use a NOP_EXPR instead. Strip inner integral conversions that do not change the precision.
Convert fabs((double)float) into (double)fabsf(float).
ABS_EXPR<ABS_EXPR<x>> = ABS_EXPR<x> even if flag_wrapv is on.
Strip sign ops from argument.
Convert ~ (-A) to A - 1.
Convert ~ (A - 1) or ~ (A + -1) to -A.
Convert ~(X ^ Y) to ~X ^ Y or X ^ ~Y if ~X or ~Y simplify.
Perform BIT_NOT_EXPR on each element individually.
Note that the operand of this must be an int
and its values must be 0 or 1.
("true" is a fixed value perhaps depending on the language,
but we don't handle values other than 1 correctly yet.) Fold *&X to X if X is an lvalue.
Referenced by vec_cst_ctor_to_array().
| tree fold_unary_to_constant | ( | ) |
Given the components of a unary expression CODE, TYPE and OP0, attempt to fold the expression to a constant without modifying TYPE or OP0.
If the expression could be simplified to a constant, then return the constant. If the expression would not be simplified to a constant, then return NULL_TREE.
| void fold_undefer_and_ignore_overflow_warnings | ( | void | ) |
Stop deferring overflow warnings, ignoring any deferred warnings.
References NULL.
| void fold_undefer_overflow_warnings | ( | ) |
Stop deferring overflow warnings. If there is a pending warning, and ISSUE is true, then issue the warning if appropriate. STMT is the statement with which the warning should be associated (used for location information); STMT may be NULL. CODE is the level of the warning–a warn_strict_overflow_code value. This function will use the smaller of CODE and the deferred code when deciding whether to issue the warning. CODE may be zero to mean to always use the deferred code.
Use the smallest code level when deciding to issue the warning.
Referenced by bit_value_binop(), create_bb(), and rhs_to_tree().
|
static |
Attempt to fold vector permutation of ARG0 and ARG1 vectors using SEL selector. Return the folded VECTOR_CST or CONSTRUCTOR if successful, NULL_TREE otherwise.
|
static |
Fold a VIEW_CONVERT_EXPR of a constant expression EXPR to type TYPE at compile-time. If we're unable to perform the conversion return NULL_TREE.
We support up to 512-bit values (for V8DFmode).
Check that the host and target are sane.
|
static |
Fold comparison ARG0 CODE ARG1 (with result in TYPE), where ARG0 is extended to a wider type.
If possible, express the comparison in the shorter mode.
If we are comparing with the integer that does not fit into the range of the shorter type, the result is known.
References TREE_OPERAND.
|
static |
Subroutine of fold_binary. If P is the value of EXPR, computes power-of-two M and (arbitrary) N such that M divides (P-N). This condition guarantees that P and N have the same least significant log2(M) bits. N is not otherwise constrained. In particular, N is not normalized to 0 <= N < M as is common. In general, the precise value of P is unknown. M is chosen as large as possible such that constant N can be determined.
Returns M and sets *RESIDUE to N.
If ALLOW_FUNC_ALIGN is true, do take functions' DECL_ALIGN_UNIT into account. This is not always possible due to PR 35705.
Compute the greatest power-of-2 divisor of op1.
If align is non-zero and less than *modulus, replace
*modulus with align., If align is 0, then either op1 is 0
or the greatest power-of-2 divisor of op1 doesn't fit in an
unsigned HOST_WIDE_INT. In either case, no additional
constraint is imposed. If we get here, we were unable to determine anything useful about the expression.
|
static |
Check whether TYPE1 and TYPE2 are equivalent integer types, suitable for use in int_const_binop, size_binop and size_diffop.
| tree int_const_binop | ( | ) |
|
static |
Combine two integer constants ARG1 and ARG2 under operation CODE to produce a new constant. Return NULL_TREE if we don't know how to evaluate CODE at compile-time.
It's unclear from the C standard whether shifts can overflow.
The following code ignores overflow; perhaps a C standard
interpretation ruling is needed.
MULT_HIGHPART_EXPR can't ever oveflow, as the multiplication
is performed in twice the precision of arguments. This is a shortcut for a common special case.
... fall through ...
This is a shortcut for a common special case.
... fall through ...
References HOST_BITS_PER_WIDE_INT, NULL_TREE, TYPE_PRECISION, and double_int::wide_mul_with_sign().
| enum tree_code invert_tree_comparison | ( | ) |
Given a tree comparison code, return the code that is the logical inverse. It is generally not safe to do this for floating-point comparisons, except for EQ_EXPR, NE_EXPR, ORDERED_EXPR and UNORDERED_EXPR, so we return ERROR_MARK in this case.
References COMPCODE_EQ, COMPCODE_GE, COMPCODE_GT, COMPCODE_LE, COMPCODE_LT, COMPCODE_LTGT, COMPCODE_NE, COMPCODE_ORD, COMPCODE_UNEQ, COMPCODE_UNGE, COMPCODE_UNGT, COMPCODE_UNLE, COMPCODE_UNLT, COMPCODE_UNORD, and gcc_unreachable.
Referenced by forward_propagate_addr_expr(), get_base_value(), get_cmp_code(), is_norm_cond_subset_of(), and simplify_using_outer_evolutions().
| tree invert_truthvalue_loc | ( | ) |
Return a simplified tree node for the truth-negation of ARG. This never alters ARG itself. We assume that ARG is an operation that returns a truth value (0 or 1 for scalars, 0 or -1 for vectors).
References get_inner_reference(), and TREE_CODE.
|
static |
Return a BIT_FIELD_REF of type TYPE to refer to BITSIZE bits of INNER starting at BITPOS. The field is unsigned if UNSIGNEDP is nonzero.
Given EXP, a logical expression, set the range it is testing into variables denoted by PIN_P, PLOW, and PHIGH. Return the expression actually being tested. *PLOW and *PHIGH will be made of the same type as the returned expression. If EXP is not a comparison, we will most likely not be returning a useful value and range. Set *STRICT_OVERFLOW_P to true if the return value is only valid because signed overflow is undefined; otherwise, do not change *STRICT_OVERFLOW_P.
Start with simply saying "EXP != 0" and then look at the code of EXP and see if we can refine the range. Some of the cases below may not happen, but it doesn't seem worth worrying about this. We "continue" the outer loop when we've changed something; otherwise we "break" the switch, which will "break" the while.
If EXP is a constant, we can evaluate whether this is true or false.
| tree make_range_step | ( | location_t | loc, |
| enum tree_code | code, | ||
| tree | arg0, | ||
| tree | arg1, | ||
| tree | exp_type, | ||
| tree * | p_low, | ||
| tree * | p_high, | ||
| int * | p_in_p, | ||
| bool * | strict_overflow_p | ||
| ) |
Helper routine for make_range. Perform one step for it, return new expression if the loop should continue or NULL_TREE if it should stop.
We can only do something if the range is testing for zero.
We can only do something if the range is testing for zero
and if the second operand is an integer constant. Note that
saying something is "in" the range we make is done by
complementing IN_P since it will set in the initial case of
being not equal to zero; "out" is leaving it alone. If this is an unsigned comparison, we also know that EXP is
greater than or equal to zero. We base the range tests we make
on that fact, so we record it here so we can parse existing
range tests. We test arg0_type since often the return type
of, e.g. EQ_EXPR, is boolean. If the high bound is missing, but we have a nonzero low
bound, reverse the range so it goes from zero to the low bound
minus 1. If flag_wrapv and ARG0_TYPE is signed, make sure
low and high are non-NULL, then normalize will DTRT. (-x) IN [a,b] -> x in [-b, -a]
~ X -> -X - 1
If flag_wrapv and ARG0_TYPE is signed, then we cannot
move a constant to the other side. If EXP is signed, any overflow in the computation is undefined,
so we don't worry about it so long as our computations on
the bounds don't overflow. For unsigned, overflow is defined
and this is exactly the right thing. Check for an unsigned range which has wrapped around the maximum
value thus making n_high < n_low, and normalize it. If the range is of the form +/- [ x+1, x ], we won't
be able to normalize it. But then, it represents the
whole range or the empty set, so make it
+/- [ -, - ]. If we're converting arg0 from an unsigned type, to exp,
a signed type, we will be doing the comparison as unsigned.
The tests above have already verified that LOW and HIGH
are both positive.
So we have to ensure that we will handle large unsigned
values the same way that the current signed bounds treat
negative values. For fixed-point modes, we need to pass the saturating flag
as the 2nd parameter. A range without an upper bound is, naturally, unbounded.
Since convert would have cropped a very large value, use
the max value for the destination type. If the low bound is specified, "and" the range with the
range for which the original unsigned value will be
positive. Otherwise, "or" the range with the range of the input
that will be interpreted as negative.
|
static |
Mask out the tz least significant bits of X of type TYPE where tz is the number of trailing zeroes in Y.
| bool may_negate_without_overflow_p | ( | ) |
Check whether we may negate an integer constant T without causing overflow.
Referenced by create_iv().
|
static |
Canonicalize the comparison ARG0 CODE ARG1 with type TYPE with undefined overflow further. Try to decrease the magnitude of constants involved by changing LE_EXPR and GE_EXPR to LT_EXPR and GT_EXPR or vice versa and put sole constants at the second argument position. Returns the canonicalized tree if changed, otherwise NULL_TREE.
Try canonicalization by simplifying arg0.
Try canonicalization by simplifying arg1 using the swapped comparison.
References integer_one_node, and omit_one_operand_loc().
|
static |
Helper that tries to canonicalize the comparison ARG0 CODE ARG1 by changing CODE to reduce the magnitude of constants involved in ARG0 of the comparison. Returns a canonicalized comparison tree if a simplification was possible, otherwise returns NULL_TREE. Set *STRICT_OVERFLOW_P to true if the canonicalization is only valid if signed overflow is undefined.
Match A +- CST code arg1 and CST code arg1. We can change the first form only if overflow is undefined.
In principle pointers also have undefined overflow behavior,
but that causes problems elsewhere. Identify the constant in arg0 and its sign.
Overflowed constants and zero will cause problems.
See if we can reduce the magnitude of the constant in arg0 by changing the comparison code.
CST <= arg1 -> CST-1 < arg1.
-CST < arg1 -> -CST-1 <= arg1.
CST > arg1 -> CST-1 >= arg1.
-CST >= arg1 -> -CST-1 > arg1.
arg1 code' CST' might be more canonical.
A - CST < arg1 -> A - CST-1 <= arg1.
A + CST > arg1 -> A + CST-1 >= arg1.
A + CST <= arg1 -> A + CST-1 < arg1.
A - CST >= arg1 -> A - CST-1 > arg1.
Now build the constant reduced in magnitude. But not if that would produce one outside of its types range.
We cannot swap the comparison here as that would cause us to endlessly recurse.
If swapping might yield to a more canonical form, do so.
|
static |
Return false if expr can be assumed not to be an lvalue, true otherwise.
We only need to wrap lvalue tree codes.
Assume the worst for front-end tree codes.
| bool merge_ranges | ( | int * | pin_p, |
| tree * | plow, | ||
| tree * | phigh, | ||
| int | in0_p, | ||
| tree | low0, | ||
| tree | high0, | ||
| int | in1_p, | ||
| tree | low1, | ||
| tree | high1 | ||
| ) |
Given two ranges, see if we can merge them into one. Return 1 if we can, 0 if we can't. Set the output range into the specified parameters.
Make range 0 be the range that starts first, or ends last if they start at the same value. Swap them if it isn't.
Now flag two cases, whether the ranges are disjoint or whether the second range is totally subsumed in the first. Note that the tests below are simplified by the ones above.
We now have four cases, depending on whether we are including or excluding the two ranges.
If they don't overlap, the result is false. If the second range
is a subset it is the result. Otherwise, the range is from the start
of the second to the end of the first. If they don't overlap, the result is the first range. If they are
equal, the result is false. If the second range is a subset of the
first, and the ranges begin at the same place, we go from just after
the end of the second range to the end of the first. If the second
range is not a subset of the first, or if it is a subset and both
ranges end at the same place, the range starts at the start of the
first range and ends just before the second range.
Otherwise, we can't describe this as a single range. We are in the weird situation where high0 > high1 but
high1 has no successor. Punt. low0 < low1 but low1 has no predecessor. Punt.
If they don't overlap, the result is the second range. If the second
is a subset of the first, the result is false. Otherwise,
the range starts just after the first range and ends at the
end of the second. high1 > high0 but high0 has no successor. Punt.
The case where we are excluding both ranges. Here the complex case
is if they don't overlap. In that case, the only time we have a
range is if they are adjacent. If the second is a subset of the
first, the result is the first. Otherwise, the range to exclude
starts at the beginning of the first range and ends at the end of the
second. Canonicalize - [min, x] into - [-, x].
FALLTHROUGH
Canonicalize - [x, max] into - [x, -].
FALLTHROUGH
The ranges might be also adjacent between the maximum and
minimum values of the given type. For
- [{min,-}, x] and - [y, {max,-}] ranges where x + 1 < y
return + [x + 1, y - 1].
References fold_build1_loc, fold_convert_loc(), gcc_assert, negate_expr(), pedantic_non_lvalue_loc(), signed_type_for(), tcc_comparison, TREE_CODE_CLASS, TREE_TYPE, and TYPE_UNSIGNED.
Referenced by optimize_range_tests_diff(), and sign_bit_p().
|
static |
For an expression that has the form (A && B) || ~B or (A || B) && ~B, we can drop one of the inner expressions and simplify to A || ~B or A && ~B LOC is the location of the resulting expression. OP is the inner logical operation; the left-hand side in the examples above, while CMPOP is the right-hand side. RHS_ONLY is used to prevent us from accidentally removing a condition that guards another, as in (A != NULL && A->...) || A == NULL which we must not transform. If RHS_ONLY is true, only eliminate the right-most operand of the inner logical operation.
| int multiple_of_p | ( | ) |
Determine if first argument is a multiple of second argument. Return 0 if it is not, or we cannot easily determined it to be.
An example of the sort of thing we care about (at this point; this routine could surely be made more general, and expanded to do what the *_DIV_EXPR's fold cases do now) is discovering that
SAVE_EXPR (I) * SAVE_EXPR (J * 8)
is a multiple of
SAVE_EXPR (J * 8)
when we know that the two SAVE_EXPR (J * 8) nodes are the same node.
This code also handles discovering that
SAVE_EXPR (I) * SAVE_EXPR (J * 8)
is a multiple of 8 so we don't have to worry about dealing with a possible remainder.
Note that we look inside a SAVE_EXPR only to determine how it was calculated; it is not safe for fold to do much of anything else with the internals of a SAVE_EXPR, since it cannot know when it will be evaluated at run time. For example, the latter example above cannot be implemented as SAVE_EXPR (I) * J or any variant thereof, since the value of J at evaluation time of the original SAVE_EXPR is not necessarily the same at the time the new expression is evaluated. The only optimization of this sort that would be valid is changing
SAVE_EXPR (I) * SAVE_EXPR (SAVE_EXPR (J) * 8)
divided by 8 to
SAVE_EXPR (I) * SAVE_EXPR (J)
(where the same SAVE_EXPR (J) is used in the original and the transformed version).
Bitwise and provides a power of two multiple. If the mask is
a multiple of BOTTOM then TOP is a multiple of BOTTOM.
FALLTHRU
const_binop may not detect overflow correctly,
so check for it explicitly here. Can't handle conversions from non-integral or wider integral type.
.. fall through ...
|
static |
Subroutine of native_encode_expr. Encode the COMPLEX_CST specified by EXPR into the buffer PTR of length LEN bytes. Return the number of bytes placed in the buffer, or zero upon failure.
| int native_encode_expr | ( | ) |
Subroutine of fold_view_convert_expr. Encode the INTEGER_CST, REAL_CST, COMPLEX_CST or VECTOR_CST specified by EXPR into the buffer PTR of length LEN bytes. Return the number of bytes placed in the buffer, or zero upon failure.
Referenced by vn_reference_lookup_3().
|
static |
Subroutine of native_encode_expr. Encode the FIXED_CST specified by EXPR into the buffer PTR of length LEN bytes. Return the number of bytes placed in the buffer, or zero upon failure.
|
static |
Subroutine of native_encode_expr. Encode the INTEGER_CST specified by EXPR into the buffer PTR of length LEN bytes. Return the number of bytes placed in the buffer, or zero upon failure.
Referenced by try_move_mult_to_index().
|
static |
Subroutine of native_encode_expr. Encode the REAL_CST specified by EXPR into the buffer PTR of length LEN bytes. Return the number of bytes placed in the buffer, or zero upon failure.
There are always 32 bits in each long, no matter the size of the hosts long. We handle floating point representations with up to 192 bits.
References CONVERT_EXPR_CODE_P, STRIP_NOPS, and STRIP_SIGN_NOPS.
|
static |
Subroutine of native_encode_expr. Encode the STRING_CST specified by EXPR into the buffer PTR of length LEN bytes. Return the number of bytes placed in the buffer, or zero upon failure.
|
static |
Subroutine of native_encode_expr. Encode the VECTOR_CST specified by EXPR into the buffer PTR of length LEN bytes. Return the number of bytes placed in the buffer, or zero upon failure.
|
static |
Subroutine of native_interpret_expr. Interpret the contents of the buffer PTR of length LEN as a COMPLEX_CST of type TYPE. If the buffer cannot be interpreted, return NULL_TREE.
References fold_convert_loc(), HOST_WIDE_INT_M1U, tree_low_cst(), TYPE_MODE, TYPE_PRECISION, and unsigned_type_for().
| tree native_interpret_expr | ( | ) |
Subroutine of fold_view_convert_expr. Interpret the contents of the buffer PTR of length LEN as a constant of type TYPE. For INTEGRAL_TYPE_P we return an INTEGER_CST, for SCALAR_FLOAT_TYPE_P we return a REAL_CST, etc... If the buffer cannot be interpreted, return NULL_TREE.
References fold_build1_loc, fold_convert_loc(), INTEGRAL_TYPE_P, TREE_OPERAND, and TYPE_PRECISION.
Referenced by vn_reference_lookup_3().
|
static |
Subroutine of native_interpret_expr. Interpret the contents of the buffer PTR of length LEN as a FIXED_CST of type TYPE. If the buffer cannot be interpreted, return NULL_TREE.
|
static |
Subroutine of native_interpret_expr. Interpret the contents of the buffer PTR of length LEN as an INTEGER_CST of type TYPE. If the buffer cannot be interpreted, return NULL_TREE.
|
static |
Subroutine of native_interpret_expr. Interpret the contents of the buffer PTR of length LEN as a REAL_CST of type TYPE. If the buffer cannot be interpreted, return NULL_TREE.
There are always 32 bits in each long, no matter the size of the hosts long. We handle floating point representations with up to 192 bits.
References build_fold_addr_expr_loc(), fold_convert_loc(), get_inner_reference(), offset, TREE_OPERAND, TYPE_MAIN_VARIANT, TYPE_QUALS, and TYPE_UNQUALIFIED.
|
static |
Subroutine of native_interpret_expr. Interpret the contents of the buffer PTR of length LEN as a VECTOR_CST of type TYPE. If the buffer cannot be interpreted, return NULL_TREE.
References fold_build2_loc, fold_convert_loc(), force_fit_type_double(), TREE_OVERFLOW, and tree_to_double_int().
Referenced by build_fold_addr_expr_loc(), fold_negate_expr(), merge_ranges(), and optimize_minmax_comparison().
|
static |
Like fold_negate_expr, but return a NEGATE_EXPR tree, if T can not be negated in a simpler way. Also allow for T to be NULL_TREE, in which case return NULL_TREE.
|
static |
Determine whether an expression T can be cheaply negated using the function negate_expr without introducing undefined overflow.
Check that -CST will not overflow type.
We want to canonicalize to positive real constants. Pretend
that only negative ones can be easily negated. -(A + B) -> (-B) - A.
-(A + B) -> (-A) - B.
We can't turn -(A-B) into B-A when we honor signed zeros.
Fall through.
In general we can't negate A / B, because if A is INT_MIN and
B is 1, we may turn this into INT_MIN / -1 which is undefined
and actually traps on some architectures. But if overflow is
undefined, we can negate, because - (INT_MIN / 1) is an
overflow. If overflow is undefined then we have to be careful because
we ask whether it's ok to associate the negate with the
division which is not ok for example for
-((a - b) / c) where (-(a - b)) / c may invoke undefined
overflow because of negating INT_MIN. So do not use
negate_expr_p here but open-code the two important cases. Negate -((double)float) as (double)(-float).
Negate -f(x) as f(-x).
Optimize -((int)x >> 31) into (unsigned)x >> 31.
|
static |
|
static |
Return true if the built-in mathematical function specified by CODE is odd, i.e. -f(x) == f(-x).
| tree non_lvalue_loc | ( | ) |
Return an expr equal to X but certainly not valid as an lvalue.
While we are in GIMPLE, NON_LVALUE_EXPR doesn't mean anything to us.
| tree omit_one_operand_loc | ( | ) |
Return a tree for the case when the result of an expression is RESULT converted to TYPE and OMITTED was previously an operand of the expression but is now not needed (e.g., we folded OMITTED * 0).
If OMITTED has side effects, we must evaluate it. Otherwise, just do the conversion of RESULT to TYPE.
If the resulting operand is an empty statement, just return the omitted statement casted to void.
Referenced by fold_builtin_fputs(), fold_builtin_n(), and maybe_canonicalize_comparison().
Return a tree for the case when the result of an expression is RESULT converted to TYPE and OMITTED1 and OMITTED2 were previously operands of the expression but are now not needed.
If OMITTED1 or OMITTED2 has side effects, they must be evaluated. If both OMITTED1 and OMITTED2 have side effects, OMITTED1 is evaluated before OMITTED2. Otherwise, if neither has side effects, just do the conversion of RESULT to TYPE.
Referenced by build_call_expr().
|
static |
Similar to operand_equal_p, but see if ARG0 might have been made by shorten_compare from ARG1 when ARG1 was being compared with OTHER.
When in doubt, return 0.
Discard any conversions that don't change the modes of ARG0 and ARG1 and see if the inner values are the same. This removes any signedness comparison, which doesn't matter here.
Duplicate what shorten_compare does to ARG1 and see if that gives the actual comparison operand, ARG0. First throw away any conversions to wider types already present in the operands.
Make sure shorter operand is extended the right way
to match the longer operand.
References eval_subst(), fold_build1_loc, fold_build2_loc, fold_build3_loc, operand_equal_p(), tcc_binary, tcc_comparison, tcc_expression, tcc_unary, and TREE_OPERAND.
| int operand_equal_p | ( | ) |
Return nonzero if two operands (typically of the same tree node) are necessarily equal. If either argument has side-effects this function returns zero. FLAGS modifies behavior as follows:
If OEP_ONLY_CONST is set, only return nonzero for constants. This function tests whether the operands are indistinguishable; it does not test whether they are equal using C's == operation. The distinction is important for IEEE floating point, because (1) -0.0 and 0.0 are distinguishable, but -0.0==0.0, and (2) two NaNs may be indistinguishable, but NaN!=NaN.
If OEP_ONLY_CONST is unset, a VAR_DECL is considered equal to itself even though it may hold multiple values during a function. This is because a GCC tree node guarantees that nothing else is executed between the evaluation of its "operands" (which may often be evaluated in arbitrary order). Hence if the operands themselves don't side-effect, the VAR_DECLs, PARM_DECLs etc... must hold the same value in each operand/subexpression. Hence leaving OEP_ONLY_CONST unset means assuming isochronic (or instantaneous) tree equivalence. Unless comparing arbitrary expression trees, such as from different statements, this flag can usually be left unset.
If OEP_PURE_SAME is set, then pure functions with identical arguments are considered the same. It is used when the caller has other ways to ensure that global memory is unchanged in between.
If either is ERROR_MARK, they aren't equal.
Similar, if either does not have a type (like a released SSA name), they aren't equal.
Check equality of integer constants before bailing out due to precision differences.
If both types don't have the same signedness, then we can't consider them equal. We must check this before the STRIP_NOPS calls because they may change the signedness of the arguments. As pointers strictly don't have a signedness, require either two pointers or two non-pointers as well.
We cannot consider pointers to different address space equal.
If both types don't have the same precision, then it is not safe to strip NOPs.
In case both args are comparisons but with different comparison code, try to swap the comparison operands of one arg to produce a match and compare that variant.
NOP_EXPR and CONVERT_EXPR are considered equal.
This is needed for conversions and for COMPONENT_REF. Might as well play it safe and always test this.
If ARG0 and ARG1 are the same SAVE_EXPR, they are necessarily equal. We don't care about side effects in that case because the SAVE_EXPR takes care of that for us. In all other cases, two expressions are equal if they have no side effects. If we have two identical expressions with side effects that should be treated the same due to the only side effects being identical SAVE_EXPR's, that will be detected in the recursive calls below. If we are taking an invariant address of two identical objects they are necessarily equal as well.
Next handle constant cases, those for which we can return 1 even if ONLY_CONST is set.
If we do not distinguish between signed and unsigned zero,
consider them equal. Define macros to test an operand from arg0 and arg1 for equality and a variant that allows null and views null as being different from any non-null value. In the latter case, if either is null, the both must be; otherwise, do the normal comparison.
Two conversions are equal only if signedness and modes match.
For commutative ops, allow the other order.
If either of the pointer (or reference) expressions we are
dereferencing contain a side effect, these cannot be equal,
but their addresses can be. Require equal extra operands and then fall through to MEM_REF
handling of the two common operands. Fallthru.
Require equal access sizes, and similar pointer types.
We can have incomplete types for array references of
variable-sized arrays from the Fortran frontend
though. Also verify the types are compatible. Operands 2 and 3 may be null.
Compare the array index by value if it is constant first as we
may have different types but same value here. Handle operand 2 the same as for ARRAY_REF. Operand 0
may be NULL when we're called to compare MEM_EXPRs. The multiplcation operands are commutative.
FALLTHRU
Otherwise take into account this is a commutative operation.
If the CALL_EXPRs call different functions, then they
clearly can not be equal. Now see if all the arguments are the same.
If we get here and both argument lists are exhausted
then the CALL_EXPRs are equal. Consider __builtin_sqrt equal to sqrt.
Referenced by allocate_vn_table(), analyze_increments(), analyze_overlapping_iterations(), dse_optimize_stmt(), dump_value_range(), final_range_test_p(), fold_builtin_constant_p(), fold_builtin_fputs(), fold_builtin_logarithm(), fold_range_test(), fold_stmt_inplace(), asan_mem_ref_hasher::hash(), initialize_hash_element_from_expr(), make_invariant_chain(), move_pointer_to_base(), ncd_of_cand_and_phis(), operand_equal_for_comparison_p(), paradoxical_subreg_p(), profitable_increment_p(), rewrite_use_nonlinear_expr(), same_phi_args_p(), valueize_val(), vect_build_slp_tree_1(), vn_reference_lookup_3(), vrp_meet_1(), vrp_val_max(), and vrp_val_min().
|
static |
Optimize a bit-field compare.
There are two cases: First is a compare against a constant and the second is a comparison of two items where the fields are at the same bit position relative to the start of a chunk (byte, halfword, word) large enough to contain it. In these cases we can avoid the shift implicit in bitfield extractions.
For constants, we emit a compare of the shifted constant with the BIT_AND_EXPR of a mask and a byte, halfword, or word of the operand being compared. For two fields at the same position, we do the ANDs with the similar mask and compare the result of the ANDs.
CODE is the comparison code, known to be either NE_EXPR or EQ_EXPR. COMPARE_TYPE is the type of the comparison, and LHS and RHS are the left and right operands of the comparison, respectively.
If the optimization described above can be done, we return the resulting tree. Otherwise we return zero.
Get all the information about the extractions being done. If the bit size if the same as the size of the underlying object, we aren't doing an extraction at all and so can do nothing. We also don't want to do anything if the inner expression is a PLACEHOLDER_EXPR since we then will no longer be able to replace it.
If this is not a constant, we can only do something if bit positions,
sizes, and signedness are the same. See if we can find a mode to refer to this field. We should be able to, but fail if we can't.
Set signed and unsigned types of the precision of this mode for the shifts below.
Compute the bit position and size for the new reference and our offset within it. If the new reference is the same size as the original, we won't optimize anything, so return zero.
Make the mask to be used against the extracted field.
If not comparing with constant, just rework the comparison and return.
Otherwise, we are handling the constant case. See if the constant is too big for the field. Warn and return a tree of for 0 (false) if so. We do this not only for its own sake, but to avoid having to test for this error case below. If we didn't, we might generate wrong code. For unsigned fields, the constant shifted right by the field length should be all zero. For signed fields, the high-order bits should agree with the sign bit.
Single-bit compares should always be against zero.
Make a new bitfield reference, shift the constant over the appropriate number of bits and mask it with the computed mask (in case this was a signed field). If we changed it, make a new one.
References STRIP_NOPS, TREE_CODE, and TREE_OPERAND.
|
static |
Optimize T, which is a comparison of a MIN_EXPR or MAX_EXPR with a constant.
If something does not permit us to optimize, return the original tree.
Now handle all the various comparison codes. We only handle EQ_EXPR and GT_EXPR, doing the rest with recursive calls using logical simplifications.
MAX (X, 0) == 0 -> X <= 0
MAX (X, 0) == 5 -> X == 5
MAX (X, 0) == -1 -> false
MIN (X, 0) == 0 -> X >= 0
MIN (X, 0) == 5 -> false
MIN (X, 0) == -1 -> X == -1
MAX (X, 0) > 0 -> X > 0
MAX (X, 0) > 5 -> X > 5 MAX (X, 0) > -1 -> true
MIN (X, 0) > 0 -> false
MIN (X, 0) > 5 -> false MIN (X, 0) > -1 -> X > -1
References negate_expr(), and TREE_CODE.
|
static |
When pedantic, return an expr equal to X but certainly not valid as a pedantic lvalue. Otherwise, return X.
Referenced by merge_ranges().
|
static |
Similar, but call pedantic_non_lvalue instead of non_lvalue.
If the resulting operand is an empty statement, just return the omitted statement casted to void.
|
static |
Return whether BASE + OFFSET + BITPOS may wrap around the address space. This is used to avoid issuing overflow warnings for expressions like &p->x which can not wrap.
We can do slightly better for SIZE if we have an ADDR_EXPR of an array.
|
inlinestatic |
Similar to protected_set_expr_location, but never modify x in place, if location can and needs to be set, unshare it.
References copy_node, and SET_EXPR_LOCATION.
| bool ptr_difference_const | ( | ) |
Returns true if addresses of E1 and E2 differ by a constant, false otherwise. If they do, E1 - E2 is stored in *DIFF.
If only one of the offsets is non-constant, the difference cannot be a constant.
|
static |
The following functions are subroutines to fold_range_test and allow it to try to change a logical combination of comparisons into a range test.
For example, both X == 2 || X == 3 || X == 4 || X == 5 and X >= 2 && X <= 5 are converted to (unsigned) (X - 2) <= 3
We describe each set of comparisons as being either inside or outside a range, using a variable named like IN_P, and then describe the range with a lower and upper bound. If one of the bounds is omitted, it represents either the highest or lowest value of the type.
In the comments below, we represent a range by two numbers in brackets preceded by a "+" to designate being inside that range, or a "-" to designate being outside that range, so the condition can be inverted by flipping the prefix. An omitted bound is represented by a "-". For example, "- [-, 10]" means being outside the range starting at the lowest possible value and ending at 10, in other words, being greater than 10. The range "+ [-, -]" is always true and hence the range "- [-, -]" is always false.
We set up things so that the missing bounds are handled in a consistent manner so neither a missing bound nor "true" and "false" need to be handled using a special case. Return the result of applying CODE to ARG0 and ARG1, but handle the case of ARG0 and/or ARG1 being omitted, meaning an unlimited range. UPPER0_P and UPPER1_P are nonzero if the respective argument is an upper bound and zero for a lower. TYPE, if nonzero, is the type of the result; it must be specified for a comparison. ARG1 will be converted to ARG0's type if both are specified.
If neither arg represents infinity, do the normal operation. Else, if not a comparison, return infinity. Else handle the special comparison rules. Note that most of the cases below won't occur, but are handled for consistency.
Set SGN[01] to -1 if ARG[01] is a lower bound, 1 for upper, and 0 for neither. In real maths, we cannot assume open ended ranges are the same. But, this is computer arithmetic, where numbers are finite. We can therefore make the transformation of any unbounded range with the value Z, Z being greater than any representable number. This permits us to treat unbounded ranges as equal.
Referenced by simple_operand_p().
|
static |
Return the predecessor of VAL in its type, handling the infinite case.
|
static |
Return the successor of VAL in its type, handling the infinite case.
|
static |
|
static |
Check whether we are allowed to reorder operands arg0 and arg1, such that the evaluation of arg1 occurs before arg0.
| tree round_down_loc | ( | ) |
Likewise, but round down.
See if VALUE is already a multiple of DIVISOR. If so, we don't have to do anything. Only do this when we are not given a const, because in that case, this check is more expensive than just doing it.
If divisor is a power of two, simplify this to bit manipulation.
| tree round_up_loc | ( | ) |
Return the value of VALUE, rounded up to a multiple of DIVISOR. This can only be applied to objects of a sizetype.
See if VALUE is already a multiple of DIVISOR. If so, we don't have to do anything. Only do this when we are not given a const, because in that case, this check is more expensive than just doing it.
If divisor is a power of two, simplify this to bit manipulation.
|
static |
|
static |
Subroutine for fold: determine if VAL is the INTEGER_CONST that represents the sign bit of EXP's type. If EXP represents a sign or zero extension, also test VAL against the unextended type. The return value is the (sub)expression whose sign bit is VAL, or NULL_TREE otherwise.
Tree EXP must have an integral type.
Tree VAL must be an integer constant.
We mask off those bits beyond TREE_TYPE (exp) so that we can treat VAL as if it were unsigned.
Handle extension from a narrower type.
References gcc_unreachable, integer_zerop(), merge_ranges(), NULL_TREE, TREE_CODE, and TYPE_UNSIGNED.
|
static |
|
static |
Subroutine for fold_truth_andor_1: determine if an operand is simple enough to be evaluated unconditionally.
Strip any conversions that don't change the machine mode.
Don't regard global variables as simple. They may be
allocated in ways unknown to the compiler (shared memory,
#pragma weak, etc). Weakrefs are not safe to be read, since they can be NULL.
They are !TREE_PUBLIC && !DECL_EXTERNAL but still
have DECL_WEAK flag set. Loading a static variable is unduly expensive, but global
registers aren't expensive.
References build_int_cst(), integer_one_node, NULL_TREE, and range_binop().
|
static |
Subroutine for fold_truth_andor: determine if an operand is simple enough to be evaluated unconditionally. I addition to simple_operand_p, we assume that comparisons, conversions, and logic-not operations are simple, if their operands are simple, too.
| tree size_binop_loc | ( | ) |
Combine operands OP1 and OP2 with arithmetic operation CODE. CODE is a tree code. The type of the result is taken from the operands. Both must be equivalent integer types, ala int_binop_types_match_p. If the operands are constant, so is the result.
Handle the special case of two integer constants faster.
And some specific cases even faster than that.
Handle general case of two integer constants. For sizetype
constant calculations we always want to know about overflow,
even in the unsigned case.
| tree size_diffop_loc | ( | ) |
Given two values, either both of sizetype or both of bitsizetype, compute the difference between the two values. Return the value in signed type corresponding to the type of the operands.
If the type is already signed, just do the simple thing.
If either operand is not a constant, do the conversions to the signed type and subtract. The hardware will do the right thing with any overflow in the subtraction.
If ARG0 is larger than ARG1, subtract and return the result in CTYPE. Otherwise, subtract the other way, convert to CTYPE (we know that can't overflow) and negate (which can't either). Special-case a result of zero while we're here.
Referenced by expand_expr_real_1().
| tree size_int_kind | ( | ) |
Create a sizetype INT_CST node with NUMBER sign extended. KIND indicates which particular sizetype to create.
|
static |
Returns the pointer to the base of the object addressed by EXP and extracts the information about the offset of the access, storing it to PBITPOS and POFFSET.
|
static |
Split a tree IN into a constant, literal and variable parts that could be combined with CODE to make IN. "constant" means an expression with TREE_CONSTANT but that isn't an actual constant. CODE must be a commutative arithmetic operation. Store the constant part into *CONP, the literal in *LITP and return the variable part. If a part isn't present, set it to null. If the tree does not decompose in this way, return the entire tree as the variable part and the other parts as null.
If CODE is PLUS_EXPR we also split trees that use MINUS_EXPR. In that case, we negate an operand that was subtracted. Except if it is a literal for which we use *MINUS_LITP instead.
If NEGATE_P is true, we are negating all of IN, again except a literal for which we use *MINUS_LITP instead.
If IN is itself a literal or constant, return it as appropriate.
Note that we do not guarantee that any of the three values will be the same type as IN, but they will have the same signedness and mode.
Strip any conversions that don't change the machine mode or signedness.
We can associate addition and subtraction together (even
though the C standard doesn't say so) for integers because
the value is not affected. For reals, the value might be
affected, so we can't. First see if either of the operands is a literal, then a constant.
If we haven't dealt with either operand, this is not a case we can
decompose. Otherwise, VAR is either of the ones remaining, if any. Now do any needed negations.
-X - 1 is folded to ~X, undo that here.
| enum tree_code swap_tree_comparison | ( | ) |
Similar, but return the comparison that results if the operands are swapped. This is safe for floating-point.
References comparison_to_compcode(), COMPCODE_EQ, COMPCODE_LTGT, COMPCODE_NE, COMPCODE_ORD, COMPCODE_TRUE, COMPCODE_UNORD, HONOR_NANS, NULL_TREE, TREE_TYPE, and TYPE_MODE.
Referenced by dump_affine_iv(), fold_range_test(), range_includes_zero_p(), and tree_single_nonnegative_warnv_p().
| bool tree_binary_nonnegative_warnv_p | ( | enum tree_code | code, |
| tree | type, | ||
| tree | op0, | ||
| tree | op1, | ||
| bool * | strict_overflow_p | ||
| ) |
Return true if (CODE OP0 OP1) is known to be non-negative. If the return value is based on the assumption that signed overflow is undefined, set *STRICT_OVERFLOW_P to true; otherwise, don't change *STRICT_OVERFLOW_P.
zero_extend(x) + zero_extend(y) is non-negative if x and y are
both unsigned and at least 2 bits shorter than the result.
x * x is always non-negative for floating point x
or without overflow. zero_extend(x) * zero_extend(y) is non-negative if x and y are
both unsigned and their total bits is shorter than the result. We don't know sign of `t', so be conservative and return false.
| bool tree_binary_nonzero_warnv_p | ( | enum tree_code | code, |
| tree | type, | ||
| tree | op0, | ||
| tree | op1, | ||
| bool * | strict_overflow_p | ||
| ) |
Return true when (CODE OP0 OP1) is an address and is known to be nonzero. For floating point we further ensure that T is not denormal. Similar logic is present in nonzero_address in rtlanal.h.
If the return value is based on the assumption that signed overflow is undefined, set *STRICT_OVERFLOW_P to true; otherwise, don't change *STRICT_OVERFLOW_P.
With the presence of negative values it is hard
to say something.
One of operands must be positive and the other non-negative.
We don't set *STRICT_OVERFLOW_P here: even if this value
overflows, on a twos-complement machine the sum of two
nonnegative numbers can never be zero. When both operands are nonzero, then MAX must be too.
MAX where operand 0 is positive is positive.
MAX where operand 1 is positive is positive.
| bool tree_call_nonnegative_warnv_p | ( | tree | type, |
| tree | fndecl, | ||
| tree | arg0, | ||
| tree | arg1, | ||
| bool * | strict_overflow_p | ||
| ) |
Return true if T is known to be non-negative. If the return value is based on the assumption that signed overflow is undefined, set *STRICT_OVERFLOW_P to true; otherwise, don't change *STRICT_OVERFLOW_P.
Always true.
sqrt(-0.0) is -0.0.
True if the 1st argument is nonnegative.
True if the 1st OR 2nd arguments are nonnegative.
True if the 1st AND 2nd arguments are nonnegative.
True if the 2nd argument is nonnegative.
True if the 1st argument is nonnegative or the second
argument is an even integer. True if the 1st argument is nonnegative or the second
argument is an even integer valued real.
Referenced by symbolic_range_p().
| bool tree_expr_nonnegative_p | ( | ) |
Return true if `t' is known to be non-negative. Handle warnings about undefined signed overflow.
References double_int::high.
Referenced by fold_builtin_int_roundingfn(), and integer_valued_real_p().
| bool tree_expr_nonnegative_warnv_p | ( | ) |
Return true if T is known to be non-negative. If the return value is based on the assumption that signed overflow is undefined, set *STRICT_OVERFLOW_P to true; otherwise, don't change *STRICT_OVERFLOW_P.
Referenced by create_iv().
| bool tree_expr_nonzero_p | ( | ) |
Return true when T is an address and is known to be nonzero. Handle warnings about undefined signed overflow.
| bool tree_expr_nonzero_warnv_p | ( | ) |
Return true when T is an address and is known to be nonzero. For floating point we further ensure that T is not denormal. Similar logic is present in nonzero_address in rtlanal.h.
If the return value is based on the assumption that signed overflow is undefined, set *STRICT_OVERFLOW_P to true; otherwise, don't change *STRICT_OVERFLOW_P.
Doing something useful for floating point would need more work.
Referenced by fold_build2_stat_loc().
| bool tree_invalid_nonnegative_warnv_p | ( | ) |
Return true if T is known to be non-negative. If the return value is based on the assumption that signed overflow is undefined, set *STRICT_OVERFLOW_P to true; otherwise, don't change *STRICT_OVERFLOW_P.
If the initializer is non-void, then it's a normal expression
that will be assigned to the slot.
Otherwise, the initializer sets the slot in some way. One common
way is an assignment statement at the end of the initializer. We don't know sign of `t', so be conservative and return false.
|
static |
Return true if CODE or TYPE is known to be non-negative.
Truth values evaluate to 0 or 1, which is nonnegative unless we have a signed:1 type (where the value is -1 and 0).
| bool tree_single_nonnegative_warnv_p | ( | ) |
Return true if T is known to be non-negative. If the return value is based on the assumption that signed overflow is undefined, set *STRICT_OVERFLOW_P to true; otherwise, don't change *STRICT_OVERFLOW_P.
We don't know sign of `t', so be conservative and return false.
References swap_tree_comparison().
| bool tree_single_nonzero_warnv_p | ( | ) |
Return true when T is an address and is known to be nonzero. For floating point we further ensure that T is not denormal. Similar logic is present in nonzero_address in rtlanal.h.
If the return value is based on the assumption that signed overflow is undefined, set *STRICT_OVERFLOW_P to true; otherwise, don't change *STRICT_OVERFLOW_P.
Weak declarations may link to NULL. Other things may also be NULL so protect with -fdelete-null-pointer-checks; but not variables allocated on the stack.
Constants are never weak.
| bool tree_swap_operands_p | ( | ) |
Test whether it is preferable two swap two operands, ARG0 and ARG1, for example because ARG0 is an integer constant and ARG1 isn't. If REORDER is true, only recommend swapping if we can evaluate the operands in reverse order.
It is preferable to swap two SSA_NAME to ensure a canonical form for commutative and comparison operators. Ensuring a canonical form allows the optimizers to find additional redundancies without having to explicitly check for both orderings.
Put SSA_NAMEs last.
Put variables last.
Referenced by vn_reference_lookup().
| bool tree_unary_nonnegative_warnv_p | ( | enum tree_code | code, |
| tree | type, | ||
| tree | op0, | ||
| bool * | strict_overflow_p | ||
| ) |
Return true if (CODE OP0) is known to be non-negative. If the return value is based on the assumption that signed overflow is undefined, set *STRICT_OVERFLOW_P to true; otherwise, don't change *STRICT_OVERFLOW_P.
We can't return 1 if flag_wrapv is set because ABS_EXPR<INT_MIN> = INT_MIN.
We don't know sign of `t', so be conservative and return false.
References build_fixed(), build_real(), fixed_arithmetic(), FIXED_VALUE_TYPE, force_fit_type_double(), gcc_unreachable, double_int::neg_with_overflow(), NULL, real_value_negate(), TREE_FIXED_CST, TREE_OVERFLOW, TREE_REAL_CST, tree_to_double_int(), TYPE_SATURATING, and TYPE_UNSIGNED.
Referenced by range_is_null().
| bool tree_unary_nonzero_warnv_p | ( | enum tree_code | code, |
| tree | type, | ||
| tree | op0, | ||
| bool * | strict_overflow_p | ||
| ) |
Return true when (CODE OP0) is an address and is known to be nonzero. For floating point we further ensure that T is not denormal. Similar logic is present in nonzero_address in rtlanal.h.
If the return value is based on the assumption that signed overflow is undefined, set *STRICT_OVERFLOW_P to true; otherwise, don't change *STRICT_OVERFLOW_P.
|
static |
Tries to replace &a[idx] p+ s * delta with &a[idx + delta], if s is step of the array. Reconstructs s and delta in the case of s * delta being an integer constant (and thus already folded). ADDR is the address. MULT is the multiplicative expression. If the function succeeds, the new address expression is returned. Otherwise NULL_TREE is returned. LOC is the location of the resulting expression.
Strip the nops that might be added when converting op1 to sizetype.
Canonicalize op1 into a possibly non-constant delta and an INTEGER_CST s.
Simulate we are delta * 1.
Handle &x.array the same as we would handle &x.array[0].
Remember if this was a multi-dimensional array.
Try if delta is a multiple of step.
Only fold here if we can verify we do not overflow one
dimension of a multi-dimensional array. We found a suitable component reference.
Remember if this was a multi-dimensional array.
Try if delta is a multiple of step.
Only fold here if we can verify we do not overflow one
dimension of a multi-dimensional array. We found the suitable array reference. So copy everything up to it, and replace the index.
References double_int_to_tree(), FIXED_VALUE_TYPE, GET_MODE_BITSIZE, GET_MODE_SIZE, HOST_BITS_PER_DOUBLE_INT, native_encode_int(), TREE_FIXED_CST, TREE_TYPE, lang_hooks_for_types::type_for_size, TYPE_MODE, TYPE_PRECISION, and lang_hooks::types.
|
static |
See if ARG is an expression that is either a comparison or is performing arithmetic on comparisons. The comparisons must only be comparing two different values, which will be stored in *CVAL1 and *CVAL2; if they are nonzero it means that some operands have already been found. No variables may be used anywhere else in the expression except in the comparisons. If SAVE_P is true it means we removed a SAVE_EXPR around the expression and save_expr needs to be called with CVAL1 and CVAL2.
If this is true, return 1. Otherwise, return zero.
We can handle some of the tcc_expression cases here.
If we've already found a CVAL1 or CVAL2, this expression is
two complex to handle. First see if we can handle the first operand, then the second. For
the second operand, we know *CVAL1 can't be zero. It must be that
one side of the comparison is each of the values; test for the
case where this isn't true by failing if the two operands
are the same.
|
static |
Subroutine for fold_truth_andor_1: C is an INTEGER_CST interpreted as a P bit value. Arrange things so the extra bits will be set to zero if and only if C is signed-extended to its full width. If MASK is nonzero, it is an INTEGER_CST that should be AND'ed with the extra bits.
We work by getting just the sign bit into the low-order bit, then into the high-order bit, then sign-extend. We then XOR that value with C.
We must use a signed type in order to get an arithmetic right shift. However, we must also avoid introducing accidental overflows, so that a subsequent call to integer_zerop will work. Hence we must do the type conversion here. At this point, the constant is either zero or one, and the conversion to a signed type can never overflow. We could get an overflow if this conversion is done anywhere else.
If necessary, convert the type back to match the type of C.
References integer_pow2p(), and integer_zerop().
|
static |
Helper function for fold_vec_perm. Store elements of VECTOR_CST or CONSTRUCTOR ARG into array ELTS and return true if successful.
References build1, fold_build2_loc, fold_unary_loc(), real_zerop(), and TREE_TYPE.
Variable Documentation
|
static |
If a warning about undefined overflow is deferred, this is the level at which the warning should be emitted.
|
static |
If a warning about undefined overflow is deferred, this is the warning. Note that this may cause us to turn two warnings into one, but that is fine since it is sufficient to only give one warning per expression.
|
static |
This is nonzero if we should defer warnings about undefined overflow. This facility exists because these warnings are a special case. The code to estimate loop iterations does not want to issue any warnings, since it works with expressions which do not occur in user code. Various bits of cleanup code call fold(), but only use the result if it has certain characteristics (e.g., is a constant); that code only wants to issue a warning if the result is used.
Referenced by fold_defer_overflow_warnings().
| int folding_initializer = 0 |
Fold a constant sub-tree into a single node for C-compiler Copyright (C) 1987-2013 Free Software Foundation, Inc.
This file is part of GCC.
GCC is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation; either version 3, or (at your option) any later version.
GCC is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.
You should have received a copy of the GNU General Public License along with GCC; see the file COPYING3. If not see http://www.gnu.org/licenses/. The entry points in this file are fold, size_int_wide and size_binop.
fold takes a tree as argument and returns a simplified tree.
size_binop takes a tree code for an arithmetic operation and two operands that are trees, and produces a tree for the result, assuming the type comes from `sizetype'.
size_int takes an integer value, and creates a tree constant with type from `sizetype'.
Note: Since the folders get called on non-gimple code as well as gimple code, we need to handle GIMPLE tuples as well as their corresponding tree equivalents. Nonzero if we are folding constants inside an initializer; zero otherwise.
| int pedantic_lvalues |
Nonzero means lvalues are limited to those valid in pedantic ANSI C. Zero means allow extended lvalues.
