#include "config.h"#include "system.h"#include "coretypes.h"#include "tm.h"#include "ggc.h"#include "flags.h"#include "tree.h"#include "basic-block.h"#include "gimple.h"#include "gimple-ssa.h"#include "tree-cfg.h"#include "tree-phinodes.h"#include "ssa-iterators.h"#include "tree-ssanames.h"#include "tree-ssa-loop-manip.h"#include "tree-ssa-loop-niter.h"#include "tree-ssa-loop.h"#include "tree-into-ssa.h"#include "tree-ssa.h"#include "tree-pass.h"#include "tree-dump.h"#include "gimple-pretty-print.h"#include "diagnostic-core.h"#include "intl.h"#include "cfgloop.h"#include "tree-scalar-evolution.h"#include "tree-ssa-propagate.h"#include "tree-chrec.h"#include "tree-ssa-threadupdate.h"#include "expr.h"#include "optabs.h"#include "tree-ssa-threadedge.h"
Data Structures | |
| struct | value_range_d |
| struct | assert_locus_d |
| struct | switch_update |
| struct | case_info |
Macros | |
| #define | VR_INITIALIZER { VR_UNDEFINED, NULL_TREE, NULL_TREE, NULL } |
Typedefs | |
| typedef struct value_range_d | value_range_t |
| typedef struct assert_locus_d * | assert_locus_t |
Variables | |
| static sbitmap * | live |
| static bitmap | need_assert_for |
| static assert_locus_t * | asserts_for |
| static unsigned | num_vr_values |
| static value_range_t ** | vr_value |
| static bool | values_propagated |
| static int * | vr_phi_edge_counts |
| static vec< edge > | to_remove_edges |
| static vec< switch_update > | to_update_switch_stmts |
| static vec< tree > | equiv_stack |
Macro Definition Documentation
| #define VR_INITIALIZER { VR_UNDEFINED, NULL_TREE, NULL_TREE, NULL } |
Referenced by check_array_ref(), range_fits_type_p(), and vrp_intersect_ranges_1().
Typedef Documentation
| typedef struct assert_locus_d* assert_locus_t |
| typedef struct value_range_d value_range_t |
Function Documentation
|
static |
If abs (min) < abs (max), set VR to [-max, max], if abs (min) >= abs (max), set VR to [-min, min].
|
static |
Add VAR and VAR's equivalence set to EQUIV. This is the central point where equivalence processing can be turned on/off.
References gcc_assert, is_overflow_infinity(), value_range_d::max, value_range_d::min, TREE_TYPE, value_range_d::type, TYPE_OVERFLOW_UNDEFINED, and VR_RANGE.
|
static |
Given a range VR, a LOOP and a variable VAR, determine whether it would be profitable to adjust VR using scalar evolution information for VAR. If so, update VR with the new limits.
TODO. Don't adjust anti-ranges. An anti-range may provide better opportunities than a regular range, but I'm not sure.
Like in PR19590, scev can return a constant function.
If STEP is symbolic, we can't know whether INIT will be the minimum or maximum value in the range. Also, unless INIT is a simple expression, compare_values and possibly other functions in tree-vrp won't be able to handle it.
... or if it may wrap.
We use TYPE_MIN_VALUE and TYPE_MAX_VALUE here instead of negative_overflow_infinity and positive_overflow_infinity, because we have concluded that the loop probably does not wrap.
Try to use estimated number of iterations for the loop to constrain the final value in the evolution.
We are only entering here for loop header PHI nodes, so using
the number of latch executions is the correct thing to use. If the multiplication overflowed we can't do a meaningful
adjustment. Likewise if the result doesn't fit in the type
of the induction variable. For a signed type we have to
check whether the result has the expected signedness which
is that of the step as number of iterations is unsigned. Likewise if the addition did.
For VARYING or UNDEFINED ranges, just about anything we get
from scalar evolutions should be better. If we would create an invalid range, then just assume we
know absolutely nothing. This may be over-conservative,
but it's clearly safe, and should happen only in unreachable
parts of code, or for invalid programs. INIT is the maximum value. If INIT is lower than VR->MAX
but no smaller than VR->MIN, set VR->MAX to INIT. According to the loop information, the variable does not
overflow. If we think it does, probably because of an
overflow due to arithmetic on a different INF value,
reset now. If INIT is bigger than VR->MIN, set VR->MIN to INIT.
If we just created an invalid range with the minimum
greater than the maximum, we fail conservatively.
This should happen only in unreachable
parts of code, or for invalid programs.
|
inlinestatic |
If VAL is now an overflow infinity, return VAL. Otherwise, return the same value with TREE_OVERFLOW clear. This can be used to avoid confusing a regular value with an overflow value.
|
static |
Given a COND_EXPR COND of the form 'V OP W', and an SSA name V, create a new SSA name N and return the assertion assignment 'V = ASSERT_EXPR <V, V OP W>'.
The new ASSERT_EXPR, creates a new SSA name that replaces the operand of the ASSERT_EXPR. Create it so the new name and the old one are registered in the replacement table so that we can fix the SSA web after adding all the ASSERT_EXPRs.
|
static |
Walk over all statements of all reachable BBs and call check_array_bounds on them.
Skip blocks that were found to be unreachable.
References bitmap_clear_bit(), NULL_TREE, and SSA_NAME_VERSION.
|
static |
walk_tree() callback that checks if *TP is an ARRAY_REF inside an ADDR_EXPR (in which an array subscript one outside the valid range is allowed). Call check_array_ref for each ARRAY_REF found. The location is passed in DATA.
References bitmap_clear_bit(), boolean_false_node, boolean_true_node, and SSA_NAME_VERSION.
|
static |
Checks one ARRAY_REF in REF, located at LOCUS. Ignores flexible arrays and "struct" hacks. If VRP can determine that the array subscript is a constant, check if it is outside valid range. If the array subscript is a RANGE, warn if it is non-overlapping with valid range. IGNORE_OFF_BY_ONE is true if the ARRAY_REF is inside a ADDR_EXPR.
Can not check flexible arrays.
Accesses to trailing arrays via pointers may access storage beyond the types array bounds.
If this is the last field in a struct type or a field in a
union type do not warn.
References dump_file, dump_flags, dump_value_range(), extract_range_basic(), extract_range_from_assignment(), FOR_EACH_SSA_TREE_OPERAND, get_value_range(), gimple_fold_stmt_to_constant(), gimple_get_lhs(), INTEGRAL_TYPE_P, is_overflow_infinity(), NULL, POINTER_TYPE_P, print_generic_expr(), set_value_range(), set_value_range_to_varying(), SSA_OP_DEF, SSA_PROP_INTERESTING, SSA_PROP_NOT_INTERESTING, SSA_PROP_VARYING, TDF_DETAILS, TREE_CODE, TREE_TYPE, value_range_d::type, TYPE_MAX_VALUE, TYPE_MIN_VALUE, update_value_range(), VR_INITIALIZER, VR_RANGE, VR_VARYING, and vrp_valueize().
|
static |
Compare two case labels sorting first by the destination bb index and then by the case value.
Make sure the default label is first in a group.
|
static |
Compare all the value ranges for names equivalent to VAR with VAL using comparison code COMP. Return the same value returned by compare_range_with_value, including the setting of *STRICT_OVERFLOW_P.
Get the set of equivalences for VAR.
Start at -1. Set it to 0 if we do a comparison without relying on overflow, or 1 if all comparisons rely on overflow.
Compare vars' value range with val.
If the equiv set is empty we have done all work we need to do.
If we get different answers from different members
of the equivalence set this check must be in a dead
code region. Folding it to a trap representation
would be correct here. For now just return don't-know.
|
static |
Given a comparison code COMP and names N1 and N2, compare all the ranges equivalent to N1 against all the ranges equivalent to N2 to determine the value of N1 COMP N2. Return the same value returned by compare_ranges. Set *STRICT_OVERFLOW_P to indicate whether we relied on an overflow infinity in the comparison.
Compare the ranges of every name equivalent to N1 against the ranges of every name equivalent to N2.
Use the fake bitmaps if e1 or e2 are not available.
Add N1 and N2 to their own set of equivalences to avoid duplicating the body of the loop just to check N1 and N2 ranges.
If the equivalence sets have a common intersection, then the two names can be compared without checking their ranges.
Start at -1. Set it to 0 if we do a comparison without relying on overflow, or 1 if all comparisons rely on overflow.
Otherwise, compare all the equivalent ranges. First, add N1 and N2 to their own set of equivalences to avoid duplicating the body of the loop just to check N1 and N2 ranges.
If we get different answers from different members
of the equivalence set this check must be in a dead
code region. Folding it to a trap representation
would be correct here. For now just return don't-know. None of the equivalent ranges are useful in computing this comparison.
References CASE_HIGH, CASE_LOW, find_case_label_range(), gcc_checking_assert, gimple_switch_label(), gimple_switch_num_labels(), value_range_d::max, value_range_d::min, NULL_TREE, tree_int_cst_compare(), value_range_d::type, VR_ANTI_RANGE, and VR_RANGE.
|
static |
Given a value range VR, a value VAL and a comparison code COMP, return BOOLEAN_TRUE_NODE if VR COMP VAL always returns true for all the values in VR. Return BOOLEAN_FALSE_NODE if the comparison always returns false. Return NULL_TREE if it is not always possible to determine the value of the comparison. Also set *STRICT_OVERFLOW_P to indicate whether a range with an overflow infinity was used in the test.
Anti-ranges need to be handled separately.
For anti-ranges, the only predicates that we can compute at
compile time are equality and inequality. ~[VAL_1, VAL_2] OP VAL is known if VAL_1 <= VAL <= VAL_2.
EQ_EXPR may only be computed if VR represents exactly
one value. If VAL is not inside VR, then they are always different.
If VR represents exactly one value equal to VAL, then return
false. Otherwise, they may or may not be different.
If VR is to the left of VAL, return true.
If VR is to the right of VAL, return false.
Otherwise, we don't know.
If VR is to the right of VAL, return true.
If VR is to the left of VAL, return false.
Otherwise, we don't know.
References build_int_cst(), EDGE_COUNT, infer_nonnull_range(), NULL_TREE, SSA_NAME_OCCURS_IN_ABNORMAL_PHI, stmt_could_throw_p(), stmt_ends_bb_p(), and TREE_TYPE.
|
static |
Given two numeric value ranges VR0, VR1 and a comparison code COMP:
- Return BOOLEAN_TRUE_NODE if VR0 COMP VR1 always returns true for all the values in the ranges.
- Return BOOLEAN_FALSE_NODE if the comparison always returns false.
- Return NULL_TREE if it is not always possible to determine the value of the comparison.
Also set *STRICT_OVERFLOW_P to indicate whether a range with an overflow infinity was used in the test.
VARYING or UNDEFINED ranges cannot be compared.
Anti-ranges need to be handled separately.
If both are anti-ranges, then we cannot compute any
comparison. These comparisons are never statically computable.
Equality can be computed only between a range and an
anti-range. ~[VAL1, VAL2] == [VAL1, VAL2] is always false. To simplify processing, make VR0 the anti-range.
Simplify processing. If COMP is GT_EXPR or GE_EXPR, switch the operands around and change the comparison code.
Equality may only be computed if both ranges represent
exactly one value. If [V0_MIN, V1_MAX] < [V1_MIN, V1_MAX] then V0 != V1.
If VR0 is completely to the left or completely to the right
of VR1, they are always different. Notice that we need to
make sure that both comparisons yield similar results to
avoid comparing values that cannot be compared at
compile-time. If VR0 and VR1 represent a single value and are identical,
return false. Otherwise, they may or may not be different.
If VR0 is to the left of VR1, return true.
If VR0 is to the right of VR1, return false.
Otherwise, we don't know.
Local functions.
Referenced by vrp_meet_1().
|
static |
Compare values like compare_values_warnv, but treat comparisons of nonconstants which rely on undefined overflow as incomparable.
|
static |
Compare two values VAL1 and VAL2. Return
-2 if VAL1 and VAL2 cannot be compared at compile-time, -1 if VAL1 < VAL2, 0 if VAL1 == VAL2, +1 if VAL1 > VAL2, and +2 if VAL1 != VAL2
This is similar to tree_int_cst_compare but supports pointer values and values that cannot be compared at compile time.
If STRICT_OVERFLOW_P is not NULL, then set *STRICT_OVERFLOW_P to true if the return value is only valid if we assume that signed overflow is undefined.
Below we rely on the fact that VAL1 and VAL2 are both pointers or both integers.
Convert the two values into the same type. This is needed because sizetype causes sign extension even for unsigned types.
If VAL1 and VAL2 are of the form 'NAME [+-] CST' or 'NAME',
return -1 or +1 accordingly. If VAL1 and VAL2 don't use the
same name, return -2. Both values must use the same name.
NAME == NAME
If overflow is defined we cannot simplify more.
NAME < NAME + CST
NAME > NAME - CST
NAME + CST > NAME
NAME + CST1 > NAME + CST2, if CST1 > CST2
NAME + CST1 > NAME - CST2
NAME - CST < NAME
NAME - CST1 < NAME + CST2
NAME - CST1 > NAME - CST2, if CST1 < CST2. Notice that
C1 and C2 are swapped in the call to compare_values. We cannot compare non-constants.
We cannot compare overflowed values, except for overflow
infinities. First see if VAL1 and VAL2 are not the same.
If VAL1 is a lower address than VAL2, return -1.
If VAL1 is a higher address than VAL2, return +1.
If VAL1 is different than VAL2, return +2.
For integer constants we either have already returned -1 or 1
or they are equivalent. We still might succeed in proving
something about non-trivial operands.
|
inlinestatic |
Copy value range FROM into value range TO.
References build_int_cst(), value_range_d::equiv, and set_value_range_to_value().
| DEBUG_FUNCTION void debug_all_asserts | ( | ) |
Dump all the registered assertions for all the names to stderr.
| DEBUG_FUNCTION void debug_all_value_ranges | ( | ) |
Dump all value ranges to stderr.
| void debug_asserts_for | ( | tree | ) |
| DEBUG_FUNCTION void debug_asserts_for | ( | ) |
Dump all the registered assertions for NAME to stderr.
| void debug_value_range | ( | value_range_t * | ) |
| DEBUG_FUNCTION void debug_value_range | ( | ) |
Dump value range VR to stderr.
| void debug_vr_equiv | ( | bitmap | ) |
| void dump_all_asserts | ( | FILE * | ) |
Referenced by find_switch_asserts().
| void dump_all_asserts | ( | ) |
Dump all the registered assertions for all the names to FILE.
References FOR_EACH_IMM_USE_STMT, is_gimple_assign(), edge_def::src, and ui.
| void dump_all_value_ranges | ( | FILE * | ) |
| void dump_all_value_ranges | ( | ) |
Dump value ranges of all SSA_NAMEs to FILE.
| void dump_asserts_for | ( | FILE * | , |
| tree | |||
| ) |
| void dump_asserts_for | ( | ) |
Dump all the registered assertions for NAME to FILE.
References dump_file, and print_generic_expr().
| void dump_value_range | ( | FILE * | , |
| value_range_t * | |||
| ) |
Debugging dumps.
Referenced by check_array_ref().
| void dump_value_range | ( | ) |
Dump value range VR to FILE.
References assert_locus_d::bb, CDI_DOMINATORS, assert_locus_d::comp_code, dominated_by_p(), assert_locus_d::e, EDGE_CRITICAL_P, assert_locus_d::expr, assert_locus_d::next, NULL, operand_equal_p(), si, assert_locus_d::si, and assert_locus_d::val.
| void dump_vr_equiv | ( | FILE * | , |
| bitmap | |||
| ) |
|
static |
Main entry point to VRP (Value Range Propagation). This pass is loosely based on J. R. C. Patterson, ``Accurate Static Branch Prediction by Value Range Propagation,'' in SIGPLAN Conference on Programming Language Design and Implementation, pp. 67-78, 1995. Also available at http://citeseer.ist.psu.edu/patterson95accurate.html
This is essentially an SSA-CCP pass modified to deal with ranges instead of constants.
While propagating ranges, we may find that two or more SSA name have equivalent, though distinct ranges. For instance,
1 x_9 = p_3->a; 2 p_4 = ASSERT_EXPR <p_3, p_3 != 0> 3 if (p_4 == q_2) 4 p_5 = ASSERT_EXPR <p_4, p_4 == q_2>; 5 endif 6 if (q_2)
In the code above, pointer p_5 has range [q_2, q_2], but from the code we can also determine that p_5 cannot be NULL and, if q_2 had a non-varying range, p_5's range should also be compatible with it.
These equivalences are created by two expressions: ASSERT_EXPR and copy operations. Since p_5 is an assertion on p_4, and p_4 was the result of another assertion, then we can use the fact that p_5 and p_4 are equivalent when evaluating p_5's range.
Together with value ranges, we also propagate these equivalences between names so that we can take advantage of information from multiple ranges when doing final replacement. Note that this equivalency relation is transitive but not symmetric.
In the example above, p_5 is equivalent to p_4, q_2 and p_3, but we cannot assert that q_2 is equivalent to p_5 because q_2 may be used in contexts where that assertion does not hold (e.g., in line 6).
TODO, the main difference between this pass and Patterson's is that we do not propagate edge probabilities. We only compute whether edges can be taken or not. That is, instead of having a spectrum of jump probabilities between 0 and 1, we only deal with 0, 1 and DON'T KNOW. In the future, it may be worthwhile to propagate probabilities to aid branch prediction.
??? This ends up using stale EDGE_DFS_BACK for liveness computation. Inserting assertions may split edges which will invalidate EDGE_DFS_BACK.
For visiting PHI nodes we need EDGE_DFS_BACK computed.
ASSERT_EXPRs must be removed before finalizing jump threads as finalizing jump threads calls the CFG cleanup code which does not properly handle ASSERT_EXPRs.
If we exposed any new variables, go ahead and put them into SSA form now, before we handle jump threading. This simplifies interactions between rewriting of _DECL nodes into SSA form and rewriting SSA_NAME nodes into SSA form after block duplication and CFG manipulation.
Remove dead edges from SWITCH_EXPR optimization. This leaves the CFG in a broken state and requires a cfg_cleanup run.
Update SWITCH_EXPR case label vector.
As we may have replaced the default label with a regular one
make sure to make it a real default label again. This ensures
optimal expansion.
|
static |
(COND_OP0 COND_CODE COND_OP1) is a predicate which uses NAME. Extract a suitable test code and value and store them into *CODE_P and *VAL_P so the predicate is normalized to NAME *CODE_P *VAL_P.
If no extraction was possible, return FALSE, otherwise return TRUE.
If INVERT is true, then we invert the result stored into *CODE_P.
Otherwise, we have a comparison of the form NAME COMP VAL or VAL COMP NAME.
If the predicate is of the form VAL COMP NAME, flip
COMP around because we need to register NAME as the
first operand in the predicate. The comparison is of the form NAME COMP VAL, so the
comparison code remains unchanged. Invert the comparison code as necessary.
VRP does not handle float types.
Do not register always-false predicates.
FIXME: this works around a limitation in fold() when dealing with
enumerations. Given 'enum { N1, N2 } x;', fold will not
fold 'if (x > N2)' to 'if (0)'.
|
static |
Try to derive a nonnegative or nonzero range out of STMT relying primarily on generic routines in fold in conjunction with range data. Store the result in *VR
If the call is __builtin_constant_p and the argument is a
function parameter resolve it to false. This avoids bogus
array bound warnings.
??? We could do this as early as inlining is finished.
Both __builtin_ffs* and __builtin_popcount return
[0, prec]. If arg is non-zero, then ffs or popcount
are non-zero. If some high bits are known to be zero,
we can decrease the maximum. __builtin_parity* returns [0, 1].
__builtin_c[lt]z* return [0, prec-1], except for
when the argument is 0, but that is undefined behavior.
On many targets where the CLZ RTL or optab value is defined
for 0 the value is prec, so include that in the range
by default. Handle only the single common value.
Magic value to give up, unless vr0 proves
arg is non-zero. From clz of VR_RANGE minimum we can compute
result maximum. From clz of VR_RANGE maximum we can compute
result minimum. __builtin_ctz* return [0, prec-1], except for
when the argument is 0, but that is undefined behavior.
If there is a ctz optab for this mode and
CTZ_DEFINED_VALUE_AT_ZERO, include that in the range,
otherwise just assume 0 won't be seen. Handle only the two common values.
Magic value to give up, unless vr0 proves
arg is non-zero. If arg is non-zero, then use [0, prec - 1].
If some high bits are known to be zero,
we can decrease the result maximum. For vr0 [0, 0] give up.
__builtin_clrsb* returns [0, prec-1].
Referenced by check_array_ref(), and extract_range_from_cond_expr().
|
static |
Extract value range information from an ASSERT_EXPR EXPR and store it in *VR_P.
Find VAR in the ASSERT_EXPR conditional.
If the predicate is of the form VAR COMP LIMIT, then we just
take LIMIT from the RHS and use the same comparison code. If the predicate is of the form LIMIT COMP VAR, then we need
to flip around the comparison code to create the proper range
for VAR. For pointer arithmetic, we only keep track of pointer equality and inequality.
If LIMIT is another SSA name and LIMIT has a range of its own, try to use LIMIT's range to avoid creating symbolic ranges unnecessarily.
LIMIT's range is only interesting if it has any useful information.
Initially, the new range has the same set of equivalences of VAR's range. This will be revised before returning the final value. Since assertions may be chained via mutually exclusive predicates, we will need to trim the set of equivalences before we are done.
Extract a new range based on the asserted comparison for VAR and LIMIT's value range. Notice that if LIMIT has an anti-range, we will only use it for equality comparisons (EQ_EXPR). For any other kind of assertion, we cannot derive a range from LIMIT's anti-range that can be used to describe the new range. For instance, ASSERT_EXPR <x_2, x_2 <= b_4>. If b_4 is ~[2, 10], then b_4 takes on the ranges [-INF, 1] and [11, +INF]. There is no single range for x_2 that could describe LE_EXPR, so we might as well build the range [b_4, +INF] for it. One special case we handle is extracting a range from a range test encoded as (unsigned)var + CST <= limit.
Make sure to not set TREE_OVERFLOW on the final type
conversion. We are willingly interpreting large positive
unsigned values as negative singed values here. We can transform a max, min range to an anti-range or
vice-versa. Use set_and_canonicalize_value_range which does
this for us. When asserting the equality VAR == LIMIT and LIMIT is another
SSA name, the new range will also inherit the equivalence set
from LIMIT. As described above, when LIMIT's range is an anti-range and
this assertion is an inequality (NE_EXPR), then we cannot
derive anything from the anti-range. For instance, if
LIMIT's range was ~[0, 0], the assertion 'VAR != LIMIT' does
not imply that VAR's range is [0, 0]. So, in the case of
anti-ranges, we just assert the inequality using LIMIT and
not its anti-range.
If LIMIT_VR is a range, we can only use it to build a new
anti-range if LIMIT_VR is a single-valued range. For
instance, if LIMIT_VR is [0, 1], the predicate
VAR != [0, 1] does not mean that VAR's range is ~[0, 1].
Rather, it means that for value 0 VAR should be ~[0, 0]
and for value 1, VAR should be ~[1, 1]. We cannot
represent these ranges.
The only situation in which we can build a valid
anti-range is when LIMIT_VR is a single-valued range
(i.e., LIMIT_VR->MIN == LIMIT_VR->MAX). In that case,
build the anti-range ~[LIMIT_VR->MIN, LIMIT_VR->MAX]. In any other case, we cannot use LIMIT's range to build a
valid anti-range. If MIN and MAX cover the whole range for their type, then
just use the original LIMIT. If LIMIT_VR is of the form [N1, N2], we need to build the
range [MIN, N2] for LE_EXPR and [MIN, N2 - 1] for
LT_EXPR. If the maximum value forces us to be out of bounds, simply punt.
It would be pointless to try and do anything more since this
all should be optimized away above us. For LT_EXPR, we create the range [MIN, MAX - 1].
If LIMIT_VR is of the form [N1, N2], we need to build the
range [N1, MAX] for GE_EXPR and [N1 + 1, MAX] for
GT_EXPR. If the minimum value forces us to be out of bounds, simply punt.
It would be pointless to try and do anything more since this
all should be optimized away above us. For GT_EXPR, we create the range [MIN + 1, MAX].
Finally intersect the new range with what we already know about var.
Referenced by extract_range_from_cond_expr().
|
static |
Try to compute a useful range out of assignment STMT and store it in *VR.
Referenced by check_array_ref(), and range_fits_type_p().
|
static |
Extract range information from a binary expression OP0 CODE OP1 based on the ranges of each of its operands with resulting type EXPR_TYPE. The resulting range is stored in *VR.
Get value ranges for each operand. For constant operands, create a new value range with the operand to simplify processing.
References set_value_range_to_varying().
Referenced by extract_range_from_cond_expr().
|
static |
Extract range information from a binary operation CODE based on the ranges of each of its operands, *VR0 and *VR1 with resulting type EXPR_TYPE. The resulting range is stored in *VR.
Not all binary expressions can be applied to ranges in a meaningful way. Handle only arithmetic operations.
If both ranges are UNDEFINED, so is the result.
If one of the ranges is UNDEFINED drop it to VARYING for the following code. At some point we may want to special-case operations that have UNDEFINED result for all or some value-ranges of the not UNDEFINED operand.
Now canonicalize anti-ranges to ranges when they are not symbolic and express ~[] op X as ([]' op X) U ([]'' op X).
Likewise for X op ~[].
The type of the resulting value range defaults to VR0.TYPE.
Refuse to operate on VARYING ranges, ranges of different kinds and symbolic ranges. As an exception, we allow BIT_AND_EXPR because we may be able to derive a useful range even if one of the operands is VR_VARYING or symbolic range. Similarly for divisions. TODO, we may be able to derive anti-ranges in some cases.
Now evaluate the expression to determine the new range.
For MIN/MAX expressions with pointers, we only care about
nullness, if both are non null, then the result is nonnull.
If both are null, then the result is null. Otherwise they
are varying. For pointer types, we are really only interested in asserting
whether the expression evaluates to non-NULL. For pointer types, we are really only interested in asserting
whether the expression evaluates to non-NULL. For integer ranges, apply the operation to each end of the range and see what we end up with.
If we have a PLUS_EXPR with two VR_RANGE integer constant
ranges compute the precise range for such case if possible. We need as many bits as the possibly unsigned inputs.
Check for overflow in double_int.
For non-wrapping arithmetic look at possibly smaller
value-ranges of the type. Check for type overflow.
If overflow wraps, truncate the values and adjust the
range kind and bounds appropriately. No overflow or both overflow or underflow. The
range kind stays VR_RANGE. Underflow and overflow, drop to VR_VARYING.
Min underflow or max overflow. The range kind
changes to VR_ANTI_RANGE. If the anti-range would cover nothing, drop to varying.
Likewise if the anti-range bounds are outside of the
types values. If overflow does not wrap, saturate to the types min/max
value. For other cases, for example if we have a PLUS_EXPR with two
VR_ANTI_RANGEs, drop to VR_VARYING. It would take more effort
to compute a precise range for such a case.
??? General even mixed range kind operations can be expressed
by for example transforming ~[3, 5] + [1, 2] to range-only
operations and a union primitive:
[-INF, 2] + [1, 2] U [5, +INF] + [1, 2]
[-INF+1, 4] U [6, +INF(OVF)]
though usually the union is not exactly representable with
a single range or anti-range as the above is
[-INF+1, +INF(OVF)] intersected with ~[5, 5]
but one could use a scheme similar to equivalences for this. For operations that make the resulting range directly
proportional to the original ranges, apply the operation to
the same end of each range. Fancy code so that with unsigned, [-3,-1]*[-3,-1] does not
drop to varying. Canonicalize the intervals.
Sort the 4 products.
Max - min.
the range covers all values.
The following should handle the wrapping and selecting
VR_ANTI_RANGE for us. If we have an unsigned MULT_EXPR with two VR_ANTI_RANGEs,
drop to VR_VARYING. It would take more effort to compute a
precise range for such a case. For example, if we have
op0 == 65536 and op1 == 65536 with their ranges both being
~[0,0] on a 32-bit machine, we would have op0 * op1 == 0, so
we cannot claim that the product is in ~[0,0]. Note that we
are guaranteed to have vr0.type == vr1.type at this
point. If we have a RSHIFT_EXPR with any shift values outside [0..prec-1],
then drop to VR_VARYING. Outside of this range we get undefined
behavior from the shift operation. We cannot even trust
SHIFT_COUNT_TRUNCATED at this stage, because that applies to rtl
shifts, and the operation at the tree level may be widened. We can map lshifts by constants to MULT_EXPR handling.
We have to use a wrapping multiply though as signed overflow
on lshifts is implementation defined in C89. If bound_shift == HOST_BITS_PER_DOUBLE_INT, the llshift can
overflow. However, for that to happen, vr1.max needs to be
zero, which means vr1 is a singleton range of zero, which
means it should be handled by the previous LSHIFT_EXPR
if-clause. [5, 6] << [1, 2] == [10, 24].
We're shifting out only zeroes, the value increases
monotonically. [0xffffff00, 0xffffffff] << [1, 2]
== [0xfffffc00, 0xfffffffe]. We're shifting out only ones, the value decreases
monotonically. [-1, 1] << [1, 2] == [-4, 4].
For non-negative numbers, we're shifting out only
zeroes, the value increases monotonically.
For negative numbers, we're shifting out only ones, the
value decreases monotomically. For division, if op1 has VR_RANGE but op0 does not, something
can be deduced just from that range. Say [min, max] / [4, max]
gives [min / 4, max / 4] range. For divisions, if flag_non_call_exceptions is true, we must
not eliminate a division by zero. For divisions, if op0 is VR_RANGE, we can deduce a range
even if op1 is VR_VARYING, VR_ANTI_RANGE, symbolic or can
include 0. For unsigned division or when divisor is known
to be non-negative, the range has to cover
all numbers from 0 to max for positive max
and all numbers from min to 0 for negative min. Otherwise the range is -max .. max or min .. -min
depending on which bound is bigger in absolute value,
as the division can change the sign. Compute MAX <|vr1.min|, |vr1.max|> - 1.
If the dividend is non-negative the modulus will be
non-negative as well. If both input ranges contain only negative values we can
truncate the result range maximum to the minimum of the
input range maxima. If either input range contains only non-negative values
we can truncate the result range maximum to the respective
maximum of the input range. If the input ranges contain only positive values we can
truncate the minimum of the result range to the maximum
of the input range minima. If either input range contains only negative values
we can truncate the minimum of the result range to the
respective minimum range. If the range has all positive or all negative values the
result is better than VARYING. If either MIN or MAX overflowed, then set the resulting range to VARYING. But we do accept an overflow infinity representation.
We punt if: 1) [-INF, +INF] 2) [-INF, +-INF(OVF)] 3) [+-INF(OVF), +INF] 4) [+-INF(OVF), +-INF(OVF)] We learn nothing when we have INF and INF(OVF) on both sides. Note that we do accept [-INF, -INF] and [+INF, +INF] without overflow.
If the new range has its limits swapped around (MIN > MAX),
then the operation caused one of them to wrap around, mark
the new range VARYING.
|
static |
Extract range information from a comparison expression EXPR based on the range of its operand and the expression code.
A disadvantage of using a special infinity as an overflow representation is that we lose the ability to record overflow when we don't have an infinity. So we have to ignore a result which relies on overflow.
Since this expression was found on the RHS of an assignment,
its type may be different from _Bool. Convert VAL to EXPR's
type. The result of a comparison is always true or false.
References analyze_scalar_evolution(), value_range_d::equiv, evolution_part_in_loop_num(), initial_condition_in_loop_num(), instantiate_parameters(), is_gimple_min_invariant(), value_range_d::max, value_range_d::min, loop::num, op_with_constant_singleton_value_range(), set_value_range_to_value(), TREE_CODE, value_range_d::type, type(), and VR_ANTI_RANGE.
Referenced by extract_range_from_cond_expr().
|
static |
Extract range information from a conditional expression STMT based on the ranges of each of its operands and the expression code.
Get value ranges for each operand. For constant operands, create a new value range with the operand to simplify processing.
The resulting value range is the union of the operand ranges
References extract_range_basic(), extract_range_from_assert(), extract_range_from_binary_expr(), extract_range_from_comparison(), extract_range_from_ssa_name(), extract_range_from_unary_expr(), get_gimple_rhs_class(), gimple_assign_rhs1(), gimple_assign_rhs2(), gimple_assign_rhs_code(), gimple_expr_type(), GIMPLE_SINGLE_RHS, is_gimple_min_invariant(), NULL, set_value_range_to_value(), set_value_range_to_varying(), tcc_binary, tcc_comparison, tcc_unary, TREE_CODE_CLASS, value_range_d::type, and VR_VARYING.
|
static |
Helper to extract a value-range *VR for a multiplicative operation *VR0 CODE *VR1.
Multiplications, divisions and shifts are a bit tricky to handle, depending on the mix of signs we have in the two ranges, we need to operate on different values to get the minimum and maximum values for the new range. One approach is to figure out all the variations of range combinations and do the operations. However, this involves several calls to compare_values and it is pretty convoluted. It's simpler to do the 4 operations (MIN0 OP MIN1, MIN0 OP MAX1, MAX0 OP MIN1 and MAX0 OP MAX0 OP MAX1) and then figure the smallest and largest values to form the new range.
Compute the 4 cross operations.
Set MIN to the minimum of VAL[i] and MAX to the maximum of VAL[i].
If we found an overflowed value, set MIN and MAX
to it so that we set the resulting range to
VARYING. If either MIN or MAX overflowed, then set the resulting range to VARYING. But we do accept an overflow infinity representation.
We punt if: 1) [-INF, +INF] 2) [-INF, +-INF(OVF)] 3) [+-INF(OVF), +INF] 4) [+-INF(OVF), +-INF(OVF)] We learn nothing when we have INF and INF(OVF) on both sides. Note that we do accept [-INF, -INF] and [+INF, +INF] without overflow.
If the new range has its limits swapped around (MIN > MAX),
then the operation caused one of them to wrap around, mark
the new range VARYING.
|
static |
Extract range information from SSA name VAR and store it in VR. If VAR has an interesting range, use it. Otherwise, create the range [VAR, VAR] and return it. This is useful in situations where we may have conditionals testing values of VARYING names. For instance,
x_3 = y_5; if (x_3 > y_5) ...
Even if y_5 is deemed VARYING, we can determine that x_3 > y_5 is always false.
Referenced by extract_range_from_cond_expr().
|
static |
Extract range information from a unary expression CODE OP0 based on the range of its operand with resulting type TYPE. The resulting range is stored in *VR.
Get value ranges for the operand. For constant operands, create a new value range with the operand to simplify processing.
Referenced by extract_range_from_cond_expr().
|
static |
Extract range information from a unary operation CODE based on the range of its operand *VR0 with type OP0_TYPE with resulting type TYPE. The The resulting range is stored in *VR.
VRP only operates on integral and pointer types.
If VR0 is UNDEFINED, so is the result.
Handle operations that we express in terms of others.
PAREN_EXPR is a simple copy.
-X is simply 0 - X, so re-use existing code that also handles
anti-ranges fine. ~X is simply -1 - X, so re-use existing code that also handles
anti-ranges fine. Now canonicalize anti-ranges to ranges when they are not symbolic and express op ~[] as (op []') U (op []'').
If the expression evaluates to a pointer, we are only interested in
determining if it evaluates to NULL [0, 0] or non-NULL (~[0, 0]). If VR0 is varying and we increase the type precision, assume
a full range for the following transformation. If VR0 is a constant range or anti-range and the conversion is
not truncating we can convert the min and max values and
canonicalize the resulting range. Otherwise we can do the
conversion if the size of the range is less than what the
precision of the target type can represent and the range is
not an anti-range. Pass through vr0 in the easy cases.
For the remaining varying or symbolic ranges we can't do anything
useful. -TYPE_MIN_VALUE = TYPE_MIN_VALUE with flag_wrapv so we can't get a
useful range. ABS_EXPR may flip the range around, if the original range
included negative values. We shouldn't generate [+INF, +INF] as set_value_range
doesn't like this and ICEs. If a VR_ANTI_RANGEs contains zero, then we have
~[-INF, min(MIN, MAX)]. Take the lower of the two values.
Create ~[-INF, min (abs(MIN), abs(MAX))]
or ~[-INF + 1, min (abs(MIN), abs(MAX))] when
flag_wrapv is set and the original anti-range doesn't include
TYPE_MIN_VALUE, remember -TYPE_MIN_VALUE = TYPE_MIN_VALUE. All else has failed, so create the range [0, INF], even for
flag_wrapv since TYPE_MIN_VALUE is in the original
anti-range. If the range contains zero then we know that the minimum value in the
range will be zero. If the range was reversed, swap MIN and MAX.
If the new range has its limits swapped around (MIN > MAX),
then the operation caused one of them to wrap around, mark
the new range VARYING. For unhandled operations fall back to varying.
|
static |
We identified all the jump threading opportunities earlier, but could not transform the CFG at that time. This routine transforms the CFG and arranges for the dominator tree to be rebuilt if necessary.
Note the SSA graph update will occur during the normal TODO processing by the pass manager.
|
static |
Do an RPO walk over the function computing SSA name liveness on-the-fly and deciding on assert expressions to insert. Returns true if there are assert expressions to be inserted.
Process BB and update the live information with uses in
this block.
Merge liveness into the predecessor blocks and free it.
Record the RPO number of the last visited block that needs
live information from this block. We can free all successors live bitmaps if all their
predecessors have been visited already.
References walk_stmt_info::info.
|
static |
Traverse all the statements in block BB looking for statements that may generate useful assertions for the SSA names in their operand. If a statement produces a useful assertion A for name N_i, then the list of assertions already generated for N_i is scanned to determine if A is actually needed.
If N_i already had the assertion A at a location dominating the current location, then nothing needs to be done. Otherwise, the new location for A is recorded instead.
1- For every statement S in BB, all the variables used by S are added to bitmap FOUND_IN_SUBGRAPH.
2- If statement S uses an operand N in a way that exposes a known value range for N, then if N was not already generated by an ASSERT_EXPR, create a new assert location for N. For instance, if N is a pointer and the statement dereferences it, we can assume that N is not NULL.
3- COND_EXPRs are a special case of #2. We can derive range information from the predicate but need to insert different ASSERT_EXPRs for each of the sub-graphs rooted at the conditional block. If the last statement of BB is a conditional expression of the form 'X op Y', then
a) Remove X and Y from the set FOUND_IN_SUBGRAPH.
b) If the conditional is the only entry point to the sub-graph corresponding to the THEN_CLAUSE, recurse into it. On return, if X and/or Y are marked in FOUND_IN_SUBGRAPH, then an ASSERT_EXPR is added for the corresponding variable.
c) Repeat step (b) on the ELSE_CLAUSE.
d) Mark X and Y in FOUND_IN_SUBGRAPH.
For instance,
if (a == 9)
b = a;
else
b = c + 1;
In this case, an assertion on the THEN clause is useful to determine that 'a' is always 9 on that edge. However, an assertion on the ELSE clause would be unnecessary.
4- If BB does not end in a conditional expression, then we recurse into BB's dominator children.
At the end of the recursive traversal, every SSA name will have a list of locations where ASSERT_EXPRs should be added. When a new location for name N is found, it is registered by calling register_new_assert_for. That function keeps track of all the registered assertions to prevent adding unnecessary assertions. For instance, if a pointer P_4 is dereferenced more than once in a dominator tree, only the location dominating all the dereference of P_4 will receive an ASSERT_EXPR.
If this function returns true, then it means that there are names for which we need to generate ASSERT_EXPRs. Those assertions are inserted by process_assert_insertions.
If BB's last statement is a conditional statement involving integer operands, determine if we need to add ASSERT_EXPRs.
If BB's last statement is a switch statement involving integer operands, determine if we need to add ASSERT_EXPRs.
Traverse all the statements in BB marking used names and looking for statements that may infer assertions for their used operands.
See if we can derive an assertion for any of STMT's operands.
If op is not live beyond this stmt, do not bother to insert
asserts for it. If OP is used in such a way that we can infer a value
range for it, and we don't find a previous assertion for
it, create a new assertion location node for OP. If we are able to infer a nonzero value range for OP,
then walk backwards through the use-def chain to see if OP
was set via a typecast.
If so, then we can also infer a nonzero value range
for the operand of the NOP_EXPR. Note we want to register the assert for the
operand of the NOP_EXPR after SI, not after the
conversion. Update live.
Traverse all PHI nodes in BB, updating live.
References TREE_NO_WARNING, and warning_at().
|
static |
Searches the case label vector VEC for the index *IDX of the CASE_LABEL that includes the value VAL. The search is restricted to the range [START_IDX, n - 1] where n is the size of VEC.
If there is a CASE_LABEL for VAL, its index is placed in IDX and true is returned.
If there is no CASE_LABEL for VAL and there is one that is larger than VAL, it is placed in IDX and false is returned.
If VAL is larger than any CASE_LABEL, n is placed on IDX and false is returned.
Find case label for minimum of the value range or the next one. At each iteration we are searching in [low, high - 1].
Note that i != high, so we never ask for n.
Cache the result of comparing CASE_LOW and val.
Ranges cannot be empty.
References int_const_binop(), integer_one_node, and TREE_CODE.
|
static |
Searches the case label vector VEC for the range of CASE_LABELs that is used for values between MIN and MAX. The first index is placed in MIN_IDX. The last index is placed in MAX_IDX. If the range of CASE_LABELs is empty then MAX_IDX < MIN_IDX. Returns true if the default label is not needed.
Only the default case label reached. Return an empty range.
If the case label range is continuous, we do not need
the default case label. Verify that.
Referenced by compare_names().
|
static |
Searches the case label vector VEC for the ranges of CASE_LABELs that are used in range VR. The indices are placed in MIN_IDX1, MAX_IDX, MIN_IDX2 and MAX_IDX2. If the ranges of CASE_LABELs are empty then MAX_IDX1 < MIN_IDX1. Returns true if the default label is not needed.
Set second range to emtpy.
Set first range to all case labels.
Make sure all the values of case labels [i , j] are contained in range [MIN, MAX].
If the range spans case labels [i, j], the corresponding anti-range spans the labels [1, i - 1] and [j + 1, n - 1].
|
static |
Determine whether the outgoing edges of BB should receive an ASSERT_EXPR for each of the operands of BB's LAST statement. The last statement of BB must be a COND_EXPR.
If any of the sub-graphs rooted at BB have an interesting use of the predicate operands, an assert location node is added to the list of assertions for the corresponding operands.
Look for uses of the operands in each of the sub-graphs rooted at BB. We need to check each of the outgoing edges separately, so that we know what kind of ASSERT_EXPR to insert.
Register the necessary assertions for each operand in the
conditional predicate.
Referenced by register_edge_assert_for_1().
|
static |
Determine whether the outgoing edges of BB should receive an ASSERT_EXPR for each of the operands of BB's LAST statement. The last statement of BB must be a SWITCH_EXPR.
If any of the sub-graphs rooted at BB have an interesting use of the predicate operands, an assert location node is added to the list of assertions for the corresponding operands.
Work around GCC 3.4 bug (PR 37086).
Build a vector of case labels sorted by destination label.
If there are multiple case labels with the same destination
we need to combine them to a single value range for the edge. Skip labels until the last of the group.
Pick up the maximum of the case label range.
Nothing to do if the range includes the default label until we
can register anti-ranges. Find the edge to register the assert expr on.
Register the necessary assertions for the operand in the
SWITCH_EXPR.
References cfun, dump_all_asserts(), dump_file, dump_flags, EXECUTE_IF_SET_IN_BITMAP, gcc_assert, gsi_commit_edge_inserts(), assert_locus_d::next, process_assert_insertions_for(), ssa_name, statistics_counter_event(), and TDF_DETAILS.
Referenced by register_edge_assert_for_1().
|
static |
If the statement pointed by SI has a predicate whose value can be computed using the value range information computed by VRP, compute its value and return true. Otherwise, return false.
|
inlinestatic |
Return false if EXPR is a predicate expression involving floating point values.
Referenced by register_edge_assert_for_1().
|
static |
|
static |
Return value range information for VAR.
If we have no values ranges recorded (ie, VRP is not running), then return NULL. Otherwise create an empty range if none existed for VAR.
If we have no recorded ranges, then return NULL.
If we query the range for a new SSA name return an unmodifiable VARYING. We should get here at most from the substitute-and-fold stage which will never try to change values.
After propagation finished do not allocate new value-ranges.
Create a default value range.
Defer allocating the equivalence set.
If VAR is a default definition of a parameter, the variable can take any value in VAR's type.
Try to use the "nonnull" attribute to create ~[0, 0]
anti-ranges for pointers. Note that this is only valid with
default definitions of PARM_DECLs.
Referenced by check_array_ref(), gimple_stmt_nonnegative_warnv_p(), insert_range_assertions(), value_inside_range(), and vrp_intersect_ranges_1().
|
inlinestatic |
Helper that gets the value range of the SSA_NAME with version I or a symbolic range containing the SSA_NAME only if the value range is varying or undefined.
If name N_i does not have a valid range, use N_i as its own range. This allows us to compare against names that may have N_i in their ranges.
References CASE_LOW, gimple_switch_label(), gimple_switch_num_labels(), and tree_int_cst_compare().
|
static |
Return true if the result of assignment STMT is know 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.
|
static |
Return true if the result of assignment STMT is know to be non-zero. 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.
References is_gimple_min_invariant(), TREE_CODE, and TREE_OPERAND.
|
static |
Return true if return value of call STMT is know 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.
|
static |
Return true if STMT is know to to compute a non-negative value. 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.
References get_base_address(), get_value_range(), gimple_assign_rhs1(), NULL_TREE, range_is_nonnull(), TREE_CODE, and TREE_OPERAND.
|
static |
Return true if STMT is known to compute a non-zero value. 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 |
Blocks which have more than one predecessor and more than one successor present jump threading opportunities, i.e., when the block is reached from a specific predecessor, we may be able to determine which of the outgoing edges will be traversed. When this optimization applies, we are able to avoid conditionals at runtime and we may expose secondary optimization opportunities.
This routine is effectively a driver for the generic jump threading code. It basically just presents the generic code with edges that may be suitable for jump threading.
Unlike DOM, we do not iterate VRP if jump threading was successful. While iterating may expose new opportunities for VRP, it is expected those opportunities would be very limited and the compile time cost to expose those opportunities would be significant.
As jump threading opportunities are discovered, they are registered for later realization.
Ugh. When substituting values earlier in this pass we can wipe the dominance information. So rebuild the dominator information as we need it within the jump threading code.
We do not allow VRP information to be used for jump threading across a back edge in the CFG. Otherwise it becomes too difficult to avoid eliminating loop exit tests. Of course EDGE_DFS_BACK is not accurate at this time so we have to recompute it.
Do not thread across edges we are about to remove. Just marking them as EDGE_DFS_BACK will do.
Allocate our unwinder stack to unwind any temporary equivalences that might be recorded.
To avoid lots of silly node creation, we create a single conditional and just modify it in-place when attempting to thread jumps.
Walk through all the blocks finding those which present a potential jump threading opportunity. We could set this up as a dominator walker and record data during the walk, but I doubt it's worth the effort for the classes of jump threading opportunities we are trying to identify at this point in compilation.
If the generic jump threading code does not find this block
interesting, then there is nothing to do. We only care about blocks ending in a COND_EXPR. While there
may be some value in handling SWITCH_EXPR here, I doubt it's
terribly important. We're basically looking for a switch or any kind of conditional with
integral or pointer type arguments. Note the type of the second
argument will be the same as the first argument, so no need to
check it explicitly. We've got a block with multiple predecessors and multiple
successors which also ends in a suitable conditional or
switch statement. For each predecessor, see if we can thread
it to a specific successor. Do not thread across back edges or abnormal edges
in the CFG. We do not actually update the CFG or SSA graphs at this point as ASSERT_EXPRs are still in the IL and cfg cleanup code does not yet handle ASSERT_EXPRs gracefully.
|
static |
If OP can be inferred to be non-zero after STMT executes, return true.
We can only assume that a pointer dereference will yield non-NULL if -fdelete-null-pointer-checks is enabled.
If "nonnull" wasn't specified, we know nothing about
the argument. If "nonnull" applies to all the arguments, then ARG
is non-null. Now see if op appears in the nonnull list.
References double_int_one.
Referenced by compare_range_with_value().
|
static |
If the range of values taken by OP can be inferred after STMT executes, return the comparison code (COMP_CODE_P) and value (VAL_P) that describes the inferred range. Return true if a range could be inferred.
Do not attempt to infer anything in names that flow through abnormal edges.
Similarly, don't infer anything from statements that may throw exceptions. ??? Relax this requirement?
If STMT is the last statement of a basic block with no successors, there is no point inferring anything about any of its operands. We would not be able to find a proper insertion point for the assertion, anyway.
References build1, build2, CONVERT_EXPR_CODE_P, dump_file, gimple_assign_cast_p(), gimple_assign_rhs1(), gimple_assign_rhs2(), gimple_assign_rhs_code(), gimple_expr_type(), has_single_use(), INTEGRAL_TYPE_P, is_gimple_assign(), live_on_edge(), NULL_TREE, SSA_NAME_DEF_STMT, TREE_CODE, TREE_TYPE, TYPE_PRECISION, and TYPE_UNSIGNED.
Referenced by register_edge_assert_for_1().
|
static |
Traverse the flowgraph looking for conditional jumps to insert range expressions. These range expressions are meant to provide information to optimizations that need to reason in terms of value ranges. They will not be expanded into RTL. For instance, given:
x = ... y = ... if (x < y) y = x - 2; else x = y + 3;
this pass will transform the code into:
x = ... y = ... if (x < y) { x = ASSERT_EXPR <x, x < y> y = x - 2 } else { y = ASSERT_EXPR <y, x <= y> x = y + 3 }
The idea is that once copy and constant propagation have run, other optimizations will be able to determine what ranges of values can 'x' take in different paths of the code, simply by checking the reaching definition of 'x'.
References FOR_EACH_SSA_TREE_OPERAND, get_value_range(), prop_set_simulate_again(), set_value_range_to_varying(), and SSA_OP_DEF.
|
static |
Intersect the two value-ranges { *VR0TYPE, *VR0MIN, *VR0MAX } and { VR1TYPE, VR0MIN, VR0MAX } and store the result in { *VR0TYPE, *VR0MIN, *VR0MAX }. This may not be the smallest possible such range. The resulting range is not canonicalized.
[] is vr0, () is vr1 in the following classification comments.
[( )]
Nothing to do for equal ranges.
For anti-range with range intersection the result is empty.
[ ] ( ) or ( ) [ ]
If the ranges have an empty intersection, the result of the
intersect operation is the range for intersecting an
anti-range with a range or empty when intersecting two ranges. If the anti-ranges are adjacent to each other merge them.
Else arbitrarily take VR0.
[ ( ) ] or [( ) ] or [ ( )]
If both are ranges the result is the inner one.
Choose the right gap if the left one is empty.
Choose the left gap if the right one is empty.
Choose the anti-range if the range is effectively varying.
Else choose the range.
If both are anti-ranges the result is the outer one.
The intersection is empty.
( [ ] ) or ([ ] ) or ( [ ])
Choose the inner range.
Choose the right gap if the left is empty.
Choose the left gap if the right is empty.
Choose the anti-range if the range is effectively varying.
Else choose the range.
If both are anti-ranges the result is the outer one.
The intersection is empty.
[ ( ] ) or [ ]( )
( [ ) ] or ( )[ ]
As a fallback simply use { *VRTYPE, *VR0MIN, *VR0MAX } as
result for the intersection. That's always a conservative
correct estimate.
|
inlinestatic |
Return whether VAL is a negative overflow infinity.
References get_gimple_rhs_class(), gimple_assign_rhs1(), gimple_assign_rhs_code(), GIMPLE_SINGLE_RHS, is_gimple_assign(), and is_overflow_infinity().
|
inlinestatic |
Return whether VAL is a positive or negative overflow infinity.
References gcc_checking_assert, TREE_TYPE, vrp_val_is_min(), and vrp_val_min().
Referenced by add_equivalence(), check_array_ref(), is_negative_overflow_infinity(), is_positive_overflow_infinity(), and vrp_meet_1().
|
inlinestatic |
Return whether VAL is a positive overflow infinity.
References is_overflow_infinity(), TREE_TYPE, vrp_val_is_max(), and vrp_val_max().
|
static |
Return true if the SSA name NAME is live on the edge E.
Referenced by infer_value_range().
|
inlinestatic |
VAL is the maximum or minimum value of a type. Return a corresponding overflow infinity.
References CONSTANT_CLASS_P, needs_overflow_infinity(), TREE_OVERFLOW, TREE_TYPE, and vrp_val_is_min().
Referenced by supports_overflow_infinity().
| gimple_opt_pass* make_pass_vrp | ( | ) |
|
static |
Find out smallest RES where RES > VAL && (RES & MASK) == RES, if any (otherwise return VAL). VAL and MASK must be zero-extended for precision PREC. If SGNBIT is non-zero, first xor VAL with SGNBIT (to transform signed values into unsigned) and at the end xor SGNBIT back.
Referenced by register_edge_assert_for_2().
|
inlinestatic |
Return whether TYPE should use an overflow infinity distinct from TYPE_{MIN,MAX}_VALUE. We use an overflow infinity value to represent a signed overflow during VRP computations. An infinity is distinct from a half-range, which will go from some number to TYPE_{MIN,MAX}_VALUE.
References CONSTANT_CLASS_P, copy_node, gcc_checking_assert, NULL_TREE, and TREE_OVERFLOW.
Referenced by make_overflow_infinity(), negative_overflow_infinity(), positive_overflow_infinity(), and vrp_val_is_min().
|
inlinestatic |
Return a negative overflow infinity for TYPE.
References CONSTANT_CLASS_P, needs_overflow_infinity(), TREE_OVERFLOW, TREE_TYPE, and vrp_val_is_max().
|
static |
Return true if ARG is marked with the nonnull attribute in the current function signature.
The static chain decl is always non null.
If "nonnull" wasn't specified, we know nothing about the argument.
If "nonnull" applies to all the arguments, then ARG is non-null.
Get the position number for ARG in the function signature.
Now see if ARG_NUM is mentioned in the nonnull list.
References compare_tree_int(), and TREE_VALUE.
|
static |
Return true if op is in a boolean [0, 1] value-range.
|
static |
If OP has a value range with a single constant value return that, otherwise return NULL_TREE. This returns OP itself if OP is a constant.
Referenced by extract_range_from_comparison().
|
inlinestatic |
Return 1 if VAL < VAL2 0 if !(VAL < VAL2) -2 if those are incomparable.
LT is folded faster than GE and others. Inline the common case.
val >= val2, not considering overflow infinity.
|
inlinestatic |
Return true if value range VR uses an overflow infinity.
|
inlinestatic |
Return a positive overflow infinity for TYPE.
References CONSTANT_CLASS_P, needs_overflow_infinity(), TREE_OVERFLOW, TREE_TYPE, vrp_val_is_max(), and vrp_val_is_min().
|
static |
Process all the insertions registered for every name N_i registered in NEED_ASSERT_FOR. The list of assertions to be inserted are found in ASSERTS_FOR[i].
References gimple_phi_result(), INTEGRAL_TYPE_P, POINTER_TYPE_P, TREE_TYPE, and virtual_operand_p().
|
static |
Create an ASSERT_EXPR for NAME and insert it in the location indicated by LOC. Return true if we made any edge insertions.
Build the comparison expression NAME_i COMP_CODE VAL.
If we have X <=> X do not insert an assert expr for that.
We have been asked to insert the assertion on an edge. This
is used only by COND_EXPR and SWITCH_EXPR assertions. Otherwise, we can insert right after LOC->SI iff the statement must not be the last statement in the block.
If STMT must be the last statement in BB, we can only insert new assertions on the non-abnormal edge out of BB. Note that since STMT is not control flow, there may only be one non-abnormal edge out of BB.
References ASSERT_EXPR_COND, ASSERT_EXPR_VAR, boolean_false_node, fold(), FOR_EACH_BB, FOR_EACH_IMM_USE_ON_STMT, FOR_EACH_IMM_USE_STMT, gcc_assert, gimple_assign_lhs(), gimple_assign_rhs1(), gimple_assign_rhs_code(), gsi_end_p(), gsi_next(), gsi_remove(), gsi_start_bb(), gsi_stmt(), is_gimple_assign(), release_defs(), SET_USE, si, and TREE_CODE.
Referenced by find_switch_asserts().
|
static |
Some quadruple precision helpers.
|
static |
|
static |
Return whether the value range *VR fits in an integer type specified by PRECISION and UNSIGNED_P.
We can only handle integral and pointer types.
An extension is fine unless VR is signed and unsigned_p, and so is an identity transform.
Now we can only handle ranges with constant bounds.
For sign changes, the MSB of the double_int has to be clear. An unsigned value with its MSB set cannot be represented by a signed double_int, while a negative value cannot be represented by an unsigned double_int.
Then we can perform the conversion on both ends and compare the result for equality.
References extract_range_from_assignment(), gimple_assign_lhs(), INTEGRAL_TYPE_P, value_range_d::min, POINTER_TYPE_P, range_int_cst_singleton_p(), TREE_CODE, TREE_TYPE, and VR_INITIALIZER.
|
inlinestatic |
Return 1 if [MIN, MAX] includes the value zero, 0 if it does not include the value zero, -2 if we cannot tell.
References swap_tree_comparison(), TREE_CODE, and TREE_OPERAND.
|
inlinestatic |
Return true if max and min of VR are INTEGER_CST. It's not necessary a singleton.
Referenced by quad_int_pair_sort().
|
inlinestatic |
Return true if VR is a INTEGER_CST singleton.
Referenced by range_fits_type_p().
|
inlinestatic |
Return true if VR is ~[0, 0].
Referenced by gimple_stmt_nonnegative_warnv_p().
|
inlinestatic |
Return true if VR is [0, 0].
References get_gimple_rhs_class(), gimple_assign_rhs1(), gimple_assign_rhs_code(), GIMPLE_BINARY_RHS, gimple_expr_type(), GIMPLE_UNARY_RHS, and tree_unary_nonnegative_warnv_p().
|
static |
Create two value-ranges in *VR0 and *VR1 from the anti-range *AR so that *VR0 U *VR1 == *AR. Returns true if that is possible, false otherwise. If *AR can be represented with a single range *VR1 will be VR_UNDEFINED.
References set_value_range_to_varying().
|
static |
Try to register an edge assertion for SSA name NAME on edge E for the condition COND contributing to the conditional jump pointed to by SI. Return true if an assertion for NAME could be registered.
Do not attempt to infer anything in names that flow through abnormal edges.
Register ASSERT_EXPRs for name.
If COND is effectively an equality test of an SSA_NAME against the value zero or one, then we may be able to assert values for SSA_NAMEs which flow into COND.
In the case of NAME == 1 or NAME != 0, for BIT_AND_EXPR defining statement of NAME we can assert both operands of the BIT_AND_EXPR have nonzero value.
In the case of NAME == 0 or NAME != 1, for BIT_IOR_EXPR defining statement of NAME we can assert both operands of the BIT_IOR_EXPR have zero value.
For BIT_IOR_EXPR only if NAME == 0 both operands have
necessarily zero value, or if type-precision is one.
References bitmap_clear_bit(), bitmap_set_bit, FOR_EACH_PHI_ARG, gimple_phi_result(), gsi_stmt(), SSA_NAME_VERSION, SSA_OP_USE, TREE_CODE, USE_FROM_PTR, and virtual_operand_p().
|
static |
OP is an operand of a truth value expression which is known to have a particular value. Register any asserts for OP and for any operands in OP's defining statement.
If CODE is EQ_EXPR, then we want to register OP is zero (false), if CODE is NE_EXPR, then we want to register OP is nonzero (true).
We only care about SSA_NAMEs.
We know that OP will have a zero or nonzero value. If OP is used more than once go ahead and register an assert for OP. The FOUND_IN_SUBGRAPH support is not helpful in this situation as it will always be set for OP (because OP is used in a COND_EXPR in the subgraph).
Now look at how OP is set. If it's set from a comparison, a truth operation or some bit operations, then we may be able to register information about the operands of that assignment.
Recurse on each operand.
Recurse, flipping CODE.
Recurse through the copy.
Recurse through the type conversion.
References bitmap_bit_p, bitmap_set_bit, find_conditional_asserts(), find_switch_asserts(), FOR_EACH_SSA_TREE_OPERAND, fp_predicate(), gimple_assign_rhs1(), gimple_assign_rhs_code(), gsi_end_p(), gsi_last_bb(), gsi_prev(), gsi_stmt(), has_single_use(), infer_value_range(), integer_zerop(), is_gimple_assign(), is_gimple_debug(), last, last_stmt(), NULL, POINTER_TYPE_P, register_new_assert_for(), si, SSA_NAME_DEF_STMT, SSA_NAME_VERSION, SSA_OP_DEF, SSA_OP_USE, TREE_CODE, TREE_TYPE, and ZERO_SSA_OPERANDS.
|
static |
Try to register an edge assertion for SSA name NAME on edge E for the condition COND contributing to the conditional jump pointed to by BSI. Invert the condition COND if INVERT is true. Return true if an assertion for NAME could be registered.
Only register an ASSERT_EXPR if NAME was found in the sub-graph reachable from E.
In the case of NAME <= CST and NAME being defined as NAME = (unsigned) NAME2 + CST2 we can assert NAME2 >= -CST2 and NAME2 <= CST - CST2. We can do the same for NAME > CST. This catches range and anti-range tests.
Extract CST2 from the (optional) addition.
Extract NAME2 from the (optional) sign-changing cast.
If name3 is used later, create an ASSERT_EXPR for it.
Build an expression for the range test.
If name2 is used later, create an ASSERT_EXPR for it.
Build an expression for the range test.
In the case of post-in/decrement tests like if (i++) ... and uses of the in/decremented value on the edge the extra name we want to assert for is not on the def chain of the name compared. Instead it is in the set of use stmts.
Cut off to use-stmts that are in the predecessor.
Add asserts for NAME cmp CST and NAME being defined
as NAME = (int) NAME2. Build an expression for the range test.
Add asserts for NAME cmp CST and NAME being defined as
NAME = NAME2 >> CST2.
Extract CST2 from the right shift. Add asserts for NAME cmp CST and NAME being defined as
NAME = NAME2 & CST2.
Extract CST2 from the and.
Also handle
NAME = (unsigned) NAME2;
casts where NAME's type is unsigned and has smaller precision
than NAME2's type as if it was NAME = NAME2 & MASK. If CST2 doesn't have most significant bit set,
but VAL is negative, we have comparison like
if ((x & 0x123) > -4) (always true). Just give up. Minimum unsigned value for equality is VAL & CST2
(should be equal to VAL, otherwise we probably should
have folded the comparison into false) and
maximum unsigned value is VAL | ~CST2. If VAL is 0, handle (X & CST2) != 0 as (X & CST2) > 0U.
If (VAL | ~CST2) is all ones, handle it as
(X & CST2) < VAL. Minimum unsigned value for >= if (VAL & CST2) == VAL
is VAL and maximum unsigned value is ~0. For signed
comparison, if CST2 doesn't have most significant bit
set, handle it similarly. If CST2 has MSB set,
the minimum is the same, and maximum is ~0U/2. If (VAL & CST2) != VAL, X & CST2 can't be equal to
VAL. Find out smallest MINV where MINV > VAL
&& (MINV & CST2) == MINV, if any. If VAL is signed and
CST2 has MSB set, compute it biased by 1 << (nprec - 1). Minimum unsigned value for <= is 0 and maximum
unsigned value is VAL | ~CST2 if (VAL & CST2) == VAL.
Otherwise, find smallest VAL2 where VAL2 > VAL
&& (VAL2 & CST2) == VAL2 and use (VAL2 - 1) | ~CST2
as maximum.
For signed comparison, if CST2 doesn't have most
significant bit set, handle it similarly. If CST2 has
MSB set, the maximum is the same and minimum is INT_MIN. Minimum unsigned value for < is 0 and maximum
unsigned value is (VAL-1) | ~CST2 if (VAL & CST2) == VAL.
Otherwise, find smallest VAL2 where VAL2 > VAL
&& (VAL2 & CST2) == VAL2 and use (VAL2 - 1) | ~CST2
as maximum.
For signed comparison, if CST2 doesn't have most
significant bit set, handle it similarly. If CST2 has
MSB set, the maximum is the same and minimum is INT_MIN.
References double_int_one, double_int_zero, double_int::is_zero(), double_int::llshift(), double_int::mask(), masked_increment(), and double_int::zext().
|
static |
If NAME doesn't have an ASSERT_EXPR registered for asserting 'EXPR COMP_CODE VAL' at a location that dominates block BB or E->DEST, then register this location as a possible insertion point for ASSERT_EXPR <NAME, EXPR COMP_CODE VAL>.
BB, E and SI provide the exact insertion point for the new ASSERT_EXPR. If BB is NULL, then the ASSERT_EXPR is to be inserted on edge E. Otherwise, if E is NULL, the ASSERT_EXPR is inserted on BB. If SI points to a COND_EXPR or a SWITCH_EXPR statement, then E must not be NULL.
Never build an assert comparing against an integer constant with TREE_OVERFLOW set. This confuses our undefined overflow warning machinery.
The new assertion A will be inserted at BB or E. We need to determine if the new location is dominated by a previously registered location for A. If we are doing an edge insertion, assume that A will be inserted at E->DEST. Note that this is not necessarily true. If E is a critical edge, it will be split. But even if E is split, the new block will dominate the same set of blocks that E->DEST dominates. The reverse, however, is not true, blocks dominated by E->DEST will not be dominated by the new block created to split E. So, if the insertion location is on a critical edge, we will not use the new location to move another assertion previously registered at a block dominated by E->DEST.
If NAME already has an ASSERT_EXPR registered for COMP_CODE and VAL at a block dominating DEST_BB, then we don't need to insert a new one. Similarly, if the same assertion already exists at a block dominated by DEST_BB and the new location is not on a critical edge, then update the existing location for the assertion (i.e., move the assertion up in the dominance tree). Note, this is implemented as a simple linked list because there should not be more than a handful of assertions registered per name. If this becomes a performance problem, a table hashed by COMP_CODE and VAL could be implemented.
If E is not a critical edge and DEST_BB
dominates the existing location for the assertion, move
the assertion up in the dominance tree by updating its
location information. Update the last node of the list and move to the next one.
If we didn't find an assertion already registered for NAME COMP_CODE VAL, add a new one at the end of the list of assertions associated with NAME.
Referenced by register_edge_assert_for_1().
|
static |
Convert range assertion expressions into the implied copies and copy propagate away the copies. Doing the trivial copy propagation here avoids the need to run the full copy propagation pass after VRP.
FIXME, this will eventually lead to copy propagation removing the names that had useful range information attached to them. For instance, if we had the assertion N_i = ASSERT_EXPR <N_j, N_j > 3>, then N_i will have the range [3, +INF].
However, by converting the assertion into the implied copy operation N_i = N_j, we will then copy-propagate N_j into the uses of N_i and lose the range information. We may want to hold on to ASSERT_EXPRs a little while longer as the ranges could be used in things like jump threading.
The problem with keeping ASSERT_EXPRs around is that passes after VRP need to handle them appropriately.
Another approach would be to make the range information a first class property of the SSA_NAME so that it can be queried from any pass. This is made somewhat more complex by the need for multiple ranges to be associated with one SSA_NAME.
Note that the BSI iterator bump happens at the bottom of the loop and no bump is necessary if we're removing the statement referenced by the current BSI.
Propagate the RHS into every use of the LHS.
And finally, remove the copy, it is not needed.
|
static |
Searches if the expr T, located at LOCATION computes address of an ARRAY_REF, and call check_array_ref on it.
We are only interested in addresses of ARRAY_REF's.
Check each ARRAY_REFs in the reference chain.
|
static |
Set value range VR to the canonical form of {T, MIN, MAX, EQUIV}. This means adjusting T, MIN and MAX representing the case of a wrapping range with MAX < MIN covering [MIN, type_max] U [type_min, MAX] as anti-rage ~[MAX+1, MIN-1]. Likewise for wrapping anti-ranges. In corner cases where MAX+1 or MIN-1 wraps this will fall back to varying. This routine exists to ease canonicalization in the case where we extract ranges from var + CST op limit.
Use the canonical setters for VR_UNDEFINED and VR_VARYING.
Nothing to canonicalize for symbolic ranges.
Wrong order for min and max, to swap them and the VR type we need to adjust them.
For one bit precision if max < min, then the swapped
range covers all values, so for VR_RANGE it is varying and
for VR_ANTI_RANGE empty range, so drop to varying as well. There's one corner case, if we had [C+1, C] before we now have
that again. But this represents an empty value range, so drop
to varying in this case. Anti-ranges that can be represented as ranges should be so.
We cannot deal with empty ranges, drop to varying.
??? This could be VR_UNDEFINED instead. Non-empty boolean ranges can always be represented
as a singleton range. As a special exception preserve non-null ranges.
Drop [-INF(OVF), +INF(OVF)] to varying.
References set_value_range_to_varying().
|
static |
Set value range VR to {T, MIN, MAX, EQUIV}.
Since updating the equivalence set involves deep copying the bitmaps, only do it if absolutely necessary.
Referenced by check_array_ref().
|
inlinestatic |
Set value range VR to a non-negative range of type TYPE. OVERFLOW_INFINITY indicates whether to use an overflow infinity rather than TYPE_MAX_VALUE; this should be true if we determine that the range is nonnegative based on the assumption that signed overflow does not occur.
|
inlinestatic |
Set value range VR to a non-NULL range of type TYPE.
|
inlinestatic |
Set value range VR to a NULL range of type TYPE.
References NULL, NULL_TREE, SSA_NAME_VERSION, and VR_VARYING.
|
inlinestatic |
Set value range VR to a range of a truthvalue of type TYPE.
|
inlinestatic |
Set value range VR to VR_UNDEFINED.
|
inlinestatic |
Set value range VR to a single value. This function is only called with values we get from statements, and exists to clear the TREE_OVERFLOW flag so that we don't think we have an overflow infinity when we shouldn't.
Referenced by copy_value_range(), extract_range_from_comparison(), extract_range_from_cond_expr(), and vrp_intersect_ranges_1().
|
inlinestatic |
Set value range VR to VR_VARYING.
Referenced by check_array_ref(), extract_range_from_binary_expr(), extract_range_from_cond_expr(), insert_range_assertions(), ranges_from_anti_range(), and set_and_canonicalize_value_range().
|
static |
If the operand to an ABS_EXPR is >= 0, then eliminate the ABS_EXPR. If the operand is <= 0, then simplify the ABS_EXPR into a NEGATE_EXPR.
|
static |
Optimize away redundant BIT_AND_EXPR and BIT_IOR_EXPR. If all the bits that are being cleared by & are already known to be zero from VR, or all the bits that are being set by | are already known to be one from VR, the bit operation is redundant.
|
static |
Simplify a conditional using a relational operator to an equality test if the range information indicates only one value can satisfy the original conditional.
If we have range information for OP0, then we might be
able to simplify this conditional.
Try again after inverting the condition. We only deal
with integral types here, so no need to worry about
issues with inverting FP comparisons. If we have a comparison of an SSA_NAME (OP0) against a constant, see if OP0 was set by a type conversion where the source of the conversion is another SSA_NAME with a range that fits into the range of OP0's type. If so, the conversion is redundant as the earlier SSA_NAME can be used for the comparison directly if we just massage the constant in the comparison.
The range must not have overflowed, or if it did overflow
we must not be wrapping/trapping overflow and optimizing
with strict overflow semantics. If the range overflowed and the user has asked for warnings
when strict overflow semantics were used to optimize code,
issue an appropriate warning.
References EDGE_COMPLEX, edge_def::flags, FOR_EACH_EDGE, gimple_cond_lhs(), gimple_cond_rhs(), gsi_last_bb(), gsi_stmt(), INTEGRAL_TYPE_P, is_gimple_min_invariant(), last, POINTER_TYPE_P, potentially_threadable_block(), simplify_stmt_for_jump_threading(), thread_across_edge(), TREE_CODE, and TREE_TYPE.
|
static |
Simplify an integral conversion from an SSA name in STMT.
Get the value-range of the inner operand.
Simulate the conversion chain to check if the result is equal if the middle conversion is removed.
If the first conversion is not injective, the second must not be widening.
We also want a medium value so that we can track the effect that narrowing conversions with sign change have.
Require that the final conversion applied to both the original and the intermediate range produces the same result.
|
static |
Simplify a division or modulo operator to a right shift or bitwise and if the first operand is unsigned or is greater than zero and the second operand is an exact power of two.
|
static |
Simplify a conversion from integral SSA name to float in STMT.
We can only handle constant ranges.
First check if we can use a signed type in place of an unsigned.
If we can do the conversion in the current input mode do nothing.
Otherwise search for a mode we can use, starting from the narrowest integer mode available.
If we cannot do a signed conversion to float from mode
or if the value-range does not fit in the signed type
try with a wider mode. But do not widen the input. Instead leave that to the
optabs expansion code. It works, insert a truncation or sign-change before the float conversion.
|
static |
A trivial wrapper so that we can present the generic jump threading code with a simple API for simplifying statements. STMT is the statement we want to simplify, WITHIN_STMT provides the location for any overflow warnings.
Referenced by simplify_cond_using_ranges().
|
static |
Simplify STMT using ranges if possible.
Transform EQ_EXPR, NE_EXPR into BIT_XOR_EXPR or identity if the RHS is zero or one, and the LHS are known to be boolean values.
Transform TRUNC_DIV_EXPR and TRUNC_MOD_EXPR into RSHIFT_EXPR
and BIT_AND_EXPR respectively if the first operand is greater
than zero and the second operand is an exact power of two.
Transform ABS (X) into X or -X as appropriate. Optimize away BIT_AND_EXPR and BIT_IOR_EXPR
if all the bits being cleared are already cleared or
all the bits being set are already set.
|
static |
Simplify a switch statement using the value range of the switch argument.
We can only handle integer ranges.
Find case label for min/max of the value range.
Bail out if this is just all edges taken.
Build a new vector of taken case labels.
Add the default edge, if necessary.
Mark needed edges.
Queue not needed edges for later removal.
And queue an update for the stmt.
|
static |
Simplify boolean operations if the source is known to be already a boolean.
We handle only !=/== case here.
Reduce number of cases to handle to NE_EXPR. As there is no BIT_XNOR_EXPR we cannot replace A == B with a single statement.
Make sure to not sign-extend a 1-bit 1 when converting the result.
For A != 0 we can substitute A itself.
For A != B we substitute A ^ B. Either with conversion.
Or without.
| bool ssa_name_nonnegative_p | ( | ) |
Return true if T, an SSA_NAME, is known to be nonnegative. Return false otherwise or if no value range information is available.
|
static |
Return true if STMT is interesting for VRP.
In general, assignments with virtual operands are not useful for deriving ranges, with the obvious exception of calls to builtin functions.
|
inlinestatic |
Return whether STMT has a constant rhs that is_overflow_infinity.
References cfun, gcc_assert, HOST_WIDE_INT, lookup_attribute(), POINTER_TYPE_P, TREE_CHAIN, TREE_CODE, TREE_TYPE, and TYPE_ATTRIBUTES.
|
inlinestatic |
Return whether TYPE can support our overflow infinity representation: we use the TREE_OVERFLOW flag, which only exists for constants. If TYPE doesn't support this, we don't optimize cases which would require signed overflow–we drop them to VARYING.
References gcc_checking_assert, make_overflow_infinity(), and vrp_val_min().
|
inlinestatic |
Return true if value range VR involves at least one symbol.
References gimple_call_arg(), gimple_call_fndecl(), gimple_call_num_args(), gimple_expr_type(), NULL_TREE, and tree_call_nonnegative_warnv_p().
|
static |
We are comparing trees OP0 and OP1 using COND_CODE. OP0 has a known value range VR.
If there is one and only one value which will satisfy the conditional, then return that value. Else return NULL.
Extract minimum/maximum values which satisfy the the conditional as it was written.
This should not be negative infinity; there is no overflow
here. This should not be positive infinity; there is no overflow
here. Now refine the minimum and maximum values using any value range information we have for op0.
If the new min/max values have converged to a single value,
then there is only one value which can satisfy the condition,
return that value.
References dump_file, fold_convert, gcc_assert, gcc_unreachable, gimple_assign_set_rhs_from_tree(), gimple_cond_make_false(), gimple_cond_make_true(), gimple_expr_type(), integer_onep(), integer_zerop(), is_gimple_assign(), print_generic_expr(), and print_gimple_expr().
|
static |
Union the two value-ranges { *VR0TYPE, *VR0MIN, *VR0MAX } and { VR1TYPE, VR0MIN, VR0MAX } and store the result in { *VR0TYPE, *VR0MIN, *VR0MAX }. This may not be the smallest possible such range. The resulting range is not canonicalized.
[] is vr0, () is vr1 in the following classification comments.
[( )]
Nothing to do for equal ranges.
For anti-range with range union the result is varying.
[ ] ( ) or ( ) [ ]
If the ranges have an empty intersection, result of the union
operation is the anti-range or if both are anti-ranges
it covers all. The result is the convex hull of both ranges.
If the result can be an anti-range, create one.
If the result can be an anti-range, create one.
[ ( ) ] or [( ) ] or [ ( )]
Arbitrarily choose the right or left gap.
The result covers everything.
( [ ] ) or ([ ] ) or ( [ ])
The result covers everything.
[ ( ] ) or [ ]( )
( [ ) ] or ( )[ ]
|
inlinestatic |
Update the value range and equivalence set for variable VAR to NEW_VR. Return true if NEW_VR is different from VAR's previous value.
NOTE: This function assumes that NEW_VR is a temporary value range object created for the sole purpose of updating VAR's range. The storage used by the equivalence set from NEW_VR will be freed by this function. Do not call update_value_range when NEW_VR is the range object associated with another SSA name.
Update the value range, if necessary.
Do not allow transitions up the lattice. The following
is slightly more awkward than just new_vr->type < old_vr->type
because VR_RANGE and VR_ANTI_RANGE need to be considered
the same. We may not have is_new when transitioning to
UNDEFINED or from VARYING.
References value_range_d::max, value_range_d::min, TREE_CODE, value_range_d::type, and VR_RANGE.
Referenced by check_array_ref().
|
static |
Return false if we can not make a valid comparison based on VR; this will be the case if it uses an overflow infinity and overflow is not undefined (i.e., -fno-strict-overflow is in effect). Otherwise return true, and set *STRICT_OVERFLOW_P to true if VR uses an overflow infinity.
|
static |
Returns true if EXPR is a valid value (as expected by compare_values) – a gimple invariant, or SSA_NAME +- CST.
References NULL_TREE, TREE_CODE, and TREE_OPERAND.
|
inlinestatic |
Return 1 if VAL is inside value range MIN <= VAL <= MAX, 0 if VAL is not inside [MIN, MAX], -2 if we cannot tell either way.
Benchmark compile/20001226-1.c compilation time after changing this function.
References get_value_range(), integer_onep(), integer_zerop(), value_range_d::max, value_range_d::min, TREE_CODE, TREE_TYPE, value_range_d::type, TYPE_PRECISION, and VR_RANGE.
|
static |
If *VR has a value rante that is a single constant value return that, otherwise return NULL_TREE.
References fold_build1, int_const_binop(), TREE_CODE, TREE_OPERAND, and TREE_TYPE.
|
inlinestatic |
Return true if *VR is know to only contain nonnegative values.
Testing for VR_ANTI_RANGE is not useful here as any anti-range which would return a useful value should be encoded as a VR_RANGE.
|
inlinestatic |
Return true if value ranges VR0 and VR1 have a non-empty intersection.
Benchmark compile/20001226-1.c compilation time after changing this function.
The value ranges do not intersect if the maximum of the first range is less than the minimum of the second range or vice versa. When those relations are unknown, we can't do any better.
|
inlinestatic |
Return true, if the bitmaps B1 and B2 are equal.
References integer_zerop(), value_range_d::max, value_range_d::min, value_range_d::type, and VR_ANTI_RANGE.
|
static |
Given (CODE OP0 OP1) within STMT, try to simplify it based on value range information. Return NULL if the conditional can not be evaluated. The ranges of all the names equivalent with the operands in COND will be used when trying to compute the value. If the result is based on undefined signed overflow, issue a warning if appropriate.
Some passes and foldings leak constants with overflow flag set into the IL. Avoid doing wrong things with these and bail out.
If the comparison is being folded and the operand on the LHS
is being compared against a constant value that is outside of
the natural range of OP0's type, then the predicate will
always fold regardless of the value of OP0. If -Wtype-limits
was specified, emit a warning.
References dump_file, dump_flags, and print_gimple_stmt().
|
static |
Helper function for vrp_evaluate_conditional_warnv.
We only deal with integral and pointer types.
|
static |
Helper function for vrp_evaluate_conditional_warnv.
|
static |
Traverse all the blocks folding conditionals with known ranges.
We must identify jump threading opportunities before we release the datastructures built by VRP.
Set value range to non pointer SSA_NAMEs.
VR_ANTI_RANGE ~[min, max] is encoded compactly as
[max + 1, min - 1] without additional attributes.
When min value > max value, we know that it is
VR_ANTI_RANGE; it is VR_RANGE otherwise. ~[0,0] anti-range is represented as
range. Free allocated memory.
So that we can distinguish between VRP data being available and not available.
|
static |
Callback for substitute_and_fold folding the stmt at *SI.
|
static |
Initialize local data structures for VRP.
If the statement is a control insn, then we do not want to avoid simulating the statement once. Failure to do so means that those edges will never get added.
References G_, gimple_has_location(), gimple_location(), input_location, integer_zerop(), and warning_at().
|
static |
Wrapper around int_const_binop. If the operation overflows and we are not using wrapping arithmetic, then adjust the result to be -INF or +INF depending on CODE, VAL1 and VAL2. This can return NULL_TREE if we need to use an overflow infinity representation but the type does not support it.
If we are using unsigned arithmetic, operate symbolically on -INF and +INF as int_const_binop only handles signed overflow.
Ensure that res = val1 [+*] val2 >= val1
or that res = val1 - val2 <= val1. Checking for multiplication overflow is done by dividing the
output of the multiplication by the first input of the
multiplication. If the result of that division operation is
not equal to the second input of the multiplication, then the
multiplication overflowed. If the singed operation wraps then int_const_binop has done everything we want.
If the operation overflowed but neither VAL1 nor VAL2 are
overflown, return -INF or +INF depending on the operation
and the combination of signs of the operands. We have to punt on adding infinities of different signs,
since we can't tell what the sign of the result should be.
Likewise for subtracting infinities of the same sign. Don't try to handle division or shifting of infinities.
Notice that we only need to handle the restricted set of
operations handled by extract_range_from_binary_expr.
Among them, only multiplication, addition and subtraction
can yield overflow without overflown operands because we
are working with integral types only... except in the
case VAL1 = -INF and VAL2 = -1 which overflows to +INF
for division too. For multiplication, the sign of the overflow is given
by the comparison of the signs of the operands. For addition, the operands must be of the same sign
to yield an overflow. Its sign is therefore that
of one of the operands, for example the first. For
infinite operands X + -INF is negative, not positive. For subtraction, non-infinite operands must be of
different signs to yield an overflow. Its sign is
therefore that of the first operand or the opposite of
that of the second operand. A first operand of 0 counts
as positive here, for the corner case 0 - (-INF), which
overflows, but must yield +INF. For infinite operands 0
- INF is negative, not positive. We only get in here with positive shift count, so the
overflow direction is the same as the sign of val1.
Actually rshift does not overflow at all, but we only
handle the case of shifting overflowed -INF and +INF. For division, the only case is -INF / -1 = +INF.
References value_range_d::min, and tree_to_double_int().
|
static |
|
static |
|
static |
Intersect the two value-ranges *VR0 and *VR1 and store the result in *VR0. This may not be the smallest possible such range.
If either range is VR_VARYING the other one wins.
When either range is VR_UNDEFINED the resulting range is VR_UNDEFINED, too.
Save the original vr0 so we can return it as conservative intersection result when our worker turns things to varying.
Make sure to canonicalize the result though as the inversion of a VR_RANGE can still be a VR_RANGE.
If that failed, use the saved original VR0.
If the result is VR_UNDEFINED there is no need to mess with the equivalencies.
The resulting set of equivalences for range intersection is the union of the two sets.
References double_int::and_not(), get_value_range(), gimple_assign_rhs1(), gimple_assign_rhs2(), gimple_assign_rhs_code(), is_gimple_min_invariant(), double_int::is_zero(), NULL, NULL_TREE, set_value_range_to_value(), TREE_CODE, VR_INITIALIZER, and zero_nonzero_bits_from_vr().
|
static |
|
static |
|
static |
Meet operation for value ranges. Given two value ranges VR0 and VR1, store in VR0 a range that contains both VR0 and VR1. This may not be the smallest possible such range.
VR0 already has the resulting range.
Nothing to do. VR0 already has the resulting range.
Failed to find an efficient meet. Before giving up and setting
the result to VARYING, see if we can at least derive a useful
anti-range. FIXME, all this nonsense about distinguishing
anti-ranges from ranges is necessary because of the odd
semantics of range_includes_zero_p and friends. Since this meet operation did not result from the meeting of
two equivalent names, VR0 cannot have any equivalences. The resulting set of equivalences is always the intersection of the two sets.
References build_int_cst(), compare_values(), EXPR_P, fold_build2, is_gimple_min_invariant(), is_overflow_infinity(), value_range_d::max, value_range_d::min, NULL, operand_equal_p(), TREE_NO_WARNING, TREE_TYPE, TYPE_MAX_VALUE, and TYPE_MIN_VALUE.
|
inlinestatic |
Return true, if VAL1 and VAL2 are equal values for VRP purposes.
References BITMAP_ALLOC, bitmap_ior_into(), bitmap_set_bit, value_range_d::equiv, NULL, and SSA_NAME_VERSION.
|
static |
Like tree_expr_nonzero_warnv_p, but this function uses value ranges obtained so far.
If we have an expression of the form &X->a, then the expression is nonnull if X is nonnull.
References fold_convert, gcc_assert, POINTER_TYPE_P, STRIP_USELESS_TYPE_CONVERSION, TREE_CODE, and TREE_TYPE.
|
inlinestatic |
Return whether VAL is equal to the maximum value of its type. This will be true for a positive overflow infinity. We can't do a simple equality comparison with TYPE_MAX_VALUE because C typedefs and Ada subtypes can produce types whose TYPE_MAX_VALUE is not == to the integer constant with the same value in the type.
References INTEGRAL_TYPE_P, and TYPE_OVERFLOW_WRAPS.
Referenced by is_positive_overflow_infinity(), negative_overflow_infinity(), and positive_overflow_infinity().
|
inlinestatic |
Return whether VAL is equal to the minimum value of its type. This will be true for a negative overflow infinity.
References CONSTANT_CLASS_P, gcc_assert, needs_overflow_infinity(), NULL_TREE, vrp_val_max(), and vrp_val_min().
Referenced by is_overflow_infinity(), make_overflow_infinity(), and positive_overflow_infinity().
|
inlinestatic |
Return the maximum value for TYPE.
References NULL_TREE, operand_equal_p(), and TREE_TYPE.
Referenced by is_positive_overflow_infinity(), and vrp_val_is_min().
|
inlinestatic |
Return the minimum value for TYPE.
References NULL_TREE, operand_equal_p(), and TREE_TYPE.
Referenced by is_overflow_infinity(), supports_overflow_infinity(), and vrp_val_is_min().
|
inlinestatic |
Return the singleton value-range for NAME or NAME.
Referenced by check_array_ref().
|
static |
Return true if VAR may overflow at STMT. This checks any available loop information to see if we can determine that VAR does not overflow.
If we get here, we know something useful about VAR based on the loop information. If it wraps, it may overflow.
|
static |
Visit assignment STMT. If it produces an interesting range, record the SSA name in *OUTPUT_P.
We only keep track of ranges in integral and pointer types.
It is valid to have NULL MIN/MAX values on a type. See
build_range_type. Try folding the statement to a constant first.
Then dispatch to value-range extracting functions.
Every other statement produces no useful ranges.
|
static |
Visit conditional statement STMT. If we can determine which edge will be taken out of STMT's basic block, record it in *TAKEN_EDGE_P and return SSA_PROP_INTERESTING. Otherwise, return SSA_PROP_VARYING.
Compute the value of the predicate COND by checking the known ranges of each of its operands.
Note that we cannot evaluate all the equivalent ranges here because those ranges may not yet be final and with the current propagation strategy, we cannot determine when the value ranges of the names in the equivalence set have changed.
For instance, given the following code fragment
i_5 = PHI <8, i_13> ... i_14 = ASSERT_EXPR <i_5, i_5 != 0> if (i_14 == 1) ...
Assume that on the first visit to i_14, i_5 has the temporary range [8, 8] because the second argument to the PHI function is not yet executable. We derive the range ~[0, 0] for i_14 and the equivalence set { i_5 }. So, when we visit 'if (i_14 == 1)' for the first time, since i_14 is equivalent to the range [8, 8], we determine that the predicate is always false.
On the next round of propagation, i_13 is determined to be VARYING, which causes i_5 to drop down to VARYING. So, another visit to i_14 is scheduled. In this second visit, we compute the exact same range and equivalence set for i_14, namely ~[0, 0] and { i_5 }. But we did not have the previous range for i_5 registered, so vrp_visit_assignment thinks that the range for i_14 has not changed. Therefore, the predicate 'if (i_14 == 1)' is not visited again, which stops propagation from visiting statements in the THEN clause of that if().
To properly fix this we would need to keep the previous range value for the names in the equivalence set. This way we would've discovered that from one visit to the other i_5 changed from range [8, 8] to VR_VARYING.
However, fixing this apparent limitation may not be worth the additional checking. Testing on several code bases (GCC, DLV, MICO, TRAMP3D and SPEC2000) showed that doing this results in 4 more predicates folded in SPEC.
|
static |
Visit all arguments for PHI node PHI that flow through executable edges. If a valid value range can be derived from all the incoming value ranges, set a new range for the LHS of PHI.
Do not allow equivalences or symbolic ranges to leak in from
backedges. That creates invalid equivalencies.
See PR53465 and PR54767.
To prevent infinite iterations in the algorithm, derive ranges
when the new value is slightly bigger or smaller than the
previous one. We don't do this if we have seen a new executable
edge; this helps us avoid an overflow infinity for conditionals
which are not in a loop. If the old value-range was VR_UNDEFINED
use the updated range and iterate one more time.
For non VR_RANGE or for pointers fall back to varying if
the range changed. If the new minimum is smaller or larger than the previous
one, go all the way to -INF. In the first case, to avoid
iterating millions of times to reach -INF, and in the
other case to avoid infinite bouncing between different
minimums. Similarly, if the new maximum is smaller or larger than
the previous one, go all the way to +INF. If we dropped either bound to +-INF then if this is a loop
PHI node SCEV may known more about its value-range. If we will end up with a (-INF, +INF) range, set it to
VARYING. Same if the previous max value was invalid for
the type and we end up with vr_result.min > vr_result.max. If the new range is different than the previous value, keep iterating.
Nothing changed, don't add outgoing edges.
No match found. Set the LHS to VARYING.
|
static |
Evaluate statement STMT. If the statement produces a useful range, return SSA_PROP_INTERESTING and record the SSA name with the interesting range into *OUTPUT_P.
If STMT is a conditional branch and we can determine its truth value, the taken edge is recorded in *TAKEN_EDGE_P.
If STMT produces a varying value, return SSA_PROP_VARYING.
All other statements produce nothing of interest for VRP, so mark their outputs varying and prevent further simulation.
|
static |
Visit switch statement STMT. If we can determine which edge will be taken out of STMT's basic block, record it in *TAKEN_EDGE_P and return SSA_PROP_INTERESTING. Otherwise, return SSA_PROP_VARYING.
Find the single edge that is taken from the switch expression.
Check if the range spans no CASE_LABEL. If so, we only reach the default label
Check if labels with index i to j and maybe the default label
are all reaching the same label.
|
static |
For range VR compute two double_int bitmasks. In *MAY_BE_NONZERO bitmask if some bit is unset, it means for all numbers in the range the bit is 0, otherwise it might be 0 or 1. In *MUST_BE_NONZERO bitmask if some bit is set, it means for all numbers in the range the bit is 1, otherwise it might be 0 or 1.
Referenced by vrp_intersect_ranges_1().
Variable Documentation
|
static |
Array of locations lists where to insert assertions. ASSERTS_FOR[I] holds a list of ASSERT_LOCUS_T nodes that describe where ASSERT_EXPRs for SSA name N_I should be inserted.
Stack of dest,src equivalency pairs that need to be restored after each attempt to thread a block's incoming edge to an outgoing edge.
A NULL entry is used to mark the end of pairs which need to be restored.
|
static |
Set of SSA names found live during the RPO traversal of the function for still active basic-blocks.
Referenced by split_all_insns().
|
static |
If bit I is present, it means that SSA name N_i has a list of assertions that should be inserted in the IL.
|
static |
Value range array. After propagation, VR_VALUE[I] holds the range of values that SSA name N_I may take.
|
static |
|
static |
|
static |
For a PHI node which sets SSA name N_I, VR_COUNTS[I] holds the number of executable edges we saw the last time we visited the node.
|
static |
