mirror of
https://github.com/github/codeql.git
synced 2026-04-27 17:55:19 +02:00
Merge pull request #440 from asger-semmle/range-analysis
JS: Range analysis for dead code detection
This commit is contained in:
47
javascript/ql/src/Statements/UselessComparisonTest.qhelp
Normal file
47
javascript/ql/src/Statements/UselessComparisonTest.qhelp
Normal file
@@ -0,0 +1,47 @@
|
||||
<!DOCTYPE qhelp PUBLIC
|
||||
"-//Semmle//qhelp//EN"
|
||||
"qhelp.dtd">
|
||||
<qhelp>
|
||||
<overview>
|
||||
<p>
|
||||
If a condition always evaluates to true or always evaluates to false, this often indicates
|
||||
incomplete code or a latent bug, and it should be examined carefully.
|
||||
</p>
|
||||
|
||||
</overview>
|
||||
<recommendation>
|
||||
|
||||
<p>
|
||||
Examine the surrounding code to determine why the condition is redundant. If it is no
|
||||
longer needed, remove it.
|
||||
</p>
|
||||
|
||||
<p>
|
||||
If the check is needed to guard against <code>NaN</code> values, insert a comment explaning the possibility of <code>NaN</code>.
|
||||
</p>
|
||||
|
||||
</recommendation>
|
||||
<example>
|
||||
|
||||
<p>
|
||||
The following example finds the index of an element in a given slice of the array:
|
||||
</p>
|
||||
|
||||
<sample src="examples/UselessComparisonTest.js" />
|
||||
|
||||
<p>
|
||||
The condition <code>i < end</code> at the end is always false, however. The code can be clarified if the
|
||||
redundant condition is removed:
|
||||
</p>
|
||||
|
||||
<sample src="examples/UselessComparisonTestGood.js" />
|
||||
|
||||
</example>
|
||||
<references>
|
||||
|
||||
|
||||
<li>Mozilla Developer Network: <a href="https://developer.mozilla.org/en-US/docs/Glossary/Truthy">Truthy</a>,
|
||||
<a href="https://developer.mozilla.org/en-US/docs/Glossary/Falsy">Falsy</a>.</li>
|
||||
|
||||
</references>
|
||||
</qhelp>
|
||||
60
javascript/ql/src/Statements/UselessComparisonTest.ql
Normal file
60
javascript/ql/src/Statements/UselessComparisonTest.ql
Normal file
@@ -0,0 +1,60 @@
|
||||
/**
|
||||
* @name Useless comparison test
|
||||
* @description A comparison that always evaluates to true or always evaluates to false may
|
||||
* indicate faulty logic and dead code.
|
||||
* @kind problem
|
||||
* @problem.severity warning
|
||||
* @id js/useless-comparison-test
|
||||
* @tags correctness
|
||||
* @precision high
|
||||
*/
|
||||
|
||||
import javascript
|
||||
|
||||
/**
|
||||
* Holds if there are any contradictory guard nodes in `container`.
|
||||
*
|
||||
* We use this to restrict reachability analysis to a small set of containers.
|
||||
*/
|
||||
predicate hasContradictoryGuardNodes(StmtContainer container) {
|
||||
exists (ConditionGuardNode guard |
|
||||
RangeAnalysis::isContradictoryGuardNode(guard) and
|
||||
container = guard.getContainer())
|
||||
}
|
||||
|
||||
/**
|
||||
* Holds if `block` is reachable and is in a container with contradictory guard nodes.
|
||||
*/
|
||||
predicate isReachable(BasicBlock block) {
|
||||
exists (StmtContainer container |
|
||||
hasContradictoryGuardNodes(container) and
|
||||
block = container.getEntryBB())
|
||||
or
|
||||
isReachable(block.getAPredecessor()) and
|
||||
not RangeAnalysis::isContradictoryGuardNode(block.getANode())
|
||||
}
|
||||
|
||||
/**
|
||||
* Holds if `block` is unreachable, but could be reached if `guard` was not contradictory.
|
||||
*/
|
||||
predicate isBlockedByContradictoryGuardNodes(BasicBlock block, ConditionGuardNode guard) {
|
||||
RangeAnalysis::isContradictoryGuardNode(guard) and
|
||||
isReachable(block.getAPredecessor()) and // the guard itself is reachable
|
||||
block = guard.getBasicBlock()
|
||||
or
|
||||
isBlockedByContradictoryGuardNodes(block.getAPredecessor(), guard) and
|
||||
not isReachable(block)
|
||||
}
|
||||
|
||||
/**
|
||||
* Holds if the given guard node is contradictory and causes an expression or statement to be unreachable.
|
||||
*/
|
||||
predicate isGuardNodeWithDeadCode(ConditionGuardNode guard) {
|
||||
exists (BasicBlock block |
|
||||
isBlockedByContradictoryGuardNodes(block, guard) and
|
||||
block.getANode() instanceof ExprOrStmt)
|
||||
}
|
||||
|
||||
from ConditionGuardNode guard
|
||||
where isGuardNodeWithDeadCode(guard)
|
||||
select guard.getTest(), "The condition '" + guard.getTest() + "' is always " + guard.getOutcome().booleanNot() + "."
|
||||
@@ -0,0 +1,12 @@
|
||||
function findValue(values, x, start, end) {
|
||||
let i;
|
||||
for (i = start; i < end; ++i) {
|
||||
if (values[i] === x) {
|
||||
return i;
|
||||
}
|
||||
}
|
||||
if (i < end) {
|
||||
return i;
|
||||
}
|
||||
return -1;
|
||||
}
|
||||
@@ -0,0 +1,8 @@
|
||||
function findValue(values, x, start, end) {
|
||||
for (let i = start; i < end; ++i) {
|
||||
if (values[i] === x) {
|
||||
return i;
|
||||
}
|
||||
}
|
||||
return -1;
|
||||
}
|
||||
@@ -36,6 +36,7 @@ import semmle.javascript.NPM
|
||||
import semmle.javascript.Paths
|
||||
import semmle.javascript.Promises
|
||||
import semmle.javascript.CanonicalNames
|
||||
import semmle.javascript.RangeAnalysis
|
||||
import semmle.javascript.Regexp
|
||||
import semmle.javascript.SSA
|
||||
import semmle.javascript.StandardLibrary
|
||||
|
||||
631
javascript/ql/src/semmle/javascript/RangeAnalysis.qll
Normal file
631
javascript/ql/src/semmle/javascript/RangeAnalysis.qll
Normal file
@@ -0,0 +1,631 @@
|
||||
import javascript
|
||||
|
||||
/*
|
||||
* The range analysis is based on Difference Bound constraints, that is, inequalities of form:
|
||||
*
|
||||
* a - b <= c
|
||||
*
|
||||
* or equivalently,
|
||||
*
|
||||
* a <= b + c
|
||||
*
|
||||
* where a and b are variables in the constraint system, and c is an integer constant.
|
||||
*
|
||||
* Such constraints obey a transitive law. Given two constraints,
|
||||
*
|
||||
* a - x <= c1
|
||||
* x - b <= c2
|
||||
*
|
||||
* adding the two inequalities yields the obvious transitive conclusion:
|
||||
*
|
||||
* a - b <= c1 + c2
|
||||
*
|
||||
* We view the system of constraints as a weighted graph, where `a - b <= c`
|
||||
* corresponds to the edge `a -> b` with weight `c`.
|
||||
*
|
||||
* Paths in this graph corresponds to the additional inequalities we can derive from the constraint set.
|
||||
* A negative-weight cycle represents a contradiction, such as `a <= a - 1`.
|
||||
*
|
||||
*
|
||||
* CONTROL FLOW:
|
||||
*
|
||||
* Each constraint is associated with a CFG node where that constraint is known to be valid.
|
||||
* The constraint is only valid within the dominator subtree of that node.
|
||||
*
|
||||
* The transitive rule additionally requires that, in order to compose two edges, one of
|
||||
* their CFG nodes must dominate the other, and the resulting edge becomes associated with the
|
||||
* dominated CFG node (i.e. the most restrictive scope). This ensures constraints
|
||||
* cannot be taken out of context.
|
||||
*
|
||||
* If a negative-weight cycle can be constructed from the edges "in scope" at a given CFG node
|
||||
* (i.e. associated with a dominator of the node), that node is unreachable.
|
||||
*
|
||||
*
|
||||
* DUAL CONSTRAINTS:
|
||||
*
|
||||
* For every data flow node `a` we have two constraint variables, `+a` and `-a` (or just `a` and `-a`)
|
||||
* representing the numerical value of `a` and its negation. Negations let us reason about the sum of
|
||||
* two variables. For example:
|
||||
*
|
||||
* a + b <= 10 becomes a - (-b) <= 10
|
||||
*
|
||||
* It also lets us reason about the upper and lower bounds of a single variable:
|
||||
*
|
||||
* a <= 10 becomes a + a <= 20 becomes a - (-a) <= 20
|
||||
* a >= 10 becomes -a <= -10 becomes (-a) - a <= -20
|
||||
*
|
||||
* For the graph analogy to include the relationship between `a` and `-a`, all constraints
|
||||
* imply their dual constraint:
|
||||
*
|
||||
* a - b <= c implies (-b) - (-a) <= c
|
||||
*
|
||||
* That is, for every edge from a -> b, there is an edge with the same weight from (-b) -> (-a).
|
||||
*
|
||||
*
|
||||
* PATH FINDING:
|
||||
*
|
||||
* See `extendedEdge` predicate for details about how we find negative-weight paths in the graph.
|
||||
*
|
||||
*
|
||||
* CAVEATS:
|
||||
*
|
||||
* - We assume !(x <= y) means x > y, ignoring NaN, unless a nearby comment or identifier mentions NaN.
|
||||
*
|
||||
* - We assume integer arithmetic is exact, ignoring values above 2^53.
|
||||
*
|
||||
*/
|
||||
|
||||
/**
|
||||
* Contains predicates for reasoning about the relative numeric value of expressions.
|
||||
*/
|
||||
module RangeAnalysis {
|
||||
/**
|
||||
* Holds if the given node is relevant for range analysis.
|
||||
*/
|
||||
private predicate isRelevant(DataFlow::Node node) {
|
||||
node = any(Comparison cmp).getAnOperand().flow()
|
||||
or
|
||||
node = any(ConditionGuardNode guard).getTest().flow()
|
||||
or
|
||||
exists (DataFlow::Node succ | isRelevant(succ) |
|
||||
succ = node.getASuccessor()
|
||||
or
|
||||
linearDefinitionStep(succ, node, _, _)
|
||||
or
|
||||
exists (BinaryExpr bin | bin instanceof AddExpr or bin instanceof SubExpr |
|
||||
succ.asExpr() = bin and
|
||||
bin.getAnOperand().flow() = node))
|
||||
}
|
||||
|
||||
/**
|
||||
* Holds if the given node has a unique data flow predecessor.
|
||||
*/
|
||||
pragma[noinline]
|
||||
private predicate hasUniquePredecessor(DataFlow::Node node) {
|
||||
isRelevant(node) and
|
||||
strictcount(node.getAPredecessor()) = 1
|
||||
}
|
||||
|
||||
/**
|
||||
* Gets the definition of `node`, without unfolding phi nodes.
|
||||
*/
|
||||
DataFlow::Node getDefinition(DataFlow::Node node) {
|
||||
if hasUniquePredecessor(node) then
|
||||
result = getDefinition(node.getAPredecessor())
|
||||
else
|
||||
result = node
|
||||
}
|
||||
|
||||
/**
|
||||
* Gets a data flow node holding the result of the add/subtract operation in
|
||||
* the given increment/decrement expression.
|
||||
*/
|
||||
private DataFlow::Node updateExprResult(UpdateExpr expr) {
|
||||
exists (SsaExplicitDefinition def | def.getDef() = expr |
|
||||
result = DataFlow::ssaDefinitionNode(def))
|
||||
or
|
||||
expr.isPrefix() and
|
||||
result = expr.flow()
|
||||
}
|
||||
|
||||
/**
|
||||
* Gets a data flow node holding the result of the given componund assignment.
|
||||
*/
|
||||
private DataFlow::Node compoundAssignResult(CompoundAssignExpr expr) {
|
||||
exists (SsaExplicitDefinition def | def.getDef() = expr |
|
||||
result = DataFlow::ssaDefinitionNode(def))
|
||||
or
|
||||
result = expr.flow()
|
||||
}
|
||||
|
||||
/**
|
||||
* A 30-bit integer.
|
||||
*
|
||||
* Adding two such integers is guaranteed not to overflow. We simply omit constraints
|
||||
* whose parameters would exceed this range.
|
||||
*/
|
||||
private class Bias extends int {
|
||||
bindingset[this]
|
||||
Bias() {
|
||||
-536870912 < this and this < 536870912
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* Holds if `r` can be modelled as `r = root * sign + bias`.
|
||||
*
|
||||
* Only looks "one step", that is, does not follow data flow and does not recursively
|
||||
* unfold nested arithmetic expressions.
|
||||
*/
|
||||
private predicate linearDefinitionStep(DataFlow::Node r, DataFlow::Node root, int sign, Bias bias) {
|
||||
not exists(r.asExpr().getIntValue()) and
|
||||
(
|
||||
exists (AddExpr expr | r.asExpr() = expr |
|
||||
// r = root + k
|
||||
root = expr.getLeftOperand().flow() and
|
||||
bias = expr.getRightOperand().getIntValue() and
|
||||
sign = 1
|
||||
or
|
||||
// r = k + root
|
||||
bias = expr.getLeftOperand().getIntValue() and
|
||||
root = expr.getRightOperand().flow() and
|
||||
sign = 1)
|
||||
or
|
||||
exists (SubExpr expr | r.asExpr() = expr |
|
||||
// r = root - k
|
||||
root = expr.getLeftOperand().flow() and
|
||||
bias = -expr.getRightOperand().getIntValue() and
|
||||
sign = 1
|
||||
or
|
||||
// r = k - root
|
||||
bias = expr.getLeftOperand().getIntValue() and
|
||||
root = expr.getRightOperand().flow() and
|
||||
sign = -1)
|
||||
or
|
||||
exists (NegExpr expr | r.asExpr() = expr |
|
||||
// r = -root
|
||||
root = expr.getOperand().flow() and
|
||||
bias = 0 and
|
||||
sign = -1)
|
||||
or
|
||||
exists (UpdateExpr update | r = updateExprResult(update) |
|
||||
// r = ++root
|
||||
root = update.getOperand().flow() and
|
||||
sign = 1 and
|
||||
if update instanceof IncExpr then
|
||||
bias = 1
|
||||
else
|
||||
bias = -1)
|
||||
or
|
||||
exists (UpdateExpr update | r.asExpr() = update | // Return value of x++ is just x (coerced to an int)
|
||||
// r = root++
|
||||
root = update.getOperand().flow() and
|
||||
not update.isPrefix() and
|
||||
sign = 1 and
|
||||
bias = 0)
|
||||
or
|
||||
exists (CompoundAssignExpr assign | r = compoundAssignResult(assign) |
|
||||
root = assign.getLhs().flow() and
|
||||
sign = 1 and
|
||||
(
|
||||
// r = root += k
|
||||
assign instanceof AssignAddExpr and
|
||||
bias = assign.getRhs().getIntValue()
|
||||
or
|
||||
// r = root -= k
|
||||
assign instanceof AssignSubExpr and
|
||||
bias = -assign.getRhs().getIntValue()
|
||||
))
|
||||
)
|
||||
}
|
||||
|
||||
/**
|
||||
* Holds if `r` can be modelled as `r = root * sign + bias`.
|
||||
*/
|
||||
predicate linearDefinition(DataFlow::Node r, DataFlow::Node root, int sign, Bias bias) {
|
||||
if hasUniquePredecessor(r) then
|
||||
linearDefinition(r.getAPredecessor(), root, sign, bias)
|
||||
else if linearDefinitionStep(r, _, _, _) then
|
||||
exists (DataFlow::Node pred, int sign1, int bias1, int sign2, int bias2 |
|
||||
// r = pred * sign1 + bias1
|
||||
linearDefinitionStep(r, pred, sign1, bias1) and
|
||||
// pred = root * sign2 + bias2
|
||||
linearDefinition(pred, root, sign2, bias2) and
|
||||
// r = (root * sign2 + bias2) * sign1 + bias1
|
||||
sign = sign1 * sign2 and
|
||||
bias = bias1 + sign1 * bias2)
|
||||
else (
|
||||
isRelevant(r) and
|
||||
root = r and
|
||||
sign = 1 and
|
||||
bias = 0
|
||||
)
|
||||
}
|
||||
|
||||
/**
|
||||
* Holds if `r` can be modelled as `r = xroot * xsign + yroot * ysign + bias`.
|
||||
*/
|
||||
predicate linearDefinitionSum(DataFlow::Node r, DataFlow::Node xroot, int xsign, DataFlow::Node yroot, int ysign, Bias bias) {
|
||||
if hasUniquePredecessor(r) then
|
||||
linearDefinitionSum(r.getAPredecessor(), xroot, xsign, yroot, ysign, bias)
|
||||
else if exists(r.asExpr().getIntValue()) then
|
||||
none() // do not model constants as sums
|
||||
else (
|
||||
exists (AddExpr add, int bias1, int bias2 | r.asExpr() = add |
|
||||
// r = r1 + r2
|
||||
linearDefinition(add.getLeftOperand().flow(), xroot, xsign, bias1) and
|
||||
linearDefinition(add.getRightOperand().flow(), yroot, ysign, bias2) and
|
||||
bias = bias1 + bias2)
|
||||
or
|
||||
exists (SubExpr sub, int bias1, int bias2 | r.asExpr() = sub |
|
||||
// r = r1 - r2
|
||||
linearDefinition(sub.getLeftOperand().flow(), xroot, xsign, bias1) and
|
||||
linearDefinition(sub.getRightOperand().flow(), yroot, -ysign, -bias2) and // Negate right-hand operand
|
||||
bias = bias1 + bias2)
|
||||
or
|
||||
linearDefinitionSum(r.asExpr().(NegExpr).getOperand().flow(), xroot, -xsign, yroot, -ysign, -bias)
|
||||
)
|
||||
}
|
||||
|
||||
/**
|
||||
* Holds if the given comparison can be modelled as `A <op> B + bias` where `<op>` is the comparison operator,
|
||||
* and `A` is `a * asign` and likewise `B` is `b * bsign`.
|
||||
*/
|
||||
predicate linearComparison(Comparison comparison, DataFlow::Node a, int asign, DataFlow::Node b, int bsign, Bias bias) {
|
||||
exists(Expr left, Expr right, int bias1, int bias2 | left = comparison.getLeftOperand() and right = comparison.getRightOperand() |
|
||||
// A <= B + c
|
||||
linearDefinition(left.flow(), a, asign, bias1) and
|
||||
linearDefinition(right.flow(), b, bsign, bias2) and
|
||||
bias = bias2 - bias1
|
||||
or
|
||||
// A - B + c1 <= c2 becomes A <= B + (c2 - c1)
|
||||
linearDefinitionSum(left.flow(), a, asign, b, -bsign, bias1) and
|
||||
right.getIntValue() = bias2 and
|
||||
bias = bias2 - bias1
|
||||
or
|
||||
// c1 <= -A + B + c2 becomes A <= B + (c2 - c1)
|
||||
left.getIntValue() = bias1 and
|
||||
linearDefinitionSum(right.flow(), a, -asign, b, bsign, bias2) and
|
||||
bias = bias2 - bias1
|
||||
)
|
||||
}
|
||||
|
||||
/**
|
||||
* Holds if the given container has a comment or identifier mentioning `NaN`.
|
||||
*/
|
||||
predicate hasNaNIndicator(StmtContainer container) {
|
||||
exists (Comment comment |
|
||||
comment.getText().regexpMatch("(?s).*N[aA]N.*") and
|
||||
comment.getFile() = container.getFile() and
|
||||
(
|
||||
comment.getLocation().getStartLine() >= container.getLocation().getStartLine() and
|
||||
comment.getLocation().getEndLine() <= container.getLocation().getEndLine()
|
||||
or
|
||||
comment.getNextToken() = container.getFirstToken()
|
||||
))
|
||||
or
|
||||
exists (Identifier id | id.getName() = "NaN" or id.getName() = "isNaN" |
|
||||
id.getContainer() = container)
|
||||
}
|
||||
|
||||
/**
|
||||
* Holds if `guard` asserts that the outcome of `A <op> B + bias` is true, where `<op>` is a comparison operator.
|
||||
*/
|
||||
predicate linearComparisonGuard(ConditionGuardNode guard, DataFlow::Node a, int asign, string operator, DataFlow::Node b, int bsign, Bias bias) {
|
||||
exists (Comparison compare | compare = getDefinition(guard.getTest().flow()).asExpr() |
|
||||
linearComparison(compare, a, asign, b, bsign, bias) and
|
||||
(
|
||||
guard.getOutcome() = true and operator = compare.getOperator()
|
||||
or
|
||||
not hasNaNIndicator(guard.getContainer()) and
|
||||
guard.getOutcome() = false and operator = negateOperator(compare.getOperator())
|
||||
)
|
||||
)
|
||||
}
|
||||
|
||||
/**
|
||||
* Gets the binary operator whose return value is the opposite of `operator` (excluding NaN comparisons).
|
||||
*/
|
||||
private string negateOperator(string operator) {
|
||||
operator = "==" and result = "!=" or
|
||||
operator = "===" and result = "!==" or
|
||||
operator = "<" and result = ">=" or
|
||||
operator = ">" and result = "<=" or
|
||||
operator = negateOperator(result)
|
||||
}
|
||||
|
||||
/**
|
||||
* Holds if immediately after `cfg` it becomes known that `A <= B + c`.
|
||||
*
|
||||
* These are the initial inputs to the difference bound constraint system.
|
||||
*
|
||||
* The dual constraint `-B <= -A + c` is not included in this predicate.
|
||||
*/
|
||||
predicate comparisonEdge(ControlFlowNode cfg, DataFlow::Node a, int asign, DataFlow::Node b, int bsign, Bias bias, boolean sharp) {
|
||||
// A <= B + c
|
||||
linearComparisonGuard(cfg, a, asign, "<=", b, bsign, bias) and
|
||||
sharp = false
|
||||
or
|
||||
// A < B + c
|
||||
linearComparisonGuard(cfg, a, asign, "<", b, bsign, bias) and
|
||||
sharp = true
|
||||
or
|
||||
// A <= B + c iff B >= A - c
|
||||
linearComparisonGuard(cfg, b, bsign, ">=", a, asign, -bias) and
|
||||
sharp = false
|
||||
or
|
||||
// A < B + c iff B > A - c
|
||||
linearComparisonGuard(cfg, b, bsign, ">", a, asign, -bias) and
|
||||
sharp = true
|
||||
or
|
||||
sharp = false and
|
||||
exists (string operator | operator = "==" or operator = "===" |
|
||||
// A == B + c iff A <= B + c and B <= A - c
|
||||
linearComparisonGuard(cfg, a, asign, operator, b, bsign, bias)
|
||||
or
|
||||
linearComparisonGuard(cfg, b, bsign, operator, a, asign, -bias)
|
||||
)
|
||||
}
|
||||
|
||||
/**
|
||||
* Holds if `node` is a phi node with `left` and `right` has the only two inputs.
|
||||
*
|
||||
* Note that this predicate is symmetric: when it holds for (left, right) it also holds for (right, left).
|
||||
*/
|
||||
predicate binaryPhiNode(DataFlow::Node node, DataFlow::Node left, DataFlow::Node right) {
|
||||
exists (SsaPhiNode phi | node = DataFlow::ssaDefinitionNode(phi) |
|
||||
isRelevant(node) and
|
||||
strictcount(phi.getAnInput()) = 2 and
|
||||
left = DataFlow::ssaDefinitionNode(phi.getAnInput()) and
|
||||
right = DataFlow::ssaDefinitionNode(phi.getAnInput()) and
|
||||
left != right)
|
||||
}
|
||||
|
||||
/**
|
||||
* Holds if `A <= B + c` can be determined based on a phi node.
|
||||
*/
|
||||
predicate phiEdge(ControlFlowNode cfg, DataFlow::Node a, int asign, DataFlow::Node b, int bsign, Bias c) {
|
||||
exists (DataFlow::Node phi, DataFlow::Node left, DataFlow::Node right |
|
||||
binaryPhiNode(phi, left, right) and
|
||||
cfg = phi.getBasicBlock()
|
||||
|
|
||||
// Both inputs are defined in terms of the same root:
|
||||
// phi = PHI(root + bias1, root + bias2)
|
||||
exists (DataFlow::Node root, int sign, Bias bias1, Bias bias2 |
|
||||
linearDefinition(left, root, sign, bias1) and
|
||||
linearDefinition(right, root, sign, bias2) and
|
||||
bias1 < bias2 and
|
||||
// root + bias1 <= phi <= root + bias2
|
||||
(
|
||||
// root <= phi - bias1
|
||||
a = root and asign = 1 and
|
||||
b = phi and bsign = 1 and
|
||||
c = -bias1
|
||||
or
|
||||
// phi <= root + bias2
|
||||
a = phi and asign = 1 and
|
||||
b = root and bsign = 1 and
|
||||
c = bias2
|
||||
)
|
||||
)
|
||||
or
|
||||
// One input is defined in terms of the phi node itself:
|
||||
// phi = PHI(phi + increment, x)
|
||||
exists (int increment, DataFlow::Node root, int sign, Bias bias |
|
||||
linearDefinition(left, phi, 1, increment) and
|
||||
linearDefinition(right, root, sign, bias) and
|
||||
(
|
||||
// If increment is positive (or zero):
|
||||
// phi >= right' + bias
|
||||
increment >= 0 and
|
||||
a = root and asign = sign and
|
||||
b = phi and bsign = 1 and
|
||||
c = -bias
|
||||
or
|
||||
// If increment is negative (or zero):
|
||||
// phi <= right' + bias
|
||||
increment <= 0 and
|
||||
a = phi and asign = 1 and
|
||||
b = root and bsign = sign and
|
||||
c = bias
|
||||
)
|
||||
)
|
||||
)
|
||||
}
|
||||
|
||||
/**
|
||||
* Holds if a comparison implies that `A <= B + c`.
|
||||
*
|
||||
* Comparisons where one operand is really a constant are converted into a unary constraint.
|
||||
*/
|
||||
predicate foldedComparisonEdge(ControlFlowNode cfg, DataFlow::Node a, int asign, DataFlow::Node b, int bsign, Bias c, boolean sharp) {
|
||||
// A <= B + c (where A and B are not constants)
|
||||
comparisonEdge(cfg, a, asign, b, bsign, c, sharp) and
|
||||
not exists(a.asExpr().getIntValue()) and
|
||||
not exists(b.asExpr().getIntValue())
|
||||
or
|
||||
// A - k <= c becomes A - (-A) <= 2*(k + c)
|
||||
exists (DataFlow::Node k, int ksign, Bias kbias, Bias value |
|
||||
comparisonEdge(cfg, a, asign, k, ksign, kbias, sharp) and
|
||||
value = k.asExpr().getIntValue() and
|
||||
b = a and
|
||||
bsign = -asign and
|
||||
c = 2 * (value * ksign + kbias))
|
||||
or
|
||||
// k - A <= c becomes -A - A <= 2(-k + c)
|
||||
exists (DataFlow::Node k, int ksign, Bias kbias, Bias value |
|
||||
comparisonEdge(cfg, k, ksign, b, bsign, kbias, sharp) and
|
||||
value = k.asExpr().getIntValue() and
|
||||
a = b and
|
||||
asign = -bsign and
|
||||
c = 2 * (-value * ksign + kbias))
|
||||
or
|
||||
// For completeness, generate a contradictory constraint for trivially false conditions.
|
||||
exists (DataFlow::Node k, int ksign, Bias bias, int avalue, int kvalue |
|
||||
comparisonEdge(cfg, a, asign, k, ksign, bias, sharp) and
|
||||
avalue = a.asExpr().getIntValue() * asign and
|
||||
kvalue = k.asExpr().getIntValue() * ksign and
|
||||
(avalue > kvalue + bias or sharp = true and avalue = kvalue + bias) and
|
||||
a = b and
|
||||
asign = bsign and
|
||||
c = -1)
|
||||
}
|
||||
|
||||
/**
|
||||
* The set of initial edges including those from dual constraints.
|
||||
*/
|
||||
predicate seedEdge(ControlFlowNode cfg, DataFlow::Node a, int asign, DataFlow::Node b, int bsign, Bias c, boolean sharp) {
|
||||
foldedComparisonEdge(cfg, a, asign, b, bsign, c, sharp)
|
||||
or
|
||||
phiEdge(cfg, a, asign, b, bsign, c) and sharp = false
|
||||
}
|
||||
|
||||
private predicate seedEdgeWithDual(ControlFlowNode cfg, DataFlow::Node a, int asign, DataFlow::Node b, int bsign, Bias c, boolean sharp) {
|
||||
// A <= B + c
|
||||
seedEdge(cfg, a, asign, b, bsign, c, sharp)
|
||||
or
|
||||
// -B <= -A + c (dual constraint)
|
||||
seedEdge(cfg, b, -bsign, a, -asign, c, sharp)
|
||||
}
|
||||
|
||||
/**
|
||||
* Adds a negative and positive integer, but only if they are within in the same
|
||||
* order of magnitude.
|
||||
*/
|
||||
bindingset[x, sharpx, y, sharpy]
|
||||
private int wideningAddition(int x, boolean sharpx, int y, boolean sharpy) {
|
||||
(x < 0 or x = 0 and sharpx = true) and
|
||||
(y > 0 or y = 0 and sharpy = false) and
|
||||
(
|
||||
x <= 0 and x >= 0
|
||||
or
|
||||
y <= 0 and y >= 0
|
||||
or
|
||||
// If non-zero, check that the values are within a factor 16 of each other
|
||||
x.abs().bitShiftRight(4) < y.abs() and
|
||||
y.abs().bitShiftRight(4) < x.abs()
|
||||
) and
|
||||
result = x + y
|
||||
}
|
||||
|
||||
/**
|
||||
* Applies a restricted transitive rule to the edge set.
|
||||
*
|
||||
* In particular, we apply the transitive rule only where a negative edge followed by a non-negative edge.
|
||||
* For example:
|
||||
*
|
||||
* A --(-1)--> B --(+3)--> C
|
||||
*
|
||||
* yields:
|
||||
*
|
||||
* A --(+2)--> C
|
||||
*
|
||||
* In practice, the restriction to edges of different sign prevent the
|
||||
* quadratic blow-up you would normally get from a transitive closure.
|
||||
*
|
||||
* It also prevents the relation from becoming infinite in case
|
||||
* there are negative-weight cycles, where the transitive weights would
|
||||
* otherwise diverge towards minus infinity.
|
||||
*
|
||||
* Moreover, the rule is enough to guarantee the following property:
|
||||
*
|
||||
* A negative-weight path from X to Y exists iff a path of negative-weight edges exists from X to Y.
|
||||
*
|
||||
* This means negative-weight cycles (contradictions) can be detected using simple cycle detection.
|
||||
*/
|
||||
pragma[noopt]
|
||||
private predicate extendedEdge(DataFlow::Node a, int asign, DataFlow::Node b, int bsign, Bias c, boolean sharp, ControlFlowNode cfg) {
|
||||
seedEdgeWithDual(cfg, a, asign, b, bsign, c, sharp)
|
||||
or
|
||||
// One of the two CFG nodes must dominate the other, and `cfg` must be bound to the dominated one.
|
||||
exists (ControlFlowNode cfg1, ControlFlowNode cfg2 |
|
||||
// They are in the same basic block
|
||||
extendedEdgeCandidate(a, asign, b, bsign, c, sharp, cfg1, cfg2) and
|
||||
exists (BasicBlock bb, int i, int j |
|
||||
bb.getNode(i) = cfg1 and
|
||||
bb.getNode(j) = cfg2 and
|
||||
if i < j then
|
||||
cfg = cfg2
|
||||
else
|
||||
cfg = cfg1)
|
||||
or
|
||||
// They are in different basic blocks
|
||||
extendedEdgeCandidate(a, asign, b, bsign, c, sharp, cfg1, cfg2) and
|
||||
exists (BasicBlock cfg1BB, ReachableBasicBlock cfg1RBB, BasicBlock cfg2BB, ReachableBasicBlock cfg2RBB |
|
||||
cfg1BB = cfg1.getBasicBlock() and
|
||||
cfg1RBB = cfg1BB.(ReachableBasicBlock) and
|
||||
cfg2BB = cfg2.getBasicBlock() and
|
||||
cfg2RBB = cfg2BB.(ReachableBasicBlock) and
|
||||
(
|
||||
cfg1RBB.strictlyDominates(cfg2BB) and
|
||||
cfg = cfg2
|
||||
or
|
||||
cfg2RBB.strictlyDominates(cfg1RBB) and
|
||||
cfg = cfg1
|
||||
))
|
||||
) and
|
||||
cfg instanceof ControlFlowNode
|
||||
}
|
||||
|
||||
/**
|
||||
* Holds if an extended edge from `A` to `B` can potentially be generates from two edges, from `cfg1` and `cfg2`, respectively.
|
||||
*
|
||||
* This does not check for dominance between `cfg1` and `cfg2`.
|
||||
*/
|
||||
pragma[nomagic]
|
||||
private predicate extendedEdgeCandidate(DataFlow::Node a, int asign, DataFlow::Node b, int bsign, Bias c, boolean sharp, ControlFlowNode cfg1, ControlFlowNode cfg2) {
|
||||
exists (DataFlow::Node mid, int midx, Bias c1, Bias c2, boolean sharp1, boolean sharp2 |
|
||||
extendedEdge(a, asign, mid, midx, c1, sharp1, cfg1) and
|
||||
extendedEdge(mid, midx, b, bsign, c2, sharp2, cfg2) and
|
||||
(a != mid or asign != midx) and
|
||||
(b != mid or bsign != midx) and
|
||||
sharp = sharp1.booleanOr(sharp2) and
|
||||
c = wideningAddition(c1, sharp1, c2, sharp2))
|
||||
}
|
||||
|
||||
/**
|
||||
* Holds if there is a negative-weight edge from src to dst.
|
||||
*/
|
||||
private predicate negativeEdge(DataFlow::Node a, int asign, DataFlow::Node b, int bsign, ControlFlowNode cfg) {
|
||||
exists (int weight, boolean sharp | extendedEdge(a, asign, b, bsign, weight, sharp, cfg) |
|
||||
weight = 0 and sharp = true // a strict "< 0" edge counts as negative
|
||||
or
|
||||
weight < 0)
|
||||
}
|
||||
|
||||
/**
|
||||
* Holds if `src` can reach `dst` using only negative-weight edges.
|
||||
*
|
||||
* The initial outgoing edge from `src` must be derived at `cfg`.
|
||||
*/
|
||||
pragma[noopt]
|
||||
private predicate reachableByNegativeEdges(DataFlow::Node a, int asign, DataFlow::Node b, int bsign, ControlFlowNode cfg) {
|
||||
negativeEdge(a, asign, b, bsign, cfg)
|
||||
or
|
||||
exists(DataFlow::Node mid, int midx, ControlFlowNode midcfg |
|
||||
reachableByNegativeEdges(a, asign, mid, midx, cfg) and
|
||||
negativeEdge(mid, midx, b, bsign, midcfg) and
|
||||
exists (BasicBlock bb, int i, int j |
|
||||
bb.getNode(i) = midcfg and
|
||||
bb.getNode(j) = cfg and
|
||||
i <= j))
|
||||
or
|
||||
// Same as above, but where CFG nodes are in different basic blocks
|
||||
exists(DataFlow::Node mid, int midx, ControlFlowNode midcfg, BasicBlock midBB, ReachableBasicBlock midRBB, BasicBlock cfgBB |
|
||||
reachableByNegativeEdges(a, asign, mid, midx, cfg) and
|
||||
negativeEdge(mid, midx, b, bsign, midcfg) and
|
||||
midBB = midcfg.getBasicBlock() and
|
||||
midRBB = midBB.(ReachableBasicBlock) and
|
||||
cfgBB = cfg.getBasicBlock() and
|
||||
midRBB.strictlyDominates(cfgBB)
|
||||
)
|
||||
}
|
||||
|
||||
/**
|
||||
* Holds if the condition asserted at `guard` is contradictory, that is, its condition always has the
|
||||
* opposite of the expected outcome.
|
||||
*/
|
||||
predicate isContradictoryGuardNode(ConditionGuardNode guard) {
|
||||
exists (DataFlow::Node a, int asign | reachableByNegativeEdges(a, asign, a, asign, guard))
|
||||
}
|
||||
}
|
||||
Reference in New Issue
Block a user