Summary:
Llair.Func.mk makes two passes over the CFG to resolve block parents,
jump destinations, and eliminate jumps to jumps. This is not
economical, but more importantly the current code mistakenly uses the
`retreating` metadata before it is set correctly. This diff combines
these passes into a single one, which also incorporates setting the
retreating field from Llair.Program.mk.
This avoids nontermination on code that contains immediate self-jumps
such as `L: goto L;` that LLVM 11 can now generate.
Reviewed By: jvillard
Differential Revision: D27262512
fbshipit-source-id: 0543ba669
Summary:
This warning (68) triggers when a function argument pattern depends on
mutable state, which prevents the remaining arguments from being
uncurried, causing additional closure allocations.
Reviewed By: jvillard
Differential Revision: D27188311
fbshipit-source-id: a43354e15
Summary:
It is possible for the filename of a source location to be the empty
string. Fpath.v asserts when passed an empty string.
Reviewed By: jvillard
Differential Revision: D27188304
fbshipit-source-id: a7d73444b
Summary:
The only difference between `program` and `identified` variables is
terminology, technically they are redundant.
Reviewed By: jvillard
Differential Revision: D26451308
fbshipit-source-id: eb4e7be43
Summary:
Negating the ids of program variables leads to inverting the order on
them. This is logically fine, the order is still a valid total order.
But it can lead to choosing younger variables as equality class
representatives over older variables, and thereby lead to more churn
as adding an equality is more likely to cause a change of
representative, and hence additional normalizing rewrites.
Reviewed By: jvillard
Differential Revision: D26451304
fbshipit-source-id: eb20d1901
Summary:
Theory.solved is a list of pairs of terms representing solved
equalities. The order of the pairs is very important, which is not
apparent from the type. This diff introduces an oriented_equality type
to make this more clear.
Reviewed By: jvillard
Differential Revision: D26451303
fbshipit-source-id: 56a49e601
Summary:
Generalize the existing find_or_add and find_and_remove operations to
find_update. Slightly simplify the interfaces of change, update,
find_or_add, and find_and_remove, reducing the gap to the natural
underlying functionality.
Reviewed By: jvillard
Differential Revision: D26451305
fbshipit-source-id: 89f67c84d
Summary:
The normalization and then extension of the carrier can be combined
into one pass. This weakens the property that this normalization needs
to achieve, which yields a small simplification, and combining the
passes is a minor optimization.
Reviewed By: jvillard
Differential Revision: D26400406
fbshipit-source-id: 8a3cbb2de
Summary:
An invariant of `And` and `Or` formulas is that they are flattened,
that is, `And` formulas do not have positive immediate subformulas of
form `And` nor negative immediate subformulas of form `Or`, and
similarly for `Or`. This invariant is ensured by the formula
constructors, which scan the immediate subformulas for cases that need
to be flattened.
When mapping over formulas, this repeated scanning is a performance
bottleneck. Most of the work of flattening is wasted since the input
formulas are necessarily already in flattened form, and the common
case is that the transformation preserves the flattened form. This
diff optimizes this case by detecting violations of flattening during
the transformation, performing the needed flattening, and avoiding the
general constructor that scans for flattening violations.
Reviewed By: jvillard
Differential Revision: D26338014
fbshipit-source-id: 9f15cca58
Summary:
Update the first-order equality solver for the sequence theory to
avoid searching the whole representation now that the super-term index
is maintained.
Reviewed By: jvillard
Differential Revision: D26338015
fbshipit-source-id: 24a9a19b6
Summary:
Similarly to terms, use the Comparar interface of Sets to defined
formulas using recursive types instead of recursive modules.
Reviewed By: jvillard
Differential Revision: D26250512
fbshipit-source-id: 84f0ae8c0
Summary: No functional change, just to make the interface easier to read
Reviewed By: jvillard
Differential Revision: D26250513
fbshipit-source-id: f3b07bccc
Summary:
Terms include Arithmetic terms, which are polynomials over terms
themselves. Monomials are represented as maps from
terms (multiplicative factors) to integers (their powers). Polynomials
are represented as maps from monomials (indeterminates) to
rationals (coefficients). In particular, terms are represented using
maps whose keys are terms themselves. This is currently implemented
using recursive modules.
This diff uses the Comparer-based interface of Maps to express this
cycle as recursive *types* rather than recursive *modules*, see the
very beginning of trm.ml. The rest of the changes are driven by the
need to expose the Arithmetic.t type at toplevel, outside the functor
that defines the arithmetic operations, and changes to stage the
definition of term and polynomial operations to remove unnecessary
recursion.
One might hope that these changes are just moving code around, but due
to how recursive modules are implemented, this refactoring is
motivated by performance profiling. In every cycle between recursive
modules, at least one of the modules must be "safe". A "safe" module
is one where all exposed values have function type. This allows the
compiler to initialize that module with functions that immediately
raise an exception, define the other modules using it, and then tie
the recursive knot by backpatching the safe module with the actual
functions at the end. This implementation works, but has the
consequence that the compiler must treat calls to functions of safe
recursive modules as indirect calls to unknown functions. This means
that they are not inlined or even called by symbol, and instead
calling them involves spilling registers if needed, loading their
address from memory, calling them by address, and restoring any
spilled registers. For operations like Trm.compare that are a handful
of instructions on the hot path, this is a significant
difference. Since terms are the keys of maps and sets in the core of
the first-order equality solver, those map operations are very very
hot.
Reviewed By: jvillard
Differential Revision: D26250533
fbshipit-source-id: f79334c68
Summary:
No functional change, apart from a minor change in invariant checking
for core constructors. Only to reduce diffs of upcoming changes.
Reviewed By: jvillard
Differential Revision: D26250518
fbshipit-source-id: a731e4fd6
Summary:
Inline functor applications, principally those involved in the
representation of (arithmetic) terms. This enables direct calls to
functions in the result of functor applications, instead of indirect
via an offset from the start of the module block resulting from the
functor appication. This also then enables further inlining
opportunities.
Reviewed By: jvillard
Differential Revision: D26338016
fbshipit-source-id: 27b770fa3
Summary:
Currently there is a direct cycle of dependencies Var -> Trm ->
Var. Inside Trm, these two modules are defined as mutually recursive
modules, so at some level this is ok. Due to internals of how
recursive modules are compiled, recursive modules that are
"unsafe" (that is, export some value of non-function type) are
compiled more efficiently. This diff moves Var from being defined
recursively with Trm to being a submodule of Trm. This eliminates that
Var -> Trm -> Var cycle and enables making Trm an "unsafe"
module. (Trm is still mutually recursive with Arith (and Arith0), but
they are "safe".)
The difference between how "safe" and "unsafe" recursive modules are
compiled, in this context, is that the safe ones are first initialized
to a block containing dummy function closures that immediately raise,
then the unsafe ones are initialized as normal modules, and then the
safe ones are back-patched. A consequence of this is that calls to
functions in safe recursive modules are "unknown" calls, and they are
implemented as loading from a pointer to a closure and calling the
generic caml_apply function. In contrast, calls to functions in normal
or "unsafe" modules can be known, and get compiled to direct assembly
calls to the statically resolved callee. The second case is
faster. Additionally, in the indirect case, the callee is unknown, and
so some register spilling and restoring is also potentially
involved. Normally (I guess), this difference in performance is not
significant, but it can be significant for e.g. compare or hash
functions of modules used as container keys / elements, where the
container data structure is very hot.
Reviewed By: ngorogiannis
Differential Revision: D26250517
fbshipit-source-id: ecef49b32
Summary:
Variable ids are currently unique non-negative integers, and register
ids are unique positive integers, so using register ids negated as
variable ids yields a situation where all variable ids are unique.
Reviewed By: jvillard
Differential Revision: D26250544
fbshipit-source-id: 9e47e9776
Summary:
`Sh.and_ b q` normalizes `b` using the equality context of `q` and
then conjoins the result to `q`. This is incorrect in case normalizing
`b` results in expressing it using existentials of `q`, which takes
the existentials out of their scope. So this diff changes from
essentially
`(∃x.Q) ∧ B = (∃x.Q) ∧ (∃x.Bρ)`
to
`(∃x.Q) ∧ B = (∃x'.Q[x'/x] ∧ Bρ)`
where `ρ` is the substitution that normalizes with respect to the
equality context.
Reviewed By: jvillard
Differential Revision: D26250536
fbshipit-source-id: 05f5c48c0
Summary:
Add assertion to detect if the variable a function call modifies
appears in the actual parameters. In this case, the implementation is
incomplete.
Reviewed By: jvillard
Differential Revision: D26250527
fbshipit-source-id: d2e81a467
Summary:
Instructions and terminators are not themselves unique, there can be
two instructions that compare equal in distinct blocks, and likewise
for terminators. Such clashes make the coverage statistics report
incorrectly low coverage. This diff fixes this by also considering the
parent blocks.
Reviewed By: jvillard
Differential Revision: D26250514
fbshipit-source-id: 93f916eab
Summary:
Instead of passing the Var.strength across interfaces, pass the Trm.pp
partially applied to the Var.strength. This hides the type of
Var.strength, and thereby removes the need for `Arithmetic.S.var`.
Reviewed By: jvillard
Differential Revision: D26250516
fbshipit-source-id: 64dbd3870
Summary: The wrapper functions do not get inlined enough to avoid excess allocations.
Reviewed By: jvillard
Differential Revision: D25946110
fbshipit-source-id: 8dc3d8259
Summary:
This diff replaces Fml.fold_dnf, which eagerly enumerates a
disjunctive-normal form expansion, with iter_dnf which can be iterated
lazily. Fml.iter_dnf is then used in Context.dnf. This means that the
full DNF expansion does not need to be allocated by
e.g. Smtlib.check_sat, which yields huge peak memory savings.
Reviewed By: jvillard
Differential Revision: D25946111
fbshipit-source-id: e6d179e9f
Summary:
Currently detection and propagation of consequences of newly added
equality constraints is implemented using some operations that
traverse the entire representation of the equality relation. This
implementation strategy makes it relatively easy and robust to ensure
that the consequences of constraints are treated in a sound and
complete way. It does have the drawback that adding a constraint
involves work that is proportional (nonlinearly even) to the number of
all constraints. This can become a performance bottleneck.
This diff reimplements the core operations using an additional data
structure that essentially tracks the "super-term" relation between
terms, which can be used to look up the existing constraints that
might be involved in deriving new consequences when adding
equalities. This means that adding new equalities no longer takes time
proportional to the size of the entire equality context
representation, but only to the number of existing constraints that
might interact with the new constraint. Also, rewriting terms
into canonical form no longer takes time proportional to size of the
whole context, but only performs table lookups on the subterms of the
term to canonize.
Reviewed By: jvillard
Differential Revision: D25883705
fbshipit-source-id: b3f266533
Summary:
This diff changes Context.extend to not only add terms and their
subterms to the carrier, but to simultaneously normalize the added
term. The normalized form is currently ignored, but it will be needed
when adding use lists.
Reviewed By: jvillard
Differential Revision: D25883709
fbshipit-source-id: bfe06f2a8
Josh Berdine