- where size of integer and floating point numbers matters (overflow behavior and interpretation of conversions)
+ cons
- perf: increases size of representation of Exp, perhaps a lot
- code complexity: need to plumb through target-specific data in order to e.g. be able to create equalities at intptr type
- instructions and globals could use accurate types to replace len fields with static sizeof type
- load instructions would need accurate types on reg to create equalities between it and its value in Exec
- memcpy and memmov would need types to create equality between src and dst in Exec
- formals would need types, to create equalities between formals and actuals in Domain
- types could be useful for approximate human-readable printing for general expressions
+ to print p+o as p.f, will likely need to consult what p is equal to, to find some meaningful type, and it could easily take much more work than this to produce reliably readable results
- target-specific types and layout
+ change Typ.target into a separate module
+ construct an instance in frontend as first step
+ use it during translation
+ return it as part of program
+ pass it from Control to Domain, etc.
- function types could include the types of throw continuation args
but they are currently the same for all functions: i8*
** ? change blocks to take all free variables as args
+ currently the scope of an identifier bound by e.g. Load is the continuation of the inst as well as all the conts that it dominates, this is somewhat messy
+ build a table from blocks to conts
+ build a table from blocks to free vars
+ need a fixed-point computation for blocks to vars table
+ to xlate a block
- get the terminator
- if all the destination blocks except the current block are already in the table
* then
- xlate block itself like now
+ when get to the terminal
+ look up free vars vector of the jump destinaton in table
+ map over the vector
* if the var is the name of a PHI instr
- find and translate the arg for the src block of the jmp instr
use the find_map of find_jump_args
* else use the var
+ use this vector for the jump args
- compute the free vars of its code
- use this vector for the cont params
- add free vars to table
- add block to cont mapping to table
* else recurse over the destination blocks except the current block
+ after entry block (and recursively everything reachable from it) is xlated, map over the function block list looking up from the table to get order of conts to match order of blocks
** hoist alloca's to the beginning of the entry block whenever possible
** clean up translation of intrinsics
separation between xlate_intrinsic (which translates an intrinsic function name to an expression constructor) and the Call case of xlate_instr (which translates calls to intrinsic functions to instructions) is not clear
** extract struct field names from llvm debug info
** normalize cfg
- remove unreachable blocks
- combine blocks with cmnd= []; term= Unreachable into one
** support variadic functions
- lower by implementing in terms of the core
- implement the va_list type as a pair or pointers into a stack represented as a linked-list, one pointer to the current element and one to the head
- a call to a variadic function pushes the args in reverse order, so that the first arg is at the top of the stack, and passes a pointer to the top as the last arg to the callee
- va_start intrinsic returns a pointer to the first va arg, by just projecting the current pointer from the last arg
- va_arg instruction returns the current va arg using argument va_list pointer to the stack, and sets the argument va_list current pointer to the next stack element
- va_copy is just a pointer copy of the source to destination va_list arguments, creating another pointer into the stack of va args, the head pointer of copies is null
- va_end deallocates the list starting from the head pointer
** support dynamic sized stack allocation (alloca in non-entry blocks)
- lower by implementing in terms of the core
- add a linked list of stack slots data structure
- each element contains
+ a pointer to some memory allocated for that slot's contents
+ a pointer to the next older slot
+ a pointer to the beginning of the function's stack frame
- add a global variable that always points to the head of the stack
- alloca in non-entry blocks adds an element and stores the result of alloc in it, sets next, and uses the frame pointer of the previous head
- function call adds a 'frame sentinel' element whose frame pointer points to itself, slot pointer is null (but used for va_arg below)
- function return (and other popping terminators) traverses the stack, popping elements, calling free on the slot pointers, until the element pointed to by the frame pointer is encountered
- stacksave intrinsic returns a pointer to a stack element
- stackrestore intrinsic pops the stack like return but only back to the argument pointer
** handle inline asm enough to over-approximate control-flow
- inline asm can take addresses of blocks as args, that can be jumped to
- treating inline asm conservatively requires considering these control flows
** support missing intrinsics
** support vector operations
- by lowering into multiple scalar operations
- most cases handled by Frontend.transform
- tests have a few exceptions, possibly for only unrealistic code
- one solution: pre-process llvm to remove [optnone] attributes before running scalarizer pass
** ? remove Exp.Nondet, replace with free variables
it is not obvious whether it will be simpler to use free variables instead of Nondet in the frontend, or to treat Nondet as a single-occurrence existential variable in the analyzer
** llvm bugs?
- Why aren't shufflevector instructions with zeroinitializer masks eliminated by the scalarizer pass?
** optimize: can identity mappings in lkp be removed?
* symbolic heap
** Congruence should handle equalities of equalities to integers
currently handled by Sh.pure
** normalize exps in terms of reps
- add operation to normalize by rewriting in terms of reps
- check for unsat
- call it in Exec.assume
** eliminate existentials
by changing Congruence reps to avoid existentials if possible and then normalizing Sh ito reps
** add exps in pure and pto (including memory siz and arr) to carrier
** Sh.with_pure is an underspeced, tightly coupled, API: replace
Sh.with_pure assumes that the replaced pure part is defined in the same vocabulary, induces the same congruence, etc. This API is fragile, and ought to be replaced with something that has simpler assumptions without imposing an excessive pessimization.
** optimize Sh.and_ with direct implementation
** perhaps it would be better to allow us and xs to intersect
but to rename xs when binding them or otherwise operating under the quantifier. But it might be an unnecessary complication to always have to deal with the potential for shadowing.
** consider how to detect unsat formulas
in relation to also wanting to express formulas in terms of congruence
class representatives in order to perform quantifier elimination. Is
there a way to detect unsat at the same time / as part of the same
normalization?
** consider hoisting existentials over disjunction:
#+BEGIN_SRC ocaml
| _ ->
let us = Set.union q1.us q2.us in
let xs1, xs, xs2 = Set.diff_inter_diff q1.xs q2.xs in
let us1 = Set.union q1.us xs in
let us2 = Set.union q2.us xs in
{ us
; xs
; cong= Congruence.true_
; pure= []
; heap= []
; djns= [[{q1 with us= us1; xs= xs1}; {q2 with us= us2; xs= xs2}]] }
| _ ->
let xs1, vs1 = Set.inter_diff q1.xs q2.us in
let xs2, vs2 = Set.inter_diff q2.xs q1.us in
let us1 = Set.union q1.us vs1 in
let us2 = Set.union q2.us vs2 in
let us = Set.union q1.us q2.us in
let xs = Set.union vs1 vs2 in
{ us
; xs
; cong= Congruence.true_
; pure= []
; heap= []
; djns= [[{q1 with us= us1; xs= xs1}; {q2 with us= us2; xs= xs2}]] }
#+END_SRC
** consider how to arrange to have a complete set of variables
at the top of formulas so that freshening wrt them is guaranteed not to clash with subformulas. This would allow removing the call to freshen_xs in rename, which is called on every subformula for every freshen/rename operation. Is it complicated to make us always include xs, as well as the us of the subformulas? That would allow the top-level us to serve as such a complete set of vars. How often would we need to compute us - xs?
** think about how to avoid having to manipulate disjunct formulas
unnecessarily, e.g. freshening, etc.
** ? should star strengthen djns with stem's cong
** optimize: refactor Sh.pure to avoid `Congruence.(and_eq true_ ...)`
** consider strengthening cong of or_ at price of freshening existentials
** consider using the append case when freshening existentials is needed
** strengthen Sh.pure_approx
* solver
** solve more existential equations in excise_exp
If sub.pure contains an equation involving an existential, add equation to min, remove the var from xs, continue. If all pure atoms normalize to true, added equations induce good existential witnesses, and excise will return them as part of min.
in calls to exec_spec, only vars in post need appear in xs, others can be existential in foot
* domain
** implement resolve_virtual to not skip virtual calls
** consider lazy renaming
- instead of eagerly constructing renaming substitutions, traverse the formula and lazily construct the renaming substitution map
- may be better in case there are many variables that do not occur in the formula
* control
** change Depths.t from environment- to state-like treatment
- currently each waiting state has an associated depths map
- the depths of all edges into a destination are joined
- could the depths be just threaded through Work.run instead?
- this would involve changing type x to Depths.t -> t -> Depths.t * t, and removing Depths.t from waiting_states
- separate joining depths from joining states
- i.e. Change to repeatedly pop edges as long as the dst is the same, and only join the states for those. This would involve keeping the waiting states in the priority queue, and removing the waiting states map entirely.
** change Work.run to move Domain.join into ~f
** canonicalize renamings in stacks
It seems possible that two edges will be distinct only due to differences between choice of fresh variable names for shadowed variables. It is not obvious that this could not lead to an infinite number of Edge.t values even without recursion. Using predictable names for local variables, such as a pair of the declared name and the depth of the stack, would avoid these difficulties.
** Control uses Var.Set for locals, but could benefit from a set with constant-time union
* roadmap
** lazy tracing
- define a [Trace.t], move global [fs] into it, and thread through code
- add a parent-pointing tree/dag of printing thunks to [Trace.t]
- use "event" and "history" terminology
- change from immediately printing to creating a closure that prints when called, and add it to the dag
- add [fork] and [join] operations on [Trace.t]
- use [Trace.fork] in [Control.exec_term], and [Trace.join] in sync with [Domain.join] (in [Control.Work.run] or wherever)
- add a form of "terminal" trace events, which prints all the ancestor events
- change [Report] (and elsewhere?) to use Trace.terminal
- support ex postfacto trace exploration
+ add a global list of terminals
+ add to terminals list instead of eagerly printing ancestors of terminals
+ dump/Marshal trace state at exit
+ add subcommand for querying dumped traces
- list terminals
- print ancestors of given terminal
+ support changing enabled status ex postfacto
- record module and function names with printing thunks
- when printing, recheck [enabled]
- support incrementally writing trace data to file
- support incrementally printing history as requested, in reverse
- ? support more advanced queries
** parallelize frontend
- make a scan_types pass over all types to populate anon_struct_name, and change struct_name to only find, not add
see http://llvm.org/doxygen/ValueEnumerator_8cpp_source.html#l00321
- [Trace.fork] a trace for each function
- replace calls to fold_left_globals and fold_left_functions with calls to parmap
- memo_type and memo_value could be put in shared memory instead
+ better sharing (as much as with sequential translation)
+ all their contents will live forever anyway
+ would need to handle concurrent accesses
+ maybe better to put entire Llair.t into shared memory
+ ? shared memory = reancient + locks
** parallelize backend
- change exec_* functions to instead of transforming the worklist, to return the new jobs (each job is an edge, depth(s?), and state)
+ also, change tracing so that they return new events rather than transform the whole event dag
- adapt infer's ProcessPool
+ When a worker finishes its task, it writes to the "up" pipe, a message indicating that it is done, which includes the worker's id and a list of discovered jobs. Then it reads another task from its "down" pipe, which might block. Maybe it should do a slice of gc before reading.
+ The orc sits in a select waiting for the "up" pipe to be non-empty. Once it receives a message that a worker has finished, it reads responses from the "up" pipe, adding the jobs sent by the workers to the queue and add the now-idle workers to the back of the queue. When the "up" pipe is empty, it iterates through the idle workers, popping the next task from the queue and writing it to the worker's "down" pipe. Then the orc loops back to waiting on the "up" pipe. If the queue empties while there are still idle workers, keep the queue and add to it on the next finish message. Maybe the orc should check the "up" pipe between writes to worker "down" pipes.
+ Actually, repeatedly pop all the jobs for the same block from the queue, and send the list of states to the worker to join and execute from.
+ Currently in infer the operation of selecting the task to send to the child is trivial, but IIUC it does not have to be, and the list of tasks does not need to be computed beforehand. So, leaving the basic communication structure the same, it does not seem like a big change to extend the messages from worker to orc to also include a list of tasks to add to the queue, and to have the orc receive them, add them to a priority queue, pop the highest priority task from the queue and send it to the worker. Plus some check to see if there was an idle worker that could be given one of the tasks just returned to the orc.
- initial inefficient version
+ communicate blocks
- by forking workers after frontend finishes, thereby giving each worker a copy of the program
- then passing block parent name/index and block index
+ but could instead, with some manual serialization code, pass blocks to/from workers over pipes
- receiver must perform a lookup to find their local copy
+ communicate states using Marshal
- likely to be slow
- will proactively lose sharing of the representation
+ communicate trace events by forcing printing thunks to strings
- optimize by storing program in shared memory (reancient?)
+ don't need to finish translation before starting analysis
+ pass block address in reancient heap instead of indices
+ receiver no longer needs to perform a lookup
+ saves memory, and time to copy it, and time to futilely GC it in all workers
- optimize by communicating states without Marshal
+ could store them in a reancient heap and then communicate their index
- probably fast, but leaky
+ could use a reancient heap for each worker, where it would store its jobs, until there is not enough space, at which point it would delete the heap and allocate a new one, passing the heap to the orc over the pipe
- this would need make a deep copy of every entry, or else deleting the heap is unsafe since there could be sharing between entries
+ could perhaps have immortal heap of states appearing in function specs, try to keep sharing between communicated states and immortal ones, and take advantage of how Marshal won't follow pointers out of the GC heap to make communicated states small
+ really ought to have a global hash-cons structure which workers add states to in order to communicate them
+ check what flow/hack/zonc do
see fbcode/hphp/hack/src/heap/hh_shared.c
+ store trace events in shared memory
- to avoid forcing them eagerly
- need a way to Marshal them from shared memory to write to file
+ perhaps serially at exit: copy to GC heap and Marshal as normal
+ perhaps incrementally copy oldest events from shared memory and Marshal to file
** relax global topological ordering
:PROPERTIES:
:ID: 6D6A0AF5-F68F-4726-95E5-178145A4CB9B
:END:
- needed for lazy translation and bottom-up analysis
- compute call graph (perhaps from ThinLTO info)
- topsort call graph (callee smaller number than caller)
+ possible alternative might be to translate functions leaving their sort_index unset
+ then set it when first encountered during analysis
+ this relies on the assumption that the analysis will perform an appropriately ordered search
+ this assumption needs to be checked
+ this is probably only applicable for top-down analysis
- add sort_index field to func like block
- change to topsort blocks intraprocedurally
- change priority queue to use lexicographically sorted pair of func and block indices, that is, (block.parent.sort_index, block.sort_index)
- if intraprocedural top orders are insufficient
+ change use of block sort_index for priority in queue
+ instead of choosing a total order (represented by ints), represent the partial order itself
+ build a graph with blocks as vertices and edges for non-retreating jumps
+ then a < b iff there is a path from a to b
+ perhaps keep the graph transitively-closed, and then a < b iff b is a successor of a
+ extending such a graph can only add new ordering relationships, never change existing ones, the partial order is stable under extension, so translating code while analyzing will not break the queue
+ is Fheap compatible with a partial order, rather than a total order?
+ when adding just-translated code, need to add edges for all existing (non-retreating?) Call sites of added functions: will need to index them
** lazy translation
- need to [[id:6D6A0AF5-F68F-4726-95E5-178145A4CB9B][generalize to partial weak topological order]] to enable adding code during analysis without breaking the priority queue
- translate function when analyzing a Call to a declared but untranslated function
- if in ThinLTO mode, will need to worry about finding/loading bitcode: will need an index from function names to bitcode modules where they are defined (ThinLTO should have this info)
** summarization
- ? standard over-approximation, or something more in tune with refutation
- ? procedures
- ? code segments between function entry and call sites
- common points:
+ summary includes
- precondition
- postcondition
- depth for which summary is "sound" assuming every worklist item has higher depth
+ a summary for a given pre and depth may be incomplete (if there is an item in the worklist)
+ a summary for a pre and depth may be extended with another for the same pre and depth, by disjoining the posts
- Currently expressions are shared between the IR for code and the analyzer's formulas. In order to strengthen the analyzer to detect overflow, it will likely be necessary to distinguish these, letting the analyzer normalize with polynomials, while keeping the same order of operations as the source for the IR. Plus another level of transformers for the IR expressions where their abstract values could be checked for being in range, etc.
