| Age | Commit message (Collapse) | Author |
|---|
|
|
|
Effectively reverts 7280e7b with a different method that isn't slow,
although it's not totally clear that this is correct - it could just
be more subtly wrong than before commit 7280e7b.
Following is mostly so I can remember what I did to document & write
up properly at some point:
What we do is compute 'pi' conditions as before by traversing the
dominator tree, we each node having a pi condition defined as the
conjunction of all guards between it and the start node in the
dominator tree. This is imprecise because we have situations like
1
/ \
2 3
| |
| 4
| |\
5 6 9
\ / |
7 10
|
8
where 8 = match_failure, 1 = start and 10 = return.
2, 3, 6 and 9 are guards as they come directly after a control flow
split, which always follows a conditional jump.
Here the path through the dominator tree for the match_failure is
1->7->8 which contains no guards so the pi condition would be empty.
What we do now is walk backwards (CFG must be acyclic at this point)
until we hit the join point prior to where we require a path
condition. We then take the disjunction of the pi conditions for the
join point's predecessors, so 5 and 6 in this case. Which gives us a
path condition of 2 | (3 & 6) as the dominator chains are 1->2->5 and
1->3->4->6.
I think this works as any split in the control flow must have been
caused by a conditional jump followed by distinct guards, so each of
the nodes immediately prior to a join point must be dominated by at
least one unique guard. It also explains why the pi conditions seem
sufficient to choose outcomes of phi functions.
If we hit a guard before a join (such as 9 for return's path
conditional) we just return the pi condition for that guard, i.e.
(3 & 9) for 10. If we reach start then the path condition is simply
true.
|
|
Makes the graph slightly cleaner, and means we can represent
conditionals in a way that shoud allow computing path conditions much
easier.
Essentially rather than a basic block being a list of instructions
where the last instruction is a jump (or other terminator) it is now a
list of instructions combined with an explict terminator, i.e.
CF_block (instrs @ I_jump (cval, label) ==> CF_block (instrs, T_jump (n, label))
Rather than storing the cval in the T_jump it just has a number that
links to a mapping from numbers to cvals, and we represent the
negation of any cval in that table by negating its number. Therefore
at any join point in the CFG, we can efficiently check when the
joining path conditionals contain both n and minus n and remove both.
|
|
Previously path conditionals for a node were defined as the path
conditional of the immediate dominator (+ a guard for explicit guard
nodes after conditional branches), whereas now they are the path
conditional of the immediate dominator plus an expression
encapsulating all the guards between the immediate dominator and the
node. This is needed as the previous method was incorrect for certain
control flow graphs.
This slows down the generated SMT massively, because it causes the
path conditionals to become huge when the immediate dominator is far
away from the node in question. It also changes computing path
conditionals from O(n) to O(n^2) which is not ideal as our inlined
graphs can become massive. Need to figure out a better way to generate
minimal path conditionals between the immediate dominator and the
node.
I upped the timeout for the SMT tests from 20s to 300s each but this
may still cause a failure in Jenkins because that machine is slow.
|
|
Jib_compile now has an option that lets it generate real value
literals (VL_real), which we don't want for backends (i.e. C), which
don't support them. Reals are encoded as actual reals in SMT, as there
isn't really any nice way to encode them as bitvectors. Currently we
just have the pure real functions, functions between integers and
reals (i.e. floor, to_real, etc) are not supported for now.
Strings are likewise encoded as SMTLIB strings, for similar reasons.
Jib_smt has ctx.use_real and ctx.use_string which are set when we
generate anything real or string related, so we can keep the logic as
Arrays+Bitvectors for most Sail that doesn't require either.
|
|
|
|
Get rid of separate V_op and V_unary constructors. jib.ott now defines
the valid operations for V_call including zero/sign extension, in such
a way that the operation ctyp can be inferred. Overall this makes the
IR less ad-hoc, and means we can share more code between SMT and C.
string_of_cval no longer used by c_backend, which now uses sgen_cval
following other sgen_ functions in the code generator, meaning
string_of_cval doesn't have to produce valid C code anymore and so can
be used for backend-agnostic debug and error messages.
|
|
|
|
Add a new AE_write_ref constructor in the ANF representation to
make writes to register references explicit in Jib_compile
|
|
Generates much better SMT that assigning each field one-by-one
starting with an undefined struct.
|
|
Had to change the hundreds and hundreds of places such values were
used. However this now lets us automatically prove cheri-concentrate
properties. Such as showing
function prop_cap_round_trip(cap: bits(128)) -> bool = {
let cap_rt = capToBits(capBitsToCapability(true, cap));
cap == cap_rt
}
is always true.
|
|
Currently only works with CVC4, test cases are in test/smt. Can prove
that RISC-V add instruction actually adds values in registers and
that's about it for now.
|
|
- Rename DeIid to Operator. It corresponds to operator <string> in the
syntax. The previous name is from when it was called deinfix in
sail1.
- Removed things that weren't actually common from
pretty_print_common.ml, e.g. printing identifiers is backend
specific. The doc_id function here was only used for a very specific
use case in pretty_print_lem, so I simplified it and renamed it to
doc_sia_id, as it is always used for a SIA.Id whatever that is.
- There is some support for anonymous records in constructors, e.g.
union Foo ('a : Type) = {
MkFoo : { field1 : 'a, field2 : int }
}
somewhat similar to the enum syntax in Rust. I'm not sure when this
was added, but there were a few odd things about it. It was
desugared in the preprocessor, rather than initial_check, and the
desugaring generated incorrect code for polymorphic anonymous
records as above.
I moved the code to initial_check, so the pre-processor now just
deals with pre-processor things and not generating types, and I
fixed the code to work with polymorphic types. This revealed some
issues in the C backend w.r.t. polymorphic structs, which is the
bulk of this commit. I also added some tests for this feature.
- OCaml backend can now generate a valid string_of function for
polymorphic structs, previously this would cause the ocaml to fail
to compile.
- Some cleanup in the Sail ott definition
- Add support for E_var in interpreter previously this would just
cause the interpreter to fail
|
|
Previously the specialization would remove any polymorphic union
constructor that was never created anywhere in the
specification. While this wasn't usually problematic, it does leave an
edge case where such a constructor could be matched upon in a pattern,
and then the resulting match would fail to compile as it would be
matching on a constructor kind that doesn't exists.
This should fix that issue by chaging the V_ctor_kind value into an
F_ctor_kind fragment. Previously a polymorphic constructor-kind would
have been represented by its mangled name, e.g.
V_ctor_kind "zSome_unit"
would now be represented as
V_ctor_kind ("Some", unifiers, ty)
where ty is a monomorphic version of the original constructor's type
such that
ctyp_unify original_ty ty = unifiers
and the mangled name we generate is
zencode_string ("Some_" ^ string_of_list "_" string_of_ctyp unifiers)
|
|
Allows us to track the last version of the return variable when the AST
in in SSA form.
|
|
Add a test case to ensure variable types in l-expressions remain the
same with flow-sensitive constraints.
|
|
Avoids duplication between l-expressions and expressions. Also means that
special variables like current_exception and have_exception are treated
normally by functions such as instr_reads and instr_writes etc. Furthermore
we can now easily annotate Jib identifiers in ways that were not previously
possible with plain sail ids.
|
|
:def <definition> evaluates a top-level definition
:(b)ind <id> : <type> creates an identifier within the interactive type-checking environment
:let <id> = <expression> defines an identifier
Using :def the following now works and brings the correct vector
operations into scope.
:def default Order dec
:load lib/prelude.sail
Also fix a type-variable shadowing bug
|
|
|
|
Add a CL_void l-expression so we don't have redundant unit-typed
variables everywhere, and add an optimization in Jib_optimize called
optimize_unit which introduces these.
Remove the basic control-flow graph in Jib_util and add a new mutable
control-flow graph type in Jib_ssa which allows the IR to be converted
into SSA form. The mutable graph allows for more efficient updates,
and includes both back and forwards references making it much more
convenient to traverse.
Having an SSA representation should make some optimizations much
simpler, and is also probably more natural for SMT generation where
variables have to be defined once using declare-const anyway.
Debug option -ddump_flow_graphs now outputs SSA'd graphs of the
functions in a specification.
|
