Summary:
Use the new LLVM bindings to handle internalization.
We would like to use global dead code elimination (gdce) to remove all the code not reachable from an entry point. However in normal compilation most functions aren't globally dead (they can be linked to). At the point of running sledge we won't be linking anymore, therefore we can internalize (make invisible outside of the compilation unit) all the symbols. In that case the whole program is dead with respect to gdce, therefore we need to preserve the entry point as external.
Previously we could only preserve `main` as the entry point (through a boolean flag). This patch uses a newer API that lets us preserve functions that satisfy a given predicate function. This enables us to have arbitrary entry points (not just main). Currently only the 3 entry points from `src/control.ml` are used, but this patch makes it easy to change.
Reviewed By: ngorogiannis
Differential Revision: D15561461
fbshipit-source-id: 88e054411
Summary:
LLVM.global_initializer casues a cast assertion failure if the value
passed to it is not a GlobalVariable. So we first check if it is a
GlobalVariable and only then ask for an initiliazer.
This is hidden if LLVM is built without assertions.
Reviewed By: jberdine
Differential Revision: D15601632
fbshipit-source-id: e9db23a12
Summary:
Sledge does not terminate on programs with recursion, because
functions get "infinitely inlined" and therefore recursion is not
treated as retreating edge.
This patch bounds the number of times the same function can "inlined"
to respect the bound (`-b` option). On each call we check the number of
occurances of the called function in the call stack. If that is higher
than the bound, we skip it.
Reviewed By: jvillard
Differential Revision: D15577134
fbshipit-source-id: 4cd3b62c6
Summary:
This diff changes the translation of global variables to translate the
initializer whenever it exists in LLVM, rather than relying on
linkage. Previously code such as
```
char *mutable_string = "hahaha";
```
would lead to LLVM code
```
nutritious_string = global i8* getelementptr inbounds ([7 x i8], [7 x i8]* @.str, i32 0, i32 0), align 8, !dbg !0 ; [#uses=2 type=i8**]
```
in which `mutable_string` had `External` linkage according to
`Llvm.linkage` even though it has an initializer. This could cause
sledge to drop the initializer.
Reviewed By: kren1
Differential Revision: D15577755
fbshipit-source-id: 50aa06c5e
Summary:
Global variables have pointer type. The size needed by the backend is
of the element type, not of the pointer itself. This diff corrects
this.
Reviewed By: kren1
Differential Revision: D15577756
fbshipit-source-id: 948ecf3cd
Summary:
The root cause is not clear, but it seems that not calling
Llvm.dispose_context avoids segfaults in the GC.
Reviewed By: kren1
Differential Revision: D15535434
fbshipit-source-id: 280e44d0b
Summary:
The name and loc tables are added-to almost exactly in sync, so
combine them to amortize the overhead.
Reviewed By: kren1
Differential Revision: D15535435
fbshipit-source-id: 801da75bb
Summary:
Format is slow. Especially Format.sprintf, which has to allocate and
initialize a buffer every time.
Reviewed By: kren1
Differential Revision: D15535437
fbshipit-source-id: ea43f44e1
Summary:
This isn't free and is expected to hold of bitcode produced by
clang/llvm. There are tests that fail verification, so keep it in
debug mode.
Reviewed By: kren1
Differential Revision: D15535438
fbshipit-source-id: 9390a8363
Summary:
If the input file has a .bc or .ll suffix, treat it as a pre-linked
bitcode file. Otherwise, treat it as before, as a file containing a
list of bitcode files to be linked. Also, perform global dead-code
elimination only when linking multiple files.
Reviewed By: kren1
Differential Revision: D15513345
fbshipit-source-id: 4c80ff9c3
Summary:
Change the sledge input format from a bitcode file to a newline
separated list of paths to LLVM bitcode files.
Reviewed By: jberdine
Differential Revision: D15470082
fbshipit-source-id: 8860f947c
Summary:
Llvm.string_of_llvalue, which just calls llvm::Value::print, is
extremely slow when called on instructions or functions. In these
cases, it initializes metadata slots for everything in the *parent
module* of the instruction or function being printed, on every
call. This is ridiculously slow, don't do it.
Reviewed By: kren1
Differential Revision: D15511376
fbshipit-source-id: 658eeccab
Summary:
Add shortcut code paths to return early in some cases guaranteed to be
the identity function.
Reviewed By: ngorogiannis
Differential Revision: D15468704
fbshipit-source-id: f137049c6
Summary:
It is pointless to track membership of atomic exps in the congruence
relation, as they cannot have any subexps which might later become
equal to something else.
Reviewed By: jvillard
Differential Revision: D15424820
fbshipit-source-id: 048dbc9e1
Summary:
For some Exp forms, Exp.solve is not complete, and this is necessary
since the result of solve is a substitution that needs to encode the
input equation as a conjunction of equations each between a variable
and an exp. This is tantamount to, stronger even than, the theory
being convex. So Exp.solve is not complete for some exps, and some of
those have constructors that perform some simplification. For example,
`(1 ≠ 0)` simplifies to `-1` (i.e. true). To enable deductions such as
`((x ≠ 0) = b) && x = 1 |- b = -1` needs the equality solver to
substitute through subexps of simplifiable exps like = and ≠, as it
does for interpreted exps like + and ×. At the same time, since
Exp.solve for non-interpreted exps cannot be complete, to enable
deductions such as `((x ≠ 0) = (y ≠ 0)) && x = 1 |- y ≠ 0` needs the
equality solver to congruence-close over subexps of simplifiable exps
such as = and ≠, as it does for uninterpreted exps.
To strengthen the equality solver in these sorts of cases, this diff
adds a new class of exps for = and ≠, and revises the equality solver
to handle them in a hybrid fashion between interpreted and
uninterpreted.
I am not currently sure whether or not this breaks the termination
proof, but I have also not managed to adapt usual divergent cases to
break this. One notable point is that simplifying = and ≠ exps always
produces genuinely simpler and smaller exps, in contrast to
e.g. polynomial simplification and gaussian elimination.
Note that the real solution to this problem is likely to be to
eliminate the i1 type in favor or a genuine boolean type, and
translate all integer operations on i1 to boolean/logical ones. Then
the disjunction implicit in e.g. equations between disequations would
appear as actual disjunction, and could be dealt with as such.
Reviewed By: jvillard
Differential Revision: D15424823
fbshipit-source-id: 67d62df1f
Summary:
* Adds compilation of cxxabi.bc
* Includes the cxxabi.bc into the sledge executable via opam-crunch
* Links cxxabi.bc in sledge frontend
Reviewed By: jberdine
Differential Revision: D15415190
fbshipit-source-id: cc42f09fb
Summary:
There are two motivations for this:
1. Distinguish between `Unreachable`, which silently terminates
execution a la `assume false`; and `Abort`, which vocally
terminates execution a la `assert false`.
2. Distinguish between undefined functions, which have `Unreachable`
bodies, and bomb functions such as:
```
define void bomb() {
tail call void llvm.trap()
unreachable
}
```
Reviewed By: ngorogiannis
Differential Revision: D15408246
fbshipit-source-id: b64354cdb
Summary:
Require a final `()` argument to explicitly indicate the end of the
arguments to the printf-like functions. For `warn` this is not
any safer because the return type is `unit` anyhow, but for `fail` the
return type is polymorphic so the final `()` prevents unintentionally
forgetting an argument.
Reviewed By: ngorogiannis
Differential Revision: D15403367
fbshipit-source-id: ce3fe4035
Summary:
llvm.trap is noreturn nounwind so calls to it are always succeeded by
Unreachable, therefore unless an alarm is desired for reaching it,
translating it as nop suffices.
Reviewed By: ngorogiannis
Differential Revision: D15328302
fbshipit-source-id: 54efe6c21
Summary:
A previous rebase fumble sometimes led to the wrong arguments being
passed to retpolines of Invoke instructions.
Reviewed By: ngorogiannis
Differential Revision: D15323950
fbshipit-source-id: f6eb6fbe2
Summary:
It is (now?) the case that `[%sexp_of: type_name]` generates just
`sexp_of_type_name`, rather than expanding `type_name` to its
definition and generating a conversion function for that. Hence, when
such cases appear within `let rec sexp_of_type_name`, they get
captured, sometimes leading to divergence.
This diff fixes this by manually expanding such types into their
definitions.
Reviewed By: ngorogiannis
Differential Revision: D15314736
fbshipit-source-id: 716fff7cc
Summary: There was a missing case for singleton monomials of degree > 1.
Reviewed By: mbouaziz
Differential Revision: D15098810
fbshipit-source-id: e6d17d899
Summary:
Boolean Xor expressions are treated directly by the solver, so the 'no
new subexps' invariant for congruence closure is not needed.
Reviewed By: mbouaziz
Differential Revision: D15098817
fbshipit-source-id: d118c5881
Summary: Shifting by too many bits is undefined, so do not normalize them.
Reviewed By: mbouaziz
Differential Revision: D15098812
fbshipit-source-id: 96f4606d7
Summary:
When constructing division expressions from rational coefficients of
polynomials, the constant numerator and denominator should be clamped
to the number of bits in the result type.
Reviewed By: mbouaziz
Differential Revision: D15098825
fbshipit-source-id: fa448c39a
Summary:
Some constants are used during Exp normalization, to encode
subtraction using addition, units of addition and multiplication,
etc. Previously these were unconditionally integers, but need to be
floats for float arithmetic.
Reviewed By: mbouaziz
Differential Revision: D15098815
fbshipit-source-id: 13d1be142
Josh Berdine
Timotej Kapus