You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

1512 lines
57 KiB

(*
* Copyright (c) Facebook, Inc. and its affiliates.
*
* This source code is licensed under the MIT license found in the
* LICENSE file in the root directory of this source tree.
*)
(** Translate LLVM to LLAIR *)
let pp_lltype fs t = Format.pp_print_string fs (Llvm.string_of_lltype t)
(* WARNING: SLOW on instructions and functions *)
let pp_llvalue fs t = Format.pp_print_string fs (Llvm.string_of_llvalue t)
let pp_llblock fs t =
Format.pp_print_string fs (Llvm.string_of_llvalue (Llvm.value_of_block t))
;;
Reg.demangle :=
let open Ctypes in
let cxa_demangle =
(* char *__cxa_demangle(const char *, char *, size_t *, int * ) *)
Foreign.foreign "__cxa_demangle"
( string
@-> ptr char
@-> ptr size_t
@-> ptr int
@-> returning string_opt )
in
let null_ptr_char = from_voidp char null in
let null_ptr_size_t = from_voidp size_t null in
let status = allocate int 0 in
fun mangled ->
let demangled =
cxa_demangle mangled null_ptr_char null_ptr_size_t status
in
if !@status = 0 then demangled else None
exception Invalid_llvm of string
let invalid_llvm : string -> 'a =
fun msg ->
let first_line =
Option.value_map ~default:msg ~f:(String.prefix msg)
(String.index msg '\n')
in
Format.printf "@\n%s@\n" msg ;
raise (Invalid_llvm first_line)
(* gather names and debug locations *)
let sym_tbl : (Llvm.llvalue, string * Loc.t) Hashtbl.t =
Hashtbl.Poly.create ~size:4_194_304 ()
let scope_tbl :
( [`Fun of Llvm.llvalue | `Mod of Llvm.llmodule]
, int ref * (string, int) Hashtbl.t )
Hashtbl.t =
Hashtbl.Poly.create ~size:32_768 ()
open struct
open struct
let loc_of_global g =
Loc.mk
?dir:(Llvm.get_debug_loc_directory g)
?file:(Llvm.get_debug_loc_filename g)
~line:(Llvm.get_debug_loc_line g)
?col:None
let loc_of_function f =
Loc.mk
?dir:(Llvm.get_debug_loc_directory f)
?file:(Llvm.get_debug_loc_filename f)
~line:(Llvm.get_debug_loc_line f)
?col:None
let loc_of_instr i =
Loc.mk
?dir:(Llvm.get_debug_loc_directory i)
?file:(Llvm.get_debug_loc_filename i)
~line:(Llvm.get_debug_loc_line i)
~col:(Llvm.get_debug_loc_column i)
let add_sym llv loc =
let maybe_scope =
match Llvm.classify_value llv with
| Argument -> Some (`Fun (Llvm.param_parent llv))
| BasicBlock ->
Some (`Fun (Llvm.block_parent (Llvm.block_of_value llv)))
| Instruction _ ->
Some (`Fun (Llvm.block_parent (Llvm.instr_parent llv)))
| GlobalVariable | Function -> Some (`Mod (Llvm.global_parent llv))
| UndefValue -> None
| ConstantExpr -> None
| ConstantPointerNull -> None
| _ ->
warn "Unexpected type of llv, might crash: %a" pp_llvalue llv () ;
Some (`Mod (Llvm.global_parent llv))
in
match maybe_scope with
| None -> ()
| Some scope ->
let next, void_tbl =
Hashtbl.find_or_add scope_tbl scope ~default:(fun () ->
(ref 0, Hashtbl.Poly.create ()) )
in
let name =
match Llvm.classify_type (Llvm.type_of llv) with
| Void -> (
let fname =
match Llvm.classify_value llv with
| Instruction (Call | Invoke) -> (
match
Llvm.value_name
(Llvm.operand llv (Llvm.num_operands llv - 1))
with
| "" -> Int.to_string (!next - 1)
| s -> s )
| _ -> "void"
in
match Hashtbl.find void_tbl fname with
| None ->
Hashtbl.set void_tbl ~key:fname ~data:1 ;
fname ^ ".void"
| Some count ->
Hashtbl.set void_tbl ~key:fname ~data:(count + 1) ;
String.concat_array
[|fname; ".void."; Int.to_string count|] )
| _ -> (
match Llvm.value_name llv with
| "" ->
(* anonymous values take the next SSA name *)
let name = !next in
next := name + 1 ;
Int.to_string name
| name -> (
match Int.of_string name with
| _ ->
(* escape to avoid clash with names of anonymous values *)
String.concat_array [|"\""; name; "\""|]
| exception _ -> name ) )
in
Hashtbl.set sym_tbl ~key:llv ~data:(name, loc)
end
let scan_names_and_locs : Llvm.llmodule -> unit =
fun m ->
let scan_global g = add_sym g (loc_of_global g) in
let scan_instr i =
let loc = loc_of_instr i in
add_sym i loc ;
match Llvm.instr_opcode i with
| Call -> (
match Llvm.(value_name (operand i (num_arg_operands i))) with
| "llvm.dbg.declare" ->
let md = Llvm.(get_mdnode_operands (operand i 0)) in
if not (Array.is_empty md) then add_sym md.(0) loc
else
warn
"could not find variable for debug info at %a with \
metadata %a"
Loc.pp loc (List.pp ", " pp_llvalue) (Array.to_list md) ()
| _ -> () )
| _ -> ()
in
let scan_block b =
add_sym (Llvm.value_of_block b) Loc.none ;
Llvm.iter_instrs scan_instr b
in
let scan_function f =
Llvm.iter_params (fun prm -> add_sym prm Loc.none) f ;
add_sym f (loc_of_function f) ;
Llvm.iter_blocks scan_block f
in
Llvm.iter_globals scan_global m ;
Llvm.iter_functions scan_function m
let find_name : Llvm.llvalue -> string =
fun v -> fst (Hashtbl.find_exn sym_tbl v)
let find_loc : Llvm.llvalue -> Loc.t =
fun v -> snd (Hashtbl.find_exn sym_tbl v)
end
let label_of_block : Llvm.llbasicblock -> string =
fun blk -> find_name (Llvm.value_of_block blk)
let anon_struct_name : (Llvm.lltype, string) Hashtbl.t =
Hashtbl.Poly.create ()
let struct_name : Llvm.lltype -> string =
fun llt ->
match Llvm.struct_name llt with
| Some name -> name
| None ->
Hashtbl.find_or_add anon_struct_name llt ~default:(fun () ->
Int.to_string (Hashtbl.length anon_struct_name) )
type x = {llcontext: Llvm.llcontext; lldatalayout: Llvm_target.DataLayout.t}
let ptr_siz : x -> int =
fun x -> Llvm_target.DataLayout.pointer_size x.lldatalayout
let size_of, bit_size_of =
let size_to_int size_of x llt =
if Llvm.type_is_sized llt then
match Int64.to_int (size_of llt x.lldatalayout) with
| Some n -> n
| None -> fail "type size too large: %a" pp_lltype llt ()
else fail "types with undetermined size: %a" pp_lltype llt ()
in
( size_to_int Llvm_target.DataLayout.abi_size
, size_to_int Llvm_target.DataLayout.size_in_bits )
let memo_type : (Llvm.lltype, Typ.t) Hashtbl.t = Hashtbl.Poly.create ()
let rec xlate_type : x -> Llvm.lltype -> Typ.t =
fun x llt ->
let xlate_type_ llt =
if Llvm.type_is_sized llt then
let byts = size_of x llt in
let bits = bit_size_of x llt in
match Llvm.classify_type llt with
| Half | Float | Double | Fp128 -> Typ.float ~bits ~byts ~enc:`IEEE
| X86fp80 -> Typ.float ~bits ~byts ~enc:`Extended
| Ppc_fp128 -> Typ.float ~bits ~byts ~enc:`Pair
| Integer -> Typ.integer ~bits ~byts
| X86_mmx -> Typ.integer ~bits ~byts
| Pointer ->
if byts <> ptr_siz x then
todo "non-integral pointer types: %a" pp_lltype llt () ;
let elt = xlate_type x (Llvm.element_type llt) in
Typ.pointer ~elt
| Vector ->
let elt = xlate_type x (Llvm.element_type llt) in
let len = Llvm.vector_size llt in
Typ.array ~elt ~len ~bits ~byts
| Array ->
let elt = xlate_type x (Llvm.element_type llt) in
let len = Llvm.array_length llt in
Typ.array ~elt ~len ~bits ~byts
| Struct ->
let llelts = Llvm.struct_element_types llt in
let len = Array.length llelts in
let packed = Llvm.is_packed llt in
if Llvm.is_literal llt then
let elts =
IArray.map ~f:(xlate_type x) (IArray.of_array llelts)
in
Typ.tuple elts ~bits ~byts ~packed
else
let name = struct_name llt in
let elts =
IArray.init len ~f:(fun i -> lazy (xlate_type x llelts.(i)))
in
Typ.struct_ ~name elts ~bits ~byts ~packed
| Function -> fail "expected to be unsized: %a" pp_lltype llt ()
| Void | Label | Metadata | Token -> assert false
else
match Llvm.classify_type llt with
| Function ->
let return = xlate_type_opt x (Llvm.return_type llt) in
let llargs = Llvm.param_types llt in
let len = Array.length llargs in
let args =
IArray.init len ~f:(fun i -> xlate_type x llargs.(i))
in
Typ.function_ ~return ~args
| Struct when Llvm.is_opaque llt -> Typ.opaque ~name:(struct_name llt)
| Token -> Typ.opaque ~name:"token"
| Vector | Array | Struct ->
todo "unsized non-opaque aggregate types: %a" pp_lltype llt ()
| Half | Float | Double | X86fp80 | Fp128 | Ppc_fp128 | Integer
|X86_mmx | Pointer ->
fail "expected to be sized: %a" pp_lltype llt ()
| Void | Label | Metadata -> assert false
in
Hashtbl.find_or_add memo_type llt ~default:(fun () ->
[%Trace.call fun {pf} -> pf "%a" pp_lltype llt]
;
xlate_type_ llt
|>
[%Trace.retn fun {pf} -> pf "%a" Typ.pp_defn] )
and xlate_type_opt : x -> Llvm.lltype -> Typ.t option =
fun x llt ->
match Llvm.classify_type llt with
| Void -> None
| _ -> Some (xlate_type x llt)
let i32 x = xlate_type x (Llvm.i32_type x.llcontext)
let suffix_after_last_space : string -> string =
fun str -> String.drop_prefix str (String.rindex_exn str ' ' + 1)
let xlate_int : x -> Llvm.llvalue -> Exp.t =
fun x llv ->
let llt = Llvm.type_of llv in
let typ = xlate_type x llt in
let data =
match Llvm.int64_of_const llv with
| Some n -> Z.of_int64 n
| None ->
Z.of_string (suffix_after_last_space (Llvm.string_of_llvalue llv))
in
Exp.integer typ data
let xlate_float : x -> Llvm.llvalue -> Exp.t =
fun x llv ->
let llt = Llvm.type_of llv in
let typ = xlate_type x llt in
let data = suffix_after_last_space (Llvm.string_of_llvalue llv) in
Exp.float typ data
let xlate_name x ?global : Llvm.llvalue -> Reg.t =
fun llv ->
let typ = xlate_type x (Llvm.type_of llv) in
Reg.program ?global typ (find_name llv)
let xlate_name_opt : x -> Llvm.llvalue -> Reg.t option =
fun x instr ->
let llt = Llvm.type_of instr in
match Llvm.classify_type llt with
| Void -> None
| _ -> Some (xlate_name x instr)
let memo_value : (bool * Llvm.llvalue, Exp.t) Hashtbl.t =
Hashtbl.Poly.create ()
let memo_global : (Llvm.llvalue, Global.t) Hashtbl.t =
Hashtbl.Poly.create ()
let should_inline : Llvm.llvalue -> bool =
fun llv ->
match Llvm.use_begin llv with
| Some use -> (
match Llvm.use_succ use with
| Some _ -> (
match Llvm.classify_value llv with
| Instruction
( Trunc | ZExt | SExt | FPToUI | FPToSI | UIToFP | SIToFP
| FPTrunc | FPExt | PtrToInt | IntToPtr | BitCast | AddrSpaceCast
) ->
true (* inline casts *)
| _ -> false (* do not inline if >= 2 uses *) )
| None -> true )
| None -> true
let ptr_fld x ~ptr ~fld ~lltyp =
let offset =
Llvm_target.DataLayout.offset_of_element lltyp fld x.lldatalayout
in
Exp.add ~typ:Typ.ptr ptr (Exp.integer Typ.siz (Z.of_int64 offset))
let ptr_idx x ~ptr ~idx ~llelt =
let stride = Llvm_target.DataLayout.abi_size llelt x.lldatalayout in
Exp.add ~typ:Typ.ptr ptr
(Exp.mul ~typ:Typ.siz (Exp.integer Typ.siz (Z.of_int64 stride)) idx)
[sledge] Simplify type conversions Summary: The treatment of type conversions is too complicated, non-uniform, etc. This diff attempts to simplify things by separating integer to integer conversions, which are interpreted, from others, which are essentially just uninterpreted functions. Integer conversions are now handled using two expression and term forms: Signed and Unsigned. These each interpret their argument as either a signed or unsigned number of a given bitwidth: ``` | Signed of {bits: int} (** [Ap1 (Signed {bits= n}, dst, arg)] is [arg] interpreted as an [n]-bit signed integer and injected into the [dst] type. That is, it two's-complement--decodes the low [n] bits of the infinite two's-complement encoding of [arg]. The injection into [dst] is a no-op, so [dst] must be an integer type with bitwidth at least [n]. *) | Unsigned of {bits: int} (** [Ap1 (Unsigned {bits= n}, dst, arg)] is [arg] interpreted as an [n]-bit unsigned integer and injected into the [dst] type. That is, it unsigned-binary--decodes the low [n] bits of the infinite two's-complement encoding of [arg]. The injection into [dst] is a no-op, so [dst] must be an integer type with bitwidth greater than [n]. *) | Convert of {src: Typ.t} (** [Ap1 (Convert {src}, dst, arg)] is [arg] converted from type [src] to type [dst], possibly with loss of information. The [src] and [dst] types must be [Typ.convertible] and must not both be [Integer] types. *) ``` Reviewed By: ngorogiannis Differential Revision: D18298140 fbshipit-source-id: 690f065b4
5 years ago
let convert_to_siz =
let siz_bits = Typ.bit_size_of Typ.siz in
fun typ arg ->
match (typ : Typ.t) with
| Integer {bits} ->
if siz_bits < bits then Exp.signed siz_bits arg ~to_:Typ.siz
else if siz_bits > bits then Exp.signed bits arg ~to_:Typ.siz
else arg
| _ -> fail "convert_to_siz: %a" Typ.pp typ ()
let xlate_llvm_eh_typeid_for : x -> Typ.t -> Exp.t -> Exp.t =
[sledge] Simplify type conversions Summary: The treatment of type conversions is too complicated, non-uniform, etc. This diff attempts to simplify things by separating integer to integer conversions, which are interpreted, from others, which are essentially just uninterpreted functions. Integer conversions are now handled using two expression and term forms: Signed and Unsigned. These each interpret their argument as either a signed or unsigned number of a given bitwidth: ``` | Signed of {bits: int} (** [Ap1 (Signed {bits= n}, dst, arg)] is [arg] interpreted as an [n]-bit signed integer and injected into the [dst] type. That is, it two's-complement--decodes the low [n] bits of the infinite two's-complement encoding of [arg]. The injection into [dst] is a no-op, so [dst] must be an integer type with bitwidth at least [n]. *) | Unsigned of {bits: int} (** [Ap1 (Unsigned {bits= n}, dst, arg)] is [arg] interpreted as an [n]-bit unsigned integer and injected into the [dst] type. That is, it unsigned-binary--decodes the low [n] bits of the infinite two's-complement encoding of [arg]. The injection into [dst] is a no-op, so [dst] must be an integer type with bitwidth greater than [n]. *) | Convert of {src: Typ.t} (** [Ap1 (Convert {src}, dst, arg)] is [arg] converted from type [src] to type [dst], possibly with loss of information. The [src] and [dst] types must be [Typ.convertible] and must not both be [Integer] types. *) ``` Reviewed By: ngorogiannis Differential Revision: D18298140 fbshipit-source-id: 690f065b4
5 years ago
fun x typ arg -> Exp.convert typ ~to_:(i32 x) arg
[sledge] Represent recursive records non-recursively Summary: In LLVM it is possible for struct constant values to be directly recursive, with no pointer dereference to close the cycle. These appear for example as the values of vtables from C++ code. Currently such recursive records in the Exp and Term languages are represented as genuinely cyclic values. Compared to a standard term representation, the presence of cyclic values is a significant complication everywhere. Since the backend solver does not do anything such as induction over these, they have to be treated as essentially atomic. This patch changes the representation to a standard non-recursive tree term structure. Instead of cyclic references, an explicit constructor is added for the "non-tree edges", which simply indicates which ancestor record value to which the recursive reference points. There is a potential issue with this representation, since for mutually recursive records, the representation is not canonical: it chooses one of the records in the cycle to start from and expresses the cycles relative to that. Currently the choice of representation is dictated by the frontend. For the case of vtables, the frontend translates globals in the same order they appear in the LLVM IR, so the representation choice is fixed. It may turn out that other potential uses require more reasoning support in the backend solver, which would involve a theory of equality of record values induced by equating the representations resulting from different rotations of the cycle of records. Reviewed By: jvillard Differential Revision: D21441533 fbshipit-source-id: 0c5a11378
5 years ago
let rec xlate_intrinsic_exp stk :
string -> (x -> Llvm.llvalue -> Exp.t) option =
fun name ->
match name with
| "llvm.eh.typeid.for" ->
Some
(fun x llv ->
let rand = Llvm.operand llv 0 in
[sledge] Represent recursive records non-recursively Summary: In LLVM it is possible for struct constant values to be directly recursive, with no pointer dereference to close the cycle. These appear for example as the values of vtables from C++ code. Currently such recursive records in the Exp and Term languages are represented as genuinely cyclic values. Compared to a standard term representation, the presence of cyclic values is a significant complication everywhere. Since the backend solver does not do anything such as induction over these, they have to be treated as essentially atomic. This patch changes the representation to a standard non-recursive tree term structure. Instead of cyclic references, an explicit constructor is added for the "non-tree edges", which simply indicates which ancestor record value to which the recursive reference points. There is a potential issue with this representation, since for mutually recursive records, the representation is not canonical: it chooses one of the records in the cycle to start from and expresses the cycles relative to that. Currently the choice of representation is dictated by the frontend. For the case of vtables, the frontend translates globals in the same order they appear in the LLVM IR, so the representation choice is fixed. It may turn out that other potential uses require more reasoning support in the backend solver, which would involve a theory of equality of record values induced by equating the representations resulting from different rotations of the cycle of records. Reviewed By: jvillard Differential Revision: D21441533 fbshipit-source-id: 0c5a11378
5 years ago
let arg = xlate_value stk x rand in
let src = xlate_type x (Llvm.type_of rand) in
xlate_llvm_eh_typeid_for x src arg )
| _ -> None
[sledge] Represent recursive records non-recursively Summary: In LLVM it is possible for struct constant values to be directly recursive, with no pointer dereference to close the cycle. These appear for example as the values of vtables from C++ code. Currently such recursive records in the Exp and Term languages are represented as genuinely cyclic values. Compared to a standard term representation, the presence of cyclic values is a significant complication everywhere. Since the backend solver does not do anything such as induction over these, they have to be treated as essentially atomic. This patch changes the representation to a standard non-recursive tree term structure. Instead of cyclic references, an explicit constructor is added for the "non-tree edges", which simply indicates which ancestor record value to which the recursive reference points. There is a potential issue with this representation, since for mutually recursive records, the representation is not canonical: it chooses one of the records in the cycle to start from and expresses the cycles relative to that. Currently the choice of representation is dictated by the frontend. For the case of vtables, the frontend translates globals in the same order they appear in the LLVM IR, so the representation choice is fixed. It may turn out that other potential uses require more reasoning support in the backend solver, which would involve a theory of equality of record values induced by equating the representations resulting from different rotations of the cycle of records. Reviewed By: jvillard Differential Revision: D21441533 fbshipit-source-id: 0c5a11378
5 years ago
and xlate_value ?(inline = false) stk : x -> Llvm.llvalue -> Exp.t =
fun x llv ->
let xlate_value_ llv =
match Llvm.classify_value llv with
| Instruction Call -> (
let func = Llvm.operand llv (Llvm.num_arg_operands llv) in
let fname = Llvm.value_name func in
[sledge] Represent recursive records non-recursively Summary: In LLVM it is possible for struct constant values to be directly recursive, with no pointer dereference to close the cycle. These appear for example as the values of vtables from C++ code. Currently such recursive records in the Exp and Term languages are represented as genuinely cyclic values. Compared to a standard term representation, the presence of cyclic values is a significant complication everywhere. Since the backend solver does not do anything such as induction over these, they have to be treated as essentially atomic. This patch changes the representation to a standard non-recursive tree term structure. Instead of cyclic references, an explicit constructor is added for the "non-tree edges", which simply indicates which ancestor record value to which the recursive reference points. There is a potential issue with this representation, since for mutually recursive records, the representation is not canonical: it chooses one of the records in the cycle to start from and expresses the cycles relative to that. Currently the choice of representation is dictated by the frontend. For the case of vtables, the frontend translates globals in the same order they appear in the LLVM IR, so the representation choice is fixed. It may turn out that other potential uses require more reasoning support in the backend solver, which would involve a theory of equality of record values induced by equating the representations resulting from different rotations of the cycle of records. Reviewed By: jvillard Differential Revision: D21441533 fbshipit-source-id: 0c5a11378
5 years ago
match xlate_intrinsic_exp stk fname with
| Some intrinsic when inline || should_inline llv -> intrinsic x llv
| _ -> Exp.reg (xlate_name x llv) )
| Instruction (Invoke | Alloca | Load | PHI | LandingPad | VAArg)
|Argument ->
Exp.reg (xlate_name x llv)
[sledge] Represent recursive records non-recursively Summary: In LLVM it is possible for struct constant values to be directly recursive, with no pointer dereference to close the cycle. These appear for example as the values of vtables from C++ code. Currently such recursive records in the Exp and Term languages are represented as genuinely cyclic values. Compared to a standard term representation, the presence of cyclic values is a significant complication everywhere. Since the backend solver does not do anything such as induction over these, they have to be treated as essentially atomic. This patch changes the representation to a standard non-recursive tree term structure. Instead of cyclic references, an explicit constructor is added for the "non-tree edges", which simply indicates which ancestor record value to which the recursive reference points. There is a potential issue with this representation, since for mutually recursive records, the representation is not canonical: it chooses one of the records in the cycle to start from and expresses the cycles relative to that. Currently the choice of representation is dictated by the frontend. For the case of vtables, the frontend translates globals in the same order they appear in the LLVM IR, so the representation choice is fixed. It may turn out that other potential uses require more reasoning support in the backend solver, which would involve a theory of equality of record values induced by equating the representations resulting from different rotations of the cycle of records. Reviewed By: jvillard Differential Revision: D21441533 fbshipit-source-id: 0c5a11378
5 years ago
| Function | GlobalVariable -> Exp.reg (xlate_global stk x llv).reg
| GlobalAlias -> xlate_value stk x (Llvm.operand llv 0)
| ConstantInt -> xlate_int x llv
| ConstantFP -> xlate_float x llv
| ConstantPointerNull -> Exp.null
| ConstantAggregateZero -> (
let typ = xlate_type x (Llvm.type_of llv) in
match typ with
| Integer _ -> Exp.integer typ Z.zero
| Pointer _ -> Exp.null
| Array _ | Tuple _ | Struct _ ->
Exp.splat typ (Exp.integer Typ.byt Z.zero)
| _ -> fail "ConstantAggregateZero of type %a" Typ.pp typ () )
| ConstantVector | ConstantArray ->
let typ = xlate_type x (Llvm.type_of llv) in
let len = Llvm.num_operands llv in
[sledge] Represent recursive records non-recursively Summary: In LLVM it is possible for struct constant values to be directly recursive, with no pointer dereference to close the cycle. These appear for example as the values of vtables from C++ code. Currently such recursive records in the Exp and Term languages are represented as genuinely cyclic values. Compared to a standard term representation, the presence of cyclic values is a significant complication everywhere. Since the backend solver does not do anything such as induction over these, they have to be treated as essentially atomic. This patch changes the representation to a standard non-recursive tree term structure. Instead of cyclic references, an explicit constructor is added for the "non-tree edges", which simply indicates which ancestor record value to which the recursive reference points. There is a potential issue with this representation, since for mutually recursive records, the representation is not canonical: it chooses one of the records in the cycle to start from and expresses the cycles relative to that. Currently the choice of representation is dictated by the frontend. For the case of vtables, the frontend translates globals in the same order they appear in the LLVM IR, so the representation choice is fixed. It may turn out that other potential uses require more reasoning support in the backend solver, which would involve a theory of equality of record values induced by equating the representations resulting from different rotations of the cycle of records. Reviewed By: jvillard Differential Revision: D21441533 fbshipit-source-id: 0c5a11378
5 years ago
let f i = xlate_value stk x (Llvm.operand llv i) in
Exp.record typ (IArray.init len ~f)
| ConstantDataVector ->
let typ = xlate_type x (Llvm.type_of llv) in
let len = Llvm.vector_size (Llvm.type_of llv) in
[sledge] Represent recursive records non-recursively Summary: In LLVM it is possible for struct constant values to be directly recursive, with no pointer dereference to close the cycle. These appear for example as the values of vtables from C++ code. Currently such recursive records in the Exp and Term languages are represented as genuinely cyclic values. Compared to a standard term representation, the presence of cyclic values is a significant complication everywhere. Since the backend solver does not do anything such as induction over these, they have to be treated as essentially atomic. This patch changes the representation to a standard non-recursive tree term structure. Instead of cyclic references, an explicit constructor is added for the "non-tree edges", which simply indicates which ancestor record value to which the recursive reference points. There is a potential issue with this representation, since for mutually recursive records, the representation is not canonical: it chooses one of the records in the cycle to start from and expresses the cycles relative to that. Currently the choice of representation is dictated by the frontend. For the case of vtables, the frontend translates globals in the same order they appear in the LLVM IR, so the representation choice is fixed. It may turn out that other potential uses require more reasoning support in the backend solver, which would involve a theory of equality of record values induced by equating the representations resulting from different rotations of the cycle of records. Reviewed By: jvillard Differential Revision: D21441533 fbshipit-source-id: 0c5a11378
5 years ago
let f i = xlate_value stk x (Llvm.const_element llv i) in
Exp.record typ (IArray.init len ~f)
| ConstantDataArray ->
let typ = xlate_type x (Llvm.type_of llv) in
let len = Llvm.array_length (Llvm.type_of llv) in
[sledge] Represent recursive records non-recursively Summary: In LLVM it is possible for struct constant values to be directly recursive, with no pointer dereference to close the cycle. These appear for example as the values of vtables from C++ code. Currently such recursive records in the Exp and Term languages are represented as genuinely cyclic values. Compared to a standard term representation, the presence of cyclic values is a significant complication everywhere. Since the backend solver does not do anything such as induction over these, they have to be treated as essentially atomic. This patch changes the representation to a standard non-recursive tree term structure. Instead of cyclic references, an explicit constructor is added for the "non-tree edges", which simply indicates which ancestor record value to which the recursive reference points. There is a potential issue with this representation, since for mutually recursive records, the representation is not canonical: it chooses one of the records in the cycle to start from and expresses the cycles relative to that. Currently the choice of representation is dictated by the frontend. For the case of vtables, the frontend translates globals in the same order they appear in the LLVM IR, so the representation choice is fixed. It may turn out that other potential uses require more reasoning support in the backend solver, which would involve a theory of equality of record values induced by equating the representations resulting from different rotations of the cycle of records. Reviewed By: jvillard Differential Revision: D21441533 fbshipit-source-id: 0c5a11378
5 years ago
let f i = xlate_value stk x (Llvm.const_element llv i) in
Exp.record typ (IArray.init len ~f)
[sledge] Represent recursive records non-recursively Summary: In LLVM it is possible for struct constant values to be directly recursive, with no pointer dereference to close the cycle. These appear for example as the values of vtables from C++ code. Currently such recursive records in the Exp and Term languages are represented as genuinely cyclic values. Compared to a standard term representation, the presence of cyclic values is a significant complication everywhere. Since the backend solver does not do anything such as induction over these, they have to be treated as essentially atomic. This patch changes the representation to a standard non-recursive tree term structure. Instead of cyclic references, an explicit constructor is added for the "non-tree edges", which simply indicates which ancestor record value to which the recursive reference points. There is a potential issue with this representation, since for mutually recursive records, the representation is not canonical: it chooses one of the records in the cycle to start from and expresses the cycles relative to that. Currently the choice of representation is dictated by the frontend. For the case of vtables, the frontend translates globals in the same order they appear in the LLVM IR, so the representation choice is fixed. It may turn out that other potential uses require more reasoning support in the backend solver, which would involve a theory of equality of record values induced by equating the representations resulting from different rotations of the cycle of records. Reviewed By: jvillard Differential Revision: D21441533 fbshipit-source-id: 0c5a11378
5 years ago
| ConstantStruct -> (
let typ = xlate_type x (Llvm.type_of llv) in
[sledge] Represent recursive records non-recursively Summary: In LLVM it is possible for struct constant values to be directly recursive, with no pointer dereference to close the cycle. These appear for example as the values of vtables from C++ code. Currently such recursive records in the Exp and Term languages are represented as genuinely cyclic values. Compared to a standard term representation, the presence of cyclic values is a significant complication everywhere. Since the backend solver does not do anything such as induction over these, they have to be treated as essentially atomic. This patch changes the representation to a standard non-recursive tree term structure. Instead of cyclic references, an explicit constructor is added for the "non-tree edges", which simply indicates which ancestor record value to which the recursive reference points. There is a potential issue with this representation, since for mutually recursive records, the representation is not canonical: it chooses one of the records in the cycle to start from and expresses the cycles relative to that. Currently the choice of representation is dictated by the frontend. For the case of vtables, the frontend translates globals in the same order they appear in the LLVM IR, so the representation choice is fixed. It may turn out that other potential uses require more reasoning support in the backend solver, which would involve a theory of equality of record values induced by equating the representations resulting from different rotations of the cycle of records. Reviewed By: jvillard Differential Revision: D21441533 fbshipit-source-id: 0c5a11378
5 years ago
match List.findi llv stk with
| Some i -> Exp.rec_record i typ
| None ->
let stk = llv :: stk in
Exp.record typ
(IArray.init (Llvm.num_operands llv) ~f:(fun i ->
xlate_value stk x (Llvm.operand llv i) )) )
| BlockAddress ->
let parent = find_name (Llvm.operand llv 0) in
let name = find_name (Llvm.operand llv 1) in
Exp.label ~parent ~name
| UndefValue ->
let typ = xlate_type x (Llvm.type_of llv) in
Exp.nondet typ (Llvm.string_of_llvalue llv)
| Instruction
( ( Trunc | ZExt | SExt | FPToUI | FPToSI | UIToFP | SIToFP
| FPTrunc | FPExt | PtrToInt | IntToPtr | BitCast | AddrSpaceCast
| Add | FAdd | Sub | FSub | Mul | FMul | UDiv | SDiv | FDiv | URem
| SRem | FRem | Shl | LShr | AShr | And | Or | Xor | ICmp | FCmp
| Select | GetElementPtr | ExtractElement | InsertElement
| ShuffleVector | ExtractValue | InsertValue ) as opcode ) ->
[sledge] Represent recursive records non-recursively Summary: In LLVM it is possible for struct constant values to be directly recursive, with no pointer dereference to close the cycle. These appear for example as the values of vtables from C++ code. Currently such recursive records in the Exp and Term languages are represented as genuinely cyclic values. Compared to a standard term representation, the presence of cyclic values is a significant complication everywhere. Since the backend solver does not do anything such as induction over these, they have to be treated as essentially atomic. This patch changes the representation to a standard non-recursive tree term structure. Instead of cyclic references, an explicit constructor is added for the "non-tree edges", which simply indicates which ancestor record value to which the recursive reference points. There is a potential issue with this representation, since for mutually recursive records, the representation is not canonical: it chooses one of the records in the cycle to start from and expresses the cycles relative to that. Currently the choice of representation is dictated by the frontend. For the case of vtables, the frontend translates globals in the same order they appear in the LLVM IR, so the representation choice is fixed. It may turn out that other potential uses require more reasoning support in the backend solver, which would involve a theory of equality of record values induced by equating the representations resulting from different rotations of the cycle of records. Reviewed By: jvillard Differential Revision: D21441533 fbshipit-source-id: 0c5a11378
5 years ago
if inline || should_inline llv then xlate_opcode stk x llv opcode
else Exp.reg (xlate_name x llv)
[sledge] Represent recursive records non-recursively Summary: In LLVM it is possible for struct constant values to be directly recursive, with no pointer dereference to close the cycle. These appear for example as the values of vtables from C++ code. Currently such recursive records in the Exp and Term languages are represented as genuinely cyclic values. Compared to a standard term representation, the presence of cyclic values is a significant complication everywhere. Since the backend solver does not do anything such as induction over these, they have to be treated as essentially atomic. This patch changes the representation to a standard non-recursive tree term structure. Instead of cyclic references, an explicit constructor is added for the "non-tree edges", which simply indicates which ancestor record value to which the recursive reference points. There is a potential issue with this representation, since for mutually recursive records, the representation is not canonical: it chooses one of the records in the cycle to start from and expresses the cycles relative to that. Currently the choice of representation is dictated by the frontend. For the case of vtables, the frontend translates globals in the same order they appear in the LLVM IR, so the representation choice is fixed. It may turn out that other potential uses require more reasoning support in the backend solver, which would involve a theory of equality of record values induced by equating the representations resulting from different rotations of the cycle of records. Reviewed By: jvillard Differential Revision: D21441533 fbshipit-source-id: 0c5a11378
5 years ago
| ConstantExpr -> xlate_opcode stk x llv (Llvm.constexpr_opcode llv)
| GlobalIFunc -> todo "ifuncs: %a" pp_llvalue llv ()
| Instruction (CatchPad | CleanupPad | CatchSwitch) ->
todo "windows exception handling: %a" pp_llvalue llv ()
| Instruction
( Invalid | Ret | Br | Switch | IndirectBr | Invalid2 | Unreachable
| Store | UserOp1 | UserOp2 | Fence | AtomicCmpXchg | AtomicRMW
| Resume | CleanupRet | CatchRet )
|NullValue | BasicBlock | InlineAsm | MDNode | MDString ->
fail "xlate_value: %a" pp_llvalue llv ()
in
Hashtbl.find_or_add memo_value (inline, llv) ~default:(fun () ->
[%Trace.call fun {pf} -> pf "%a" pp_llvalue llv]
;
xlate_value_ llv
|>
[%Trace.retn fun {pf} exp -> pf "%a" Exp.pp exp] )
[sledge] Represent recursive records non-recursively Summary: In LLVM it is possible for struct constant values to be directly recursive, with no pointer dereference to close the cycle. These appear for example as the values of vtables from C++ code. Currently such recursive records in the Exp and Term languages are represented as genuinely cyclic values. Compared to a standard term representation, the presence of cyclic values is a significant complication everywhere. Since the backend solver does not do anything such as induction over these, they have to be treated as essentially atomic. This patch changes the representation to a standard non-recursive tree term structure. Instead of cyclic references, an explicit constructor is added for the "non-tree edges", which simply indicates which ancestor record value to which the recursive reference points. There is a potential issue with this representation, since for mutually recursive records, the representation is not canonical: it chooses one of the records in the cycle to start from and expresses the cycles relative to that. Currently the choice of representation is dictated by the frontend. For the case of vtables, the frontend translates globals in the same order they appear in the LLVM IR, so the representation choice is fixed. It may turn out that other potential uses require more reasoning support in the backend solver, which would involve a theory of equality of record values induced by equating the representations resulting from different rotations of the cycle of records. Reviewed By: jvillard Differential Revision: D21441533 fbshipit-source-id: 0c5a11378
5 years ago
and xlate_opcode stk : x -> Llvm.llvalue -> Llvm.Opcode.t -> Exp.t =
fun x llv opcode ->
[%Trace.call fun {pf} -> pf "%a" pp_llvalue llv]
;
[sledge] Represent recursive records non-recursively Summary: In LLVM it is possible for struct constant values to be directly recursive, with no pointer dereference to close the cycle. These appear for example as the values of vtables from C++ code. Currently such recursive records in the Exp and Term languages are represented as genuinely cyclic values. Compared to a standard term representation, the presence of cyclic values is a significant complication everywhere. Since the backend solver does not do anything such as induction over these, they have to be treated as essentially atomic. This patch changes the representation to a standard non-recursive tree term structure. Instead of cyclic references, an explicit constructor is added for the "non-tree edges", which simply indicates which ancestor record value to which the recursive reference points. There is a potential issue with this representation, since for mutually recursive records, the representation is not canonical: it chooses one of the records in the cycle to start from and expresses the cycles relative to that. Currently the choice of representation is dictated by the frontend. For the case of vtables, the frontend translates globals in the same order they appear in the LLVM IR, so the representation choice is fixed. It may turn out that other potential uses require more reasoning support in the backend solver, which would involve a theory of equality of record values induced by equating the representations resulting from different rotations of the cycle of records. Reviewed By: jvillard Differential Revision: D21441533 fbshipit-source-id: 0c5a11378
5 years ago
let xlate_rand i = xlate_value stk x (Llvm.operand llv i) in
let typ = lazy (xlate_type x (Llvm.type_of llv)) in
[sledge] Simplify type conversions Summary: The treatment of type conversions is too complicated, non-uniform, etc. This diff attempts to simplify things by separating integer to integer conversions, which are interpreted, from others, which are essentially just uninterpreted functions. Integer conversions are now handled using two expression and term forms: Signed and Unsigned. These each interpret their argument as either a signed or unsigned number of a given bitwidth: ``` | Signed of {bits: int} (** [Ap1 (Signed {bits= n}, dst, arg)] is [arg] interpreted as an [n]-bit signed integer and injected into the [dst] type. That is, it two's-complement--decodes the low [n] bits of the infinite two's-complement encoding of [arg]. The injection into [dst] is a no-op, so [dst] must be an integer type with bitwidth at least [n]. *) | Unsigned of {bits: int} (** [Ap1 (Unsigned {bits= n}, dst, arg)] is [arg] interpreted as an [n]-bit unsigned integer and injected into the [dst] type. That is, it unsigned-binary--decodes the low [n] bits of the infinite two's-complement encoding of [arg]. The injection into [dst] is a no-op, so [dst] must be an integer type with bitwidth greater than [n]. *) | Convert of {src: Typ.t} (** [Ap1 (Convert {src}, dst, arg)] is [arg] converted from type [src] to type [dst], possibly with loss of information. The [src] and [dst] types must be [Typ.convertible] and must not both be [Integer] types. *) ``` Reviewed By: ngorogiannis Differential Revision: D18298140 fbshipit-source-id: 690f065b4
5 years ago
let convert opcode =
let dst = Lazy.force typ in
let rand = Llvm.operand llv 0 in
let src = xlate_type x (Llvm.type_of rand) in
[sledge] Represent recursive records non-recursively Summary: In LLVM it is possible for struct constant values to be directly recursive, with no pointer dereference to close the cycle. These appear for example as the values of vtables from C++ code. Currently such recursive records in the Exp and Term languages are represented as genuinely cyclic values. Compared to a standard term representation, the presence of cyclic values is a significant complication everywhere. Since the backend solver does not do anything such as induction over these, they have to be treated as essentially atomic. This patch changes the representation to a standard non-recursive tree term structure. Instead of cyclic references, an explicit constructor is added for the "non-tree edges", which simply indicates which ancestor record value to which the recursive reference points. There is a potential issue with this representation, since for mutually recursive records, the representation is not canonical: it chooses one of the records in the cycle to start from and expresses the cycles relative to that. Currently the choice of representation is dictated by the frontend. For the case of vtables, the frontend translates globals in the same order they appear in the LLVM IR, so the representation choice is fixed. It may turn out that other potential uses require more reasoning support in the backend solver, which would involve a theory of equality of record values induced by equating the representations resulting from different rotations of the cycle of records. Reviewed By: jvillard Differential Revision: D21441533 fbshipit-source-id: 0c5a11378
5 years ago
let arg = xlate_value stk x rand in
match (opcode : Llvm.Opcode.t) with
[sledge] Simplify type conversions Summary: The treatment of type conversions is too complicated, non-uniform, etc. This diff attempts to simplify things by separating integer to integer conversions, which are interpreted, from others, which are essentially just uninterpreted functions. Integer conversions are now handled using two expression and term forms: Signed and Unsigned. These each interpret their argument as either a signed or unsigned number of a given bitwidth: ``` | Signed of {bits: int} (** [Ap1 (Signed {bits= n}, dst, arg)] is [arg] interpreted as an [n]-bit signed integer and injected into the [dst] type. That is, it two's-complement--decodes the low [n] bits of the infinite two's-complement encoding of [arg]. The injection into [dst] is a no-op, so [dst] must be an integer type with bitwidth at least [n]. *) | Unsigned of {bits: int} (** [Ap1 (Unsigned {bits= n}, dst, arg)] is [arg] interpreted as an [n]-bit unsigned integer and injected into the [dst] type. That is, it unsigned-binary--decodes the low [n] bits of the infinite two's-complement encoding of [arg]. The injection into [dst] is a no-op, so [dst] must be an integer type with bitwidth greater than [n]. *) | Convert of {src: Typ.t} (** [Ap1 (Convert {src}, dst, arg)] is [arg] converted from type [src] to type [dst], possibly with loss of information. The [src] and [dst] types must be [Typ.convertible] and must not both be [Integer] types. *) ``` Reviewed By: ngorogiannis Differential Revision: D18298140 fbshipit-source-id: 690f065b4
5 years ago
| Trunc -> Exp.signed (Typ.bit_size_of dst) arg ~to_:dst
| SExt -> Exp.signed (Typ.bit_size_of src) arg ~to_:dst
| ZExt -> Exp.unsigned (Typ.bit_size_of src) arg ~to_:dst
| (BitCast | AddrSpaceCast) when Typ.equal dst src -> arg
| FPToUI | FPToSI | UIToFP | SIToFP | FPTrunc | FPExt | PtrToInt
|IntToPtr | BitCast | AddrSpaceCast ->
Exp.convert src ~to_:dst arg
| _ -> fail "convert: %a" pp_llvalue llv ()
in
let binary (mk : ?typ:_ -> _) =
if Poly.equal (Llvm.classify_type (Llvm.type_of llv)) Vector then
todo "vector operations: %a" pp_llvalue llv () ;
let typ = xlate_type x (Llvm.type_of (Llvm.operand llv 0)) in
mk ~typ (xlate_rand 0) (xlate_rand 1)
in
let unordered_or mk =
binary (fun ?typ e f ->
Exp.or_ ~typ:Typ.bool (Exp.uno ?typ e f) (mk ?typ e f) )
in
( match opcode with
[sledge] Simplify type conversions Summary: The treatment of type conversions is too complicated, non-uniform, etc. This diff attempts to simplify things by separating integer to integer conversions, which are interpreted, from others, which are essentially just uninterpreted functions. Integer conversions are now handled using two expression and term forms: Signed and Unsigned. These each interpret their argument as either a signed or unsigned number of a given bitwidth: ``` | Signed of {bits: int} (** [Ap1 (Signed {bits= n}, dst, arg)] is [arg] interpreted as an [n]-bit signed integer and injected into the [dst] type. That is, it two's-complement--decodes the low [n] bits of the infinite two's-complement encoding of [arg]. The injection into [dst] is a no-op, so [dst] must be an integer type with bitwidth at least [n]. *) | Unsigned of {bits: int} (** [Ap1 (Unsigned {bits= n}, dst, arg)] is [arg] interpreted as an [n]-bit unsigned integer and injected into the [dst] type. That is, it unsigned-binary--decodes the low [n] bits of the infinite two's-complement encoding of [arg]. The injection into [dst] is a no-op, so [dst] must be an integer type with bitwidth greater than [n]. *) | Convert of {src: Typ.t} (** [Ap1 (Convert {src}, dst, arg)] is [arg] converted from type [src] to type [dst], possibly with loss of information. The [src] and [dst] types must be [Typ.convertible] and must not both be [Integer] types. *) ``` Reviewed By: ngorogiannis Differential Revision: D18298140 fbshipit-source-id: 690f065b4
5 years ago
| Trunc | ZExt | SExt | FPToUI | FPToSI | UIToFP | SIToFP | FPTrunc
|FPExt | PtrToInt | IntToPtr | BitCast | AddrSpaceCast ->
convert opcode
| ICmp -> (
match Option.value_exn (Llvm.icmp_predicate llv) with
| Eq -> binary Exp.eq
| Ne -> binary Exp.dq
| Sgt -> binary Exp.gt
| Sge -> binary Exp.ge
| Slt -> binary Exp.lt
| Sle -> binary Exp.le
| Ugt -> binary Exp.ugt
| Uge -> binary Exp.uge
| Ult -> binary Exp.ult
| Ule -> binary Exp.ule )
| FCmp -> (
match Llvm.fcmp_predicate llv with
| None | Some False -> binary (fun ?typ:_ _ _ -> Exp.false_)
| Some Oeq -> binary Exp.eq
| Some Ogt -> binary Exp.gt
| Some Oge -> binary Exp.ge
| Some Olt -> binary Exp.lt
| Some Ole -> binary Exp.le
| Some One -> binary Exp.dq
| Some Ord -> binary Exp.ord
| Some Uno -> binary Exp.uno
| Some Ueq -> unordered_or Exp.eq
| Some Ugt -> unordered_or Exp.gt
| Some Uge -> unordered_or Exp.ge
| Some Ult -> unordered_or Exp.lt
| Some Ule -> unordered_or Exp.le
| Some Une -> unordered_or Exp.dq
| Some True -> binary (fun ?typ:_ _ _ -> Exp.true_) )
| Add | FAdd -> binary Exp.add
| Sub | FSub -> binary Exp.sub
| Mul | FMul -> binary Exp.mul
| SDiv | FDiv -> binary Exp.div
| UDiv -> binary Exp.udiv
| SRem | FRem -> binary Exp.rem
| URem -> binary Exp.urem
| Shl -> binary Exp.shl
| LShr -> binary Exp.lshr
| AShr -> binary Exp.ashr
| And -> binary Exp.and_
| Or -> binary Exp.or_
| Xor -> binary Exp.xor
| Select ->
let typ = xlate_type x (Llvm.type_of (Llvm.operand llv 1)) in
Exp.conditional ~typ ~cnd:(xlate_rand 0) ~thn:(xlate_rand 1)
~els:(xlate_rand 2)
| ExtractElement | InsertElement -> (
let typ =
let lltyp = Llvm.type_of (Llvm.operand llv 0) in
let llelt = Llvm.element_type lltyp in
let elt = xlate_type x llelt in
let len = Llvm.vector_size llelt in
let byts = size_of x lltyp in
let bits = bit_size_of x lltyp in
Typ.array ~elt ~len ~bits ~byts
in
let idx i =
match xlate_rand i with
| Integer {data} -> Z.to_int data
| _ -> todo "vector operations: %a" pp_llvalue llv ()
in
let rcd = xlate_rand 0 in
match opcode with
| ExtractElement -> Exp.select typ rcd (idx 1)
| InsertElement -> Exp.update typ ~rcd (idx 2) ~elt:(xlate_rand 1)
| _ -> assert false )
| ExtractValue | InsertValue ->
let agg = xlate_rand 0 in
let typ = xlate_type x (Llvm.type_of (Llvm.operand llv 0)) in
let indices = Llvm.indices llv in
let num = Array.length indices in
let rec xlate_indices i rcd typ =
let rcd_i, typ_i, upd =
match (typ : Typ.t) with
| Tuple {elts} | Struct {elts} ->
( Exp.select typ rcd indices.(i)
, IArray.get elts indices.(i)
, Exp.update typ ~rcd indices.(i) )
| Array {elt} ->
( Exp.select typ rcd indices.(i)
, elt
, Exp.update typ ~rcd indices.(i) )
| _ -> fail "xlate_value: %a" pp_llvalue llv ()
in
let update_or_return elt ret =
match[@warning "p"] opcode with
| InsertValue -> upd ~elt:(Lazy.force elt)
| ExtractValue -> ret
in
if i < num - 1 then
let elt = xlate_indices (i + 1) rcd_i typ_i in
update_or_return (lazy elt) elt
else
let elt = lazy (xlate_rand 1) in
update_or_return elt rcd_i
in
xlate_indices 0 agg typ
| GetElementPtr ->
if Poly.equal (Llvm.classify_type (Llvm.type_of llv)) Vector then
todo "vector operations: %a" pp_llvalue llv () ;
let len = Llvm.num_operands llv in
[sledge] Simplify type conversions Summary: The treatment of type conversions is too complicated, non-uniform, etc. This diff attempts to simplify things by separating integer to integer conversions, which are interpreted, from others, which are essentially just uninterpreted functions. Integer conversions are now handled using two expression and term forms: Signed and Unsigned. These each interpret their argument as either a signed or unsigned number of a given bitwidth: ``` | Signed of {bits: int} (** [Ap1 (Signed {bits= n}, dst, arg)] is [arg] interpreted as an [n]-bit signed integer and injected into the [dst] type. That is, it two's-complement--decodes the low [n] bits of the infinite two's-complement encoding of [arg]. The injection into [dst] is a no-op, so [dst] must be an integer type with bitwidth at least [n]. *) | Unsigned of {bits: int} (** [Ap1 (Unsigned {bits= n}, dst, arg)] is [arg] interpreted as an [n]-bit unsigned integer and injected into the [dst] type. That is, it unsigned-binary--decodes the low [n] bits of the infinite two's-complement encoding of [arg]. The injection into [dst] is a no-op, so [dst] must be an integer type with bitwidth greater than [n]. *) | Convert of {src: Typ.t} (** [Ap1 (Convert {src}, dst, arg)] is [arg] converted from type [src] to type [dst], possibly with loss of information. The [src] and [dst] types must be [Typ.convertible] and must not both be [Integer] types. *) ``` Reviewed By: ngorogiannis Differential Revision: D18298140 fbshipit-source-id: 690f065b4
5 years ago
assert (len > 0 || invalid_llvm (Llvm.string_of_llvalue llv)) ;
if len = 1 then convert BitCast
else
let rec xlate_indices i =
[%Trace.call fun {pf} ->
pf "%i %a" i pp_llvalue (Llvm.operand llv i)]
;
let idx =
[sledge] Simplify type conversions Summary: The treatment of type conversions is too complicated, non-uniform, etc. This diff attempts to simplify things by separating integer to integer conversions, which are interpreted, from others, which are essentially just uninterpreted functions. Integer conversions are now handled using two expression and term forms: Signed and Unsigned. These each interpret their argument as either a signed or unsigned number of a given bitwidth: ``` | Signed of {bits: int} (** [Ap1 (Signed {bits= n}, dst, arg)] is [arg] interpreted as an [n]-bit signed integer and injected into the [dst] type. That is, it two's-complement--decodes the low [n] bits of the infinite two's-complement encoding of [arg]. The injection into [dst] is a no-op, so [dst] must be an integer type with bitwidth at least [n]. *) | Unsigned of {bits: int} (** [Ap1 (Unsigned {bits= n}, dst, arg)] is [arg] interpreted as an [n]-bit unsigned integer and injected into the [dst] type. That is, it unsigned-binary--decodes the low [n] bits of the infinite two's-complement encoding of [arg]. The injection into [dst] is a no-op, so [dst] must be an integer type with bitwidth greater than [n]. *) | Convert of {src: Typ.t} (** [Ap1 (Convert {src}, dst, arg)] is [arg] converted from type [src] to type [dst], possibly with loss of information. The [src] and [dst] types must be [Typ.convertible] and must not both be [Integer] types. *) ``` Reviewed By: ngorogiannis Differential Revision: D18298140 fbshipit-source-id: 690f065b4
5 years ago
convert_to_siz
(xlate_type x (Llvm.type_of (Llvm.operand llv i)))
(xlate_rand i)
in
( if i = 1 then
let base = xlate_rand 0 in
let lltyp = Llvm.type_of (Llvm.operand llv 0) in
let llelt =
match Llvm.classify_type lltyp with
| Pointer -> Llvm.element_type lltyp
| _ -> fail "xlate_opcode: %i %a" i pp_llvalue llv ()
in
(* translate [gep t*, iN M] as [gep [1 x t]*, iN M] *)
(ptr_idx x ~ptr:base ~idx ~llelt, llelt)
else
let ptr, lltyp = xlate_indices (i - 1) in
match Llvm.classify_type lltyp with
| Array | Vector ->
let llelt = Llvm.element_type lltyp in
(ptr_idx x ~ptr ~idx ~llelt, llelt)
| Struct ->
let fld =
match
Option.bind ~f:Int64.to_int
(Llvm.int64_of_const (Llvm.operand llv i))
with
| Some n -> n
| None -> fail "xlate_opcode: %i %a" i pp_llvalue llv ()
in
let llelt = (Llvm.struct_element_types lltyp).(fld) in
(ptr_fld x ~ptr ~fld ~lltyp, llelt)
| _ -> fail "xlate_opcode: %i %a" i pp_llvalue llv () )
|>
[%Trace.retn fun {pf} (exp, llt) ->
pf "%a %a" Exp.pp exp pp_lltype llt]
in
fst (xlate_indices (len - 1))
| ShuffleVector -> (
(* translate shufflevector <N x t> %x, _, <N x i32> zeroinitializer to
%x *)
[sledge] Represent recursive records non-recursively Summary: In LLVM it is possible for struct constant values to be directly recursive, with no pointer dereference to close the cycle. These appear for example as the values of vtables from C++ code. Currently such recursive records in the Exp and Term languages are represented as genuinely cyclic values. Compared to a standard term representation, the presence of cyclic values is a significant complication everywhere. Since the backend solver does not do anything such as induction over these, they have to be treated as essentially atomic. This patch changes the representation to a standard non-recursive tree term structure. Instead of cyclic references, an explicit constructor is added for the "non-tree edges", which simply indicates which ancestor record value to which the recursive reference points. There is a potential issue with this representation, since for mutually recursive records, the representation is not canonical: it chooses one of the records in the cycle to start from and expresses the cycles relative to that. Currently the choice of representation is dictated by the frontend. For the case of vtables, the frontend translates globals in the same order they appear in the LLVM IR, so the representation choice is fixed. It may turn out that other potential uses require more reasoning support in the backend solver, which would involve a theory of equality of record values induced by equating the representations resulting from different rotations of the cycle of records. Reviewed By: jvillard Differential Revision: D21441533 fbshipit-source-id: 0c5a11378
5 years ago
let exp = xlate_value stk x (Llvm.operand llv 0) in
let exp_typ = xlate_type x (Llvm.type_of (Llvm.operand llv 0)) in
let llmask = Llvm.operand llv 2 in
let mask_typ = xlate_type x (Llvm.type_of llmask) in
match (exp_typ, mask_typ) with
| Array {len= m}, Array {len= n} when m = n && Llvm.is_null llmask ->
exp
| _ -> todo "vector operations: %a" pp_llvalue llv () )
| Invalid | Ret | Br | Switch | IndirectBr | Invoke | Invalid2
|Unreachable | Alloca | Load | Store | PHI | Call | UserOp1 | UserOp2
|Fence | AtomicCmpXchg | AtomicRMW | Resume | LandingPad | CleanupRet
|CatchRet | CatchPad | CleanupPad | CatchSwitch | VAArg ->
fail "xlate_opcode: %a" pp_llvalue llv () )
|>
[%Trace.retn fun {pf} exp -> pf "%a" Exp.pp exp]
[sledge] Represent recursive records non-recursively Summary: In LLVM it is possible for struct constant values to be directly recursive, with no pointer dereference to close the cycle. These appear for example as the values of vtables from C++ code. Currently such recursive records in the Exp and Term languages are represented as genuinely cyclic values. Compared to a standard term representation, the presence of cyclic values is a significant complication everywhere. Since the backend solver does not do anything such as induction over these, they have to be treated as essentially atomic. This patch changes the representation to a standard non-recursive tree term structure. Instead of cyclic references, an explicit constructor is added for the "non-tree edges", which simply indicates which ancestor record value to which the recursive reference points. There is a potential issue with this representation, since for mutually recursive records, the representation is not canonical: it chooses one of the records in the cycle to start from and expresses the cycles relative to that. Currently the choice of representation is dictated by the frontend. For the case of vtables, the frontend translates globals in the same order they appear in the LLVM IR, so the representation choice is fixed. It may turn out that other potential uses require more reasoning support in the backend solver, which would involve a theory of equality of record values induced by equating the representations resulting from different rotations of the cycle of records. Reviewed By: jvillard Differential Revision: D21441533 fbshipit-source-id: 0c5a11378
5 years ago
and xlate_global stk : x -> Llvm.llvalue -> Global.t =
fun x llg ->
Hashtbl.find_or_add memo_global llg ~default:(fun () ->
[%Trace.call fun {pf} -> pf "%a" pp_llvalue llg]
;
let g = xlate_name x ~global:() llg in
let loc = find_loc llg in
(* add to tbl without initializer in case of recursive occurrences in
its own initializer *)
Hashtbl.set memo_global ~key:llg ~data:(Global.mk g loc) ;
let init =
match Llvm.classify_value llg with
| GlobalVariable ->
Option.map (Llvm.global_initializer llg) ~f:(fun llv ->
(xlate_value stk x llv, size_of x (Llvm.type_of llv)) )
| _ -> None
in
Global.mk ?init g loc
|>
[%Trace.retn fun {pf} -> pf "%a" Global.pp_defn] )
[sledge] Represent recursive records non-recursively Summary: In LLVM it is possible for struct constant values to be directly recursive, with no pointer dereference to close the cycle. These appear for example as the values of vtables from C++ code. Currently such recursive records in the Exp and Term languages are represented as genuinely cyclic values. Compared to a standard term representation, the presence of cyclic values is a significant complication everywhere. Since the backend solver does not do anything such as induction over these, they have to be treated as essentially atomic. This patch changes the representation to a standard non-recursive tree term structure. Instead of cyclic references, an explicit constructor is added for the "non-tree edges", which simply indicates which ancestor record value to which the recursive reference points. There is a potential issue with this representation, since for mutually recursive records, the representation is not canonical: it chooses one of the records in the cycle to start from and expresses the cycles relative to that. Currently the choice of representation is dictated by the frontend. For the case of vtables, the frontend translates globals in the same order they appear in the LLVM IR, so the representation choice is fixed. It may turn out that other potential uses require more reasoning support in the backend solver, which would involve a theory of equality of record values induced by equating the representations resulting from different rotations of the cycle of records. Reviewed By: jvillard Differential Revision: D21441533 fbshipit-source-id: 0c5a11378
5 years ago
let xlate_intrinsic_exp = xlate_intrinsic_exp []
let xlate_value ?inline = xlate_value ?inline []
let xlate_opcode = xlate_opcode []
let xlate_global = xlate_global []
type pop_thunk = Loc.t -> Llair.inst list
let pop_stack_frame_of_function :
x -> Llvm.llvalue -> Llvm.llbasicblock -> pop_thunk =
fun x func entry_blk ->
let append_stack_regs blk regs =
Llvm.fold_right_instrs
(fun instr regs ->
match Llvm.instr_opcode instr with
| Alloca -> xlate_name x instr :: regs
| _ -> regs )
blk regs
in
let entry_regs = append_stack_regs entry_blk [] in
Llvm.iter_blocks
(fun blk ->
if not (Poly.equal entry_blk blk) then
Llvm.iter_instrs
(fun instr ->
match Llvm.instr_opcode instr with
| Alloca ->
warn "stack allocation after function entry:@ %a" Loc.pp
(find_loc instr) ()
| _ -> () )
blk )
func ;
let pop retn_loc =
List.map entry_regs ~f:(fun reg ->
Llair.Inst.free ~ptr:(Exp.reg reg) ~loc:retn_loc )
in
pop
(** construct the types involved in landingpads: i32, std::type_info*, and
__cxa_exception *)
let landingpad_typs : x -> Llvm.llvalue -> Typ.t * Typ.t * Llvm.lltype =
fun x instr ->
let llt = Llvm.type_of instr in
let i32 = i32 x in
if
not
( Poly.(Llvm.classify_type llt = Struct)
&&
let llelts = Llvm.struct_element_types llt in
Array.length llelts = 2
&& Poly.(llelts.(0) = Llvm.pointer_type (Llvm.i8_type x.llcontext))
&& Poly.(llelts.(1) = Llvm.i32_type x.llcontext) )
then
todo "landingpad of type other than {i8*, i32}: %a" pp_llvalue instr () ;
let llcontext =
Llvm.(
module_context (global_parent (block_parent (instr_parent instr))))
in
let llpi8 = Llvm.(pointer_type (integer_type llcontext 8)) in
let ti = Llvm.(named_struct_type llcontext "class.std::type_info") in
let tip = Llvm.pointer_type ti in
let void = Llvm.void_type llcontext in
let dtor = Llvm.(pointer_type (function_type void [|llpi8|])) in
let cxa_exception = Llvm.struct_type llcontext [|tip; dtor|] in
(i32, xlate_type x tip, cxa_exception)
let exception_typs =
let pi8 = Typ.pointer ~elt:Typ.byt in
let i32 = Typ.integer ~bits:32 ~byts:4 in
let exc =
Typ.tuple ~packed:false (IArray.of_array [|pi8; i32|]) ~bits:96 ~byts:12
in
(pi8, i32, exc)
(** Translate a control transfer from instruction [instr] to block [dst] to
a jump, if necessary by extending [blocks] with a trampoline containing
the PHIs of [dst] translated to a move. *)
let xlate_jump :
x
-> ?reg_exps:(Reg.t * Exp.t) list
-> Llvm.llvalue
-> Llvm.llbasicblock
-> Loc.t
-> Llair.block list
-> Llair.jump * Llair.block list =
fun x ?(reg_exps = []) instr dst loc blocks ->
let src = Llvm.instr_parent instr in
let rec xlate_jump_ reg_exps (pos : _ Llvm.llpos) =
match pos with
| Before dst_instr -> (
match Llvm.instr_opcode dst_instr with
| PHI ->
let reg_exp =
List.find_map_exn (Llvm.incoming dst_instr)
~f:(fun (arg, pred) ->
if Poly.equal pred src then
Some (xlate_name x dst_instr, xlate_value x arg)
else None )
in
xlate_jump_ (reg_exp :: reg_exps) (Llvm.instr_succ dst_instr)
| _ -> reg_exps )
| At_end blk -> fail "xlate_jump: %a" pp_llblock blk ()
in
let dst_lbl = label_of_block dst in
let jmp = Llair.Jump.mk dst_lbl in
match xlate_jump_ reg_exps (Llvm.instr_begin dst) with
| [] -> (jmp, blocks)
| reg_exps ->
let mov =
Llair.Inst.move ~reg_exps:(IArray.of_list_rev reg_exps) ~loc
in
let lbl = find_name instr ^ ".jmp." ^ dst_lbl in
let blk =
Llair.Block.mk ~lbl
~cmnd:(IArray.of_array [|mov|])
~term:(Llair.Term.goto ~dst:jmp ~loc)
in
let blocks =
match List.find blocks ~f:(fun b -> String.equal lbl b.lbl) with
| None -> blk :: blocks
| Some blk0 ->
assert (Llair.Block.equal blk0 blk) ;
blocks
in
(Llair.Jump.mk lbl, blocks)
(** An LLVM instruction is translated to a sequence of LLAIR instructions
and a terminator, plus some additional blocks to which it may refer
(that is, essentially a function body). These are needed since LLVM and
LLAIR blocks are not in 1:1 correspondence. *)
type code = Llair.inst list * Llair.term * Llair.block list
let pp_code fs (insts, term, blocks) =
Format.fprintf fs "@[<hv>@,@[%a%t@]%t@[<hv>%a@]@]"
(List.pp "@ " Llair.Inst.pp)
insts
(fun fs ->
match term with
| Llair.Unreachable -> ()
| _ ->
Format.fprintf fs "%t%a"
(fun fs ->
if List.is_empty insts then () else Format.fprintf fs "@ " )
Llair.Term.pp term )
(fun fs -> if List.is_empty blocks then () else Format.fprintf fs "@\n")
(List.pp "@ " Llair.Block.pp)
blocks
let rec xlate_func_name x llv =
match Llvm.classify_value llv with
| Function | GlobalVariable -> Exp.reg (xlate_name x ~global:() llv)
| ConstantExpr -> xlate_opcode x llv (Llvm.constexpr_opcode llv)
| Argument | Instruction _ -> xlate_value x llv
| GlobalAlias -> xlate_func_name x (Llvm.operand llv 0)
| GlobalIFunc -> todo "ifunc: %a" pp_llvalue llv ()
| InlineAsm -> todo "inline asm: %a" pp_llvalue llv ()
| ConstantPointerNull -> todo "call null: %a" pp_llvalue llv ()
| _ -> todo "function kind in %a" pp_llvalue llv ()
let ignored_callees = Hash_set.create (module String)
let xlate_size_of x llv =
Exp.integer Typ.siz (Z.of_int (size_of x (Llvm.type_of llv)))
let xlate_instr :
pop_thunk
-> x
-> Llvm.llvalue
-> ((Llair.inst list * Llair.term -> code) -> code)
-> code =
fun pop x instr continue ->
[%Trace.call fun {pf} -> pf "%a" pp_llvalue instr]
;
let continue insts_term_to_code =
[%Trace.retn
fun {pf} () ->
pf "%a" pp_code (insts_term_to_code ([], Llair.Term.unreachable))]
() ;
continue insts_term_to_code
in
let nop () = continue (fun (insts, term) -> (insts, term, [])) in
let emit_inst inst =
continue (fun (insts, term) -> (inst :: insts, term, []))
in
let emit_term ?(prefix = []) ?(blocks = []) term =
[%Trace.retn fun {pf} () -> pf "%a" pp_code (prefix, term, blocks)] () ;
(prefix, term, blocks)
in
let name = find_name instr in
let loc = find_loc instr in
let inline_or_move xlate =
if should_inline instr then nop ()
else
let reg = xlate_name x instr in
let exp = xlate instr in
let reg_exps = IArray.of_array [|(reg, exp)|] in
emit_inst (Llair.Inst.move ~reg_exps ~loc)
in
let opcode = Llvm.instr_opcode instr in
match opcode with
| Load ->
let reg = xlate_name x instr in
let len = xlate_size_of x instr in
let ptr = xlate_value x (Llvm.operand instr 0) in
emit_inst (Llair.Inst.load ~reg ~ptr ~len ~loc)
| Store ->
let rand0 = Llvm.operand instr 0 in
let exp = xlate_value x rand0 in
let len = xlate_size_of x rand0 in
let ptr = xlate_value x (Llvm.operand instr 1) in
emit_inst (Llair.Inst.store ~ptr ~exp ~len ~loc)
| Alloca ->
let reg = xlate_name x instr in
let rand = Llvm.operand instr 0 in
let num =
[sledge] Simplify type conversions Summary: The treatment of type conversions is too complicated, non-uniform, etc. This diff attempts to simplify things by separating integer to integer conversions, which are interpreted, from others, which are essentially just uninterpreted functions. Integer conversions are now handled using two expression and term forms: Signed and Unsigned. These each interpret their argument as either a signed or unsigned number of a given bitwidth: ``` | Signed of {bits: int} (** [Ap1 (Signed {bits= n}, dst, arg)] is [arg] interpreted as an [n]-bit signed integer and injected into the [dst] type. That is, it two's-complement--decodes the low [n] bits of the infinite two's-complement encoding of [arg]. The injection into [dst] is a no-op, so [dst] must be an integer type with bitwidth at least [n]. *) | Unsigned of {bits: int} (** [Ap1 (Unsigned {bits= n}, dst, arg)] is [arg] interpreted as an [n]-bit unsigned integer and injected into the [dst] type. That is, it unsigned-binary--decodes the low [n] bits of the infinite two's-complement encoding of [arg]. The injection into [dst] is a no-op, so [dst] must be an integer type with bitwidth greater than [n]. *) | Convert of {src: Typ.t} (** [Ap1 (Convert {src}, dst, arg)] is [arg] converted from type [src] to type [dst], possibly with loss of information. The [src] and [dst] types must be [Typ.convertible] and must not both be [Integer] types. *) ``` Reviewed By: ngorogiannis Differential Revision: D18298140 fbshipit-source-id: 690f065b4
5 years ago
convert_to_siz
(xlate_type x (Llvm.type_of rand))
(xlate_value x rand)
in
assert (Poly.(Llvm.classify_type (Llvm.type_of instr) = Pointer)) ;
let len = xlate_size_of x instr in
emit_inst (Llair.Inst.alloc ~reg ~num ~len ~loc)
| Call -> (
let maybe_llfunc = Llvm.operand instr (Llvm.num_operands instr - 1) in
let lltyp = Llvm.type_of maybe_llfunc in
assert (Poly.(Llvm.classify_type lltyp = Pointer)) ;
let llfunc =
let llfunc_valuekind = Llvm.classify_value maybe_llfunc in
match llfunc_valuekind with
| Function | Instruction _ | InlineAsm | Argument -> maybe_llfunc
| ConstantExpr -> (
match Llvm.constexpr_opcode maybe_llfunc with
| BitCast -> Llvm.operand maybe_llfunc 0
| _ ->
todo "opcode kind in call instruction %a" pp_llvalue
maybe_llfunc () )
| _ ->
todo "operand kind in call instruction %a" pp_llvalue
maybe_llfunc ()
in
let fname = Llvm.value_name llfunc in
let skip msg =
( match Hash_set.strict_add ignored_callees fname with
| Ok () -> warn "ignoring uninterpreted %s %s" msg fname ()
| Error _ -> () ) ;
let reg = xlate_name_opt x instr in
emit_inst (Llair.Inst.nondet ~reg ~msg:fname ~loc)
in
(* intrinsics *)
match xlate_intrinsic_exp fname with
| Some intrinsic -> inline_or_move (intrinsic x)
| None -> (
match String.split fname ~on:'.' with
| ["__llair_throw"] ->
let exc = xlate_value x (Llvm.operand instr 0) in
emit_term ~prefix:(pop loc) (Llair.Term.throw ~exc ~loc)
| ["__llair_alloc" (* void* __llair_alloc(unsigned size) *)] ->
let reg = xlate_name x instr in
let num_operand = Llvm.operand instr 0 in
let num =
[sledge] Simplify type conversions Summary: The treatment of type conversions is too complicated, non-uniform, etc. This diff attempts to simplify things by separating integer to integer conversions, which are interpreted, from others, which are essentially just uninterpreted functions. Integer conversions are now handled using two expression and term forms: Signed and Unsigned. These each interpret their argument as either a signed or unsigned number of a given bitwidth: ``` | Signed of {bits: int} (** [Ap1 (Signed {bits= n}, dst, arg)] is [arg] interpreted as an [n]-bit signed integer and injected into the [dst] type. That is, it two's-complement--decodes the low [n] bits of the infinite two's-complement encoding of [arg]. The injection into [dst] is a no-op, so [dst] must be an integer type with bitwidth at least [n]. *) | Unsigned of {bits: int} (** [Ap1 (Unsigned {bits= n}, dst, arg)] is [arg] interpreted as an [n]-bit unsigned integer and injected into the [dst] type. That is, it unsigned-binary--decodes the low [n] bits of the infinite two's-complement encoding of [arg]. The injection into [dst] is a no-op, so [dst] must be an integer type with bitwidth greater than [n]. *) | Convert of {src: Typ.t} (** [Ap1 (Convert {src}, dst, arg)] is [arg] converted from type [src] to type [dst], possibly with loss of information. The [src] and [dst] types must be [Typ.convertible] and must not both be [Integer] types. *) ``` Reviewed By: ngorogiannis Differential Revision: D18298140 fbshipit-source-id: 690f065b4
5 years ago
convert_to_siz
(xlate_type x (Llvm.type_of num_operand))
(xlate_value x num_operand)
in
let len = Exp.integer Typ.siz (Z.of_int 1) in
emit_inst (Llair.Inst.alloc ~reg ~num ~len ~loc)
| ["_Znwm" (* operator new(size_t num) *)]
|[ "_ZnwmSt11align_val_t"
(* operator new(unsigned long, std::align_val_t) *) ] ->
let reg = xlate_name x instr in
let num = xlate_value x (Llvm.operand instr 0) in
let len = xlate_size_of x instr in
emit_inst (Llair.Inst.alloc ~reg ~num ~len ~loc)
| ["_ZdlPv" (* operator delete(void* ptr) *)]
|[ "_ZdlPvSt11align_val_t"
(* operator delete(void* ptr, std::align_val_t) *) ]
|[ "_ZdlPvmSt11align_val_t"
(* operator delete(void* ptr, unsigned long, std::align_val_t) *)
]
|["free" (* void free(void* ptr) *)] ->
let ptr = xlate_value x (Llvm.operand instr 0) in
emit_inst (Llair.Inst.free ~ptr ~loc)
| "llvm" :: "memset" :: _ ->
let dst = xlate_value x (Llvm.operand instr 0) in
let byt = xlate_value x (Llvm.operand instr 1) in
let len = xlate_value x (Llvm.operand instr 2) in
emit_inst (Llair.Inst.memset ~dst ~byt ~len ~loc)
| "llvm" :: "memcpy" :: _ ->
let dst = xlate_value x (Llvm.operand instr 0) in
let src = xlate_value x (Llvm.operand instr 1) in
let len = xlate_value x (Llvm.operand instr 2) in
emit_inst (Llair.Inst.memcpy ~dst ~src ~len ~loc)
| "llvm" :: "memmove" :: _ ->
let dst = xlate_value x (Llvm.operand instr 0) in
let src = xlate_value x (Llvm.operand instr 1) in
let len = xlate_value x (Llvm.operand instr 2) in
emit_inst (Llair.Inst.memmov ~dst ~src ~len ~loc)
| ["abort"] | ["llvm"; "trap"] -> emit_inst (Llair.Inst.abort ~loc)
(* dropped / handled elsewhere *)
| ["llvm"; "dbg"; ("declare" | "value")]
|"llvm" :: ("lifetime" | "invariant") :: ("start" | "end") :: _ ->
nop ()
(* unimplemented *)
| ["llvm"; ("stacksave" | "stackrestore")] ->
skip "dynamic stack deallocation"
| "llvm" :: "coro" :: _ ->
todo "coroutines:@ %a" pp_llvalue instr ()
| "llvm" :: "experimental" :: "gc" :: "statepoint" :: _ ->
todo "statepoints:@ %a" pp_llvalue instr ()
| ["llvm"; ("va_start" | "va_copy" | "va_end")] ->
skip "variadic function intrinsic"
| "llvm" :: _ -> skip "intrinsic"
| _ when Poly.equal (Llvm.classify_value llfunc) InlineAsm ->
skip "inline asm"
(* general function call that may not throw *)
| _ ->
let callee = xlate_func_name x llfunc in
let typ = xlate_type x lltyp in
let lbl = name ^ ".ret" in
let call =
let actuals =
let num_actuals =
if not (Llvm.is_var_arg (Llvm.element_type lltyp)) then
Llvm.num_arg_operands instr
else
let fname = Llvm.value_name llfunc in
( match Hash_set.strict_add ignored_callees fname with
| Ok () when not (Llvm.is_declaration llfunc) ->
warn
"ignoring variable arguments to variadic \
function: %a"
Exp.pp callee ()
| _ -> () ) ;
let llfty = Llvm.element_type lltyp in
( match Llvm.classify_type llfty with
| Function -> ()
| _ ->
fail "called function not of function type: %a"
pp_llvalue instr () ) ;
Array.length (Llvm.param_types llfty)
in
List.rev_init num_actuals ~f:(fun i ->
xlate_value x (Llvm.operand instr i) )
in
let areturn = xlate_name_opt x instr in
let return = Llair.Jump.mk lbl in
Llair.Term.call ~callee ~typ ~actuals ~areturn ~return
~throw:None ~loc
in
continue (fun (insts, term) ->
let cmnd = IArray.of_list insts in
([], call, [Llair.Block.mk ~lbl ~cmnd ~term]) ) ) )
| Invoke -> (
let llfunc = Llvm.operand instr (Llvm.num_operands instr - 3) in
let lltyp = Llvm.type_of llfunc in
assert (Poly.(Llvm.classify_type lltyp = Pointer)) ;
let fname = Llvm.value_name llfunc in
let return_blk = Llvm.get_normal_dest instr in
let unwind_blk = Llvm.get_unwind_dest instr in
let num_actuals =
if not (Llvm.is_var_arg (Llvm.element_type lltyp)) then
Llvm.num_arg_operands instr
else (
( match Hash_set.strict_add ignored_callees fname with
| Ok () when not (Llvm.is_declaration llfunc) ->
warn "ignoring variable arguments to variadic function: %a"
Global.pp (xlate_global x llfunc) ()
| _ -> () ) ;
assert (Poly.(Llvm.classify_type lltyp = Pointer)) ;
Array.length (Llvm.param_types (Llvm.element_type lltyp)) )
in
(* intrinsics *)
match String.split fname ~on:'.' with
| _ when Option.is_some (xlate_intrinsic_exp fname) ->
let dst, blocks = xlate_jump x instr return_blk loc [] in
emit_term (Llair.Term.goto ~dst ~loc) ~blocks
| ["__llair_throw"] ->
let dst, blocks = xlate_jump x instr unwind_blk loc [] in
emit_term (Llair.Term.goto ~dst ~loc) ~blocks
| ["abort"] ->
emit_term ~prefix:[Llair.Inst.abort ~loc] Llair.Term.unreachable
| ["_Znwm" (* operator new(size_t num) *)]
|[ "_ZnwmSt11align_val_t"
(* operator new(unsigned long num, std::align_val_t) *) ]
when num_actuals > 0 ->
let reg = xlate_name x instr in
let num = xlate_value x (Llvm.operand instr 0) in
let len = xlate_size_of x instr in
let dst, blocks = xlate_jump x instr return_blk loc [] in
emit_term
~prefix:[Llair.Inst.alloc ~reg ~num ~len ~loc]
(Llair.Term.goto ~dst ~loc)
~blocks
(* unimplemented *)
| "llvm" :: "experimental" :: "gc" :: "statepoint" :: _ ->
todo "statepoints:@ %a" pp_llvalue instr ()
(* general function call that may throw *)
| _ ->
let callee = xlate_func_name x llfunc in
let typ = xlate_type x (Llvm.type_of llfunc) in
let actuals =
List.rev_init num_actuals ~f:(fun i ->
xlate_value x (Llvm.operand instr i) )
in
let areturn = xlate_name_opt x instr in
let return, blocks = xlate_jump x instr return_blk loc [] in
let throw, blocks = xlate_jump x instr unwind_blk loc blocks in
let throw = Some throw in
emit_term
(Llair.Term.call ~callee ~typ ~actuals ~areturn ~return ~throw
~loc)
~blocks )
| Ret ->
let exp =
if Llvm.num_operands instr = 0 then None
else Some (xlate_value x (Llvm.operand instr 0))
in
emit_term ~prefix:(pop loc) (Llair.Term.return ~exp ~loc)
| Br -> (
match Option.value_exn (Llvm.get_branch instr) with
| `Unconditional blk ->
let dst, blocks = xlate_jump x instr blk loc [] in
emit_term (Llair.Term.goto ~dst ~loc) ~blocks
| `Conditional (cnd, thn, els) ->
let key = xlate_value x cnd in
let thn, blocks = xlate_jump x instr thn loc [] in
let els, blocks = xlate_jump x instr els loc blocks in
emit_term (Llair.Term.branch ~key ~nzero:thn ~zero:els ~loc) ~blocks
)
| Switch ->
let key = xlate_value x (Llvm.operand instr 0) in
let cases, blocks =
let num_cases = (Llvm.num_operands instr / 2) - 1 in
let rec xlate_cases i blocks =
if i <= num_cases then
let idx = Llvm.operand instr (2 * i) in
let blk =
Llvm.block_of_value (Llvm.operand instr ((2 * i) + 1))
in
let num = xlate_value x idx in
let jmp, blocks = xlate_jump x instr blk loc blocks in
let rest, blocks = xlate_cases (i + 1) blocks in
((num, jmp) :: rest, blocks)
else ([], blocks)
in
xlate_cases 1 []
in
let tbl = IArray.of_list cases in
let blk = Llvm.block_of_value (Llvm.operand instr 1) in
let els, blocks = xlate_jump x instr blk loc blocks in
emit_term (Llair.Term.switch ~key ~tbl ~els ~loc) ~blocks
| IndirectBr ->
let ptr = xlate_value x (Llvm.operand instr 0) in
let num_dests = Llvm.num_operands instr - 1 in
let lldests, blocks =
let rec dests i blocks =
if i <= num_dests then
let v = Llvm.operand instr i in
let blk = Llvm.block_of_value v in
let jmp, blocks = xlate_jump x instr blk loc blocks in
let rest, blocks = dests (i + 1) blocks in
(jmp :: rest, blocks)
else ([], blocks)
in
dests 1 []
in
let tbl = IArray.of_list lldests in
emit_term (Llair.Term.iswitch ~ptr ~tbl ~loc) ~blocks
| LandingPad ->
(* Translate the landingpad clauses to code to load the type_info from
the thrown exception, and test the type_info against the clauses,
eventually jumping to the handler code following the landingpad,
passing a value for the selector which the handler code tests to
e.g. either cleanup or rethrow. *)
let i32, tip, cxa_exception = landingpad_typs x instr in
let pi8, _, exc_typ = exception_typs in
let exc = Exp.reg (Reg.program pi8 (find_name instr ^ ".exc")) in
let ti = Reg.program tip (name ^ ".ti") in
(* std::type_info* ti = ((__cxa_exception* )exc - 1)->exceptionType *)
let load_ti =
let typ = cxa_exception in
(* field number of the exceptionType member of __cxa_exception *)
let fld = 0 in
(* index from exc that points to header *)
let idx = Exp.integer Typ.siz Z.minus_one in
let ptr =
ptr_fld x
~ptr:(ptr_idx x ~ptr:exc ~idx ~llelt:typ)
~fld ~lltyp:typ
in
let len = Exp.integer Typ.siz (Z.of_int (size_of x typ)) in
Llair.Inst.load ~reg:ti ~ptr ~len ~loc
in
let ti = Exp.reg ti in
let typeid = xlate_llvm_eh_typeid_for x tip ti in
let lbl = name ^ ".unwind" in
let reg = xlate_name x instr in
let jump_unwind i sel rev_blocks =
let exp = Exp.record exc_typ (IArray.of_array [|exc; sel|]) in
let mov =
Llair.Inst.move ~reg_exps:(IArray.of_array [|(reg, exp)|]) ~loc
in
let lbl_i = lbl ^ "." ^ Int.to_string i in
let blk =
Llair.Block.mk ~lbl:lbl_i
~cmnd:(IArray.of_array [|mov|])
~term:(Llair.Term.goto ~dst:(Llair.Jump.mk lbl) ~loc)
in
(Llair.Jump.mk lbl_i, blk :: rev_blocks)
in
let goto_unwind i sel blocks =
let dst, blocks = jump_unwind i sel blocks in
(Llair.Term.goto ~dst ~loc, blocks)
in
let term_unwind, rev_blocks =
if Llvm.is_cleanup instr then
goto_unwind 0 (Exp.integer i32 Z.zero) []
else
let num_clauses = Llvm.num_operands instr in
let lbl i = name ^ "." ^ Int.to_string i in
let jump i = Llair.Jump.mk (lbl i) in
let block i term =
Llair.Block.mk ~lbl:(lbl i) ~cmnd:IArray.empty ~term
in
let match_filter i rev_blocks =
jump_unwind i
(Exp.sub ~typ:i32 (Exp.integer i32 Z.zero) typeid)
rev_blocks
in
let xlate_clause i rev_blocks =
let clause = Llvm.operand instr i in
let num_tis = Llvm.num_operands clause in
if num_tis = 0 then
let dst, rev_blocks = match_filter i rev_blocks in
(Llair.Term.goto ~dst ~loc, rev_blocks)
else
match Llvm.classify_type (Llvm.type_of clause) with
| Array (* filter *) -> (
match Llvm.classify_value clause with
| ConstantArray ->
let rec xlate_filter i =
let tiI = xlate_value x (Llvm.operand clause i) in
if i < num_tis - 1 then
Exp.and_ ~typ:Typ.bool (Exp.dq ~typ:tip tiI ti)
(xlate_filter (i + 1))
else Exp.dq ~typ:tip tiI ti
in
let key = xlate_filter 0 in
let nzero, rev_blocks = match_filter i rev_blocks in
( Llair.Term.branch ~loc ~key ~nzero ~zero:(jump (i + 1))
, rev_blocks )
| _ -> fail "xlate_instr: %a" pp_llvalue instr () )
| _ (* catch *) ->
let typ = xlate_type x (Llvm.type_of clause) in
let clause = xlate_value x clause in
let key =
Exp.or_ ~typ:Typ.bool
(Exp.eq ~typ clause Exp.null)
(Exp.eq ~typ clause ti)
in
let nzero, rev_blocks = jump_unwind i typeid rev_blocks in
( Llair.Term.branch ~loc ~key ~nzero ~zero:(jump (i + 1))
, rev_blocks )
in
let rec rev_blocks i z =
if i < num_clauses then
let term, z = xlate_clause i z in
rev_blocks (i + 1) (block i term :: z)
else block i Llair.Term.unreachable :: z
in
xlate_clause 0 (rev_blocks 1 [])
in
continue (fun (insts, term) ->
( [load_ti]
, term_unwind
, List.rev_append rev_blocks
[Llair.Block.mk ~lbl ~cmnd:(IArray.of_list insts) ~term] ) )
| Resume ->
let llrcd = Llvm.operand instr 0 in
let typ = xlate_type x (Llvm.type_of llrcd) in
let rcd = xlate_value x llrcd in
let exc = Exp.select typ rcd 0 in
emit_term ~prefix:(pop loc) (Llair.Term.throw ~exc ~loc)
| Unreachable -> emit_term Llair.Term.unreachable
| Trunc | ZExt | SExt | FPToUI | FPToSI | UIToFP | SIToFP | FPTrunc
|FPExt | PtrToInt | IntToPtr | BitCast | AddrSpaceCast | Add | FAdd
|Sub | FSub | Mul | FMul | UDiv | SDiv | FDiv | URem | SRem | FRem
|Shl | LShr | AShr | And | Or | Xor | ICmp | FCmp | Select
|GetElementPtr | ExtractElement | InsertElement | ShuffleVector
|ExtractValue | InsertValue ->
inline_or_move (xlate_value ~inline:true x)
| VAArg ->
let reg = xlate_name_opt x instr in
warn "variadic function argument: %a" Loc.pp loc () ;
emit_inst (Llair.Inst.nondet ~reg ~msg:"vaarg" ~loc)
| CleanupRet | CatchRet | CatchPad | CleanupPad | CatchSwitch ->
todo "windows exception handling: %a" pp_llvalue instr ()
| Fence | AtomicCmpXchg | AtomicRMW ->
fail "xlate_instr: %a" pp_llvalue instr ()
| PHI | Invalid | Invalid2 | UserOp1 | UserOp2 -> assert false
let skip_phis : Llvm.llbasicblock -> _ Llvm.llpos =
fun blk ->
let rec skip_phis_ (pos : _ Llvm.llpos) =
match pos with
| Before instr -> (
match Llvm.instr_opcode instr with
| PHI -> skip_phis_ (Llvm.instr_succ instr)
| _ -> pos )
| _ -> pos
in
skip_phis_ (Llvm.instr_begin blk)
let rec xlate_instrs : pop_thunk -> x -> _ Llvm.llpos -> code =
fun pop x -> function
| Before instrI ->
xlate_instr pop x instrI (fun xlate_instrI ->
let instrJ = Llvm.instr_succ instrI in
let instsJ, termJ, blocksJN = xlate_instrs pop x instrJ in
let instsI, termI, blocksI = xlate_instrI (instsJ, termJ) in
(instsI, termI, blocksI @ blocksJN) )
| At_end blk -> fail "xlate_instrs: %a" pp_llblock blk ()
let xlate_block : pop_thunk -> x -> Llvm.llbasicblock -> Llair.block list =
fun pop x blk ->
[%Trace.call fun {pf} -> pf "%a" pp_llblock blk]
;
let lbl = label_of_block blk in
let pos = skip_phis blk in
let insts, term, blocks = xlate_instrs pop x pos in
Llair.Block.mk ~lbl ~cmnd:(IArray.of_list insts) ~term :: blocks
|>
[%Trace.retn fun {pf} blocks -> pf "%s" (List.hd_exn blocks).Llair.lbl]
let report_undefined func name =
if Option.is_some (Llvm.use_begin func) then
[%Trace.info "undefined function: %a" Global.pp name]
let xlate_function : x -> Llvm.llvalue -> Llair.func =
fun x llf ->
[%Trace.call fun {pf} -> pf "%a" pp_llvalue llf]
;
let name = xlate_global x llf in
let formals =
Llvm.fold_left_params
(fun rev_args param -> xlate_name x param :: rev_args)
[] llf
in
let freturn =
match Reg.typ name.reg with
| Pointer {elt= Function {return= Some typ; _}} ->
Some (Reg.program typ "freturn")
| _ -> None
in
let _, _, exc_typ = exception_typs in
let fthrow = Reg.program exc_typ "fthrow" in
( match Llvm.block_begin llf with
| Before entry_blk ->
let pop = pop_stack_frame_of_function x llf entry_blk in
let[@warning "p"] (entry_block :: entry_blocks) =
xlate_block pop x entry_blk
in
let entry =
let {Llair.lbl; cmnd; term} = entry_block in
Llair.Block.mk ~lbl ~cmnd ~term
in
let cfg =
let rec trav_blocks rev_cfg prev =
match Llvm.block_succ prev with
| Before blk ->
trav_blocks
(List.rev_append (xlate_block pop x blk) rev_cfg)
blk
| At_end _ -> IArray.of_list_rev rev_cfg
in
trav_blocks (List.rev entry_blocks) entry_blk
in
Llair.Func.mk ~name ~formals ~freturn ~fthrow ~entry ~cfg
| At_end _ ->
report_undefined llf name ;
Llair.Func.mk_undefined ~name ~formals ~freturn ~fthrow )
|>
[%Trace.retn fun {pf} -> pf "@\n%a" Llair.Func.pp]
let transform ~internalize : Llvm.llmodule -> unit =
fun llmodule ->
let pm = Llvm.PassManager.create () in
let entry_points = Config.find_list "entry-points" in
if internalize then
Llvm_ipo.add_internalize_predicate pm (fun fn ->
List.exists entry_points ~f:(String.equal fn) ) ;
Llvm_ipo.add_global_dce pm ;
[sledge] Add globalopt pass to remove globals Summary: This adds a globalopt optimization pass to sledge. Consider code like: ``` const char *a_string = "I'm a string"; int an_int = 0; int c() { return an_int; } int main() { char *c1 = a_string; return c(); } ``` When compiled there are 2 levels of indirection. For example `return an_int` Get's compiled as ``` %0 = load i32, i32* an_int1 ret i32 %0 ``` Global opt reduces this (if `an_int` is internal) to just ` ret i32 0`. Similarly and more importantly `c1 = a_string;` get's compiled into ``` @.str = private unnamed_addr constant [13 x i8] c"I'm a string\00" a_string = dso_local global i8* getelementptr inbounds ([13 x i8], [13 x i8]* @.str, i32 0, i32 0) %c1 = alloca i8*, align 8 %0 = load i8*, i8** a_string, align 8, !dbg !25 store i8* %0, i8** %c1, align 8, !dbg !24 ``` So there is a level of indirection between `c1` and `.str` where the string is stored. With global opt, this gets reduced to: ``` @.str = private unnamed_addr constant [13 x i8] c"I'm a string\00" %c1 = alloca i8*, align 8 store i8* getelementptr inbounds ([13 x i8], [13 x i8]* @.str, i64 0, i64 0), i8** %c1, align 8, !dbg !23 ``` and `a_string` variable gets deleted. On sledge this has the effect of reducing the complexity of the symbolic heap significantly. Without this optimisation, running `sledge.dbg llvm analyze -trace Domain.call global_vars.bc` Gives prints the following segments: ``` ∧ %.str -[)-> ⟨13,{}⟩ * %a_string -[)-> ⟨8,%.str⟩ * %an_int -[)-> ⟨4,0⟩ * %c1 -[)-> ⟨8,%.str⟩ * %retval -[)-> ⟨4,0⟩ ``` So there are `an_int` and `a_string` segments, which are redundant. with the optimisation, the heap looks like: `∧ %.str -[)-> ⟨13,{}⟩ * %c1 -[)-> ⟨8,%.str⟩ * %retval -[)-> ⟨4,0⟩`, Where we only have the `.str` segment and the `c1` segment, which are the two we need. Reviewed By: ngorogiannis Differential Revision: D15649195 fbshipit-source-id: 5f71e56e8
6 years ago
Llvm_ipo.add_global_optimizer pm ;
Llvm_ipo.add_merge_functions pm ;
[sledge] Add LLVM passes that reduce bitcode size Summary: : This patch adds several passes that reduce the amount of bitcode making sledge's job easier, more info: https://llvm.org/docs/Passes.html `-mergefunc` This pass merges functions that do the same thing, this can be because of templating or casts (ie. same functionality but on 32bit and 64bit ints, which is the same in machine code). More details at http://llvm.org/docs/MergeFunctions.html Note that this pass is currently not available through C/OCaml API. `-constmerge` This merges constants that have the same value, this is possible to do when the constants are internalized. `-argpromotion` ``` This pass promotes “by reference” arguments to be “by value” arguments. In practice, this means looking for internal functions that have pointer arguments. If it can prove, through the use of alias analysis, that an argument is only loaded, then it can pass the value into the function instead of the address of the value. This can cause recursive simplification of code and lead to the elimination of allocas (especially in C++ template code like the STL). ``` `-ipsccp` ``` Sparse conditional constant propagation and merging, which can be summarized as: Assumes values are constant unless proven otherwise Assumes BasicBlocks are dead unless proven otherwise Proves values to be constant, and replaces them with constants Proves conditional branches to be unconditional ``` `-deadargelim` Removes dead arguments of internal functions, good to run after other inter-procedural passes. Seems to crash llvm if run directly after `ipsccp`. Note that while this might look like doing full link-time optimisation, we are actually picking relatively cheap optimisations that mostly look at globals and walk their use chains. The main reason link-time optimisations are expensive is due to inlining and then running the full optimisation again from there. Reviewed By: jberdine Differential Revision: D15851408 fbshipit-source-id: be7191683
6 years ago
Llvm_ipo.add_constant_merge pm ;
Llvm_ipo.add_argument_promotion pm ;
Llvm_ipo.add_ipsccp pm ;
Llvm_scalar_opts.add_memory_to_register_promotion pm ;
Llvm_scalar_opts.add_dce pm ;
[sledge] Add LLVM passes that reduce bitcode size Summary: : This patch adds several passes that reduce the amount of bitcode making sledge's job easier, more info: https://llvm.org/docs/Passes.html `-mergefunc` This pass merges functions that do the same thing, this can be because of templating or casts (ie. same functionality but on 32bit and 64bit ints, which is the same in machine code). More details at http://llvm.org/docs/MergeFunctions.html Note that this pass is currently not available through C/OCaml API. `-constmerge` This merges constants that have the same value, this is possible to do when the constants are internalized. `-argpromotion` ``` This pass promotes “by reference” arguments to be “by value” arguments. In practice, this means looking for internal functions that have pointer arguments. If it can prove, through the use of alias analysis, that an argument is only loaded, then it can pass the value into the function instead of the address of the value. This can cause recursive simplification of code and lead to the elimination of allocas (especially in C++ template code like the STL). ``` `-ipsccp` ``` Sparse conditional constant propagation and merging, which can be summarized as: Assumes values are constant unless proven otherwise Assumes BasicBlocks are dead unless proven otherwise Proves values to be constant, and replaces them with constants Proves conditional branches to be unconditional ``` `-deadargelim` Removes dead arguments of internal functions, good to run after other inter-procedural passes. Seems to crash llvm if run directly after `ipsccp`. Note that while this might look like doing full link-time optimisation, we are actually picking relatively cheap optimisations that mostly look at globals and walk their use chains. The main reason link-time optimisations are expensive is due to inlining and then running the full optimisation again from there. Reviewed By: jberdine Differential Revision: D15851408 fbshipit-source-id: be7191683
6 years ago
Llvm_ipo.add_global_dce pm ;
Llvm_ipo.add_dead_arg_elimination pm ;
Llvm_scalar_opts.add_lower_atomic pm ;
Llvm_scalar_opts.add_scalar_repl_aggregation pm ;
Llvm_scalar_opts.add_scalarizer pm ;
Llvm_scalar_opts.add_unify_function_exit_nodes pm ;
Llvm_scalar_opts.add_cfg_simplification pm ;
Llvm.PassManager.run_module llmodule pm |> (ignore : bool -> _) ;
Llvm.PassManager.dispose pm
let read_and_parse llcontext bc_file =
[%Trace.call fun {pf} -> pf "%s" bc_file]
;
let llmemorybuffer =
try Llvm.MemoryBuffer.of_file bc_file
with Llvm.IoError msg -> fail "%s: %s" bc_file msg ()
in
( try Llvm_irreader.parse_ir llcontext llmemorybuffer
with Llvm_irreader.Error msg -> invalid_llvm msg )
|>
[%Trace.retn fun {pf} _ -> pf ""]
let link_in : Llvm.llcontext -> Llvm.lllinker -> string -> unit =
fun llcontext link_ctx bc_file ->
Llvm_linker.link_in link_ctx (read_and_parse llcontext bc_file)
let check_datalayout llcontext lldatalayout =
let check_size llt typ =
let llsiz =
Int64.to_int_exn (Llvm_target.DataLayout.abi_size llt lldatalayout)
in
let siz = Typ.size_of typ in
if llsiz != siz then
todo "size_of %a = %i != %i" Typ.pp typ llsiz siz ()
in
check_size (Llvm.i1_type llcontext) Typ.bool ;
check_size (Llvm.i8_type llcontext) Typ.byt ;
check_size (Llvm.i32_type llcontext) Typ.int ;
check_size (Llvm.i64_type llcontext) Typ.siz ;
check_size
(Llvm_target.DataLayout.intptr_type llcontext lldatalayout)
Typ.ptr
(* The Llvm.dispose_ functions free memory allocated off the OCaml heap. The
OCaml heap can later grow into that memory once it is freed. There are
naked pointers into the LLVM-allocated memory from various values
returned from Llvm functions. If the GC scans a block with such a naked
pointer after the heap has grown into the memory previously allocated by
Llvm, the GC will follow the pointer expecting a well-formed OCaml value,
and likely segfault. Therefore it is necessary to ensure that all the
values containing naked pointers are dead (which is the reason for
clearing the hashtbls) and then collected (which is the reason for the
Gc.full_major) before freeing the memory with Llvm.dispose_module and
Llvm.dispose_context. *)
let cleanup llmodule llcontext =
Hashtbl.clear sym_tbl ;
Hashtbl.clear scope_tbl ;
Hashtbl.clear anon_struct_name ;
Hashtbl.clear memo_type ;
Hashtbl.clear memo_global ;
Hashtbl.clear memo_value ;
Hash_set.clear ignored_callees ;
Gc.full_major () ;
Llvm.dispose_module llmodule ;
Llvm.dispose_context llcontext
let translate ~models ~fuzzer ~internalize : string list -> Llair.t =
fun inputs ->
[%Trace.call fun {pf} ->
pf "%a" (List.pp "@ " Format.pp_print_string) inputs]
;
Llvm.install_fatal_error_handler invalid_llvm ;
let llcontext = Llvm.global_context () in
let input, inputs = List.pop_exn inputs in
let llmodule = read_and_parse llcontext input in
let link_ctx = Llvm_linker.get_linker llmodule in
List.iter ~f:(link_in llcontext link_ctx) inputs ;
let link_model_file name =
Llvm_linker.link_in link_ctx
(Llvm_irreader.parse_ir llcontext
(Llvm.MemoryBuffer.of_string (Option.value_exn (Model.read name))))
in
if models then link_model_file "/cxxabi.bc" ;
if fuzzer then link_model_file "/lib_fuzzer_main.bc" ;
Llvm_linker.linker_dispose link_ctx ;
assert (
Llvm_analysis.verify_module llmodule |> Option.for_all ~f:invalid_llvm
) ;
transform ~internalize llmodule ;
scan_names_and_locs llmodule ;
let lldatalayout =
Llvm_target.DataLayout.of_string (Llvm.data_layout llmodule)
in
check_datalayout llcontext lldatalayout ;
let x = {llcontext; lldatalayout} in
let globals =
Llvm.fold_left_globals
(fun globals llg ->
if
Poly.equal (Llvm.linkage llg) Appending
&& Llvm.(array_length (element_type (type_of llg))) = 0
then globals
else xlate_global x llg :: globals )
[] llmodule
in
let functions =
Llvm.fold_left_functions
(fun functions llf ->
let name = Llvm.value_name llf in
if
String.is_prefix name ~prefix:"__llair_"
|| String.is_prefix name ~prefix:"llvm."
then functions
else xlate_function x llf :: functions )
[] llmodule
in
cleanup llmodule llcontext ;
Llair.mk ~globals ~functions
|>
[%Trace.retn fun {pf} _ ->
pf "number of globals %d, number of functions %d" (List.length globals)
(List.length functions)]