Initial forward-edge CFI implementation (#3693)

* Initial forward-edge CFI implementation

Give the user the option to start all basic blocks that are targets
of indirect branches with the BTI instruction introduced by the
Branch Target Identification extension to the Arm instruction set
architecture.

Copyright (c) 2022, Arm Limited.

* Refactor `from_artifacts` to avoid second `make_executable` (#1)

This involves "parsing" twice but this is parsing just the header of an
ELF file so it's not a very intensive operation and should be ok to do
twice.

* Address the code review feedback

Copyright (c) 2022, Arm Limited.

Co-authored-by: Alex Crichton <alex@alexcrichton.com>

cranelift/codegen/meta/src/isa/arm64.rs

@@ -5,13 +5,13 @@ use crate::shared::Definitions as SharedDefinitions;
 
 fn define_settings(_shared: &SettingGroup) -> SettingGroup {
     let mut setting = SettingGroupBuilder::new("arm64");
-    let has_lse = setting.add_bool(
+
+    setting.add_bool(
         "has_lse",
         "Has Large System Extensions (FEAT_LSE) support.",
         "",
         false,
     );
-
     setting.add_bool(
         "has_pauth",
         "Has Pointer authentication (FEAT_PAuth) support; enables the use of \
@@ -44,8 +44,13 @@ fn define_settings(_shared: &SettingGroup) -> SettingGroup {
         "",
         false,
     );
+    setting.add_bool(
+        "use_bti",
+        "Use Branch Target Identification (FEAT_BTI) instructions.",
+        "",
+        false,
+    );
 
-    setting.add_predicate("use_lse", predicate!(has_lse));
     setting.build()
 }
 

cranelift/codegen/src/alias_analysis.rs

@@ -237,7 +237,7 @@ impl<'a> AliasAnalysis<'a> {
                 trace!("after inst{}: state is {:?}", inst.index(), state);
             }
 
-            visit_block_succs(self.func, block, |_inst, succ| {
+            visit_block_succs(self.func, block, |_inst, succ, _from_table| {
                 let succ_first_inst = self
                     .func
                     .layout

cranelift/codegen/src/inst_predicates.rs

@@ -129,8 +129,15 @@ pub fn has_memory_fence_semantics(op: Opcode) -> bool {
     }
 }
 
-/// Visit all successors of a block with a given visitor closure.
+/// Visit all successors of a block with a given visitor closure. The closure
-pub(crate) fn visit_block_succs<F: FnMut(Inst, Block)>(f: &Function, block: Block, mut visit: F) {
+/// arguments are the branch instruction that is used to reach the successor,
+/// the successor block itself, and a flag indicating whether the block is
+/// branched to via a table entry.
+pub(crate) fn visit_block_succs<F: FnMut(Inst, Block, bool)>(
+    f: &Function,
+    block: Block,
+    mut visit: F,
+) {
     for inst in f.layout.block_likely_branches(block) {
         if f.dfg[inst].opcode().is_branch() {
             visit_branch_targets(f, inst, &mut visit);
@@ -138,18 +145,20 @@ pub(crate) fn visit_block_succs<F: FnMut(Inst, Block)>(f: &Function, block: Bloc
     }
 }
 
-fn visit_branch_targets<F: FnMut(Inst, Block)>(f: &Function, inst: Inst, visit: &mut F) {
+fn visit_branch_targets<F: FnMut(Inst, Block, bool)>(f: &Function, inst: Inst, visit: &mut F) {
     match f.dfg[inst].analyze_branch(&f.dfg.value_lists) {
         BranchInfo::NotABranch => {}
         BranchInfo::SingleDest(dest, _) => {
-            visit(inst, dest);
+            visit(inst, dest, false);
         }
         BranchInfo::Table(table, maybe_dest) => {
             if let Some(dest) = maybe_dest {
-                visit(inst, dest);
+                // The default block is reached via a direct conditional branch,
+                // so it is not part of the table.
+                visit(inst, dest, false);
             }
             for &dest in f.jump_tables[table].as_slice() {
-                visit(inst, dest);
+                visit(inst, dest, true);
             }
         }
     }

cranelift/codegen/src/isa/aarch64/abi.rs

@@ -67,7 +67,11 @@ fn saved_reg_stack_size(
 /// point for the trait; it is never actually instantiated.
 pub struct AArch64MachineDeps;
 
-impl IsaFlags for aarch64_settings::Flags {}
+impl IsaFlags for aarch64_settings::Flags {
+    fn is_forward_edge_cfi_enabled(&self) -> bool {
+        self.use_bti()
+    }
+}
 
 impl ABIMachineSpec for AArch64MachineDeps {
     type I = Inst;
@@ -549,13 +553,21 @@ impl ABIMachineSpec for AArch64MachineDeps {
                     },
                 });
             }
-        } else if flags.unwind_info() && call_conv.extends_apple_aarch64() {
+        } else {
-            // The macOS unwinder seems to require this.
+            if isa_flags.use_bti() {
-            insts.push(Inst::Unwind {
+                insts.push(Inst::Bti {
-                inst: UnwindInst::Aarch64SetPointerAuth {
+                    targets: BranchTargetType::C,
-                    return_addresses: false,
+                });
-                },
+            }
-            });
+
+            if flags.unwind_info() && call_conv.extends_apple_aarch64() {
+                // The macOS unwinder seems to require this.
+                insts.push(Inst::Unwind {
+                    inst: UnwindInst::Aarch64SetPointerAuth {
+                        return_addresses: false,
+                    },
+                });
+            }
         }
 
         insts

cranelift/codegen/src/isa/aarch64/inst.isle

@@ -880,6 +880,11 @@
        ;; supported.
        (Xpaclri)
 
+       ;; Branch target identification; equivalent to a no-op if Branch Target
+       ;; Identification (FEAT_BTI) is not supported.
+       (Bti
+        (targets BranchTargetType))
+
        ;; Marker, no-op in generated code: SP "virtual offset" is adjusted. This
        ;; controls how AMode::NominalSPOffset args are lowered.
        (VirtualSPOffsetAdj
@@ -1568,6 +1573,15 @@
     (B)
 ))
 
+;; Branch target types
+(type BranchTargetType
+  (enum
+    (None)
+    (C)
+    (J)
+    (JC)
+))
+
 ;; Extractors for target features ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 (decl pure sign_return_address_disabled () Unit)
 (extern constructor sign_return_address_disabled sign_return_address_disabled)

cranelift/codegen/src/isa/aarch64/inst/emit.rs

@@ -3332,6 +3332,16 @@ impl MachInstEmit for Inst {
                 sink.put4(0xd503233f | key << 6);
             }
             &Inst::Xpaclri => sink.put4(0xd50320ff),
+            &Inst::Bti { targets } => {
+                let targets = match targets {
+                    BranchTargetType::None => 0b00,
+                    BranchTargetType::C => 0b01,
+                    BranchTargetType::J => 0b10,
+                    BranchTargetType::JC => 0b11,
+                };
+
+                sink.put4(0xd503241f | targets << 6);
+            }
             &Inst::VirtualSPOffsetAdj { offset } => {
                 trace!(
                     "virtual sp offset adjusted by {} -> {}",

cranelift/codegen/src/isa/aarch64/inst/emit_tests.rs

@@ -58,6 +58,13 @@ fn test_aarch64_binemit() {
     ));
     insns.push((Inst::Pacisp { key: APIKey::B }, "7F2303D5", "pacibsp"));
     insns.push((Inst::Xpaclri, "FF2003D5", "xpaclri"));
+    insns.push((
+        Inst::Bti {
+            targets: BranchTargetType::J,
+        },
+        "9F2403D5",
+        "bti j",
+    ));
     insns.push((Inst::Nop0, "", "nop-zero-len"));
     insns.push((Inst::Nop4, "1F2003D5", "nop"));
     insns.push((Inst::Csdb, "9F2203D5", "csdb"));

cranelift/codegen/src/isa/aarch64/inst/mod.rs

@@ -36,10 +36,10 @@ mod emit_tests;
 // Instructions (top level): definition
 
 pub use crate::isa::aarch64::lower::isle::generated_code::{
-    ALUOp, ALUOp3, AMode, APIKey, AtomicRMWLoopOp, AtomicRMWOp, BitOp, FPUOp1, FPUOp2, FPUOp3,
+    ALUOp, ALUOp3, AMode, APIKey, AtomicRMWLoopOp, AtomicRMWOp, BitOp, BranchTargetType, FPUOp1,
-    FpuRoundMode, FpuToIntOp, IntToFpuOp, MInst as Inst, MoveWideOp, VecALUModOp, VecALUOp,
+    FPUOp2, FPUOp3, FpuRoundMode, FpuToIntOp, IntToFpuOp, MInst as Inst, MoveWideOp, VecALUModOp,
-    VecExtendOp, VecLanesOp, VecMisc2, VecPairOp, VecRRLongOp, VecRRNarrowOp, VecRRPairLongOp,
+    VecALUOp, VecExtendOp, VecLanesOp, VecMisc2, VecPairOp, VecRRLongOp, VecRRNarrowOp,
-    VecRRRLongModOp, VecRRRLongOp, VecShiftImmModOp, VecShiftImmOp,
+    VecRRPairLongOp, VecRRRLongModOp, VecRRRLongOp, VecShiftImmModOp, VecShiftImmOp,
 };
 
 /// A floating-point unit (FPU) operation with two args, a register and an immediate.
@@ -1072,6 +1072,7 @@ fn aarch64_get_operands<F: Fn(VReg) -> VReg>(inst: &Inst, collector: &mut Operan
             // Neither LR nor SP is an allocatable register, so there is no need
             // to do anything.
         }
+        &Inst::Bti { .. } => {}
         &Inst::VirtualSPOffsetAdj { .. } => {}
 
         &Inst::ElfTlsGetAddr { rd, .. } => {
@@ -1266,6 +1267,19 @@ impl MachInst for Inst {
     fn ref_type_regclass(_: &settings::Flags) -> RegClass {
         RegClass::Int
     }
+
+    fn gen_block_start(
+        is_indirect_branch_target: bool,
+        is_forward_edge_cfi_enabled: bool,
+    ) -> Option<Self> {
+        if is_indirect_branch_target && is_forward_edge_cfi_enabled {
+            Some(Inst::Bti {
+                targets: BranchTargetType::J,
+            })
+        } else {
+            None
+        }
+    }
 }
 
 //=============================================================================
@@ -2700,7 +2714,7 @@ impl Inst {
                         "csel {}, xzr, {}, hs ; ",
                         "csdb ; ",
                         "adr {}, pc+16 ; ",
-                        "ldrsw {}, [{}, {}, LSL 2] ; ",
+                        "ldrsw {}, [{}, {}, uxtw #2] ; ",
                         "add {}, {}, {} ; ",
                         "br {} ; ",
                         "jt_entries {:?}"
@@ -2812,6 +2826,16 @@ impl Inst {
                 "paci".to_string() + key + "sp"
             }
             &Inst::Xpaclri => "xpaclri".to_string(),
+            &Inst::Bti { targets } => {
+                let targets = match targets {
+                    BranchTargetType::None => "",
+                    BranchTargetType::C => " c",
+                    BranchTargetType::J => " j",
+                    BranchTargetType::JC => " jc",
+                };
+
+                "bti".to_string() + targets
+            }
             &Inst::VirtualSPOffsetAdj { offset } => {
                 state.virtual_sp_offset += offset;
                 format!("virtual_sp_offset_adjust {}", offset)

cranelift/codegen/src/isa/aarch64/lower/isle.rs

@@ -86,7 +86,7 @@ impl Context for IsleContext<'_, '_, MInst, Flags, IsaFlags, 6> {
     }
 
     fn use_lse(&mut self, _: Inst) -> Option<()> {
-        if self.isa_flags.use_lse() {
+        if self.isa_flags.has_lse() {
             Some(())
         } else {
             None

cranelift/codegen/src/isa/aarch64/lower_inst.rs

@@ -657,18 +657,20 @@ pub(crate) fn lower_branch(
                 //   emit_island  // this forces an island at this point
                 //                // if the jumptable would push us past
                 //                // the deadline
-                //   subs idx, #jt_size
+                //   cmp idx, #jt_size
                 //   b.hs default
+                //   csel vTmp2, xzr, idx, hs
+                //   csdb
                 //   adr vTmp1, PC+16
-                //   ldr vTmp2, [vTmp1, idx, lsl #2]
+                //   ldr vTmp2, [vTmp1, vTmp2, uxtw #2]
-                //   add vTmp2, vTmp2, vTmp1
+                //   add vTmp1, vTmp1, vTmp2
-                //   br vTmp2
+                //   br vTmp1
                 //   [jumptable offsets relative to JT base]
                 let jt_size = targets.len() - 1;
                 assert!(jt_size <= std::u32::MAX as usize);
 
                 ctx.emit(Inst::EmitIsland {
-                    needed_space: 4 * (6 + jt_size) as CodeOffset,
+                    needed_space: 4 * (8 + jt_size) as CodeOffset,
                 });
 
                 let ridx = put_input_in_reg(
@@ -707,8 +709,10 @@ pub(crate) fn lower_branch(
                 // Emit the compound instruction that does:
                 //
                 // b.hs default
+                // csel rB, xzr, rIndex, hs
+                // csdb
                 // adr rA, jt
-                // ldrsw rB, [rA, rIndex, UXTW 2]
+                // ldrsw rB, [rA, rB, uxtw #2]
                 // add rA, rA, rB
                 // br rA
                 // [jt entries]

cranelift/codegen/src/isa/aarch64/mod.rs

@@ -115,6 +115,10 @@ impl TargetIsa for AArch64Backend {
         self.isa_flags.iter().collect()
     }
 
+    fn is_branch_protection_enabled(&self) -> bool {
+        self.isa_flags.use_bti()
+    }
+
     fn dynamic_vector_bytes(&self, _dyn_ty: Type) -> u32 {
         16
     }

cranelift/codegen/src/isa/mod.rs

@@ -226,6 +226,11 @@ pub trait TargetIsa: fmt::Display + Send + Sync {
     /// Get the ISA-dependent flag values that were used to make this trait object.
     fn isa_flags(&self) -> Vec<settings::Value>;
 
+    /// Get a flag indicating whether branch protection is enabled.
+    fn is_branch_protection_enabled(&self) -> bool {
+        false
+    }
+
     /// Get the ISA-dependent maximum vector register size, in bytes.
     fn dynamic_vector_bytes(&self, dynamic_ty: ir::Type) -> u32;
 

cranelift/codegen/src/machinst/abi.rs

@@ -286,7 +286,12 @@ impl StackAMode {
 }
 
 /// Trait implemented by machine-specific backend to represent ISA flags.
-pub trait IsaFlags: Clone {}
+pub trait IsaFlags: Clone {
+    /// Get a flag indicating whether forward-edge CFI is enabled.
+    fn is_forward_edge_cfi_enabled(&self) -> bool {
+        false
+    }
+}
 
 /// Trait implemented by machine-specific backend to provide information about
 /// register assignments and to allow generating the specific instructions for
@@ -1256,6 +1261,10 @@ impl<M: ABIMachineSpec> Callee<M> {
         }
     }
 
+    pub fn is_forward_edge_cfi_enabled(&self) -> bool {
+        self.isa_flags.is_forward_edge_cfi_enabled()
+    }
+
     /// Get the calling convention implemented by this ABI object.
     pub fn call_conv(&self, sigs: &SigSet) -> isa::CallConv {
         sigs[self.sig].call_conv

cranelift/codegen/src/machinst/blockorder.rs

@@ -106,6 +106,8 @@ pub struct BlockLoweringOrder {
     /// which is used by VCode emission to sink the blocks at the last
     /// moment (when we actually emit bytes into the MachBuffer).
     cold_blocks: FxHashSet<BlockIndex>,
+    /// Lowered blocks that are indirect branch targets.
+    indirect_branch_targets: FxHashSet<BlockIndex>,
 }
 
 /// The origin of a block in the lowered block-order: either an original CLIF
@@ -230,14 +232,20 @@ impl BlockLoweringOrder {
         // Cache the block successors to avoid re-examining branches below.
         let mut block_succs: SmallVec<[(Inst, usize, Block); 128]> = SmallVec::new();
         let mut block_succ_range = SecondaryMap::with_default((0, 0));
+        let mut indirect_branch_target_clif_blocks = FxHashSet::default();
+
         for block in f.layout.blocks() {
             let block_succ_start = block_succs.len();
             let mut succ_idx = 0;
-            visit_block_succs(f, block, |inst, succ| {
+            visit_block_succs(f, block, |inst, succ, from_table| {
                 block_out_count[block] += 1;
                 block_in_count[succ] += 1;
                 block_succs.push((inst, succ_idx, succ));
                 succ_idx += 1;
+
+                if from_table {
+                    indirect_branch_target_clif_blocks.insert(succ);
+                }
             });
             let block_succ_end = block_succs.len();
             block_succ_range[block] = (block_succ_start, block_succ_end);
@@ -432,6 +440,7 @@ impl BlockLoweringOrder {
         let mut cold_blocks = FxHashSet::default();
         let mut lowered_succ_ranges = vec![];
         let mut lb_to_bindex = FxHashMap::default();
+        let mut indirect_branch_targets = FxHashSet::default();
         for (block, succ_range) in rpo.into_iter() {
             let index = BlockIndex::new(lowered_order.len());
             lb_to_bindex.insert(block, index);
@@ -445,11 +454,19 @@ impl BlockLoweringOrder {
                     if f.layout.is_cold(block) {
                         cold_blocks.insert(index);
                     }
+
+                    if indirect_branch_target_clif_blocks.contains(&block) {
+                        indirect_branch_targets.insert(index);
+                    }
                 }
                 LoweredBlock::Edge { pred, succ, .. } => {
                     if f.layout.is_cold(pred) || f.layout.is_cold(succ) {
                         cold_blocks.insert(index);
                     }
+
+                    if indirect_branch_target_clif_blocks.contains(&succ) {
+                        indirect_branch_targets.insert(index);
+                    }
                 }
             }
         }
@@ -474,6 +491,7 @@ impl BlockLoweringOrder {
             lowered_succ_ranges,
             orig_map,
             cold_blocks,
+            indirect_branch_targets,
         };
         trace!("BlockLoweringOrder: {:?}", result);
         result
@@ -494,6 +512,12 @@ impl BlockLoweringOrder {
     pub fn is_cold(&self, block: BlockIndex) -> bool {
         self.cold_blocks.contains(&block)
     }
+
+    /// Determine whether the given lowered block index is an indirect branch
+    /// target.
+    pub fn is_indirect_branch_target(&self, block: BlockIndex) -> bool {
+        self.indirect_branch_targets.contains(&block)
+    }
 }
 
 #[cfg(test)]

cranelift/codegen/src/machinst/mod.rs

@@ -168,6 +168,16 @@ pub trait MachInst: Clone + Debug {
     /// Is this a safepoint?
     fn is_safepoint(&self) -> bool;
 
+    /// Generate an instruction that must appear at the beginning of a basic
+    /// block, if any. Note that the return value must not be subject to
+    /// register allocation.
+    fn gen_block_start(
+        _is_indirect_branch_target: bool,
+        _is_forward_edge_cfi_enabled: bool,
+    ) -> Option<Self> {
+        None
+    }
+
     /// A label-use kind: a type that describes the types of label references that
     /// can occur in an instruction.
     type LabelUse: MachInstLabelUse;

cranelift/codegen/src/machinst/vcode.rs

@@ -845,6 +845,8 @@ impl<I: VCodeInst> VCode<I> {
             ra_edits_per_block.push((end_edit_idx - start_edit_idx) as u32);
         }
 
+        let is_forward_edge_cfi_enabled = self.abi.is_forward_edge_cfi_enabled();
+
         for (block_order_idx, &block) in final_order.iter().enumerate() {
             trace!("emitting block {:?}", block);
             let new_offset = I::align_basic_block(buffer.cur_offset());
@@ -902,6 +904,13 @@ impl<I: VCodeInst> VCode<I> {
                 last_offset = Some(cur_offset);
             }
 
+            if let Some(block_start) = I::gen_block_start(
+                self.block_order.is_indirect_branch_target(block),
+                is_forward_edge_cfi_enabled,
+            ) {
+                do_emit(&block_start, &[], &mut disasm, &mut buffer, &mut state);
+            }
+
             for inst_or_edit in regalloc.block_insts_and_edits(&self, block) {
                 match inst_or_edit {
                     InstOrEdit::Inst(iix) => {

cranelift/filetests/filetests/isa/aarch64/bti.clif

@@ -0,0 +1,111 @@
+test compile precise-output
+set unwind_info=false
+target aarch64 use_bti
+
+function %f1(i32) -> i32 {
+    jt0 = jump_table [block1, block2, block3]
+
+block0(v0: i32):
+    br_table v0, block4, jt0
+
+block1:
+    v1 = iconst.i32 1
+    jump block5(v1)
+
+block2:
+    v2 = iconst.i32 2
+    jump block5(v2)
+
+block3:
+    v3 = iconst.i32 3
+    jump block5(v3)
+
+block4:
+    v4 = iconst.i32 4
+    jump block5(v4)
+
+block5(v5: i32):
+    v6 = iadd.i32 v0, v5
+    return v6
+}
+
+;   bti c
+; block0:
+;   emit_island 44
+;   subs wzr, w0, #3
+;   b.hs label1 ; csel x1, xzr, x0, hs ; csdb ; adr x15, pc+16 ; ldrsw x1, [x15, x1, uxtw #2] ; add x15, x15, x1 ; br x15 ; jt_entries [Label(MachLabel(3)), Label(MachLabel(5)), Label(MachLabel(7))]
+; block1:
+;   movz x5, #4
+;   b label2
+; block2:
+;   b label9
+; block3:
+;   bti j
+;   movz x5, #1
+;   b label4
+; block4:
+;   b label9
+; block5:
+;   bti j
+;   movz x5, #2
+;   b label6
+; block6:
+;   b label9
+; block7:
+;   bti j
+;   movz x5, #3
+;   b label8
+; block8:
+;   b label9
+; block9:
+;   add w0, w0, w5
+;   ret
+
+function %f2(i64) -> i64 {
+    jt0 = jump_table [block2]
+
+block0(v0: i64):
+    v1 = ireduce.i32 v0
+    v2 = load.i64 notrap aligned table v0
+    br_table v1, block1, jt0
+
+block1:
+    return v2
+
+block2:
+    v3 = iconst.i64 42
+    v4 = iadd.i64 v2, v3
+    return v4
+}
+
+;   bti c
+; block0:
+;   ldr x6, [x0]
+;   emit_island 36
+;   subs wzr, w0, #1
+;   b.hs label1 ; csel x8, xzr, x0, hs ; csdb ; adr x7, pc+16 ; ldrsw x8, [x7, x8, uxtw #2] ; add x7, x7, x8 ; br x7 ; jt_entries [Label(MachLabel(2))]
+; block1:
+;   mov x0, x6
+;   ret
+; block2:
+;   bti j
+;   mov x0, x6
+;   add x0, x0, #42
+;   ret
+
+function %f3(i64) -> i64 {
+    fn0 = %g(i64) -> i64
+
+block0(v0: i64):
+    v1 = call fn0(v0)
+    return v1
+}
+
+;   bti c
+;   stp fp, lr, [sp, #-16]!
+;   mov fp, sp
+; block0:
+;   ldr x4, 8 ; b 12 ; data TestCase(%g) + 0
+;   blr x4
+;   ldp fp, lr, [sp], #16
+;   ret

cranelift/filetests/filetests/isa/aarch64/jumptable.clif

@@ -30,9 +30,9 @@ block5(v5: i32):
 }
 
 ; block0:
-;   emit_island 36
+;   emit_island 44
 ;   subs wzr, w0, #3
-;   b.hs label1 ; csel x1, xzr, x0, hs ; csdb ; adr x15, pc+16 ; ldrsw x1, [x15, x1, LSL 2] ; add x15, x15, x1 ; br x15 ; jt_entries [Label(MachLabel(3)), Label(MachLabel(5)), Label(MachLabel(7))]
+;   b.hs label1 ; csel x1, xzr, x0, hs ; csdb ; adr x15, pc+16 ; ldrsw x1, [x15, x1, uxtw #2] ; add x15, x15, x1 ; br x15 ; jt_entries [Label(MachLabel(3)), Label(MachLabel(5)), Label(MachLabel(7))]
 ; block1:
 ;   movz x5, #4
 ;   b label2

cranelift/filetests/filetests/runtests/br_table.clif

@@ -1,6 +1,7 @@
 test interpret
 test run
 target aarch64
+target aarch64 use_bti
 target x86_64
 target s390x
 

cranelift/jit/src/backend.rs

@@ -1,6 +1,6 @@
 //! Defines `JITModule`.
 
-use crate::{compiled_blob::CompiledBlob, memory::Memory};
+use crate::{compiled_blob::CompiledBlob, memory::BranchProtection, memory::Memory};
 use cranelift_codegen::isa::TargetIsa;
 use cranelift_codegen::settings::Configurable;
 use cranelift_codegen::{self, ir, settings, MachReloc};
@@ -480,6 +480,12 @@ impl JITModule {
             );
         }
 
+        let branch_protection =
+            if cfg!(target_arch = "aarch64") && use_bti(&builder.isa.isa_flags()) {
+                BranchProtection::BTI
+            } else {
+                BranchProtection::None
+            };
         let mut module = Self {
             isa: builder.isa,
             hotswap_enabled: builder.hotswap_enabled,
@@ -487,9 +493,10 @@ impl JITModule {
             lookup_symbols: builder.lookup_symbols,
             libcall_names: builder.libcall_names,
             memory: MemoryHandle {
-                code: Memory::new(),
+                code: Memory::new(branch_protection),
-                readonly: Memory::new(),
+                // Branch protection is not applicable to non-executable memory.
-                writable: Memory::new(),
+                readonly: Memory::new(BranchProtection::None),
+                writable: Memory::new(BranchProtection::None),
             },
             declarations: ModuleDeclarations::default(),
             function_got_entries: SecondaryMap::new(),
@@ -959,3 +966,10 @@ fn lookup_with_dlsym(name: &str) -> Option<*const u8> {
         None
     }
 }
+
+fn use_bti(isa_flags: &Vec<settings::Value>) -> bool {
+    isa_flags
+        .iter()
+        .find(|&f| f.name == "use_bti")
+        .map_or(false, |f| f.as_bool().unwrap_or(false))
+}

cranelift/jit/src/memory.rs

@@ -104,6 +104,15 @@ impl Drop for PtrLen {
 
 // TODO: add a `Drop` impl for `cfg(target_os = "windows")`
 
+/// Type of branch protection to apply to executable memory.
+#[derive(Clone, Debug, PartialEq)]
+pub(crate) enum BranchProtection {
+    /// No protection.
+    None,
+    /// Use the Branch Target Identification extension of the Arm architecture.
+    BTI,
+}
+
 /// JIT memory manager. This manages pages of suitably aligned and
 /// accessible memory. Memory will be leaked by default to have
 /// function pointers remain valid for the remainder of the
@@ -113,15 +122,17 @@ pub(crate) struct Memory {
     already_protected: usize,
     current: PtrLen,
     position: usize,
+    branch_protection: BranchProtection,
 }
 
 impl Memory {
-    pub(crate) fn new() -> Self {
+    pub(crate) fn new(branch_protection: BranchProtection) -> Self {
         Self {
             allocations: Vec::new(),
             already_protected: 0,
             current: PtrLen::new(),
             position: 0,
+            branch_protection,
         }
     }
 
@@ -157,14 +168,35 @@ impl Memory {
     pub(crate) fn set_readable_and_executable(&mut self) {
         self.finish_current();
 
+        let set_region_readable_and_executable = |ptr, len| {
+            if len != 0 {
+                if self.branch_protection == BranchProtection::BTI {
+                    #[cfg(all(target_arch = "aarch64", target_os = "linux"))]
+                    if std::arch::is_aarch64_feature_detected!("bti") {
+                        let prot = libc::PROT_EXEC | libc::PROT_READ | /* PROT_BTI */ 0x10;
+
+                        unsafe {
+                            if libc::mprotect(ptr as *mut libc::c_void, len, prot) < 0 {
+                                panic!("unable to make memory readable+executable");
+                            }
+                        }
+
+                        return;
+                    }
+                }
+
+                unsafe {
+                    region::protect(ptr, len, region::Protection::READ_EXECUTE)
+                        .expect("unable to make memory readable+executable");
+                }
+            }
+        };
+
         #[cfg(feature = "selinux-fix")]
         {
             for &PtrLen { ref map, ptr, len } in &self.allocations[self.already_protected..] {
-                if len != 0 && map.is_some() {
+                if map.is_some() {
-                    unsafe {
+                    set_region_readable_and_executable(ptr, len);
-                        region::protect(ptr, len, region::Protection::READ_EXECUTE)
-                            .expect("unable to make memory readable+executable");
-                    }
                 }
             }
         }
@@ -172,12 +204,7 @@ impl Memory {
         #[cfg(not(feature = "selinux-fix"))]
         {
             for &PtrLen { ptr, len } in &self.allocations[self.already_protected..] {
-                if len != 0 {
+                set_region_readable_and_executable(ptr, len);
-                    unsafe {
-                        region::protect(ptr, len, region::Protection::READ_EXECUTE)
-                            .expect("unable to make memory readable+executable");
-                    }
-                }
             }
         }
 

crates/cranelift/src/compiler.rs

@@ -459,6 +459,10 @@ impl wasmtime_environ::Compiler for Compiler {
             .collect()
     }
 
+    fn is_branch_protection_enabled(&self) -> bool {
+        self.isa.is_branch_protection_enabled()
+    }
+
     #[cfg(feature = "component-model")]
     fn component_compiler(&self) -> &dyn wasmtime_environ::component::ComponentCompiler {
         self

crates/environ/src/compilation.rs

@@ -266,6 +266,9 @@ pub trait Compiler: Send + Sync {
     /// Same as [`Compiler::flags`], but ISA-specific (a cranelift-ism)
     fn isa_flags(&self) -> BTreeMap<String, FlagValue>;
 
+    /// Get a flag indicating whether branch protection is enabled.
+    fn is_branch_protection_enabled(&self) -> bool;
+
     /// Returns a suitable compiler usable for component-related compliations.
     ///
     /// Note that the `ComponentCompiler` trait can also be implemented for

crates/fuzzing/src/generators/codegen_settings.rs

@@ -127,6 +127,7 @@ impl<'a> Arbitrary<'a> for CodegenSettings {
                 "aarch64" => {
                     test: is_aarch64_feature_detected,
 
+                    std: "bti" => clif: "use_bti",
                     std: "lse" => clif: "has_lse",
                     // even though the natural correspondence seems to be
                     // between "paca" and "has_pauth", the latter has no effect

crates/jit/src/code_memory.rs

@@ -40,7 +40,7 @@ pub struct Publish<'a> {
     pub obj: File<'a>,
 
     /// Reference to the entire `MmapVec` and its contents.
-    pub mmap: &'a [u8],
+    pub mmap: &'a MmapVec,
 
     /// Reference to just the text section of the object file, a subslice of
     /// `mmap`.
@@ -87,7 +87,7 @@ impl CodeMemory {
     /// After this function executes all JIT code should be ready to execute.
     /// The various parsed results of the internals of the `MmapVec` are
     /// returned through the `Publish` structure.
-    pub fn publish(&mut self) -> Result<Publish<'_>> {
+    pub fn publish(&mut self, enable_branch_protection: bool) -> Result<Publish<'_>> {
         assert!(!self.published);
         self.published = true;
 
@@ -159,7 +159,7 @@ impl CodeMemory {
             // read/execute, notably not using read/write/execute to prevent
             // modifications.
             self.mmap
-                .make_executable(text_range.clone())
+                .make_executable(text_range.clone(), enable_branch_protection)
                 .expect("unable to make memory executable");
 
             #[cfg(all(target_arch = "aarch64", target_os = "linux"))]

crates/jit/src/instantiate.rs

@@ -136,6 +136,9 @@ struct Metadata {
     /// Note that even if this flag is `true` sections may be missing if they
     /// weren't found in the original wasm module itself.
     has_wasm_debuginfo: bool,
+
+    /// Whether or not branch protection is enabled.
+    is_branch_protection_enabled: bool,
 }
 
 /// Finishes compilation of the `translation` specified, producing the final
@@ -160,6 +163,7 @@ pub fn finish_compile(
     funcs: PrimaryMap<DefinedFuncIndex, FunctionInfo>,
     trampolines: Vec<Trampoline>,
     tunables: &Tunables,
+    is_branch_protection_enabled: bool,
 ) -> Result<(MmapVec, CompiledModuleInfo)> {
     let ModuleTranslation {
         mut module,
@@ -265,6 +269,7 @@ pub fn finish_compile(
             has_unparsed_debuginfo,
             code_section_offset: debuginfo.wasm_file.code_section_offset,
             has_wasm_debuginfo: tunables.parse_wasm_debuginfo,
+            is_branch_protection_enabled,
         },
     };
     bincode::serialize_into(&mut bytes, &info)?;
@@ -398,19 +403,10 @@ impl CompiledModule {
         profiler: &dyn ProfilingAgent,
         id_allocator: &CompiledModuleIdAllocator,
     ) -> Result<Self> {
-        // Transfer ownership of `obj` to a `CodeMemory` object which will
+        let obj = File::parse(&mmap[..]).context("failed to parse internal elf file")?;
-        // manage permissions, such as the executable bit. Once it's located
+        let opt_section = |name: &str| obj.section_by_name(name).and_then(|s| s.data().ok());
-        // there we also publish it for being able to execute. Note that this
-        // step will also resolve pending relocations in the compiled image.
-        let mut code_memory = CodeMemory::new(mmap);
-        let code = code_memory
-            .publish()
-            .context("failed to publish code memory")?;
-
         let section = |name: &str| {
-            code.obj
+            opt_section(name)
-                .section_by_name(name)
-                .and_then(|s| s.data().ok())
                 .ok_or_else(|| anyhow!("missing section `{}` in compilation artifacts", name))
         };
 
@@ -422,35 +418,28 @@ impl CompiledModule {
                 .context("failed to deserialize wasmtime module info")?,
         };
 
-        let func_name_data = match code
-            .obj
-            .section_by_name(ELF_NAME_DATA)
-            .and_then(|s| s.data().ok())
-        {
-            Some(data) => subslice_range(data, code.mmap),
-            None => 0..0,
-        };
-
         let mut ret = Self {
             module: Arc::new(info.module),
             funcs: info.funcs,
             trampolines: info.trampolines,
-            wasm_data: subslice_range(section(ELF_WASM_DATA)?, code.mmap),
+            wasm_data: subslice_range(section(ELF_WASM_DATA)?, &mmap),
-            address_map_data: code
+            address_map_data: opt_section(ELF_WASMTIME_ADDRMAP)
-                .obj
+                .map(|slice| subslice_range(slice, &mmap))
-                .section_by_name(ELF_WASMTIME_ADDRMAP)
-                .and_then(|s| s.data().ok())
-                .map(|slice| subslice_range(slice, code.mmap))
                 .unwrap_or(0..0),
-            trap_data: subslice_range(section(ELF_WASMTIME_TRAPS)?, code.mmap),
+            func_name_data: opt_section(ELF_NAME_DATA)
-            code: subslice_range(code.text, code.mmap),
+                .map(|slice| subslice_range(slice, &mmap))
+                .unwrap_or(0..0),
+            trap_data: subslice_range(section(ELF_WASMTIME_TRAPS)?, &mmap),
+            code: subslice_range(section(".text")?, &mmap),
             dbg_jit_registration: None,
-            code_memory,
+            code_memory: CodeMemory::new(mmap),
             meta: info.meta,
             unique_id: id_allocator.alloc(),
             func_names: info.func_names,
-            func_name_data,
         };
+        ret.code_memory
+            .publish(ret.meta.is_branch_protection_enabled)
+            .context("failed to publish code memory")?;
         ret.register_debug_and_profiling(profiler)?;
 
         Ok(ret)

crates/runtime/src/mmap.rs

@@ -412,7 +412,11 @@ impl Mmap {
     }
 
     /// Makes the specified `range` within this `Mmap` to be read/execute.
-    pub unsafe fn make_executable(&self, range: Range<usize>) -> Result<()> {
+    pub unsafe fn make_executable(
+        &self,
+        range: Range<usize>,
+        enable_branch_protection: bool,
+    ) -> Result<()> {
         assert!(range.start <= self.len());
         assert!(range.end <= self.len());
         assert!(range.start <= range.end);
@@ -428,8 +432,15 @@ impl Mmap {
             use std::io;
             use windows_sys::Win32::System::Memory::*;
 
+            let flags = if enable_branch_protection {
+                // TODO: We use this check to avoid an unused variable warning,
+                // but some of the CFG-related flags might be applicable
+                PAGE_EXECUTE_READ
+            } else {
+                PAGE_EXECUTE_READ
+            };
             let mut old = 0;
-            let result = VirtualProtect(base, len, PAGE_EXECUTE_READ, &mut old);
+            let result = VirtualProtect(base, len, flags, &mut old);
             if result == 0 {
                 return Err(io::Error::last_os_error().into());
             }
@@ -438,8 +449,25 @@ impl Mmap {
         #[cfg(not(windows))]
         {
             use rustix::mm::{mprotect, MprotectFlags};
-            mprotect(base, len, MprotectFlags::READ | MprotectFlags::EXEC)?;
+
+            let flags = MprotectFlags::READ | MprotectFlags::EXEC;
+            let flags = if enable_branch_protection {
+                #[cfg(all(target_arch = "aarch64", target_os = "linux"))]
+                if std::arch::is_aarch64_feature_detected!("bti") {
+                    MprotectFlags::from_bits_unchecked(flags.bits() | /* PROT_BTI */ 0x10)
+                } else {
+                    flags
+                }
+
+                #[cfg(not(all(target_arch = "aarch64", target_os = "linux")))]
+                flags
+            } else {
+                flags
+            };
+
+            mprotect(base, len, flags)?;
         }
+
         Ok(())
     }
 

crates/runtime/src/mmap_vec.rs

@@ -102,9 +102,15 @@ impl MmapVec {
     }
 
     /// Makes the specified `range` within this `mmap` to be read/execute.
-    pub unsafe fn make_executable(&self, range: Range<usize>) -> Result<()> {
+    pub unsafe fn make_executable(
-        self.mmap
+        &self,
-            .make_executable(range.start + self.range.start..range.end + self.range.start)
+        range: Range<usize>,
+        enable_branch_protection: bool,
+    ) -> Result<()> {
+        self.mmap.make_executable(
+            range.start + self.range.start..range.end + self.range.start,
+            enable_branch_protection,
+        )
     }
 
     /// Returns the underlying file that this mmap is mapping, if present.

crates/wasmtime/src/component/component.rs

@@ -164,7 +164,7 @@ impl Component {
         let static_modules = static_modules?;
         let (lowerings, always_trap, transcoders, trampolines, trampoline_obj) = trampolines?;
         let mut trampoline_obj = CodeMemory::new(trampoline_obj);
-        let code = trampoline_obj.publish()?;
+        let code = trampoline_obj.publish(engine.compiler().is_branch_protection_enabled())?;
         let text = wasmtime_jit::subslice_range(code.text, code.mmap);
 
         // This map is used to register all known tramplines in the

crates/wasmtime/src/engine.rs

@@ -462,6 +462,9 @@ impl Engine {
                 "sign_return_address" => Some(true),
                 // No effect on its own.
                 "sign_return_address_with_bkey" => Some(true),
+                // The `BTI` instruction acts as a `NOP` when unsupported, so it
+                // is safe to enable it.
+                "use_bti" => Some(true),
                 // fall through to the very bottom to indicate that support is
                 // not enabled to test whether this feature is enabled on the
                 // host.

crates/wasmtime/src/module.rs

@@ -1,7 +1,7 @@
-use crate::Engine;
 use crate::{
     signatures::SignatureCollection,
     types::{ExportType, ExternType, ImportType},
+    Engine,
 };
 use anyhow::{bail, Context, Result};
 use once_cell::sync::OnceCell;
@@ -328,18 +328,15 @@ impl Module {
     ///
     /// This is where compilation actually happens of WebAssembly modules and
     /// translation/parsing/validation of the binary input occurs. The actual
-    /// result here is a triple of:
+    /// result here is a combination of:
     ///
-    /// * The index into the second field of the "main module". The "main
+    /// * The compilation artifacts for the module found within `wasm`, as
-    ///   module" in this case is the outermost module described by the `wasm`
+    ///   returned by `wasmtime_jit::finish_compile`.
-    ///   input, and is here for the module linking proposal.
+    /// * Type information about the module returned. All returned modules have
-    /// * A list of compilation artifacts for each module found within `wasm`.
+    ///   local type information with indices that refer to these returned
-    ///   Note that if module linking is disabled then this list will always
-    ///   have a size of exactly 1. These pairs are returned by
-    ///   `wasmtime_jit::finish_compile`.
-    /// * Type information about all the modules returned. All returned modules
-    ///   have local type information with indices that refer to these returned
     ///   tables.
+    /// * A boolean value indicating whether forward-edge CFI has been applied
+    ///   to the compiled module.
     #[cfg(compiler)]
     pub(crate) fn build_artifacts(
         engine: &Engine,
@@ -431,8 +428,14 @@ impl Module {
         // table lazy init.
         translation.try_func_table_init();
 
-        let (mmap, info) =
+        let (mmap, info) = wasmtime_jit::finish_compile(
-            wasmtime_jit::finish_compile(translation, obj, funcs, trampolines, tunables)?;
+            translation,
+            obj,
+            funcs,
+            trampolines,
+            tunables,
+            engine.compiler().is_branch_protection_enabled(),
+        )?;
 
         Ok((mmap, Some(info)))
     }

crates/wasmtime/src/trampoline/func.rs

@@ -118,7 +118,7 @@ where
     // Copy the results of JIT compilation into executable memory, and this will
     // also take care of unwind table registration.
     let mut code_memory = CodeMemory::new(obj);
-    let code = code_memory.publish()?;
+    let code = code_memory.publish(engine.compiler().is_branch_protection_enabled())?;
 
     register_trampolines(engine.profiler(), &code.obj);