Enable back-edge CFI by default on macOS (#4720)

Also, adjust the tests that are executed on that platform. Finally,
fix a bug with obtaining backtraces when back-edge CFI is enabled.

Copyright (c) 2022, Arm Limited.

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

@@ -780,6 +780,11 @@
        (Pacisp
         (key APIKey))
 
+       ;; Strip pointer authentication code from instruction address in LR;
+       ;; equivalent to a no-op if Pointer authentication (FEAT_PAuth) is not
+       ;; supported.
+       (Xpaclri)
+
        ;; Marker, no-op in generated code: SP "virtual offset" is adjusted. This
        ;; controls how AMode::NominalSPOffset args are lowered.
        (VirtualSPOffsetAdj
@@ -1356,6 +1361,9 @@
 ))
 
 ;; Extractors for target features ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+(decl pure sign_return_address_disabled () Unit)
+(extern constructor sign_return_address_disabled sign_return_address_disabled)
+
 (decl use_lse () Inst)
 (extern extractor use_lse use_lse)
 
@@ -2577,4 +2585,11 @@
 
 (decl aarch64_link () Reg)
 (rule (aarch64_link)
+      (if (sign_return_address_disabled))
       (mov_preg (preg_link)))
+
+(rule (aarch64_link)
+      ;; This constructor is always used for non-leaf functions, so it is safe
+      ;; to clobber LR.
+      (let ((_ Unit (emit (MInst.Xpaclri))))
+           (mov_preg (preg_link))))

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

@@ -3121,6 +3121,7 @@ impl MachInstEmit for Inst {
 
                 sink.put4(0xd503233f | key << 6);
             }
+            &Inst::Xpaclri => sink.put4(0xd50320ff),
             &Inst::VirtualSPOffsetAdj { offset } => {
                 trace!(
                     "virtual sp offset adjusted by {} -> {}",

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

@@ -57,6 +57,7 @@ fn test_aarch64_binemit() {
         "retab",
     ));
     insns.push((Inst::Pacisp { key: APIKey::B }, "7F2303D5", "pacibsp"));
+    insns.push((Inst::Xpaclri, "FF2003D5", "xpaclri"));
     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

@@ -981,12 +981,7 @@ fn aarch64_get_operands<F: Fn(VReg) -> VReg>(inst: &Inst, collector: &mut Operan
             collector.reg_def(rd);
             collector.reg_use(rn);
         }
-        &Inst::Ret { ref rets } => {
+        &Inst::Ret { ref rets } | &Inst::AuthenticatedRet { ref rets, .. } => {
-            for &ret in rets {
-                collector.reg_use(ret);
-            }
-        }
-        &Inst::AuthenticatedRet { ref rets, .. } => {
             for &ret in rets {
                 collector.reg_use(ret);
             }
@@ -1039,7 +1034,10 @@ fn aarch64_get_operands<F: Fn(VReg) -> VReg>(inst: &Inst, collector: &mut Operan
             collector.reg_def(rd);
             memarg_operands(mem, collector);
         }
-        &Inst::Pacisp { .. } => {}
+        &Inst::Pacisp { .. } | &Inst::Xpaclri => {
+            // Neither LR nor SP is an allocatable register, so there is no need
+            // to do anything.
+        }
         &Inst::VirtualSPOffsetAdj { .. } => {}
 
         &Inst::ElfTlsGetAddr { .. } => {
@@ -2715,6 +2713,7 @@ impl Inst {
 
                 "paci".to_string() + key + "sp"
             }
+            &Inst::Xpaclri => "xpaclri".to_string(),
             &Inst::VirtualSPOffsetAdj { offset } => {
                 state.virtual_sp_offset += offset;
                 format!("virtual_sp_offset_adjust {}", offset)

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

@@ -71,6 +71,14 @@ pub struct SinkableAtomicLoad {
 impl generated_code::Context for IsleContext<'_, '_, MInst, Flags, IsaFlags, 6> {
     isle_prelude_methods!();
 
+    fn sign_return_address_disabled(&mut self) -> Option<()> {
+        if self.isa_flags.sign_return_address() {
+            None
+        } else {
+            Some(())
+        }
+    }
+
     fn use_lse(&mut self, _: Inst) -> Option<()> {
         if self.isa_flags.use_lse() {
             Some(())

cranelift/native/src/lib.rs

@@ -138,6 +138,9 @@ pub fn builder_with_options(infer_native_flags: bool) -> Result<isa::Builder, &'
         }
 
         if cfg!(target_os = "macos") {
+            // Pointer authentication is always available on Apple Silicon.
+            isa_builder.enable("sign_return_address").unwrap();
+            // macOS enforces the use of the B key for return addresses.
             isa_builder.enable("sign_return_address_with_bkey").unwrap();
         }
     }

crates/fiber/src/lib.rs

@@ -234,6 +234,8 @@ mod tests {
                 .any(|s| s.contains("look_for_me"))
                 // TODO: apparently windows unwind routines don't unwind through fibers, so this will always fail. Is there a way we can fix that?
                 || cfg!(windows)
+                // TODO: the system libunwind is broken (#2808)
+                || cfg!(all(target_os = "macos", target_arch = "aarch64"))
             );
         }
 

crates/fiber/src/unix/aarch64.rs

@@ -22,17 +22,15 @@ use wasmtime_asm_macros::asm_func;
 
 cfg_if::cfg_if! {
     if #[cfg(target_os = "macos")] {
-        macro_rules! cfi_window_save { () => (""); }
+        macro_rules! paci1716 { () => ("pacib1716\n"); }
-        macro_rules! pacia1716 { () => (""); }
+        macro_rules! pacisp { () => ("pacibsp\n"); }
-        macro_rules! paciasp { () => (""); }
+        macro_rules! autisp { () => ("autibsp\n"); }
-        macro_rules! autiasp { () => (""); }
         macro_rules! sym_adrp { ($s:tt) => (concat!("_", $s, "@PAGE")); }
         macro_rules! sym_add { ($s:tt) => (concat!("_", $s, "@PAGEOFF")); }
     } else {
-        macro_rules! cfi_window_save { () => (".cfi_window_save\n"); }
+        macro_rules! paci1716 { () => ("pacia1716\n"); }
-        macro_rules! pacia1716 { () => ("pacia1716\n"); }
+        macro_rules! pacisp { () => ("paciasp\n"); }
-        macro_rules! paciasp { () => ("paciasp\n"); }
+        macro_rules! autisp { () => ("autiasp\n"); }
-        macro_rules! autiasp { () => ("autiasp\n"); }
         macro_rules! sym_adrp { ($s:tt) => (concat!($s, "")); }
         macro_rules! sym_add { ($s:tt) => (concat!(":lo12:", $s)); }
     }
@@ -44,9 +42,9 @@ asm_func!(
     "
         .cfi_startproc
     ",
-    paciasp!(),
+    pacisp!(),
-    cfi_window_save!(),
     "
+        .cfi_window_save
         // Save all callee-saved registers on the stack since we're
         // assuming they're clobbered as a result of the stack switch.
         stp x29, x30, [sp, -16]!
@@ -81,9 +79,9 @@ asm_func!(
         ldp x20, x19, [sp], 16
         ldp x29, x30, [sp], 16
     ",
-    autiasp!(),
+    autisp!(),
-    cfi_window_save!(),
     "
+        .cfi_window_save
         ret
         .cfi_endproc
     ",
@@ -121,7 +119,7 @@ asm_func!(
         adrp x17, ", sym_adrp!("wasmtime_fiber_start"), "
         add x17, x17, ", sym_add!("wasmtime_fiber_start"), "
     ",
-    pacia1716!(),
+    paci1716!(),
     "
         str x17, [x16, -0x8] // x17 => lr
         str x0, [x16, -0x18] // x0 => x19
@@ -151,9 +149,7 @@ asm_func!(
             0x23, 0xa0, 0x1  /* DW_OP_plus_uconst 0xa0 */
         .cfi_rel_offset x29, -0x10
         .cfi_rel_offset x30, -0x08
-    ",
+        .cfi_window_save
-    cfi_window_save!(),
-    "
         .cfi_rel_offset x19, -0x18
         .cfi_rel_offset x20, -0x20
         .cfi_rel_offset x21, -0x28

crates/fuzzing/src/generators/codegen_settings.rs

@@ -128,6 +128,16 @@ impl<'a> Arbitrary<'a> for CodegenSettings {
                     test: is_aarch64_feature_detected,
 
                     std: "lse" => clif: "has_lse",
+                    // even though the natural correspondence seems to be
+                    // between "paca" and "has_pauth", the latter has no effect
+                    // in isolation, so we actually use the setting that affects
+                    // code generation
+                    std: "paca" => clif: "sign_return_address",
+                    // "paca" and "pacg" check for the same underlying
+                    // architectural feature, so we use the latter to cover more
+                    // code generation settings, of which we have chosen the one
+                    // with the most significant effect
+                    std: "pacg" => clif: "sign_return_address_all" ratio: 1 in 2,
                 },
             };
             return Ok(CodegenSettings::Target {