Auto merge of #1269 - RalfJung:windows-dtor, r=RalfJung

run Windows TLS dtor function

Mostly fixes https://github.com/rust-lang/miri/issues/442

src/eval.rs

@@ -184,10 +184,9 @@ pub fn create_ecx<'mir, 'tcx: 'mir>(
 /// Returns `Some(return_code)` if program executed completed.
 /// Returns `None` if an evaluation error occured.
 pub fn eval_main<'tcx>(tcx: TyCtxt<'tcx>, main_id: DefId, config: MiriConfig) -> Option<i64> {
-    // FIXME: We always ignore leaks on some OSs where we do not
+    // FIXME: on Windows, locks and TLS dtor management allocate and leave that memory in `static`s.
-    // correctly implement TLS destructors.
+    // So we need https://github.com/rust-lang/miri/issues/940 to fix the leaks there.
-    let target_os = &tcx.sess.target.target.target_os;
+    let ignore_leaks = config.ignore_leaks || tcx.sess.target.target.target_os == "windows";
-    let ignore_leaks = config.ignore_leaks || target_os == "windows";
 
     let (mut ecx, ret_place) = match create_ecx(tcx, main_id, config) {
         Ok(v) => v,

src/helpers.rs

@@ -413,7 +413,7 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriEvalContextExt<'mir, 'tcx
     fn set_last_error_from_io_error(&mut self, e: std::io::Error) -> InterpResult<'tcx> {
         use std::io::ErrorKind::*;
         let this = self.eval_context_mut();
-        let target = &this.tcx.tcx.sess.target.target;
+        let target = &this.tcx.sess.target.target;
         let last_error = if target.options.target_family == Some("unix".to_owned()) {
             this.eval_libc(match e.kind() {
                 ConnectionRefused => "ECONNREFUSED",

src/machine.rs

@@ -102,6 +102,20 @@ impl MemoryExtra {
         }
     }
 
+    fn add_extern_static<'tcx, 'mir>(
+        this: &mut MiriEvalContext<'mir, 'tcx>,
+        name: &str,
+        ptr: Scalar<Tag>,
+    ) {
+        let ptr = ptr.assert_ptr();
+        assert_eq!(ptr.offset, Size::ZERO);
+        this.memory
+            .extra
+            .extern_statics
+            .insert(Symbol::intern(name), ptr.alloc_id)
+            .unwrap_none();
+    }
+
     /// Sets up the "extern statics" for this machine.
     pub fn init_extern_statics<'tcx, 'mir>(
         this: &mut MiriEvalContext<'mir, 'tcx>,
@@ -113,17 +127,17 @@ impl MemoryExtra {
                 let layout = this.layout_of(this.tcx.types.usize)?;
                 let place = this.allocate(layout, MiriMemoryKind::Machine.into());
                 this.write_scalar(Scalar::from_machine_usize(0, &*this.tcx), place.into())?;
-                this.memory
+                Self::add_extern_static(this, "__cxa_thread_atexit_impl", place.ptr);
-                    .extra
-                    .extern_statics
-                    .insert(Symbol::intern("__cxa_thread_atexit_impl"), place.ptr.assert_ptr().alloc_id)
-                    .unwrap_none();
                 // "environ"
-                this.memory
+                Self::add_extern_static(this, "environ", this.machine.env_vars.environ.unwrap().ptr);
-                    .extra
+            }
-                    .extern_statics
+            "windows" => {
-                    .insert(Symbol::intern("environ"), this.machine.env_vars.environ.unwrap().ptr.assert_ptr().alloc_id)
+                // "_tls_used"
-                    .unwrap_none();
+                // This is some obscure hack that is part of the Windows TLS story. It's a `u8`.
+                let layout = this.layout_of(this.tcx.types.u8)?;
+                let place = this.allocate(layout, MiriMemoryKind::Machine.into());
+                this.write_scalar(Scalar::from_u8(0), place.into())?;
+                Self::add_extern_static(this, "_tls_used", place.ptr);
             }
             _ => {} // No "extern statics" supported on this target
         }

src/shims/foreign_items.rs

@@ -19,7 +19,7 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriEvalContextExt<'mir, 'tcx
     fn min_align(&self, size: u64, kind: MiriMemoryKind) -> Align {
         let this = self.eval_context_ref();
         // List taken from `libstd/sys_common/alloc.rs`.
-        let min_align = match this.tcx.tcx.sess.target.target.arch.as_str() {
+        let min_align = match this.tcx.sess.target.target.arch.as_str() {
             "x86" | "arm" | "mips" | "powerpc" | "powerpc64" | "asmjs" | "wasm32" => 8,
             "x86_64" | "aarch64" | "mips64" | "s390x" | "sparc64" => 16,
             arch => bug!("Unsupported target architecture: {}", arch),
@@ -124,7 +124,7 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriEvalContextExt<'mir, 'tcx
         };
         // Strip linker suffixes (seen on 32-bit macOS).
         let link_name = link_name.trim_end_matches("$UNIX2003");
-        let tcx = &{ this.tcx.tcx };
+        let tcx = this.tcx.tcx;
 
         // First: functions that diverge.
         let (dest, ret) = match ret {
@@ -133,8 +133,8 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriEvalContextExt<'mir, 'tcx
                 // The implementation is provided by the function with the `#[panic_handler]` attribute.
                 "panic_impl" => {
                     this.check_panic_supported()?;
-                    let panic_impl_id = this.tcx.lang_items().panic_impl().unwrap();
+                    let panic_impl_id = tcx.lang_items().panic_impl().unwrap();
-                    let panic_impl_instance = ty::Instance::mono(*this.tcx, panic_impl_id);
+                    let panic_impl_instance = ty::Instance::mono(tcx, panic_impl_id);
                     return Ok(Some(&*this.load_mir(panic_impl_instance.def, None)?));
                 }
                 | "exit"

src/shims/foreign_items/posix.rs

@@ -17,7 +17,6 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriEvalContextExt<'mir, 'tcx
         ret: mir::BasicBlock,
     ) -> InterpResult<'tcx, bool> {
         let this = self.eval_context_mut();
-        let tcx = &{ this.tcx.tcx };
 
         match link_name {
             // Environment related shims
@@ -65,7 +64,7 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriEvalContextExt<'mir, 'tcx
             "write" => {
                 let fd = this.read_scalar(args[0])?.to_i32()?;
                 let buf = this.read_scalar(args[1])?.not_undef()?;
-                let n = this.read_scalar(args[2])?.to_machine_usize(tcx)?;
+                let n = this.read_scalar(args[2])?.to_machine_usize(this)?;
                 trace!("Called write({:?}, {:?}, {:?})", fd, buf, n);
                 let result = if fd == 1 || fd == 2 {
                     // stdout/stderr
@@ -209,7 +208,7 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriEvalContextExt<'mir, 'tcx
             }
             "pthread_getspecific" => {
                 let key = this.force_bits(this.read_scalar(args[0])?.not_undef()?, args[0].layout.size)?;
-                let ptr = this.machine.tls.load_tls(key, tcx)?;
+                let ptr = this.machine.tls.load_tls(key, this)?;
                 this.write_scalar(ptr, dest)?;
             }
             "pthread_setspecific" => {

src/shims/foreign_items/windows.rs

@@ -13,7 +13,6 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriEvalContextExt<'mir, 'tcx
         _ret: mir::BasicBlock,
     ) -> InterpResult<'tcx, bool> {
         let this = self.eval_context_mut();
-        let tcx = &{ this.tcx.tcx };
 
         match link_name {
             // Windows API stubs.
@@ -160,7 +159,7 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriEvalContextExt<'mir, 'tcx
             }
             "TlsGetValue" => {
                 let key = u128::from(this.read_scalar(args[0])?.to_u32()?);
-                let ptr = this.machine.tls.load_tls(key, tcx)?;
+                let ptr = this.machine.tls.load_tls(key, this)?;
                 this.write_scalar(ptr, dest)?;
             }
             "TlsSetValue" => {

src/shims/intrinsics.rs

@@ -24,13 +24,12 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriEvalContextExt<'mir, 'tcx
         if this.emulate_intrinsic(span, instance, args, ret)? {
             return Ok(());
         }
-        let tcx = &{ this.tcx.tcx };
         let substs = instance.substs;
 
         // All these intrinsics take raw pointers, so if we access memory directly
         // (as opposed to through a place), we have to remember to erase any tag
         // that might still hang around!
-        let intrinsic_name = &*tcx.item_name(instance.def_id()).as_str();
+        let intrinsic_name = &*this.tcx.item_name(instance.def_id()).as_str();
 
         // First handle intrinsics without return place.
         let (dest, ret) = match ret {

src/shims/tls.rs

@@ -70,7 +70,7 @@ impl<'tcx> TlsData<'tcx> {
     }
 
     pub fn load_tls(
-        &mut self,
+        &self,
         key: TlsKey,
         cx: &impl HasDataLayout,
     ) -> InterpResult<'tcx, Scalar<Tag>> {
@@ -107,7 +107,8 @@ impl<'tcx> TlsData<'tcx> {
         Ok(())
     }
 
-    /// Returns a dtor, its argument and its index, if one is supposed to run
+    /// Returns a dtor, its argument and its index, if one is supposed to run.
+    /// `key` is the last dtors that was run; we return the *next* one after that.
     ///
     /// An optional destructor function may be associated with each key value.
     /// At thread exit, if a key value has a non-NULL destructor pointer,
@@ -158,6 +159,32 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriEvalContextExt<'mir, 'tcx
         assert!(!this.machine.tls.dtors_running, "running TLS dtors twice");
         this.machine.tls.dtors_running = true;
 
+        if this.tcx.sess.target.target.target_os == "windows" {
+            // Windows has a special magic linker section that is run on certain events.
+            // Instead of searching for that section and supporting arbitrary hooks in there
+            // (that would be basically https://github.com/rust-lang/miri/issues/450),
+            // we specifically look up the static in libstd that we know is placed
+            // in that section.
+            let thread_callback = this.eval_path_scalar(&["std", "sys", "windows", "thread_local", "p_thread_callback"])?;
+            let thread_callback = this.memory.get_fn(thread_callback.not_undef()?)?.as_instance()?;
+
+            // The signature of this function is `unsafe extern "system" fn(h: c::LPVOID, dwReason: c::DWORD, pv: c::LPVOID)`.
+            let reason = this.eval_path_scalar(&["std", "sys", "windows", "c", "DLL_PROCESS_DETACH"])?;
+            let ret_place = MPlaceTy::dangling(this.layout_of(this.tcx.mk_unit())?, this).into();
+            this.call_function(
+                thread_callback,
+                &[Scalar::ptr_null(this).into(), reason.into(), Scalar::ptr_null(this).into()],
+                Some(ret_place),
+                StackPopCleanup::None { cleanup: true },
+            )?;
+
+            // step until out of stackframes
+            this.run()?;
+
+            // Windows doesn't have other destructors.
+            return Ok(());
+        }
+
         // The macOS global dtor runs "before any TLS slots get freed", so do that first.
         if let Some((instance, data)) = this.machine.tls.global_dtor {
             trace!("Running global dtor {:?} on {:?}", instance, data);
@@ -191,12 +218,13 @@ pub trait EvalContextExt<'mir, 'tcx: 'mir>: crate::MiriEvalContextExt<'mir, 'tcx
             // step until out of stackframes
             this.run()?;
 
+            // Fetch next dtor after `key`.
             dtor = match this.machine.tls.fetch_tls_dtor(Some(key)) {
                 dtor @ Some(_) => dtor,
+                // We ran each dtor once, start over from the beginning.
                 None => this.machine.tls.fetch_tls_dtor(None),
             };
         }
-        // FIXME: On a windows target, call `unsafe extern "system" fn on_tls_callback`.
         Ok(())
     }
 }