Auto merge of #78646 - tgnottingham:packed_fingerprints, r=nnethercote

Use PackedFingerprint in DepNode to reduce memory consumption

Author: bors

compiler/rustc_data_structures/src/fingerprint.rs

@@ -151,8 +151,67 @@ impl<D: rustc_serialize::Decoder> FingerprintDecoder for D {
         panic!("Cannot decode `Fingerprint` with `{}`", std::any::type_name::<D>());
     }
 }
+
 impl FingerprintDecoder for opaque::Decoder<'_> {
     fn decode_fingerprint(&mut self) -> Result<Fingerprint, String> {
         Fingerprint::decode_opaque(self)
     }
 }
+
+// `PackedFingerprint` wraps a `Fingerprint`. Its purpose is to, on certain
+// architectures, behave like a `Fingerprint` without alignment requirements.
+// This behavior is only enabled on x86 and x86_64, where the impact of
+// unaligned accesses is tolerable in small doses.
+//
+// This may be preferable to use in large collections of structs containing
+// fingerprints, as it can reduce memory consumption by preventing the padding
+// that the more strictly-aligned `Fingerprint` can introduce. An application of
+// this is in the query dependency graph, which contains a large collection of
+// `DepNode`s. As of this writing, the size of a `DepNode` decreases by ~30%
+// (from 24 bytes to 17) by using the packed representation here, which
+// noticeably decreases total memory usage when compiling large crates.
+//
+// The wrapped `Fingerprint` is private to reduce the chance of a client
+// invoking undefined behavior by taking a reference to the packed field.
+#[cfg_attr(any(target_arch = "x86", target_arch = "x86_64"), repr(packed))]
+#[derive(Eq, PartialEq, Ord, PartialOrd, Debug, Clone, Copy, Hash)]
+pub struct PackedFingerprint(Fingerprint);
+
+impl std::fmt::Display for PackedFingerprint {
+    #[inline]
+    fn fmt(&self, formatter: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        // Copy to avoid taking reference to packed field.
+        let copy = self.0;
+        copy.fmt(formatter)
+    }
+}
+
+impl<E: rustc_serialize::Encoder> Encodable<E> for PackedFingerprint {
+    #[inline]
+    fn encode(&self, s: &mut E) -> Result<(), E::Error> {
+        // Copy to avoid taking reference to packed field.
+        let copy = self.0;
+        copy.encode(s)
+    }
+}
+
+impl<D: rustc_serialize::Decoder> Decodable<D> for PackedFingerprint {
+    #[inline]
+    fn decode(d: &mut D) -> Result<Self, D::Error> {
+        Fingerprint::decode(d).map(|f| PackedFingerprint(f))
+    }
+}
+
+impl From<Fingerprint> for PackedFingerprint {
+    #[inline]
+    fn from(f: Fingerprint) -> PackedFingerprint {
+        PackedFingerprint(f)
+    }
+}
+
+impl From<PackedFingerprint> for Fingerprint {
+    #[inline]
+    fn from(f: PackedFingerprint) -> Fingerprint {
+        f.0
+    }
+}

compiler/rustc_data_structures/src/lib.rs

@@ -31,6 +31,7 @@
 #![feature(once_cell)]
 #![feature(maybe_uninit_uninit_array)]
 #![allow(rustc::default_hash_types)]
+#![deny(unaligned_references)]
 
 #[macro_use]
 extern crate tracing;

compiler/rustc_middle/src/dep_graph/dep_node.rs

@@ -193,6 +193,15 @@ macro_rules! define_dep_nodes {
 
         pub type DepNode = rustc_query_system::dep_graph::DepNode<DepKind>;
 
+        // We keep a lot of `DepNode`s in memory during compilation. It's not
+        // required that their size stay the same, but we don't want to change
+        // it inadvertently. This assert just ensures we're aware of any change.
+        #[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
+        static_assert_size!(DepNode, 17);
+
+        #[cfg(not(any(target_arch = "x86", target_arch = "x86_64")))]
+        static_assert_size!(DepNode, 24);
+
         pub trait DepNodeExt: Sized {
             /// Construct a DepNode from the given DepKind and DefPathHash. This
             /// method will assert that the given DepKind actually requires a
@@ -227,7 +236,7 @@ macro_rules! define_dep_nodes {
                 debug_assert!(kind.can_reconstruct_query_key() && kind.has_params());
                 DepNode {
                     kind,
-                    hash: def_path_hash.0,
+                    hash: def_path_hash.0.into(),
                 }
             }
 
@@ -243,7 +252,7 @@ macro_rules! define_dep_nodes {
             /// has been removed.
             fn extract_def_id(&self, tcx: TyCtxt<'tcx>) -> Option<DefId> {
                 if self.kind.can_reconstruct_query_key() {
-                    let def_path_hash = DefPathHash(self.hash);
+                    let def_path_hash = DefPathHash(self.hash.into());
                     tcx.def_path_hash_to_def_id.as_ref()?.get(&def_path_hash).cloned()
                 } else {
                     None

compiler/rustc_query_system/src/dep_graph/dep_node.rs

@@ -44,7 +44,7 @@
 
 use super::{DepContext, DepKind};
 
-use rustc_data_structures::fingerprint::Fingerprint;
+use rustc_data_structures::fingerprint::{Fingerprint, PackedFingerprint};
 use rustc_data_structures::stable_hasher::{HashStable, StableHasher};
 
 use std::fmt;
@@ -53,7 +53,7 @@ use std::hash::Hash;
 #[derive(Clone, Copy, PartialEq, Eq, PartialOrd, Ord, Hash, Encodable, Decodable)]
 pub struct DepNode<K> {
     pub kind: K,
-    pub hash: Fingerprint,
+    pub hash: PackedFingerprint,
 }
 
 impl<K: DepKind> DepNode<K> {
@@ -62,7 +62,7 @@ impl<K: DepKind> DepNode<K> {
     /// does not require any parameters.
     pub fn new_no_params(kind: K) -> DepNode<K> {
         debug_assert!(!kind.has_params());
-        DepNode { kind, hash: Fingerprint::ZERO }
+        DepNode { kind, hash: Fingerprint::ZERO.into() }
     }
 
     pub fn construct<Ctxt, Key>(tcx: Ctxt, kind: K, arg: &Key) -> DepNode<K>
@@ -71,7 +71,7 @@ impl<K: DepKind> DepNode<K> {
         Key: DepNodeParams<Ctxt>,
     {
         let hash = arg.to_fingerprint(tcx);
-        let dep_node = DepNode { kind, hash };
+        let dep_node = DepNode { kind, hash: hash.into() };
 
         #[cfg(debug_assertions)]
         {

compiler/rustc_query_system/src/dep_graph/graph.rs

@@ -976,7 +976,7 @@ impl<K: DepKind> CurrentDepGraph<K> {
             // Fingerprint::combine() is faster than sending Fingerprint
             // through the StableHasher (at least as long as StableHasher
             // is so slow).
-            hash: self.anon_id_seed.combine(hasher.finish()),
+            hash: self.anon_id_seed.combine(hasher.finish()).into(),
         };
 
         self.intern_node(target_dep_node, task_deps.reads, Fingerprint::ZERO)