Auto merge of #66013 - nnethercote:avoid-hashing-twice-in-get_query, r=<try>

Avoid hashing the key twice in `get_query()`.

For a single-threaded parallel compiler, this reduces instruction counts
across several benchmarks, by up to 2.8%.

The commit also adds documentation about `Sharded`'s use of `FxHasher`.

r? @Zoxc

Author: bors

src/librustc/lib.rs

@@ -61,6 +61,7 @@
 #![feature(log_syntax)]
 #![feature(associated_type_bounds)]
 #![feature(rustc_attrs)]
+#![feature(hash_raw_entry)]
 
 #![recursion_limit="512"]
 

src/librustc/ty/query/plumbing.rs

@@ -14,12 +14,13 @@ use errors::Level;
 use errors::Diagnostic;
 use errors::FatalError;
 use errors::Handler;
-use rustc_data_structures::fx::{FxHashMap};
+use rustc_data_structures::fx::{FxHasher, FxHashMap};
 use rustc_data_structures::sync::{Lrc, Lock};
 use rustc_data_structures::sharded::Sharded;
 use rustc_data_structures::thin_vec::ThinVec;
 #[cfg(not(parallel_compiler))]
 use rustc_data_structures::cold_path;
+use std::hash::{Hash, Hasher};
 use std::mem;
 use std::ptr;
 use std::collections::hash_map::Entry;
@@ -80,8 +81,17 @@ impl<'a, 'tcx, Q: QueryDescription<'tcx>> JobOwner<'a, 'tcx, Q> {
     pub(super) fn try_get(tcx: TyCtxt<'tcx>, span: Span, key: &Q::Key) -> TryGetJob<'a, 'tcx, Q> {
         let cache = Q::query_cache(tcx);
         loop {
-            let mut lock = cache.get_shard_by_value(key).lock();
-            if let Some(value) = lock.results.get(key) {
+            // We compute the key's hash once and then use it for both the
+            // shard lookup and the hashmap lookup. This relies on the fact
+            // that both of them use `FxHasher`.
+            let mut state = FxHasher::default();
+            key.hash(&mut state);
+            let key_hash = state.finish();
+
+            let mut lock = cache.get_shard_by_hash(key_hash).lock();
+            if let Some((_, value)) =
+                lock.results.raw_entry().from_key_hashed_nocheck(key_hash, key)
+            {
                 tcx.prof.query_cache_hit(Q::NAME);
                 let result = (value.value.clone(), value.index);
                 #[cfg(debug_assertions)]

src/librustc_data_structures/sharded.rs

@@ -22,6 +22,7 @@ const SHARD_BITS: usize = 0;
 pub const SHARDS: usize = 1 << SHARD_BITS;
 
 /// An array of cache-line aligned inner locked structures with convenience methods.
+/// Shards are selected by hashing values with `FxHasher`.
 #[derive(Clone)]
 pub struct Sharded<T> {
     shards: [CacheAligned<Lock<T>>; SHARDS],
@@ -69,6 +70,8 @@ impl<T> Sharded<T> {
         }
     }
 
+    /// Get a shard with a pre-computed hash value, which must be computed by
+    /// `FxHasher`.
     #[inline]
     pub fn get_shard_by_hash(&self, hash: u64) -> &Lock<T> {
         let hash_len = mem::size_of::<usize>();