Auto merge of rust-lang#119302 - Mark-Simulacrum:relative-spans, r=WaffleLapkin

Support encoding spans with relative offsets

The relative offset is often smaller than the absolute offset, and with
the LEB128 encoding, this ends up cutting the overall metadata size
considerably (~1.5 megabytes on libcore). We can support both relative
and absolute encodings essentially for free since we already take a full
byte to differentiate between direct and indirect encodings (so an extra
variant is quite cheap).

Author: bors

compiler/rustc_metadata/src/rmeta/decoder.rs

@@ -507,10 +507,18 @@ impl<'a, 'tcx> Decodable<DecodeContext<'a, 'tcx>> for ExpnId {
 
 impl<'a, 'tcx> Decodable<DecodeContext<'a, 'tcx>> for Span {
     fn decode(decoder: &mut DecodeContext<'a, 'tcx>) -> Span {
+        let start = decoder.position();
         let mode = SpanEncodingMode::decode(decoder);
         let data = match mode {
             SpanEncodingMode::Direct => SpanData::decode(decoder),
-            SpanEncodingMode::Shorthand(position) => decoder.with_position(position, |decoder| {
+            SpanEncodingMode::RelativeOffset(offset) => {
+                decoder.with_position(start - offset, |decoder| {
+                    let mode = SpanEncodingMode::decode(decoder);
+                    debug_assert!(matches!(mode, SpanEncodingMode::Direct));
+                    SpanData::decode(decoder)
+                })
+            }
+            SpanEncodingMode::AbsoluteOffset(addr) => decoder.with_position(addr, |decoder| {
                 let mode = SpanEncodingMode::decode(decoder);
                 debug_assert!(matches!(mode, SpanEncodingMode::Direct));
                 SpanData::decode(decoder)

compiler/rustc_metadata/src/rmeta/encoder.rs

@@ -169,7 +169,19 @@ impl<'a, 'tcx> Encodable<EncodeContext<'a, 'tcx>> for ExpnId {
 impl<'a, 'tcx> Encodable<EncodeContext<'a, 'tcx>> for Span {
     fn encode(&self, s: &mut EncodeContext<'a, 'tcx>) {
         match s.span_shorthands.entry(*self) {
-            Entry::Occupied(o) => SpanEncodingMode::Shorthand(*o.get()).encode(s),
+            Entry::Occupied(o) => {
+                // If an offset is smaller than the absolute position, we encode with the offset.
+                // This saves space since smaller numbers encode in less bits.
+                let last_location = *o.get();
+                // This cannot underflow. Metadata is written with increasing position(), so any
+                // previously saved offset must be smaller than the current position.
+                let offset = s.opaque.position() - last_location;
+                if offset < last_location {
+                    SpanEncodingMode::RelativeOffset(offset).encode(s)
+                } else {
+                    SpanEncodingMode::AbsoluteOffset(last_location).encode(s)
+                }
+            }
             Entry::Vacant(v) => {
                 let position = s.opaque.position();
                 v.insert(position);

compiler/rustc_metadata/src/rmeta/mod.rs

@@ -68,7 +68,8 @@ pub const METADATA_HEADER: &[u8] = &[b'r', b'u', b's', b't', 0, 0, 0, METADATA_V
 
 #[derive(Encodable, Decodable)]
 enum SpanEncodingMode {
-    Shorthand(usize),
+    RelativeOffset(usize),
+    AbsoluteOffset(usize),
     Direct,
 }