Combine sort keys into u64/u128

This allows the compiler to perform branch-less comparisons, which are
more efficient.

This results in ~5% fewer instructions executed.

src/ion/data_structures.rs

@@ -543,3 +543,17 @@ pub struct Stats {
     pub halfmoves_count: usize,
     pub edits_count: usize,
 }
+
+// Helper function for generating sorting keys. The order of arguments is from
+// the most significant field to the least significant one.
+//
+// These work best when the fields are stored in reverse order in memory so that
+// they can be loaded with a single u64 load.
+#[inline(always)]
+pub fn u64_key(b: u32, a: u32) -> u64 {
+    a as u64 | (b as u64) << 32
+}
+#[inline(always)]
+pub fn u128_key(d: u32, c: u32, b: u32, a: u32) -> u128 {
+    a as u128 | (b as u128) << 32 | (c as u128) << 64 | (d as u128) << 96
+}

src/ion/liveranges.rs

@@ -18,7 +18,7 @@ use super::{
     SpillSetIndex, Use, VRegData, VRegIndex, SLOT_NONE,
 };
 use crate::indexset::IndexSet;
-use crate::ion::data_structures::MultiFixedRegFixup;
+use crate::ion::data_structures::{u128_key, MultiFixedRegFixup};
 use crate::{
     Allocation, Block, Function, Inst, InstPosition, Operand, OperandConstraint, OperandKind,
     OperandPos, PReg, ProgPoint, RegAllocError, VReg,
@@ -1141,8 +1141,24 @@ impl<'a, F: Function> Env<'a, F> {
             }
         }
 
-        self.blockparam_ins.sort_unstable();
+        self.blockparam_ins
-        self.blockparam_outs.sort_unstable();
+            .sort_unstable_by_key(|(to_vreg, to_block, from_block)| {
+                u128_key(
+                    to_vreg.raw_u32(),
+                    to_block.raw_u32(),
+                    from_block.raw_u32(),
+                    0,
+                )
+            });
+        self.blockparam_outs
+            .sort_unstable_by_key(|(from_vreg, from_block, to_block, to_vreg)| {
+                u128_key(
+                    from_vreg.raw_u32(),
+                    from_block.raw_u32(),
+                    to_block.raw_u32(),
+                    to_vreg.raw_u32(),
+                )
+            });
         self.prog_move_srcs.sort_unstable_by_key(|(pos, _)| *pos);
         self.prog_move_dsts.sort_unstable_by_key(|(pos, _)| *pos);
 

src/ion/moves.rs

@@ -17,6 +17,7 @@ use super::{
     VRegIndex, SLOT_NONE,
 };
 
+use crate::ion::data_structures::u64_key;
 use crate::moves::ParallelMoves;
 use crate::{
     Allocation, Block, Edit, Function, Inst, InstPosition, OperandConstraint, OperandKind,
@@ -850,7 +851,7 @@ impl<'a, F: Function> Env<'a, F> {
         // resolve (see cases below).
         let mut i = 0;
         self.inserted_moves
-            .sort_unstable_by_key(|m| (m.pos.to_index(), m.prio));
+            .sort_unstable_by_key(|m| u64_key(m.pos.to_index(), m.prio as u32));
 
         // Redundant-move elimination state tracker.
         let mut redundant_moves = RedundantMoveEliminator::default();
@@ -1104,7 +1105,7 @@ impl<'a, F: Function> Env<'a, F> {
         // Add edits to describe blockparam locations too. This is
         // required by the checker. This comes after any edge-moves.
         self.blockparam_allocs
-            .sort_unstable_by_key(|&(block, idx, _, _)| (block, idx));
+            .sort_unstable_by_key(|&(block, idx, _, _)| u64_key(block.raw_u32(), idx));
         self.stats.blockparam_allocs_count = self.blockparam_allocs.len();
         let mut i = 0;
         while i < self.blockparam_allocs.len() {
@@ -1137,7 +1138,8 @@ impl<'a, F: Function> Env<'a, F> {
         // be a stable sort! We have to keep the order produced by the
         // parallel-move resolver for all moves within a single sort
         // key.
-        self.edits.sort_by_key(|&(pos, prio, _)| (pos, prio));
+        self.edits
+            .sort_by_key(|&(pos, prio, _)| u64_key(pos, prio as u32));
         self.stats.edits_count = self.edits.len();
 
         // Add debug annotations.

src/ion/stackmap.rs

@@ -13,7 +13,7 @@
 //! Stackmap computation.
 
 use super::{Env, ProgPoint, VRegIndex};
-use crate::Function;
+use crate::{ion::data_structures::u64_key, Function};
 
 impl<'a, F: Function> Env<'a, F> {
     pub fn compute_stackmaps(&mut self) {
@@ -64,7 +64,8 @@ impl<'a, F: Function> Env<'a, F> {
             }
         }
 
-        self.safepoint_slots.sort_unstable();
+        self.safepoint_slots
+            .sort_unstable_by_key(|(progpoint, slot)| u64_key(progpoint.to_index(), slot.bits()));
         log::trace!("final safepoint slots info: {:?}", self.safepoint_slots);
     }
 }

src/moves.rs

@@ -3,7 +3,7 @@
  * exception. See `LICENSE` for details.
  */
 
-use crate::Allocation;
+use crate::{ion::data_structures::u64_key, Allocation};
 use smallvec::{smallvec, SmallVec};
 
 pub type MoveVec<T> = SmallVec<[(Allocation, Allocation, T); 16]>;
@@ -53,7 +53,8 @@ impl<T: Clone + Copy + Default> ParallelMoves<T> {
 
         // Sort moves by source so that we can efficiently test for
         // presence.
-        self.parallel_moves.sort_by_key(|&(src, dst, _)| (src, dst));
+        self.parallel_moves
+            .sort_by_key(|&(src, dst, _)| u64_key(src.bits(), dst.bits()));
 
         // Do any dests overlap sources? If not, we can also just
         // return the list.