fix: Cache line alignment (#316)

Author: zharinov

crates/plotnik-lib/src/bytecode/dump.rs

@@ -12,6 +12,7 @@ use super::ids::TypeId;
 use super::instructions::StepId;
 use super::ir::NodeTypeIR;
 use super::module::{Instruction, Module};
+use super::nav::Nav;
 use super::type_meta::{TypeData, TypeKind};
 use super::{Call, Match, Return, Trampoline};
 
@@ -364,6 +365,30 @@ fn dump_entrypoints(out: &mut String, module: &Module, ctx: &DumpContext) {
     out.push('\n');
 }
 
+/// Check if an instruction is padding (all-zeros Match8).
+///
+/// Padding slots contain zero bytes which decode as terminal epsilon Match8
+/// with Any node type, no field constraint, and next=0.
+fn is_padding(instr: &Instruction) -> bool {
+    match instr {
+        Instruction::Match(m) => {
+            m.is_match8()
+                && m.nav == Nav::Epsilon
+                && matches!(m.node_type, NodeTypeIR::Any)
+                && m.node_field.is_none()
+                && m.is_terminal()
+        }
+        _ => false,
+    }
+}
+
+/// Format a single padding step line.
+///
+/// Output: `  07  ... ` (step number and " ... " in symbol column)
+fn format_padding_step(step: u16, step_width: usize) -> String {
+    LineBuilder::new(step_width).instruction_prefix(step, Symbol::PADDING)
+}
+
 fn dump_code(out: &mut String, module: &Module, ctx: &DumpContext) {
     let c = &ctx.colors;
     let header = module.header();
@@ -386,6 +411,14 @@ fn dump_code(out: &mut String, module: &Module, ctx: &DumpContext) {
         }
 
         let instr = module.decode_step(step);
+
+        // Check for padding (all-zeros Match8 instruction)
+        if is_padding(&instr) {
+            writeln!(out, "{}", format_padding_step(step, step_width)).unwrap();
+            step += 1;
+            continue;
+        }
+
         let line = format_instruction(step, &instr, module, ctx, step_width);
         out.push_str(&line);
         out.push('\n');

crates/plotnik-lib/src/bytecode/format.rs

@@ -62,6 +62,9 @@ impl Symbol {
     /// Epsilon symbol for unconditional transitions.
     pub const EPSILON: Symbol = Symbol::new("  ", "ε", "  ");
 
+    /// Padding indicator (centered "..." in 5-char column).
+    pub const PADDING: Symbol = Symbol::new(" ", "...", " ");
+
     /// Format as a 5-character string.
     pub fn format(&self) -> String {
         format!("{}{}{}", self.left, self.center, self.right)

crates/plotnik-lib/src/bytecode/instructions.rs

@@ -216,6 +216,12 @@ impl<'a> Match<'a> {
         self.nav == Nav::Epsilon
     }
 
+    /// Check if this is a Match8 (8-byte fast-path instruction).
+    #[inline]
+    pub fn is_match8(&self) -> bool {
+        self.is_match8
+    }
+
     /// Number of successors.
     #[inline]
     pub fn succ_count(&self) -> usize {

crates/plotnik-lib/src/emit/layout.rs

@@ -1,7 +1,7 @@
 //! Cache-aligned instruction layout.
 //!
-//! Uses Pettis-Hansen inspired greedy chain extraction to place
-//! hot paths contiguously and avoid cache line straddling.
+//! Extracts linear chains from the control flow graph and places them
+//! contiguously. Pads instructions to prevent cache line straddling.
 
 use std::collections::{BTreeMap, HashSet};
 
@@ -161,18 +161,23 @@ fn assign_step_ids(
             };
             let size = instr.size();
 
-            // Cache line alignment for large instructions
-            if size >= 48 {
-                let line_offset = current_offset % CACHE_LINE;
-                if line_offset + size > CACHE_LINE {
-                    // Would straddle cache line - pad to next line
-                    let padding_bytes = CACHE_LINE - line_offset;
-                    let padding_steps = (padding_bytes / STEP_SIZE) as u16;
-                    current_step += padding_steps;
-                    current_offset += padding_bytes;
-                }
+            // Pad if instruction would straddle cache line boundary
+            let line_offset = current_offset % CACHE_LINE;
+            if line_offset + size > CACHE_LINE {
+                let padding_bytes = CACHE_LINE - line_offset;
+                let padding_steps = (padding_bytes / STEP_SIZE) as u16;
+                current_step += padding_steps;
+                current_offset += padding_bytes;
             }
 
+            // Invariant: instruction must not straddle cache line
+            assert!(
+                current_offset % CACHE_LINE + size <= CACHE_LINE,
+                "instruction at offset {} with size {} straddles 64-byte cache line",
+                current_offset,
+                size
+            );
+
             mapping.insert(label, current_step);
             let step_count = (size / STEP_SIZE) as u16;
             current_step += step_count;

crates/plotnik-lib/src/emit/layout_tests.rs

@@ -113,45 +113,132 @@ fn layout_branch() {
 }
 
 #[test]
-fn layout_large_instruction_cache_alignment() {
-    // Large instruction (Match48 = 48 bytes = 6 steps) near cache line boundary
-    // Start at step 5 (offset 40), would straddle - should pad
-    let large_match = MatchIR::at(Label(1))
-        .nav(Nav::Down)
-        .node_type(NodeTypeIR::Named(NonZeroU16::new(10)))
-        .pre_effect(EffectIR::start_obj())
-        .pre_effect(EffectIR::start_obj())
-        .pre_effect(EffectIR::start_obj())
-        .post_effect(EffectIR::node())
-        .post_effect(EffectIR::end_obj())
-        .post_effect(EffectIR::end_obj())
-        .post_effect(EffectIR::end_obj())
-        .next_many(vec![
-            Label(100),
-            Label(101),
-            Label(102),
-            Label(103),
-            Label(104),
-            Label(105),
-            Label(106),
-            Label(107),
-        ]);
-
-    // Verify it's large enough to trigger alignment
-    assert!(large_match.size() >= 48);
+fn layout_match16_cache_alignment() {
+    // Match16 (16 bytes, 2 steps) at offset 56 would straddle (56+16=72 > 64)
+    // Place 7 Match8s (7 steps = 56 bytes) before Match16
+    // Expected: Match16 gets padded to step 8 (offset 64)
+    let mut instructions = Vec::new();
+
+    // 7 Match8 instructions in a chain: Label(0) -> Label(1) -> ... -> Label(6) -> Label(7)
+    for i in 0..7 {
+        instructions.push(
+            MatchIR::at(Label(i))
+                .nav(Nav::Down)
+                .next(Label(i + 1))
+                .into(),
+        );
+    }
+
+    // Match16 at Label(7): needs 2+ successors to become Match16
+    instructions.push(
+        MatchIR::at(Label(7))
+            .nav(Nav::Down)
+            .next_many(vec![Label(100), Label(101)])
+            .into(),
+    );
 
-    let instructions = vec![
-        // Small instruction first
-        MatchIR::epsilon(Label(0), Label(1)).into(),
-        large_match.into(),
-    ];
+    let result = CacheAligned::layout(&instructions, &[Label(0)]);
+
+    // Labels 0-6 should be at steps 0-6 (no padding needed)
+    for i in 0..7 {
+        assert_eq!(
+            result.label_to_step.get(&Label(i)),
+            Some(&(i as u16)),
+            "Label({i}) should be at step {i}"
+        );
+    }
+
+    // Label(7) would be at step 7 (offset 56) without padding
+    // But Match16 at offset 56 straddles (56+16=72 > 64), so it must be padded
+    // After padding: step 8 (offset 64)
+    let step7 = *result.label_to_step.get(&Label(7)).unwrap();
+    assert_eq!(step7, 8, "Match16 should be padded to step 8 (offset 64)");
+
+    // Total steps: 8 (padding at step 7) + 2 (Match16) = 10
+    assert_eq!(result.total_steps, 10);
+}
+
+#[test]
+fn layout_match8_no_padding_needed() {
+    // Match8 (8 bytes) never straddles: max offset 56, 56+8=64 <= 64
+    // Place 7 Match8s, then another Match8 - should NOT need padding
+    let mut instructions = Vec::new();
+
+    for i in 0..8 {
+        if i < 7 {
+            instructions.push(
+                MatchIR::at(Label(i))
+                    .nav(Nav::Down)
+                    .next(Label(i + 1))
+                    .into(),
+            );
+        } else {
+            instructions.push(MatchIR::terminal(Label(i)).nav(Nav::Down).into());
+        }
+    }
 
     let result = CacheAligned::layout(&instructions, &[Label(0)]);
 
-    // Label 0 at step 0 (offset 0)
-    assert_eq!(result.label_to_step.get(&Label(0)), Some(&0u16));
+    // All 8 Match8s should be contiguous: steps 0-7
+    for i in 0..8 {
+        assert_eq!(
+            result.label_to_step.get(&Label(i)),
+            Some(&(i as u16)),
+            "Label({i}) should be at step {i} (no padding)"
+        );
+    }
+
+    // Total steps: 8 (no padding)
+    assert_eq!(result.total_steps, 8);
+}
+
+#[test]
+fn layout_match32_cache_alignment() {
+    // Match32 (32 bytes, 4 steps) at offset 40 would straddle (40+32=72 > 64)
+    // Place 5 Match8s (5 steps = 40 bytes) before Match32
+    // Expected: Match32 gets padded to step 8 (offset 64)
+    let mut instructions: Vec<crate::bytecode::InstructionIR> = Vec::new();
+
+    // 5 Match8 instructions: Label(0) -> ... -> Label(4) -> Label(5)
+    for i in 0..5 {
+        instructions.push(
+            MatchIR::at(Label(i))
+                .nav(Nav::Down)
+                .next(Label(i + 1))
+                .into(),
+        );
+    }
+
+    // Match32 at Label(5): needs enough payload to become Match32 (9-12 slots)
+    // 3 pre + 3 post + 4 successors = 10 slots -> Match32
+    instructions.push(
+        MatchIR::at(Label(5))
+            .nav(Nav::Down)
+            .pre_effect(EffectIR::start_obj())
+            .pre_effect(EffectIR::start_obj())
+            .pre_effect(EffectIR::start_obj())
+            .post_effect(EffectIR::end_obj())
+            .post_effect(EffectIR::end_obj())
+            .post_effect(EffectIR::end_obj())
+            .next_many(vec![Label(100), Label(101), Label(102), Label(103)])
+            .into(),
+    );
+
+    // Verify it's Match32 (32 bytes)
+    assert_eq!(instructions.last().unwrap().size(), 32);
+
+    let result = CacheAligned::layout(&instructions, &[Label(0)]);
+
+    // Labels 0-4 at steps 0-4
+    for i in 0..5 {
+        assert_eq!(result.label_to_step.get(&Label(i)), Some(&(i as u16)));
+    }
+
+    // Label(5) would be at step 5 (offset 40) without padding
+    // Match32 at offset 40 straddles (40+32=72 > 64), so padded to step 8
+    let step5 = *result.label_to_step.get(&Label(5)).unwrap();
+    assert_eq!(step5, 8, "Match32 should be padded to step 8 (offset 64)");
 
-    // Label 1 should be aligned - either at step 1 or padded to cache line
-    let step1 = *result.label_to_step.get(&Label(1)).unwrap();
-    assert!(step1 >= 1);
+    // Total steps: 8 (3 padding steps at 5,6,7) + 4 (Match32) = 12
+    assert_eq!(result.total_steps, 12);
 }

crates/plotnik-lib/src/emit/snapshots/plotnik_lib__emit__emit_tests__alternations_captured.snap

@@ -34,8 +34,9 @@ _ObjWrap:
   05                                        ▶
 
 Test:
-  06   ε                                    07
-  07   ε                                    10, 12
-  09                                        ▶
-  10   !   (identifier) [Node Set(M0)]      09
-  12   !   (number) [Node Set(M0)]          09
+  06   ε                                    08
+  07  ...  
+  08   ε                                    11, 13
+  10                                        ▶
+  11   !   (identifier) [Node Set(M0)]      10
+  13   !   (number) [Node Set(M0)]          10

crates/plotnik-lib/src/emit/snapshots/plotnik_lib__emit__emit_tests__alternations_captured_tagged.snap

@@ -45,9 +45,10 @@ _ObjWrap:
   05                                        ▶
 
 Test:
-  06   ε                                    07
-  07   ε                                    12, 15
-  09                                        ▶
-  10   ε   [Set(M4)]                        09
-  12   !   [Enum(M2)] (identifier) [Node Set(M0) EndEnum]  10
-  15   !   [Enum(M3)] (number) [Node Set(M1) EndEnum]  10
+  06   ε                                    08
+  07  ...  
+  08   ε                                    13, 16
+  10                                        ▶
+  11   ε   [Set(M4)]                        10
+  13   !   [Enum(M2)] (identifier) [Node Set(M0) EndEnum]  11
+  16   !   [Enum(M3)] (number) [Node Set(M1) EndEnum]  11

crates/plotnik-lib/src/emit/snapshots/plotnik_lib__emit__emit_tests__alternations_in_quantifier.snap

@@ -53,23 +53,26 @@ Test:
   06   ε                                    07
   07   !   (object)                         08
   08   ε   [Arr]                            10
-  10   ε                                    39, 20
+  10   ε                                    42, 21
   12   ε   [EndArr Set(M5)]                 14
-  14   △   _                                19
-  15   ε   [EndObj Push]                    17
-  17   ε                                    45, 12
-  19                                        ▶
-  20   ε   [EndArr Set(M5)]                 19
-  22   ε   [Set(M4)]                        15
-  24   !   [Enum(M2)] (pair) [Node Set(M0) EndEnum]  22
-  27   !   [Enum(M3)] (shorthand_property_identifier) [Node Set(M1) EndEnum]  22
-  30   ε                                    24, 27
-  32   ε   [Obj]                            30
-  34   ▷   _                                37
-  35   ε                                    34, 12
-  37   ε                                    32, 35
-  39   ▽   _                                37
-  40   ▷   _                                43
-  41   ε                                    40, 12
-  43   ε                                    32, 41
-  45   ▷   _                                43
+  14   △   _                                20
+  15  ...  
+  16   ε   [EndObj Push]                    18
+  18   ε                                    48, 12
+  20                                        ▶
+  21   ε   [EndArr Set(M5)]                 20
+  23  ...  
+  24   ε   [Set(M4)]                        16
+  26   !   [Enum(M2)] (pair) [Node Set(M0) EndEnum]  24
+  29   !   [Enum(M3)] (shorthand_property_identifier) [Node Set(M1) EndEnum]  24
+  32   ε                                    26, 29
+  34   ε   [Obj]                            32
+  36   ▷   _                                40
+  37   ε                                    36, 12
+  39  ...  
+  40   ε                                    34, 37
+  42   ▽   _                                40
+  43   ▷   _                                46
+  44   ε                                    43, 12
+  46   ε                                    34, 44
+  48   ▷   _                                46

crates/plotnik-lib/src/emit/snapshots/plotnik_lib__emit__emit_tests__alternations_labeled.snap

@@ -45,8 +45,11 @@ _ObjWrap:
   05                                        ▶
 
 Test:
-  06   ε                                    07
-  07   ε                                    10, 13
-  09                                        ▶
-  10   !   [Enum(M2)] (identifier) [Node Set(M0) EndEnum]  09
-  13   !   [Enum(M3)] (number) [Node Set(M1) EndEnum]  09
+  06   ε                                    08
+  07  ...  
+  08   ε                                    11, 16
+  10                                        ▶
+  11   !   [Enum(M2)] (identifier) [Node Set(M0) EndEnum]  10
+  14  ...  
+  15  ...  
+  16   !   [Enum(M3)] (number) [Node Set(M1) EndEnum]  10