From 75fdc9d3a403fd63b1123d2bf018758afd4d95b6 Mon Sep 17 00:00:00 2001
From: T0b1 <T0b1-iOS@users.noreply.github.com>
Date: Sun, 16 Apr 2023 14:10:24 +0200
Subject: [PATCH] try bigger smallvec for vregs

---
 src/ion/fast_alloc.rs | 62 ++++++++++++++++++++++++++++++++++++++++---
 1 file changed, 58 insertions(+), 4 deletions(-)

diff --git a/src/ion/fast_alloc.rs b/src/ion/fast_alloc.rs
index 358f18b..38c2439 100644
--- a/src/ion/fast_alloc.rs
+++ b/src/ion/fast_alloc.rs
@@ -144,8 +144,8 @@ struct FastAllocState<'a, F: Function> {
     pub vregs: Vec<VRegData>,
     pub pregs: Vec<PRegData>,
     pub blocks: Vec<BlockData>,
-    pub liveins: Vec<BlockBitmap>,
-    pub liveouts: Vec<BlockBitmap>,
+    pub liveins: Vec<VRegBitmap>,
+    pub liveouts: Vec<VRegBitmap>,
 
     pub cur_stack_slot_idx: u32,
     pub reftype_vregs_in_pregs_count: u32,
@@ -1147,6 +1147,60 @@ impl BlockBitmap {
     }
 }
 
+#[derive(Clone, Debug)]
+struct VRegBitmap {
+    storage: SmallVec<[u64; 8]>,
+}
+
+impl VRegBitmap {
+    fn init(vreg_count: usize) -> Self {
+        let u64_count = (vreg_count + 63) / 64;
+        let mut storage = SmallVec::<[u64; 8]>::with_capacity(u64_count);
+        storage.resize(u64_count, 0);
+        Self { storage }
+    }
+
+    fn set(&mut self, idx: usize) {
+        let idx = idx / 64;
+        let bit = 1u64 << (idx % 64);
+        self.storage[idx] |= bit;
+    }
+
+    fn un_set(&mut self, idx: usize) {
+        let idx = idx / 64;
+        let bit = 1u64 << (idx % 64);
+        self.storage[idx] &= !bit;
+    }
+
+    fn is_set(&mut self, idx: usize) -> bool {
+        let idx = idx / 64;
+        let bit = 1u64 << (idx % 64);
+        (self.storage[idx] & bit) != 0
+    }
+
+    fn is_empty(&self) -> bool {
+        for &b in self.storage.iter() {
+            if b != 0 {
+                return false;
+            }
+        }
+        return true;
+    }
+
+    pub fn union_with(&mut self, other: &Self) -> bool {
+        let mut changed = 0;
+        for (word_idx, &bits) in other.storage.iter().enumerate() {
+            if bits == 0 {
+                continue;
+            }
+            let self_word = &mut self.storage[word_idx];
+            changed |= bits & !*self_word;
+            *self_word |= bits;
+        }
+        changed != 0
+    }
+}
+
 // currently, copy from liveranges.rs
 // don't inline for better perf stats
 #[inline(never)]
@@ -1156,10 +1210,10 @@ fn calc_live_bitmaps<'a, F: Function>(
 ) -> Result<(), RegAllocError> {
     state
         .liveins
-        .resize(state.blocks.len(), BlockBitmap::init(state.vregs.len()));
+        .resize(state.blocks.len(), VRegBitmap::init(state.vregs.len()));
     state
         .liveouts
-        .resize(state.blocks.len(), BlockBitmap::init(state.vregs.len()));
+        .resize(state.blocks.len(), VRegBitmap::init(state.vregs.len()));
 
     // Run a worklist algorithm to precisely compute liveins and
     // liveouts.