diff --git a/src/util/arch/arm/bitutils.h b/src/util/arch/arm/bitutils.h
index ddca35c9..498db568 100644
--- a/src/util/arch/arm/bitutils.h
+++ b/src/util/arch/arm/bitutils.h
@@ -82,11 +82,7 @@ u32 findAndClearLSB_64_impl(u64a *v) {
 
 static really_inline
 u32 findAndClearMSB_32_impl(u32 *v) {
-    u32 val = *v;
-    u32 offset = 31 - clz32_impl(val);
-    *v = val & ~(1 << offset);
-    assert(offset < 32);
-    return offset;
+    return findAndClearMSB_32_impl_c(v);
 }
 
 static really_inline
@@ -107,20 +103,19 @@ u64a compress64_impl(u64a x, u64a m) {
 static really_inline
 m128 compress128_impl(m128 x, m128 m) {
     m128 one = set1_2x64(1);
-    m128 bitset = one;
-    m128 vres = zeroes128();
+    m128 bb = one;
+    m128 res = zeroes128();
     while (isnonzero128(m)) {
-	m128 mm = sub_2x64(zeroes128(), m);
-	m128 tv = and128(x, m);
-	tv = and128(tv, mm);
-
-	m128 mask = not128(eq64_m128(tv, zeroes128()));
-	mask = vandq_s64(bitset, mask);
-        vres = or128(vres, mask);
-	m = and128(m, sub_2x64(m, one));
-        bitset = lshift64_m128(bitset, 1);
+        m128 mm = sub_2x64(zeroes128(), m);
+        m128 xm = and128(x, m);
+        xm = and128(xm, mm);
+ 
+        m128 mask = not128(eq64_m128(xm, zeroes128()));
+        res = or128(res, and128(bb, mask));
+        m = and128(m, sub_2x64(m, one));
+        bb = lshift64_m128(bb, 1);
     }
-    return vres;
+    return res;
 }
 
 static really_inline
@@ -136,20 +131,18 @@ u64a expand64_impl(u64a x, u64a m) {
 static really_inline
 m128 expand128_impl(m128 x, m128 m) {
     m128 one = set1_2x64(1);
-    m128 bitset = one;
-    m128 vres = zeroes128();
+    m128 bb = one;
+    m128 res = zeroes128();
     while (isnonzero128(m)) {
-	m128 tv = and128(x, m);
-
-	m128 mm = sub_2x64(zeroes128(), m);
-	m128 mask = not128(eq64_m128(tv, zeroes128()));
-	mask = vandq_s64(bitset, mask);
-	mask = and128(mask, mm);
-        vres = or128(vres, mask);
-	m = and128(m, sub_2x64(m, one));
-        bitset = lshift64_m128(bitset, 1);
+        m128 xm = and128(x, bb);
+        m128 mm = sub_2x64(zeroes128(), m);
+        m128 mask = not128(eq64_m128(xm, zeroes128()));
+        mask = and128(mask, and128(m, mm));
+        res = or128(res, mask);
+        m = and128(m, sub_2x64(m, one));
+        bb = lshift64_m128(bb, 1);
     }
-    return vres;
+    return res;
 }
 
 /* returns the first set bit after begin (if not ~0U). If no bit is set after
diff --git a/unit/internal/bitutils.cpp b/unit/internal/bitutils.cpp
index 3f788544..8af8f9a4 100644
--- a/unit/internal/bitutils.cpp
+++ b/unit/internal/bitutils.cpp
@@ -294,6 +294,39 @@ TEST(BitUtils, compress64) {
     }
 }
 
+TEST(BitUtils, compress128) {
+    const m128 all_zeroes = zeroes128();
+    const m128 all_ones = ones128();
+    const m128 odd_bits = set1_2x64(0x5555555555555555ull);
+    const m128 even_bits = set1_2x64(0xaaaaaaaaaaaaaaaaull);
+
+    EXPECT_EQ(0, diff128(all_zeroes, compress128(all_zeroes, all_zeroes)));
+    EXPECT_EQ(0, diff128(all_zeroes, compress128(all_zeroes, set1_4x32(1))));
+    EXPECT_EQ(0, diff128(all_zeroes, compress128(all_zeroes, all_ones)));
+    EXPECT_EQ(0, diff128(all_ones, compress128(all_ones, all_ones)));
+    EXPECT_EQ(0, diff128(set1_2x64(0xffffffffull), compress128(odd_bits, odd_bits)));
+    EXPECT_EQ(0, diff128(set1_2x64(0xffffffffull), compress128(even_bits, even_bits)));
+    EXPECT_EQ(0, diff128(all_zeroes, compress128(odd_bits, even_bits)));
+    EXPECT_EQ(0, diff128(all_zeroes, compress128(even_bits, odd_bits)));
+
+    // Some single-bit tests.
+    for (u32 i = 0; i < 64; i++) {
+        const m128 one_bit = set1_2x64(1ull << i);
+
+        EXPECT_EQ(0, diff128(all_zeroes, compress128(all_zeroes, one_bit)));
+        EXPECT_EQ(0, diff128(set1_2x64(1ull), compress128(one_bit, one_bit)));
+        EXPECT_EQ(0, diff128(one_bit, compress128(one_bit, all_ones)));
+
+        if (i % 2) {
+            EXPECT_EQ(0, diff128(set1_2x64(1ull << (i / 2)), compress128(one_bit, even_bits)));
+            EXPECT_EQ(0, diff128(all_zeroes, compress128(one_bit, odd_bits)));
+        } else {
+            EXPECT_EQ(0, diff128(set1_2x64(1ull << (i / 2)), compress128(one_bit, odd_bits)));
+            EXPECT_EQ(0, diff128(all_zeroes, compress128(one_bit, even_bits)));
+        }
+    }
+}
+
 TEST(BitUtils, expand32) {
     const u32 all_ones = 0xffffffffu;
     const u32 odd_bits = 0x55555555u;
@@ -352,6 +385,35 @@ TEST(BitUtils, expand64) {
     }
 }
 
+TEST(BitUtils, expand128) {
+    const m128 all_zeroes = zeroes128();
+    const m128 all_ones = ones128();
+    const m128 odd_bits = set1_2x64(0x5555555555555555ull);
+    const m128 even_bits = set1_2x64(0xaaaaaaaaaaaaaaaaull);
+
+    EXPECT_EQ(0, diff128(all_zeroes, expand128(all_zeroes, all_zeroes)));
+    EXPECT_EQ(0, diff128(all_zeroes, expand128(all_zeroes, set1_2x64(1ull))));
+    EXPECT_EQ(0, diff128(all_zeroes, expand128(all_zeroes, all_ones)));
+    EXPECT_EQ(0, diff128(all_ones, expand128(all_ones, all_ones)));
+    EXPECT_EQ(0, diff128(odd_bits, expand128(set1_2x64(0xffffffffull), odd_bits)));
+    EXPECT_EQ(0, diff128(even_bits, expand128(set1_2x64(0xffffffffull), even_bits)));
+    EXPECT_EQ(0, diff128(all_zeroes, expand128(set1_2x64(0xffffffff00000000ull), even_bits)));
+    EXPECT_EQ(0, diff128(all_zeroes, expand128(set1_2x64(0xffffffff00000000ull), odd_bits)));
+    EXPECT_EQ(0, diff128(set1_2x64(1u), expand128(set1_2x64(1u), odd_bits)));
+    EXPECT_EQ(0, diff128(set1_2x64(2u), expand128(set1_2x64(1u), even_bits)));
+
+    // Some single-bit tests.
+    for (u32 i = 0; i < 64; i++) {
+        const m128 one_bit = set1_2x64(1ull << i);
+
+        EXPECT_EQ(0, diff128(all_zeroes, expand128(all_zeroes, one_bit)));
+        EXPECT_EQ(0, diff128(one_bit, expand128(set1_2x64(1ull), one_bit)));
+        EXPECT_EQ(0, diff128(one_bit, expand128(one_bit, all_ones)));
+
+        EXPECT_EQ(0, diff128(one_bit, expand128(set1_2x64(1ull << (i / 2)), i % 2 ? even_bits : odd_bits)));
+    }
+}
+
 TEST(BitUtils, bf_op_1) {
     u64a a = 0;
     for (u32 i = 0; i < 64; i++) {