diff --git a/src/util/arch/arm/bitutils.h b/src/util/arch/arm/bitutils.h
index 0b579dc9..1d1e0167 100644
--- a/src/util/arch/arm/bitutils.h
+++ b/src/util/arch/arm/bitutils.h
@@ -107,102 +107,20 @@ u64a compress64_impl(u64a x, u64a m) {
 static really_inline
 m128 compress128_impl(m128 x, m128 m) {
 
-/*    x = and128(x, m); // clear irrelevant bits
-
-    // Return zero quickly on trivial cases
-    if (diff128(x, zeroes128()) == 0) {
-        return zeroes128();
-    }*/
-
-
-    u64a ALIGN_ATTR(16) xv[2];
-    u64a ALIGN_ATTR(16) mv[2];
-    u64a ALIGN_ATTR(16) res[2];
-    u64a ALIGN_ATTR(16) t[2];
-    u64a ALIGN_ATTR(16) bbv[2];
-    store128(xv, x);
-    store128(mv, m);
-    res[0] = 0;
-    res[1] = 0;
-    printf("x[%d] = %0llx\n", 0, xv[0]);
-    printf("x[%d] = %0llx\n", 1, xv[1]);
-
     m128 one = set1_2x64(1);
     m128 bitset = one;
     m128 vres = zeroes128();
-    for (u64a bb = 1; mv[0] | mv[1]; bb <<= 1) {
-        printf("bb = %lld\n", bb);
-	store128(bbv, bitset);
-        printf("bb[%d] = %0lld\n", 0, bbv[0]);
-        printf("bb[%d] = %0lld\n", 1, bbv[1]);
-        printf("m[%d] = %0llx\n", 0, mv[0]);
-        printf("m[%d] = %0llx\n", 1, mv[1]);
-        printf("scalar: -m[%d] = %0llx\n", 0, -mv[0]);
-        printf("scalar: -m[%d] = %0llx\n", 1, -mv[1]);
+    while (isnonzero128(m)) {
 	m128 mm = sub_2x64(zeroes128(), m);
-	store128(t, mm);
-        printf("vector: -m[0] = %0llx\n", t[0]);
-        printf("vector: -m[1] = %0llx\n", t[1]);
 	m128 tv = and128(x, m);
-	store128(t, tv);
-        printf("vector: x[0] & m[0] = %0llx\n", t[0]);
-        printf("vector: x[1] & m[1] = %0llx\n", t[1]);
 	tv = and128(tv, mm);
-	store128(t, tv);
-        printf("vector: x[0] & m[0] & -m[0] = %0llx\n", t[0]);
-        printf("vector: x[1] & m[1] & -m[1] = %0llx\n", t[1]);
-        t[0] = xv[0] & mv[0];
-        t[1] = xv[1] & mv[1];
-        printf("scalar: x[0] & m[0] = %0llx\n", t[0]);
-        printf("scalar: x[1] & m[1] = %0llx\n", t[1]);
-        t[0] = xv[0] & mv[0] & -mv[0];
-        t[1] = xv[1] & mv[1] & -mv[1];
-        printf("scalar: x[0] & m[0] & -m[0] = %0llx\n", t[0]);
-        printf("scalar: x[1] & m[1] & -m[1] = %0llx\n", t[1]);
-        
-        if ( t[0] ) {
-            printf("x & m & -m != 0\n");
-            res[0] |= bb;
-            printf("x[%d] = %0llx\n", 0, xv[0]);
-        }
-        if ( t[1] ) {
-            printf("x & m & -m != 0\n");
-            res[1] |= bb;
-            printf("x[%d] = %0llx\n", 1, xv[1]);
-        }
 
 	m128 mask = not128(eq64_m128(tv, zeroes128()));
-	store128(t, mask);
-        printf("mask: x[0] & m[0] & -m[0] != 0 : %0llx\n", t[0]);
-        printf("mask: x[1] & m[1] & -m[1] != 0 : %0llx\n", t[1]);
-
 	mask = vandq_s64(bitset, mask);
-	store128(t, mask);
-        printf("mask: mask[0] & bitset[1] != 0 : %0llx\n", t[0]);
-        printf("mask: mask[1] & bitset[1] != 0 : %0llx\n", t[1]);
-
         vres = or128(vres, mask);
-	store128(t, vres);
-        printf("res: res[0] != 0 : %0llx\n", t[0]);
-        printf("res: res[1] != 0 : %0llx\n", t[1]);
-	if (t[0] != res[0]) {
-            printf("mismatch: t[0] != res[0]: %0llx != %0llx\n", t[0], res[0]);
-        }
-	if (t[1] != res[1]) {
-            printf("mismatch: t[1] != res[1]: %0llx != %0llx\n", t[1], res[1]);
-        }
-
-        mv[0] &= mv[0] - 1;
-        mv[1] &= mv[1] - 1;
 	m = and128(m, sub_2x64(m, set1_2x64(1)));
-        printf("x[%d] = %0llx\n", 0, xv[0]);
-        printf("x[%d] = %0llx\n", 1, xv[1]);
         bitset = lshift64_m128(bitset, 1);
     }
-    store128(res, vres);
-    printf("final x[%d] = %0llx\n", 0, res[0]);
-    printf("final x[%d] = %0llx\n", 1, res[1]);
-//    x = load128(res);
     return vres;
 }
 
diff --git a/src/util/arch/x86/bitutils.h b/src/util/arch/x86/bitutils.h
index a0769a5e..424ad957 100644
--- a/src/util/arch/x86/bitutils.h
+++ b/src/util/arch/x86/bitutils.h
@@ -239,7 +239,6 @@ u64a expand64_impl(u64a x, u64a m) {
 #endif
 }
 
-
 /* returns the first set bit after begin (if not ~0U). If no bit is set after
  * begin returns ~0U
  */
diff --git a/src/util/state_compress.c b/src/util/state_compress.c
index 586e47f4..360ec39e 100644
--- a/src/util/state_compress.c
+++ b/src/util/state_compress.c
@@ -109,10 +109,6 @@ static really_inline
 void storecompressed128_64bit(void *ptr, m128 xvec, m128 mvec) {
     printf("storecompressed128_64bit()\n");
     // First, decompose our vectors into 64-bit chunks.
-/*    u64a x[2];
-    memcpy(x, &xvec, sizeof(xvec));
-    u64a m[2];
-    memcpy(m, &mvec, sizeof(mvec));*/
     u64a ALIGN_ATTR(16) x[2];
     u64a ALIGN_ATTR(16) m[2];
     store128(m, mvec);
@@ -121,10 +117,8 @@ void storecompressed128_64bit(void *ptr, m128 xvec, m128 mvec) {
     // Count the number of bits of compressed state we're writing out per
     // chunk.
     u32 ALIGN_ATTR(16) bits[2] = { popcount64(m[0]), popcount64(m[1]) };
-    //m128 vbits = load128(bits);
 
     // Compress each 64-bit chunk individually.
-    //u64a v[2] = { compress64(x[0], m[0]), compress64(x[1], m[1]) };
     xvec = compress128(xvec, mvec);
     store128(x, xvec);
 
@@ -169,29 +163,13 @@ m128 loadcompressed128_64bit(const void *ptr, m128 mvec) {
     // First, decompose our vectors into 64-bit chunks.
     u64a ALIGN_ATTR(16) m[2];
     store128(m, mvec);
-    printf("m[0] = %0llx\n", m[0]);
-    printf("m[1] = %0llx\n", m[1]);
-
-//    m[0] = movq(mvec);
-//    m[1] = movq(rshiftbyte_m128(mvec, 8));
-    //store128(m, mvec);
-//    printf("m[0] = %0llx\n", m[0]);
-//    printf("m[1] = %0llx\n", m[1]);
 
     u32 bits[2] = { popcount64(m[0]), popcount64(m[1]) };
     u64a ALIGN_ATTR(16) v[2];
 
-    printf("bits[0] = %0x\n", bits[0]);
-    printf("bits[1] = %0x\n", bits[1]);
-
     unpack_bits_64(v, (const u8 *)ptr, bits, 2);
-    printf("v[0] = %0llx\n", v[0]);
-    printf("v[1] = %0llx\n", v[1]);
 
     u64a x[2] = { expand64(v[0], m[0]), expand64(v[1], m[1]) };
-    printf("x[0] = %0llx\n", x[0]);
-    printf("x[1] = %0llx\n", x[1]);
-
 
     return set2x64(x[1], x[0]);
 }