diff --git a/src/util/arch/common/simd_utils.h b/src/util/arch/common/simd_utils.h
index e682e2d5..56d9dbaf 100644
--- a/src/util/arch/common/simd_utils.h
+++ b/src/util/arch/common/simd_utils.h
@@ -147,10 +147,12 @@ static really_inline int isnonzero256(m256 a) {
 }
 
 /**
- * "Rich" version of diff256(). Takes two vectors a and b and returns an 8-bit
+ * "Rich" version of diff256(). Takes two vectors a and b and returns a 8-bit
  * mask indicating which 32-bit words contain differences.
  */
-static really_inline u32 diffrich256(m256 a, m256 b) {
+static really_inline
+u32 diffrich256(m256 a, m256 b) {
+    return diffrich128(a.lo, b.lo) | (diffrich128(a.hi, b.hi) << 8);
 }
 
 /**
@@ -311,26 +313,6 @@ m256 pshufb_m256(m256 a, m256 b) {
     return rv;
 }
 
-#define cast256to128(a) _mm256_castsi256_si128(a)
-#define cast128to256(a) _mm256_castsi128_si256(a)
-#define swap128in256(a) _mm256_permute4x64_epi64(a, 0x4E)
-#define insert128to256(a, b, imm) _mm256_inserti128_si256(a, b, imm)
-#define rshift128_m256(a, count_immed) _mm256_srli_si256(a, count_immed)
-#define lshift128_m256(a, count_immed) _mm256_slli_si256(a, count_immed)
-#define extract64from256(a, imm) _mm_extract_epi64(_mm256_extracti128_si256(a, imm >> 1), imm % 2)
-#define extract32from256(a, imm) _mm_extract_epi32(_mm256_extracti128_si256(a, imm >> 2), imm % 4)
-#define extractlow64from256(a) _mm_cvtsi128_si64(cast256to128(a))
-#define extractlow32from256(a) movd(cast256to128(a))
-#define interleave256hi(a, b) _mm256_unpackhi_epi8(a, b)
-#define interleave256lo(a, b) _mm256_unpacklo_epi8(a, b)
-#define vpalignr(r, l, offset) _mm256_alignr_epi8(r, l, offset)
-
-#define extract128from512(a, imm) _mm512_extracti32x4_epi32(a, imm)
-#define interleave512hi(a, b) _mm512_unpackhi_epi8(a, b)
-#define interleave512lo(a, b) _mm512_unpacklo_epi8(a, b)
-#define set2x256(a) _mm512_broadcast_i64x4(a)
-#define mask_set2x256(src, k, a) _mm512_mask_broadcast_i64x4(src, k, a)
-
 #endif // HAVE_SIMD_256_BITS
 
 /****
@@ -402,13 +384,6 @@ static really_inline int isnonzero384(m384 a) {
     return isnonzero128(or128(or128(a.lo, a.mid), a.hi));
 }
 
-/**
- * "Rich" version of diff384(). Takes two vectors a and b and returns a 12-bit
- * mask indicating which 32-bit words contain differences.
- */
-static really_inline u32 diffrich384(m384 a, m384 b) {
-}
-
 /**
  * "Rich" version of diff384(), 64-bit variant. Takes two vectors a and b and
  * returns a 12-bit mask indicating which 64-bit words contain differences.
@@ -507,9 +482,6 @@ char testbit384(m384 val, unsigned int n) {
  ****/
 
 #if !defined(HAVE_SIMD_512_BITS)
-#define eq512mask(a, b) _mm512_cmpeq_epi8_mask((a), (b))
-#define masked_eq512mask(k, a, b) _mm512_mask_cmpeq_epi8_mask((k), (a), (b))
-#define vpermq512(idx, a) _mm512_permutexvar_epi64(idx, a)
 
 static really_inline
 m512 zeroes512(void) {
@@ -608,12 +580,6 @@ m512 lshift64_m512(m512 a, unsigned b) {
     return rv;
 }
 
-#if defined(HAVE_AVX512)
-#define rshift64_m512(a, b) _mm512_srli_epi64((a), (b))
-#define rshift128_m512(a, count_immed) _mm512_bsrli_epi128(a, count_immed)
-#define lshift128_m512(a, count_immed) _mm512_bslli_epi128(a, count_immed)
-#endif
-
 static really_inline
 int diff512(m512 a, m512 b) {
     return diff256(a.lo, b.lo) || diff256(a.hi, b.hi);
@@ -621,9 +587,9 @@ int diff512(m512 a, m512 b) {
 
 static really_inline
 int isnonzero512(m512 a) {
-    m128 x = or128(a.lo.lo, a.lo.hi);
-    m128 y = or128(a.hi.lo, a.hi.hi);
-    return isnonzero128(or128(x, y));
+    m256 x = or256(a.lo, a.lo);
+    m256 y = or256(a.hi, a.hi);
+    return isnonzero256(or256(x, y));
 }
 
 /**
diff --git a/src/util/arch/x86/simd_utils.h b/src/util/arch/x86/simd_utils.h
index 2d099f56..4a1a691e 100644
--- a/src/util/arch/x86/simd_utils.h
+++ b/src/util/arch/x86/simd_utils.h
@@ -127,22 +127,8 @@ static really_inline u32 movd(const m128 in) {
     return _mm_cvtsi128_si32(in);
 }
 
-#if defined(HAVE_AVX512)
-static really_inline u32 movd512(const m512 in) {
-    // NOTE: seems gcc doesn't support _mm512_cvtsi512_si32(in),
-    //       so we use 2-step convertions to work around.
-    return _mm_cvtsi128_si32(_mm512_castsi512_si128(in));
-}
-#endif
-
 static really_inline u64a movq(const m128 in) {
-#if defined(ARCH_X86_64)
     return _mm_cvtsi128_si64(in);
-#else // 32-bit - this is horrific
-    u32 lo = movd(in);
-    u32 hi = movd(_mm_srli_epi64(in, 32));
-    return (u64a)hi << 32 | lo;
-#endif
 }
 
 /* another form of movq */
@@ -281,36 +267,6 @@ m128 pshufb_m128(m128 a, m128 b) {
     return result;
 }
 
-static really_inline
-m256 pshufb_m256(m256 a, m256 b) {
-#if defined(HAVE_AVX2)
-    return _mm256_shuffle_epi8(a, b);
-#else
-    m256 rv;
-    rv.lo = pshufb_m128(a.lo, b.lo);
-    rv.hi = pshufb_m128(a.hi, b.hi);
-    return rv;
-#endif
-}
-
-#if defined(HAVE_AVX512)
-static really_inline
-m512 pshufb_m512(m512 a, m512 b) {
-    return _mm512_shuffle_epi8(a, b);
-}
-
-static really_inline
-m512 maskz_pshufb_m512(__mmask64 k, m512 a, m512 b) {
-    return _mm512_maskz_shuffle_epi8(k, a, b);
-}
-
-#if defined(HAVE_AVX512VBMI)
-#define vpermb512(idx, a) _mm512_permutexvar_epi8(idx, a)
-#define maskz_vpermb512(k, idx, a) _mm512_maskz_permutexvar_epi8(k, idx, a)
-#endif
-
-#endif
-
 static really_inline
 m128 variable_byte_shift_m128(m128 in, s32 amount) {
     assert(amount >= -16 && amount <= 16);
@@ -352,7 +308,12 @@ m128 set2x64(u64a hi, u64a lo) {
  **** 256-bit Primitives
  ****/
 
-#if defined(HAVE_AVX2)
+#if defined(HAVE_SIMD_256_BITS) && defined(HAVE_AVX2)
+
+static really_inline
+m256 pshufb_m256(m256 a, m256 b) {
+    return _mm256_shuffle_epi8(a, b);
+}
 
 static really_really_inline
 m256 lshift64_m256(m256 a, unsigned b) {
@@ -379,143 +340,41 @@ m256 set1_2x128(m128 a) {
     return _mm256_broadcastsi128_si256(a);
 }
 
-#else
-
-static really_really_inline
-m256 lshift64_m256(m256 a, int b) {
-    m256 rv = a;
-    rv.lo = lshift64_m128(rv.lo, b);
-    rv.hi = lshift64_m128(rv.hi, b);
-    return rv;
-}
-
-static really_inline
-m256 rshift64_m256(m256 a, int b) {
-    m256 rv = a;
-    rv.lo = rshift64_m128(rv.lo, b);
-    rv.hi = rshift64_m128(rv.hi, b);
-    return rv;
-}
-
-static really_inline
-m256 eq256(m256 a, m256 b) {
-    m256 rv;
-    rv.lo = eq128(a.lo, b.lo);
-    rv.hi = eq128(a.hi, b.hi);
-    return rv;
-}
-
-static really_inline
-u32 movemask256(m256 a) {
-    u32 lo_mask = movemask128(a.lo);
-    u32 hi_mask = movemask128(a.hi);
-    return lo_mask | (hi_mask << 16);
-}
-
-static really_inline
-m256 set1_2x128(m128 a) {
-    m256 rv = {a, a};
-    return rv;
-}
-#endif
-
 static really_inline m256 zeroes256(void) {
-#if defined(HAVE_AVX2)
     return _mm256_setzero_si256();
-#else
-    m256 rv = {zeroes128(), zeroes128()};
-    return rv;
-#endif
 }
 
 static really_inline m256 ones256(void) {
-#if defined(HAVE_AVX2)
     m256 rv = _mm256_set1_epi8(0xFF);
-#else
-    m256 rv = {ones128(), ones128()};
-#endif
     return rv;
 }
 
-#if defined(HAVE_AVX2)
 static really_inline m256 and256(m256 a, m256 b) {
     return _mm256_and_si256(a, b);
 }
-#else
-static really_inline m256 and256(m256 a, m256 b) {
-    m256 rv;
-    rv.lo = and128(a.lo, b.lo);
-    rv.hi = and128(a.hi, b.hi);
-    return rv;
-}
-#endif
 
-#if defined(HAVE_AVX2)
 static really_inline m256 or256(m256 a, m256 b) {
     return _mm256_or_si256(a, b);
 }
-#else
-static really_inline m256 or256(m256 a, m256 b) {
-    m256 rv;
-    rv.lo = or128(a.lo, b.lo);
-    rv.hi = or128(a.hi, b.hi);
-    return rv;
-}
-#endif
 
-#if defined(HAVE_AVX2)
 static really_inline m256 xor256(m256 a, m256 b) {
     return _mm256_xor_si256(a, b);
 }
-#else
-static really_inline m256 xor256(m256 a, m256 b) {
-    m256 rv;
-    rv.lo = xor128(a.lo, b.lo);
-    rv.hi = xor128(a.hi, b.hi);
-    return rv;
-}
-#endif
 
-#if defined(HAVE_AVX2)
 static really_inline m256 not256(m256 a) {
     return _mm256_xor_si256(a, ones256());
 }
-#else
-static really_inline m256 not256(m256 a) {
-    m256 rv;
-    rv.lo = not128(a.lo);
-    rv.hi = not128(a.hi);
-    return rv;
-}
-#endif
 
-#if defined(HAVE_AVX2)
 static really_inline m256 andnot256(m256 a, m256 b) {
     return _mm256_andnot_si256(a, b);
 }
-#else
-static really_inline m256 andnot256(m256 a, m256 b) {
-    m256 rv;
-    rv.lo = andnot128(a.lo, b.lo);
-    rv.hi = andnot128(a.hi, b.hi);
-    return rv;
-}
-#endif
 
 static really_inline int diff256(m256 a, m256 b) {
-#if defined(HAVE_AVX2)
     return !!(_mm256_movemask_epi8(_mm256_cmpeq_epi8(a, b)) ^ (int)-1);
-#else
-    return diff128(a.lo, b.lo) || diff128(a.hi, b.hi);
-#endif
 }
 
 static really_inline int isnonzero256(m256 a) {
-#if defined(HAVE_AVX2)
     return !!diff256(a, zeroes256());
-#else
-    return isnonzero128(or128(a.lo, a.hi));
-#endif
 }
 
 /**
@@ -523,16 +382,8 @@ static really_inline int isnonzero256(m256 a) {
  * mask indicating which 32-bit words contain differences.
  */
 static really_inline u32 diffrich256(m256 a, m256 b) {
-#if defined(HAVE_AVX2)
     a = _mm256_cmpeq_epi32(a, b);
     return ~(_mm256_movemask_ps(_mm256_castsi256_ps(a))) & 0xFF;
-#else
-    m128 z = zeroes128();
-    a.lo = _mm_cmpeq_epi32(a.lo, b.lo);
-    a.hi = _mm_cmpeq_epi32(a.hi, b.hi);
-    m128 packed = _mm_packs_epi16(_mm_packs_epi32(a.lo, a.hi), z);
-    return ~(_mm_movemask_epi8(packed)) & 0xff;
-#endif
 }
 
 /**
@@ -547,24 +398,12 @@ static really_inline u32 diffrich64_256(m256 a, m256 b) {
 // aligned load
 static really_inline m256 load256(const void *ptr) {
     assert(ISALIGNED_N(ptr, alignof(m256)));
-#if defined(HAVE_AVX2)
     return _mm256_load_si256((const m256 *)ptr);
-#else
-    m256 rv = { load128(ptr), load128((const char *)ptr + 16) };
-    return rv;
-#endif
 }
 
 // aligned load  of 128-bit value to low and high part of 256-bit value
 static really_inline m256 load2x128(const void *ptr) {
-#if defined(HAVE_AVX2)
     return set1_2x128(load128(ptr));
-#else
-    assert(ISALIGNED_N(ptr, alignof(m128)));
-    m256 rv;
-    rv.hi = rv.lo = load128(ptr);
-    return rv;
-#endif
 }
 
 static really_inline m256 loadu2x128(const void *ptr) {
@@ -574,32 +413,17 @@ static really_inline m256 loadu2x128(const void *ptr) {
 // aligned store
 static really_inline void store256(void *ptr, m256 a) {
     assert(ISALIGNED_N(ptr, alignof(m256)));
-#if defined(HAVE_AVX2)
     _mm256_store_si256((m256 *)ptr, a);
-#else
-    ptr = assume_aligned(ptr, 16);
-    *(m256 *)ptr = a;
-#endif
 }
 
 // unaligned load
 static really_inline m256 loadu256(const void *ptr) {
-#if defined(HAVE_AVX2)
     return _mm256_loadu_si256((const m256 *)ptr);
-#else
-    m256 rv = { loadu128(ptr), loadu128((const char *)ptr + 16) };
-    return rv;
-#endif
 }
 
 // unaligned store
 static really_inline void storeu256(void *ptr, m256 a) {
-#if defined(HAVE_AVX2)
     _mm256_storeu_si256((m256 *)ptr, a);
-#else
-    storeu128(ptr, a.lo);
-    storeu128((char *)ptr + 16, a.hi);
-#endif
 }
 
 // packed unaligned store of first N bytes
@@ -628,101 +452,19 @@ m256 mask1bit256(unsigned int n) {
 
 static really_inline
 m256 set1_32x8(u32 in) {
-#if defined(HAVE_AVX2)
     return _mm256_set1_epi8(in);
-#else
-    m256 rv;
-    rv.hi = set1_16x8(in);
-    rv.lo = set1_16x8(in);
-    return rv;
-#endif
 }
 
 static really_inline
 m256 set8x32(u32 hi_3, u32 hi_2, u32 hi_1, u32 hi_0, u32 lo_3, u32 lo_2, u32 lo_1, u32 lo_0) {
-#if defined(HAVE_AVX2)
     return _mm256_set_epi32(hi_3, hi_2, hi_1, hi_0, lo_3, lo_2, lo_1, lo_0);
-#else
-    m256 rv;
-    rv.hi = set4x32(hi_3, hi_2, hi_1, hi_0);
-    rv.lo = set4x32(lo_3, lo_2, lo_1, lo_0);
-    return rv;
-#endif
 }
 
 static really_inline
 m256 set4x64(u64a hi_1, u64a hi_0, u64a lo_1, u64a lo_0) {
-#if defined(HAVE_AVX2)
     return _mm256_set_epi64x(hi_1, hi_0, lo_1, lo_0);
-#else
-    m256 rv;
-    rv.hi = set2x64(hi_1, hi_0);
-    rv.lo = set2x64(lo_1, lo_0);
-    return rv;
-#endif
 }
 
-#if !defined(HAVE_AVX2)
-// switches on bit N in the given vector.
-static really_inline
-void setbit256(m256 *ptr, unsigned int n) {
-    assert(n < sizeof(*ptr) * 8);
-    m128 *sub;
-    if (n < 128) {
-        sub = &ptr->lo;
-    } else {
-        sub = &ptr->hi;
-        n -= 128;
-    }
-    setbit128(sub, n);
-}
-
-// switches off bit N in the given vector.
-static really_inline
-void clearbit256(m256 *ptr, unsigned int n) {
-    assert(n < sizeof(*ptr) * 8);
-    m128 *sub;
-    if (n < 128) {
-        sub = &ptr->lo;
-    } else {
-        sub = &ptr->hi;
-        n -= 128;
-    }
-    clearbit128(sub, n);
-}
-
-// tests bit N in the given vector.
-static really_inline
-char testbit256(m256 val, unsigned int n) {
-    assert(n < sizeof(val) * 8);
-    m128 sub;
-    if (n < 128) {
-        sub = val.lo;
-    } else {
-        sub = val.hi;
-        n -= 128;
-    }
-    return testbit128(sub, n);
-}
-
-static really_really_inline
-m128 movdq_hi(m256 x) {
-    return x.hi;
-}
-
-static really_really_inline
-m128 movdq_lo(m256 x) {
-    return x.lo;
-}
-
-static really_inline
-m256 combine2x128(m128 hi, m128 lo) {
-    m256 rv = {lo, hi};
-    return rv;
-}
-
-#else // AVX2
-
 // switches on bit N in the given vector.
 static really_inline
 void setbit256(m256 *ptr, unsigned int n) {
@@ -775,88 +517,12 @@ m256 combine2x128(m128 hi, m128 lo) {
 }
 #endif //AVX2
 
-#if defined(HAVE_AVX512)
-#define extract128from512(a, imm) _mm512_extracti32x4_epi32(a, imm)
-#define interleave512hi(a, b) _mm512_unpackhi_epi8(a, b)
-#define interleave512lo(a, b) _mm512_unpacklo_epi8(a, b)
-#define set2x256(a) _mm512_broadcast_i64x4(a)
-#define mask_set2x256(src, k, a) _mm512_mask_broadcast_i64x4(src, k, a)
-#define vpermq512(idx, a) _mm512_permutexvar_epi64(idx, a)
-#endif
-
-/****
- **** 384-bit Primitives
- ****/
-
-static really_inline m384 and384(m384 a, m384 b) {
-    m384 rv;
-    rv.lo = and128(a.lo, b.lo);
-    rv.mid = and128(a.mid, b.mid);
-    rv.hi = and128(a.hi, b.hi);
-    return rv;
-}
-
-static really_inline m384 or384(m384 a, m384 b) {
-    m384 rv;
-    rv.lo = or128(a.lo, b.lo);
-    rv.mid = or128(a.mid, b.mid);
-    rv.hi = or128(a.hi, b.hi);
-    return rv;
-}
-
-static really_inline m384 xor384(m384 a, m384 b) {
-    m384 rv;
-    rv.lo = xor128(a.lo, b.lo);
-    rv.mid = xor128(a.mid, b.mid);
-    rv.hi = xor128(a.hi, b.hi);
-    return rv;
-}
-static really_inline m384 not384(m384 a) {
-    m384 rv;
-    rv.lo = not128(a.lo);
-    rv.mid = not128(a.mid);
-    rv.hi = not128(a.hi);
-    return rv;
-}
-static really_inline m384 andnot384(m384 a, m384 b) {
-    m384 rv;
-    rv.lo = andnot128(a.lo, b.lo);
-    rv.mid = andnot128(a.mid, b.mid);
-    rv.hi = andnot128(a.hi, b.hi);
-    return rv;
-}
-
-static really_really_inline
-m384 lshift64_m384(m384 a, unsigned b) {
-    m384 rv;
-    rv.lo = lshift64_m128(a.lo, b);
-    rv.mid = lshift64_m128(a.mid, b);
-    rv.hi = lshift64_m128(a.hi, b);
-    return rv;
-}
-
-static really_inline m384 zeroes384(void) {
-    m384 rv = {zeroes128(), zeroes128(), zeroes128()};
-    return rv;
-}
-
-static really_inline m384 ones384(void) {
-    m384 rv = {ones128(), ones128(), ones128()};
-    return rv;
-}
-
-static really_inline int diff384(m384 a, m384 b) {
-    return diff128(a.lo, b.lo) || diff128(a.mid, b.mid) || diff128(a.hi, b.hi);
-}
-
-static really_inline int isnonzero384(m384 a) {
-    return isnonzero128(or128(or128(a.lo, a.mid), a.hi));
-}
-
+#if defined(HAVE_SIMD_128_BITS)
 /**
  * "Rich" version of diff384(). Takes two vectors a and b and returns a 12-bit
  * mask indicating which 32-bit words contain differences.
  */
+
 static really_inline u32 diffrich384(m384 a, m384 b) {
     m128 z = zeroes128();
     a.lo = _mm_cmpeq_epi32(a.lo, b.lo);
@@ -867,102 +533,42 @@ static really_inline u32 diffrich384(m384 a, m384 b) {
     return ~(_mm_movemask_epi8(packed)) & 0xfff;
 }
 
-/**
- * "Rich" version of diff384(), 64-bit variant. Takes two vectors a and b and
- * returns a 12-bit mask indicating which 64-bit words contain differences.
- */
-static really_inline u32 diffrich64_384(m384 a, m384 b) {
-    u32 d = diffrich384(a, b);
-    return (d | (d >> 1)) & 0x55555555;
-}
-
-// aligned load
-static really_inline m384 load384(const void *ptr) {
-    assert(ISALIGNED_16(ptr));
-    m384 rv = { load128(ptr), load128((const char *)ptr + 16),
-                load128((const char *)ptr + 32) };
-    return rv;
-}
-
-// aligned store
-static really_inline void store384(void *ptr, m384 a) {
-    assert(ISALIGNED_16(ptr));
-    ptr = assume_aligned(ptr, 16);
-    *(m384 *)ptr = a;
-}
-
-// unaligned load
-static really_inline m384 loadu384(const void *ptr) {
-    m384 rv = { loadu128(ptr), loadu128((const char *)ptr + 16),
-                loadu128((const char *)ptr + 32)};
-    return rv;
-}
-
-// packed unaligned store of first N bytes
-static really_inline
-void storebytes384(void *ptr, m384 a, unsigned int n) {
-    assert(n <= sizeof(a));
-    memcpy(ptr, &a, n);
-}
-
-// packed unaligned load of first N bytes, pad with zero
-static really_inline
-m384 loadbytes384(const void *ptr, unsigned int n) {
-    m384 a = zeroes384();
-    assert(n <= sizeof(a));
-    memcpy(&a, ptr, n);
-    return a;
-}
-
-// switches on bit N in the given vector.
-static really_inline
-void setbit384(m384 *ptr, unsigned int n) {
-    assert(n < sizeof(*ptr) * 8);
-    m128 *sub;
-    if (n < 128) {
-        sub = &ptr->lo;
-    } else if (n < 256) {
-        sub = &ptr->mid;
-    } else {
-        sub = &ptr->hi;
-    }
-    setbit128(sub, n % 128);
-}
-
-// switches off bit N in the given vector.
-static really_inline
-void clearbit384(m384 *ptr, unsigned int n) {
-    assert(n < sizeof(*ptr) * 8);
-    m128 *sub;
-    if (n < 128) {
-        sub = &ptr->lo;
-    } else if (n < 256) {
-        sub = &ptr->mid;
-    } else {
-        sub = &ptr->hi;
-    }
-    clearbit128(sub, n % 128);
-}
-
-// tests bit N in the given vector.
-static really_inline
-char testbit384(m384 val, unsigned int n) {
-    assert(n < sizeof(val) * 8);
-    m128 sub;
-    if (n < 128) {
-        sub = val.lo;
-    } else if (n < 256) {
-        sub = val.mid;
-    } else {
-        sub = val.hi;
-    }
-    return testbit128(sub, n % 128);
-}
+#endif // HAVE_SIMD_128_BITS
 
 /****
  **** 512-bit Primitives
  ****/
 
+#if defined(HAVE_SIMD_512_BITS)
+
+#define extract128from512(a, imm) _mm512_extracti32x4_epi32(a, imm)
+#define interleave512hi(a, b) _mm512_unpackhi_epi8(a, b)
+#define interleave512lo(a, b) _mm512_unpacklo_epi8(a, b)
+#define set2x256(a) _mm512_broadcast_i64x4(a)
+#define mask_set2x256(src, k, a) _mm512_mask_broadcast_i64x4(src, k, a)
+#define vpermq512(idx, a) _mm512_permutexvar_epi64(idx, a)
+
+static really_inline u32 movd512(const m512 in) {
+    // NOTE: seems gcc doesn't support _mm512_cvtsi512_si32(in),
+    //       so we use 2-step convertions to work around.
+    return _mm_cvtsi128_si32(_mm512_castsi512_si128(in));
+}
+
+static really_inline
+m512 pshufb_m512(m512 a, m512 b) {
+    return _mm512_shuffle_epi8(a, b);
+}
+
+static really_inline
+m512 maskz_pshufb_m512(__mmask64 k, m512 a, m512 b) {
+    return _mm512_maskz_shuffle_epi8(k, a, b);
+}
+
+#if defined(HAVE_AVX512VBMI)
+#define vpermb512(idx, a) _mm512_permutexvar_epi8(idx, a)
+#define maskz_vpermb512(k, idx, a) _mm512_maskz_permutexvar_epi8(k, idx, a)
+#endif
+
 #define eq512mask(a, b) _mm512_cmpeq_epi8_mask((a), (b))
 #define masked_eq512mask(k, a, b) _mm512_mask_cmpeq_epi8_mask((k), (a), (b))
 
@@ -978,16 +584,10 @@ m512 zeroes512(void) {
 
 static really_inline
 m512 ones512(void) {
-#if defined(HAVE_AVX512)
     return _mm512_set1_epi8(0xFF);
     //return _mm512_xor_si512(_mm512_setzero_si512(), _mm512_setzero_si512());
-#else
-    m512 rv = {ones256(), ones256()};
-    return rv;
-#endif
 }
 
-#if defined(HAVE_AVX512)
 static really_inline
 m512 set1_64x8(u8 a) {
     return _mm512_set1_epi8(a);
@@ -1015,69 +615,32 @@ static really_inline
 m512 set1_4x128(m128 a) {
     return _mm512_broadcast_i32x4(a);
 }
-#endif
 
 static really_inline
 m512 and512(m512 a, m512 b) {
-#if defined(HAVE_AVX512)
     return _mm512_and_si512(a, b);
-#else
-    m512 rv;
-    rv.lo = and256(a.lo, b.lo);
-    rv.hi = and256(a.hi, b.hi);
-    return rv;
-#endif
 }
 
 static really_inline
 m512 or512(m512 a, m512 b) {
-#if defined(HAVE_AVX512)
     return _mm512_or_si512(a, b);
-#else
-    m512 rv;
-    rv.lo = or256(a.lo, b.lo);
-    rv.hi = or256(a.hi, b.hi);
-    return rv;
-#endif
 }
 
 static really_inline
 m512 xor512(m512 a, m512 b) {
-#if defined(HAVE_AVX512)
     return _mm512_xor_si512(a, b);
-#else
-    m512 rv;
-    rv.lo = xor256(a.lo, b.lo);
-    rv.hi = xor256(a.hi, b.hi);
-    return rv;
-#endif
 }
 
 static really_inline
 m512 not512(m512 a) {
-#if defined(HAVE_AVX512)
     return _mm512_xor_si512(a, ones512());
-#else
-    m512 rv;
-    rv.lo = not256(a.lo);
-    rv.hi = not256(a.hi);
-    return rv;
-#endif
 }
 
 static really_inline
 m512 andnot512(m512 a, m512 b) {
-#if defined(HAVE_AVX512)
     return _mm512_andnot_si512(a, b);
-#else
-    m512 rv;
-    rv.lo = andnot256(a.lo, b.lo);
-    rv.hi = andnot256(a.hi, b.hi);
-    return rv;
-#endif
 }
 
-#if defined(HAVE_AVX512)
 static really_really_inline
 m512 lshift64_m512(m512 a, unsigned b) {
 #if defined(HAVE__BUILTIN_CONSTANT_P)
@@ -1088,21 +651,10 @@ m512 lshift64_m512(m512 a, unsigned b) {
     m128 x = _mm_cvtsi32_si128(b);
     return _mm512_sll_epi64(a, x);
 }
-#else
-static really_really_inline
-m512 lshift64_m512(m512 a, unsigned b) {
-    m512 rv;
-    rv.lo = lshift64_m256(a.lo, b);
-    rv.hi = lshift64_m256(a.hi, b);
-    return rv;
-}
-#endif
 
-#if defined(HAVE_AVX512)
 #define rshift64_m512(a, b) _mm512_srli_epi64((a), (b))
 #define rshift128_m512(a, count_immed) _mm512_bsrli_epi128(a, count_immed)
 #define lshift128_m512(a, count_immed) _mm512_bslli_epi128(a, count_immed)
-#endif
 
 #if !defined(_MM_CMPINT_NE)
 #define _MM_CMPINT_NE 0x4
@@ -1110,25 +662,12 @@ m512 lshift64_m512(m512 a, unsigned b) {
 
 static really_inline
 int diff512(m512 a, m512 b) {
-#if defined(HAVE_AVX512)
     return !!_mm512_cmp_epi8_mask(a, b, _MM_CMPINT_NE);
-#else
-    return diff256(a.lo, b.lo) || diff256(a.hi, b.hi);
-#endif
 }
 
 static really_inline
 int isnonzero512(m512 a) {
-#if defined(HAVE_AVX512)
     return diff512(a, zeroes512());
-#elif defined(HAVE_AVX2)
-    m256 x = or256(a.lo, a.hi);
-    return !!diff256(x, zeroes256());
-#else
-    m128 x = or128(a.lo.lo, a.lo.hi);
-    m128 y = or128(a.hi.lo, a.hi.hi);
-    return isnonzero128(or128(x, y));
-#endif
 }
 
 /**
@@ -1137,19 +676,7 @@ int isnonzero512(m512 a) {
  */
 static really_inline
 u32 diffrich512(m512 a, m512 b) {
-#if defined(HAVE_AVX512)
     return _mm512_cmp_epi32_mask(a, b, _MM_CMPINT_NE);
-#elif defined(HAVE_AVX2)
-    return diffrich256(a.lo, b.lo) | (diffrich256(a.hi, b.hi) << 8);
-#else
-    a.lo.lo = _mm_cmpeq_epi32(a.lo.lo, b.lo.lo);
-    a.lo.hi = _mm_cmpeq_epi32(a.lo.hi, b.lo.hi);
-    a.hi.lo = _mm_cmpeq_epi32(a.hi.lo, b.hi.lo);
-    a.hi.hi = _mm_cmpeq_epi32(a.hi.hi, b.hi.hi);
-    m128 packed = _mm_packs_epi16(_mm_packs_epi32(a.lo.lo, a.lo.hi),
-                                  _mm_packs_epi32(a.hi.lo, a.hi.hi));
-    return ~(_mm_movemask_epi8(packed)) & 0xffff;
-#endif
 }
 
 /**
@@ -1166,43 +693,22 @@ u32 diffrich64_512(m512 a, m512 b) {
 // aligned load
 static really_inline
 m512 load512(const void *ptr) {
-#if defined(HAVE_AVX512)
     return _mm512_load_si512(ptr);
-#else
-    assert(ISALIGNED_N(ptr, alignof(m256)));
-    m512 rv = { load256(ptr), load256((const char *)ptr + 32) };
-    return rv;
-#endif
 }
 
 // aligned store
 static really_inline
 void store512(void *ptr, m512 a) {
     assert(ISALIGNED_N(ptr, alignof(m512)));
-#if defined(HAVE_AVX512)
     return _mm512_store_si512(ptr, a);
-#elif defined(HAVE_AVX2)
-    m512 *x = (m512 *)ptr;
-    store256(&x->lo, a.lo);
-    store256(&x->hi, a.hi);
-#else
-    ptr = assume_aligned(ptr, 16);
-    *(m512 *)ptr = a;
-#endif
 }
 
 // unaligned load
 static really_inline
 m512 loadu512(const void *ptr) {
-#if defined(HAVE_AVX512)
     return _mm512_loadu_si512(ptr);
-#else
-    m512 rv = { loadu256(ptr), loadu256((const char *)ptr + 32) };
-    return rv;
-#endif
 }
 
-#if defined(HAVE_AVX512)
 static really_inline
 m512 loadu_maskz_m512(__mmask64 k, const void *ptr) {
     return _mm512_maskz_loadu_epi8(k, ptr);
@@ -1217,7 +723,6 @@ static really_inline
 m512 set_mask_m512(__mmask64 k) {
     return _mm512_movm_epi8(k);
 }
-#endif
 
 // packed unaligned store of first N bytes
 static really_inline
@@ -1247,91 +752,24 @@ m512 mask1bit512(unsigned int n) {
 static really_inline
 void setbit512(m512 *ptr, unsigned int n) {
     assert(n < sizeof(*ptr) * 8);
-#if !defined(HAVE_AVX2)
-    m128 *sub;
-    if (n < 128) {
-        sub = &ptr->lo.lo;
-    } else if (n < 256) {
-        sub = &ptr->lo.hi;
-    } else if (n < 384) {
-        sub = &ptr->hi.lo;
-    } else {
-        sub = &ptr->hi.hi;
-    }
-    setbit128(sub, n % 128);
-#elif defined(HAVE_AVX512)
     *ptr = or512(mask1bit512(n), *ptr);
-#else
-    m256 *sub;
-    if (n < 256) {
-        sub = &ptr->lo;
-    } else {
-        sub = &ptr->hi;
-        n -= 256;
-    }
-    setbit256(sub, n);
-#endif
 }
 
 // switches off bit N in the given vector.
 static really_inline
 void clearbit512(m512 *ptr, unsigned int n) {
     assert(n < sizeof(*ptr) * 8);
-#if !defined(HAVE_AVX2)
-    m128 *sub;
-    if (n < 128) {
-        sub = &ptr->lo.lo;
-    } else if (n < 256) {
-        sub = &ptr->lo.hi;
-    } else if (n < 384) {
-        sub = &ptr->hi.lo;
-    } else {
-        sub = &ptr->hi.hi;
-    }
-    clearbit128(sub, n % 128);
-#elif defined(HAVE_AVX512)
     *ptr = andnot512(mask1bit512(n), *ptr);
-#else
-    m256 *sub;
-    if (n < 256) {
-        sub = &ptr->lo;
-    } else {
-        sub = &ptr->hi;
-        n -= 256;
-    }
-    clearbit256(sub, n);
-#endif
 }
 
 // tests bit N in the given vector.
 static really_inline
 char testbit512(m512 val, unsigned int n) {
     assert(n < sizeof(val) * 8);
-#if !defined(HAVE_AVX2)
-    m128 sub;
-    if (n < 128) {
-        sub = val.lo.lo;
-    } else if (n < 256) {
-        sub = val.lo.hi;
-    } else if (n < 384) {
-        sub = val.hi.lo;
-    } else {
-        sub = val.hi.hi;
-    }
-    return testbit128(sub, n % 128);
-#elif defined(HAVE_AVX512)
     const m512 mask = mask1bit512(n);
     return !!_mm512_test_epi8_mask(mask, val);
-#else
-    m256 sub;
-    if (n < 256) {
-        sub = val.lo;
-    } else {
-        sub = val.hi;
-        n -= 256;
-    }
-    return testbit256(sub, n);
-#endif
 }
 
+#endif // HAVE_SIMD_512_BITS
+
 #endif // ARCH_X86_SIMD_UTILS_H