diff --git a/crypto_sign/dilithium2/avx2/invntt.S b/crypto_sign/dilithium2/avx2/invntt.S
index 17cf515a..5a5ee1ca 100644
--- a/crypto_sign/dilithium2/avx2/invntt.S
+++ b/crypto_sign/dilithium2/avx2/invntt.S
@@ -45,9 +45,9 @@ vpsrlq		$32,%ymm\h3,%ymm\h3
 .global cdecl(PQCLEAN_DILITHIUM2_AVX2_invntt_levels0t4_avx)
 cdecl(PQCLEAN_DILITHIUM2_AVX2_invntt_levels0t4_avx):
 #consts
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM2_AVX2_8xqinv)(%rip),%ymm0
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM2_AVX2_8xq)(%rip),%ymm1
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM2_AVX2_8x256q)(%rip),%ymm2
+vmovdqa		cdecl(PQCLEAN_DILITHIUM2_AVX2_8xqinv)(%rip),%ymm0
+vmovdqa		cdecl(PQCLEAN_DILITHIUM2_AVX2_8xq)(%rip),%ymm1
+vmovdqa		cdecl(PQCLEAN_DILITHIUM2_AVX2_8x256q)(%rip),%ymm2
 
 #load
 vmovdqa		(%rsi),%ymm6
@@ -165,9 +165,9 @@ ret
 .global cdecl(PQCLEAN_DILITHIUM2_AVX2_invntt_levels5t7_avx)
 cdecl(PQCLEAN_DILITHIUM2_AVX2_invntt_levels5t7_avx):
 #consts
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM2_AVX2_8xqinv)(%rip),%ymm0
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM2_AVX2_8xq)(%rip),%ymm1
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM2_AVX2_8x256q)(%rip),%ymm2
+vmovdqa		cdecl(PQCLEAN_DILITHIUM2_AVX2_8xqinv)(%rip),%ymm0
+vmovdqa		cdecl(PQCLEAN_DILITHIUM2_AVX2_8xq)(%rip),%ymm1
+vmovdqa		cdecl(PQCLEAN_DILITHIUM2_AVX2_8x256q)(%rip),%ymm2
 
 #load
 vmovdqa		(%rsi),%ymm4
@@ -237,7 +237,7 @@ vpbroadcastd	24(%rdx),%ymm3
 butterfly	4,5,6,7,8,9,10,11,3,3
 
 #consts
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM2_AVX2_8xdiv)(%rip),%ymm3
+vmovdqa		cdecl(PQCLEAN_DILITHIUM2_AVX2_8xdiv)(%rip),%ymm3
 
 vpmuludq        %ymm3,%ymm4,%ymm4
 vpmuludq        %ymm3,%ymm5,%ymm5
@@ -261,7 +261,7 @@ vpsrlq          $32,%ymm6,%ymm6
 vpsrlq          $32,%ymm7,%ymm7
 
 #store
-vmovdqa         cdecl(_PQCLEAN_DILITHIUM2_AVX2_mask)(%rip),%ymm3
+vmovdqa         cdecl(PQCLEAN_DILITHIUM2_AVX2_mask)(%rip),%ymm3
 vpermd          %ymm4,%ymm3,%ymm4
 vpermd          %ymm5,%ymm3,%ymm5
 vpermd          %ymm6,%ymm3,%ymm6
diff --git a/crypto_sign/dilithium2/avx2/ntt.S b/crypto_sign/dilithium2/avx2/ntt.S
index 4110ef4a..1e1f7754 100644
--- a/crypto_sign/dilithium2/avx2/ntt.S
+++ b/crypto_sign/dilithium2/avx2/ntt.S
@@ -44,9 +44,9 @@ vpsubd		%ymm15,%ymm\rh3,%ymm\rh3
 .global cdecl(PQCLEAN_DILITHIUM2_AVX2_ntt_levels0t2_avx)
 cdecl(PQCLEAN_DILITHIUM2_AVX2_ntt_levels0t2_avx):
 #consts
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM2_AVX2_8xqinv)(%rip),%ymm0
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM2_AVX2_8xq)(%rip),%ymm1
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM2_AVX2_8x2q)(%rip),%ymm2
+vmovdqa		cdecl(PQCLEAN_DILITHIUM2_AVX2_8xqinv)(%rip),%ymm0
+vmovdqa		cdecl(PQCLEAN_DILITHIUM2_AVX2_8xq)(%rip),%ymm1
+vmovdqa		cdecl(PQCLEAN_DILITHIUM2_AVX2_8x2q)(%rip),%ymm2
 
 level0:
 #zetas
@@ -95,9 +95,9 @@ ret
 .global cdecl(PQCLEAN_DILITHIUM2_AVX2_ntt_levels3t8_avx)
 cdecl(PQCLEAN_DILITHIUM2_AVX2_ntt_levels3t8_avx):
 #consts
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM2_AVX2_8xqinv)(%rip),%ymm0
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM2_AVX2_8xq)(%rip),%ymm1
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM2_AVX2_8x2q)(%rip),%ymm2
+vmovdqa		cdecl(PQCLEAN_DILITHIUM2_AVX2_8xqinv)(%rip),%ymm0
+vmovdqa		cdecl(PQCLEAN_DILITHIUM2_AVX2_8xq)(%rip),%ymm1
+vmovdqa		cdecl(PQCLEAN_DILITHIUM2_AVX2_8x2q)(%rip),%ymm2
 
 #load
 vmovdqa		(%rsi),%ymm4
diff --git a/crypto_sign/dilithium2/avx2/nttconsts.c b/crypto_sign/dilithium2/avx2/nttconsts.c
index 7fff311d..a351ab32 100644
--- a/crypto_sign/dilithium2/avx2/nttconsts.c
+++ b/crypto_sign/dilithium2/avx2/nttconsts.c
@@ -5,19 +5,19 @@
 #define DIV (((MONT*MONT % Q) * (Q-1) % Q) * ((Q-1) >> 8) % Q)
 
 
-const aligned_uint32x8_t _PQCLEAN_DILITHIUM2_AVX2_8xqinv = {.as_arr = {QINV, QINV, QINV, QINV, QINV, QINV, QINV, QINV}};
-const aligned_uint32x8_t _PQCLEAN_DILITHIUM2_AVX2_8xq = {.as_arr = {Q, Q, Q, Q, Q, Q, Q, Q}};
-const aligned_uint32x8_t _PQCLEAN_DILITHIUM2_AVX2_8x2q = {.as_arr = {2 * Q, 2 * Q, 2 * Q, 2 * Q, 2 * Q, 2 * Q, 2 * Q, 2 * Q}};
-const aligned_uint32x8_t _PQCLEAN_DILITHIUM2_AVX2_8x256q = {.as_arr = {256 * Q, 256 * Q, 256 * Q, 256 * Q, 256 * Q, 256 * Q, 256 * Q,
-                                                                       256 * Q
-                                                                      }
-                                                           };
-const aligned_uint32x8_t _PQCLEAN_DILITHIUM2_AVX2_mask = {.as_arr = {0, 2, 4, 6, 0, 0, 0, 0}};
-const aligned_uint32x8_t _PQCLEAN_DILITHIUM2_AVX2_8x23ones = {.as_arr = {0x7FFFFF, 0x7FFFFF, 0x7FFFFF, 0x7FFFFF, 0x7FFFFF, 0x7FFFFF,
-                                                                         0x7FFFFF, 0x7FFFFF
-                                                                        }
-                                                             };
-const aligned_uint32x8_t _PQCLEAN_DILITHIUM2_AVX2_8xdiv = { .as_arr = {DIV, DIV, DIV, DIV, DIV, DIV, DIV, DIV}};
+const aligned_uint32x8_t PQCLEAN_DILITHIUM2_AVX2_8xqinv = {.as_arr = {QINV, QINV, QINV, QINV, QINV, QINV, QINV, QINV}};
+const aligned_uint32x8_t PQCLEAN_DILITHIUM2_AVX2_8xq = {.as_arr = {Q, Q, Q, Q, Q, Q, Q, Q}};
+const aligned_uint32x8_t PQCLEAN_DILITHIUM2_AVX2_8x2q = {.as_arr = {2 * Q, 2 * Q, 2 * Q, 2 * Q, 2 * Q, 2 * Q, 2 * Q, 2 * Q}};
+const aligned_uint32x8_t PQCLEAN_DILITHIUM2_AVX2_8x256q = {.as_arr = {256 * Q, 256 * Q, 256 * Q, 256 * Q, 256 * Q, 256 * Q, 256 * Q,
+                                                                      256 * Q
+                                                                     }
+                                                          };
+const aligned_uint32x8_t PQCLEAN_DILITHIUM2_AVX2_mask = {.as_arr = {0, 2, 4, 6, 0, 0, 0, 0}};
+const aligned_uint32x8_t PQCLEAN_DILITHIUM2_AVX2_8x23ones = {.as_arr = {0x7FFFFF, 0x7FFFFF, 0x7FFFFF, 0x7FFFFF, 0x7FFFFF, 0x7FFFFF,
+                                                                        0x7FFFFF, 0x7FFFFF
+                                                                       }
+                                                            };
+const aligned_uint32x8_t PQCLEAN_DILITHIUM2_AVX2_8xdiv = { .as_arr = {DIV, DIV, DIV, DIV, DIV, DIV, DIV, DIV}};
 
 #undef QINV
 #undef MONT
diff --git a/crypto_sign/dilithium2/avx2/nttconsts.h b/crypto_sign/dilithium2/avx2/nttconsts.h
index c2dd6b51..107bdc87 100644
--- a/crypto_sign/dilithium2/avx2/nttconsts.h
+++ b/crypto_sign/dilithium2/avx2/nttconsts.h
@@ -12,13 +12,13 @@ typedef ALIGNED_UINT32(8) aligned_uint32x8_t;
 typedef ALIGNED_UINT32(N) aligned_uint32xN_t;
 
 
-extern const aligned_uint32x8_t _PQCLEAN_DILITHIUM2_AVX2_8xqinv;
-extern const aligned_uint32x8_t _PQCLEAN_DILITHIUM2_AVX2_8xq;
-extern const aligned_uint32x8_t _PQCLEAN_DILITHIUM2_AVX2_8x2q;
-extern const aligned_uint32x8_t _PQCLEAN_DILITHIUM2_AVX2_8x256q;
-extern const aligned_uint32x8_t _PQCLEAN_DILITHIUM2_AVX2_mask;
-extern const aligned_uint32x8_t _PQCLEAN_DILITHIUM2_AVX2_8x23ones;
-extern const aligned_uint32x8_t _PQCLEAN_DILITHIUM2_AVX2_8xdiv;
+extern const aligned_uint32x8_t PQCLEAN_DILITHIUM2_AVX2_8xqinv;
+extern const aligned_uint32x8_t PQCLEAN_DILITHIUM2_AVX2_8xq;
+extern const aligned_uint32x8_t PQCLEAN_DILITHIUM2_AVX2_8x2q;
+extern const aligned_uint32x8_t PQCLEAN_DILITHIUM2_AVX2_8x256q;
+extern const aligned_uint32x8_t PQCLEAN_DILITHIUM2_AVX2_mask;
+extern const aligned_uint32x8_t PQCLEAN_DILITHIUM2_AVX2_8x23ones;
+extern const aligned_uint32x8_t PQCLEAN_DILITHIUM2_AVX2_8xdiv;
 
 extern const aligned_uint32xN_t PQCLEAN_DILITHIUM2_AVX2_zetas;
 extern const aligned_uint32xN_t PQCLEAN_DILITHIUM2_AVX2_zetas_inv;
diff --git a/crypto_sign/dilithium2/avx2/pointwise.S b/crypto_sign/dilithium2/avx2/pointwise.S
index d0132791..1f638b1e 100644
--- a/crypto_sign/dilithium2/avx2/pointwise.S
+++ b/crypto_sign/dilithium2/avx2/pointwise.S
@@ -4,8 +4,8 @@
 .global cdecl(PQCLEAN_DILITHIUM2_AVX2_pointwise_avx)
 cdecl(PQCLEAN_DILITHIUM2_AVX2_pointwise_avx):
 #consts
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM2_AVX2_8xqinv)(%rip),%ymm0
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM2_AVX2_8xq)(%rip),%ymm1
+vmovdqa		cdecl(PQCLEAN_DILITHIUM2_AVX2_8xqinv)(%rip),%ymm0
+vmovdqa		cdecl(PQCLEAN_DILITHIUM2_AVX2_8xq)(%rip),%ymm1
 
 xor		%eax,%eax
 _looptop1:
@@ -136,8 +136,8 @@ vpaddq		%ymm9,%ymm5,%ymm5
 .global cdecl(PQCLEAN_DILITHIUM2_AVX2_pointwise_acc_avx)
 cdecl(PQCLEAN_DILITHIUM2_AVX2_pointwise_acc_avx):
 #consts
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM2_AVX2_8xqinv)(%rip),%ymm0
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM2_AVX2_8xq)(%rip),%ymm1
+vmovdqa		cdecl(PQCLEAN_DILITHIUM2_AVX2_8xqinv)(%rip),%ymm0
+vmovdqa		cdecl(PQCLEAN_DILITHIUM2_AVX2_8xq)(%rip),%ymm1
 
 xor		%eax,%eax
 _looptop2:
diff --git a/crypto_sign/dilithium2/avx2/poly.c b/crypto_sign/dilithium2/avx2/poly.c
index 24a6ea06..a3f4dfc9 100644
--- a/crypto_sign/dilithium2/avx2/poly.c
+++ b/crypto_sign/dilithium2/avx2/poly.c
@@ -82,7 +82,7 @@ void PQCLEAN_DILITHIUM2_AVX2_poly_add(poly *c, const poly *a, const poly *b)  {
 **************************************************/
 void PQCLEAN_DILITHIUM2_AVX2_poly_sub(poly *c, const poly *a, const poly *b) {
     __m256i vec0, vec1;
-    const __m256i twoq = _mm256_load_si256(_PQCLEAN_DILITHIUM2_AVX2_8x2q.as_vec);
+    const __m256i twoq = _mm256_load_si256(PQCLEAN_DILITHIUM2_AVX2_8x2q.as_vec);
 
     for (size_t i = 0; i < N / 8; i++) {
         vec0 = _mm256_load_si256(&a->coeffs_x8[i]);
diff --git a/crypto_sign/dilithium2/avx2/reduce.S b/crypto_sign/dilithium2/avx2/reduce.S
index c02d5973..941828d5 100644
--- a/crypto_sign/dilithium2/avx2/reduce.S
+++ b/crypto_sign/dilithium2/avx2/reduce.S
@@ -3,7 +3,7 @@
 .global cdecl(PQCLEAN_DILITHIUM2_AVX2_reduce_avx)
 cdecl(PQCLEAN_DILITHIUM2_AVX2_reduce_avx):
 #consts
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM2_AVX2_8x23ones)(%rip),%ymm0
+vmovdqa		cdecl(PQCLEAN_DILITHIUM2_AVX2_8x23ones)(%rip),%ymm0
 
 xor		%eax,%eax
 _looptop_rdc32:
@@ -51,7 +51,7 @@ ret
 .global cdecl(PQCLEAN_DILITHIUM2_AVX2_csubq_avx)
 cdecl(PQCLEAN_DILITHIUM2_AVX2_csubq_avx):
 #consts
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM2_AVX2_8xq)(%rip),%ymm0
+vmovdqa		cdecl(PQCLEAN_DILITHIUM2_AVX2_8xq)(%rip),%ymm0
 
 xor		%eax,%eax
 _looptop_csubq:
diff --git a/crypto_sign/dilithium3/avx2/invntt.S b/crypto_sign/dilithium3/avx2/invntt.S
index 6588e5ef..0dbfacf6 100644
--- a/crypto_sign/dilithium3/avx2/invntt.S
+++ b/crypto_sign/dilithium3/avx2/invntt.S
@@ -45,9 +45,9 @@ vpsrlq		$32,%ymm\h3,%ymm\h3
 .global cdecl(PQCLEAN_DILITHIUM3_AVX2_invntt_levels0t4_avx)
 cdecl(PQCLEAN_DILITHIUM3_AVX2_invntt_levels0t4_avx):
 #consts
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM3_AVX2_8xqinv)(%rip),%ymm0
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM3_AVX2_8xq)(%rip),%ymm1
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM3_AVX2_8x256q)(%rip),%ymm2
+vmovdqa		cdecl(PQCLEAN_DILITHIUM3_AVX2_8xqinv)(%rip),%ymm0
+vmovdqa		cdecl(PQCLEAN_DILITHIUM3_AVX2_8xq)(%rip),%ymm1
+vmovdqa		cdecl(PQCLEAN_DILITHIUM3_AVX2_8x256q)(%rip),%ymm2
 
 #load
 vmovdqa		(%rsi),%ymm6
@@ -165,9 +165,9 @@ ret
 .global cdecl(PQCLEAN_DILITHIUM3_AVX2_invntt_levels5t7_avx)
 cdecl(PQCLEAN_DILITHIUM3_AVX2_invntt_levels5t7_avx):
 #consts
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM3_AVX2_8xqinv)(%rip),%ymm0
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM3_AVX2_8xq)(%rip),%ymm1
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM3_AVX2_8x256q)(%rip),%ymm2
+vmovdqa		cdecl(PQCLEAN_DILITHIUM3_AVX2_8xqinv)(%rip),%ymm0
+vmovdqa		cdecl(PQCLEAN_DILITHIUM3_AVX2_8xq)(%rip),%ymm1
+vmovdqa		cdecl(PQCLEAN_DILITHIUM3_AVX2_8x256q)(%rip),%ymm2
 
 #load
 vmovdqa		(%rsi),%ymm4
@@ -237,7 +237,7 @@ vpbroadcastd	24(%rdx),%ymm3
 butterfly	4,5,6,7,8,9,10,11,3,3
 
 #consts
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM3_AVX2_8xdiv)(%rip),%ymm3
+vmovdqa		cdecl(PQCLEAN_DILITHIUM3_AVX2_8xdiv)(%rip),%ymm3
 
 vpmuludq        %ymm3,%ymm4,%ymm4
 vpmuludq        %ymm3,%ymm5,%ymm5
@@ -261,7 +261,7 @@ vpsrlq          $32,%ymm6,%ymm6
 vpsrlq          $32,%ymm7,%ymm7
 
 #store
-vmovdqa         cdecl(_PQCLEAN_DILITHIUM3_AVX2_mask)(%rip),%ymm3
+vmovdqa         cdecl(PQCLEAN_DILITHIUM3_AVX2_mask)(%rip),%ymm3
 vpermd          %ymm4,%ymm3,%ymm4
 vpermd          %ymm5,%ymm3,%ymm5
 vpermd          %ymm6,%ymm3,%ymm6
diff --git a/crypto_sign/dilithium3/avx2/ntt.S b/crypto_sign/dilithium3/avx2/ntt.S
index db959478..4cb18d8b 100644
--- a/crypto_sign/dilithium3/avx2/ntt.S
+++ b/crypto_sign/dilithium3/avx2/ntt.S
@@ -44,9 +44,9 @@ vpsubd		%ymm15,%ymm\rh3,%ymm\rh3
 .global cdecl(PQCLEAN_DILITHIUM3_AVX2_ntt_levels0t2_avx)
 cdecl(PQCLEAN_DILITHIUM3_AVX2_ntt_levels0t2_avx):
 #consts
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM3_AVX2_8xqinv)(%rip),%ymm0
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM3_AVX2_8xq)(%rip),%ymm1
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM3_AVX2_8x2q)(%rip),%ymm2
+vmovdqa		cdecl(PQCLEAN_DILITHIUM3_AVX2_8xqinv)(%rip),%ymm0
+vmovdqa		cdecl(PQCLEAN_DILITHIUM3_AVX2_8xq)(%rip),%ymm1
+vmovdqa		cdecl(PQCLEAN_DILITHIUM3_AVX2_8x2q)(%rip),%ymm2
 
 level0:
 #zetas
@@ -95,9 +95,9 @@ ret
 .global cdecl(PQCLEAN_DILITHIUM3_AVX2_ntt_levels3t8_avx)
 cdecl(PQCLEAN_DILITHIUM3_AVX2_ntt_levels3t8_avx):
 #consts
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM3_AVX2_8xqinv)(%rip),%ymm0
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM3_AVX2_8xq)(%rip),%ymm1
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM3_AVX2_8x2q)(%rip),%ymm2
+vmovdqa		cdecl(PQCLEAN_DILITHIUM3_AVX2_8xqinv)(%rip),%ymm0
+vmovdqa		cdecl(PQCLEAN_DILITHIUM3_AVX2_8xq)(%rip),%ymm1
+vmovdqa		cdecl(PQCLEAN_DILITHIUM3_AVX2_8x2q)(%rip),%ymm2
 
 #load
 vmovdqa		(%rsi),%ymm4
diff --git a/crypto_sign/dilithium3/avx2/nttconsts.c b/crypto_sign/dilithium3/avx2/nttconsts.c
index 12d9ceac..2bd3b20b 100644
--- a/crypto_sign/dilithium3/avx2/nttconsts.c
+++ b/crypto_sign/dilithium3/avx2/nttconsts.c
@@ -5,19 +5,19 @@
 #define DIV (((MONT*MONT % Q) * (Q-1) % Q) * ((Q-1) >> 8) % Q)
 
 
-const aligned_uint32x8_t _PQCLEAN_DILITHIUM3_AVX2_8xqinv = {.as_arr = {QINV, QINV, QINV, QINV, QINV, QINV, QINV, QINV}};
-const aligned_uint32x8_t _PQCLEAN_DILITHIUM3_AVX2_8xq = {.as_arr = {Q, Q, Q, Q, Q, Q, Q, Q}};
-const aligned_uint32x8_t _PQCLEAN_DILITHIUM3_AVX2_8x2q = {.as_arr = {2 * Q, 2 * Q, 2 * Q, 2 * Q, 2 * Q, 2 * Q, 2 * Q, 2 * Q}};
-const aligned_uint32x8_t _PQCLEAN_DILITHIUM3_AVX2_8x256q = {.as_arr = {256 * Q, 256 * Q, 256 * Q, 256 * Q, 256 * Q, 256 * Q, 256 * Q,
-                                                                       256 * Q
-                                                                      }
-                                                           };
-const aligned_uint32x8_t _PQCLEAN_DILITHIUM3_AVX2_mask = {.as_arr = {0, 2, 4, 6, 0, 0, 0, 0}};
-const aligned_uint32x8_t _PQCLEAN_DILITHIUM3_AVX2_8x23ones = {.as_arr = {0x7FFFFF, 0x7FFFFF, 0x7FFFFF, 0x7FFFFF, 0x7FFFFF, 0x7FFFFF,
-                                                                         0x7FFFFF, 0x7FFFFF
-                                                                        }
-                                                             };
-const aligned_uint32x8_t _PQCLEAN_DILITHIUM3_AVX2_8xdiv = { .as_arr = {DIV, DIV, DIV, DIV, DIV, DIV, DIV, DIV}};
+const aligned_uint32x8_t PQCLEAN_DILITHIUM3_AVX2_8xqinv = {.as_arr = {QINV, QINV, QINV, QINV, QINV, QINV, QINV, QINV}};
+const aligned_uint32x8_t PQCLEAN_DILITHIUM3_AVX2_8xq = {.as_arr = {Q, Q, Q, Q, Q, Q, Q, Q}};
+const aligned_uint32x8_t PQCLEAN_DILITHIUM3_AVX2_8x2q = {.as_arr = {2 * Q, 2 * Q, 2 * Q, 2 * Q, 2 * Q, 2 * Q, 2 * Q, 2 * Q}};
+const aligned_uint32x8_t PQCLEAN_DILITHIUM3_AVX2_8x256q = {.as_arr = {256 * Q, 256 * Q, 256 * Q, 256 * Q, 256 * Q, 256 * Q, 256 * Q,
+                                                                      256 * Q
+                                                                     }
+                                                          };
+const aligned_uint32x8_t PQCLEAN_DILITHIUM3_AVX2_mask = {.as_arr = {0, 2, 4, 6, 0, 0, 0, 0}};
+const aligned_uint32x8_t PQCLEAN_DILITHIUM3_AVX2_8x23ones = {.as_arr = {0x7FFFFF, 0x7FFFFF, 0x7FFFFF, 0x7FFFFF, 0x7FFFFF, 0x7FFFFF,
+                                                                        0x7FFFFF, 0x7FFFFF
+                                                                       }
+                                                            };
+const aligned_uint32x8_t PQCLEAN_DILITHIUM3_AVX2_8xdiv = { .as_arr = {DIV, DIV, DIV, DIV, DIV, DIV, DIV, DIV}};
 
 #undef QINV
 #undef MONT
diff --git a/crypto_sign/dilithium3/avx2/nttconsts.h b/crypto_sign/dilithium3/avx2/nttconsts.h
index ed8df189..caf6945d 100644
--- a/crypto_sign/dilithium3/avx2/nttconsts.h
+++ b/crypto_sign/dilithium3/avx2/nttconsts.h
@@ -12,13 +12,13 @@ typedef ALIGNED_UINT32(8) aligned_uint32x8_t;
 typedef ALIGNED_UINT32(N) aligned_uint32xN_t;
 
 
-extern const aligned_uint32x8_t _PQCLEAN_DILITHIUM3_AVX2_8xqinv;
-extern const aligned_uint32x8_t _PQCLEAN_DILITHIUM3_AVX2_8xq;
-extern const aligned_uint32x8_t _PQCLEAN_DILITHIUM3_AVX2_8x2q;
-extern const aligned_uint32x8_t _PQCLEAN_DILITHIUM3_AVX2_8x256q;
-extern const aligned_uint32x8_t _PQCLEAN_DILITHIUM3_AVX2_mask;
-extern const aligned_uint32x8_t _PQCLEAN_DILITHIUM3_AVX2_8x23ones;
-extern const aligned_uint32x8_t _PQCLEAN_DILITHIUM3_AVX2_8xdiv;
+extern const aligned_uint32x8_t PQCLEAN_DILITHIUM3_AVX2_8xqinv;
+extern const aligned_uint32x8_t PQCLEAN_DILITHIUM3_AVX2_8xq;
+extern const aligned_uint32x8_t PQCLEAN_DILITHIUM3_AVX2_8x2q;
+extern const aligned_uint32x8_t PQCLEAN_DILITHIUM3_AVX2_8x256q;
+extern const aligned_uint32x8_t PQCLEAN_DILITHIUM3_AVX2_mask;
+extern const aligned_uint32x8_t PQCLEAN_DILITHIUM3_AVX2_8x23ones;
+extern const aligned_uint32x8_t PQCLEAN_DILITHIUM3_AVX2_8xdiv;
 
 extern const aligned_uint32xN_t PQCLEAN_DILITHIUM3_AVX2_zetas;
 extern const aligned_uint32xN_t PQCLEAN_DILITHIUM3_AVX2_zetas_inv;
diff --git a/crypto_sign/dilithium3/avx2/pointwise.S b/crypto_sign/dilithium3/avx2/pointwise.S
index 4aca6373..b0085373 100644
--- a/crypto_sign/dilithium3/avx2/pointwise.S
+++ b/crypto_sign/dilithium3/avx2/pointwise.S
@@ -4,8 +4,8 @@
 .global cdecl(PQCLEAN_DILITHIUM3_AVX2_pointwise_avx)
 cdecl(PQCLEAN_DILITHIUM3_AVX2_pointwise_avx):
 #consts
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM3_AVX2_8xqinv)(%rip),%ymm0
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM3_AVX2_8xq)(%rip),%ymm1
+vmovdqa		cdecl(PQCLEAN_DILITHIUM3_AVX2_8xqinv)(%rip),%ymm0
+vmovdqa		cdecl(PQCLEAN_DILITHIUM3_AVX2_8xq)(%rip),%ymm1
 
 xor		%eax,%eax
 _looptop1:
@@ -136,8 +136,8 @@ vpaddq		%ymm9,%ymm5,%ymm5
 .global cdecl(PQCLEAN_DILITHIUM3_AVX2_pointwise_acc_avx)
 cdecl(PQCLEAN_DILITHIUM3_AVX2_pointwise_acc_avx):
 #consts
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM3_AVX2_8xqinv)(%rip),%ymm0
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM3_AVX2_8xq)(%rip),%ymm1
+vmovdqa		cdecl(PQCLEAN_DILITHIUM3_AVX2_8xqinv)(%rip),%ymm0
+vmovdqa		cdecl(PQCLEAN_DILITHIUM3_AVX2_8xq)(%rip),%ymm1
 
 xor		%eax,%eax
 _looptop2:
diff --git a/crypto_sign/dilithium3/avx2/poly.c b/crypto_sign/dilithium3/avx2/poly.c
index 3f4223c3..f6876f81 100644
--- a/crypto_sign/dilithium3/avx2/poly.c
+++ b/crypto_sign/dilithium3/avx2/poly.c
@@ -82,7 +82,7 @@ void PQCLEAN_DILITHIUM3_AVX2_poly_add(poly *c, const poly *a, const poly *b)  {
 **************************************************/
 void PQCLEAN_DILITHIUM3_AVX2_poly_sub(poly *c, const poly *a, const poly *b) {
     __m256i vec0, vec1;
-    const __m256i twoq = _mm256_load_si256(_PQCLEAN_DILITHIUM3_AVX2_8x2q.as_vec);
+    const __m256i twoq = _mm256_load_si256(PQCLEAN_DILITHIUM3_AVX2_8x2q.as_vec);
 
     for (size_t i = 0; i < N / 8; i++) {
         vec0 = _mm256_load_si256(&a->coeffs_x8[i]);
diff --git a/crypto_sign/dilithium3/avx2/reduce.S b/crypto_sign/dilithium3/avx2/reduce.S
index 1847274f..cef9a7a4 100644
--- a/crypto_sign/dilithium3/avx2/reduce.S
+++ b/crypto_sign/dilithium3/avx2/reduce.S
@@ -3,7 +3,7 @@
 .global cdecl(PQCLEAN_DILITHIUM3_AVX2_reduce_avx)
 cdecl(PQCLEAN_DILITHIUM3_AVX2_reduce_avx):
 #consts
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM3_AVX2_8x23ones)(%rip),%ymm0
+vmovdqa		cdecl(PQCLEAN_DILITHIUM3_AVX2_8x23ones)(%rip),%ymm0
 
 xor		%eax,%eax
 _looptop_rdc32:
@@ -51,7 +51,7 @@ ret
 .global cdecl(PQCLEAN_DILITHIUM3_AVX2_csubq_avx)
 cdecl(PQCLEAN_DILITHIUM3_AVX2_csubq_avx):
 #consts
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM3_AVX2_8xq)(%rip),%ymm0
+vmovdqa		cdecl(PQCLEAN_DILITHIUM3_AVX2_8xq)(%rip),%ymm0
 
 xor		%eax,%eax
 _looptop_csubq:
diff --git a/crypto_sign/dilithium4/avx2/invntt.S b/crypto_sign/dilithium4/avx2/invntt.S
index 2e8a4c02..6c94513b 100644
--- a/crypto_sign/dilithium4/avx2/invntt.S
+++ b/crypto_sign/dilithium4/avx2/invntt.S
@@ -45,9 +45,9 @@ vpsrlq		$32,%ymm\h3,%ymm\h3
 .global cdecl(PQCLEAN_DILITHIUM4_AVX2_invntt_levels0t4_avx)
 cdecl(PQCLEAN_DILITHIUM4_AVX2_invntt_levels0t4_avx):
 #consts
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM4_AVX2_8xqinv)(%rip),%ymm0
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM4_AVX2_8xq)(%rip),%ymm1
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM4_AVX2_8x256q)(%rip),%ymm2
+vmovdqa		cdecl(PQCLEAN_DILITHIUM4_AVX2_8xqinv)(%rip),%ymm0
+vmovdqa		cdecl(PQCLEAN_DILITHIUM4_AVX2_8xq)(%rip),%ymm1
+vmovdqa		cdecl(PQCLEAN_DILITHIUM4_AVX2_8x256q)(%rip),%ymm2
 
 #load
 vmovdqa		(%rsi),%ymm6
@@ -165,9 +165,9 @@ ret
 .global cdecl(PQCLEAN_DILITHIUM4_AVX2_invntt_levels5t7_avx)
 cdecl(PQCLEAN_DILITHIUM4_AVX2_invntt_levels5t7_avx):
 #consts
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM4_AVX2_8xqinv)(%rip),%ymm0
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM4_AVX2_8xq)(%rip),%ymm1
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM4_AVX2_8x256q)(%rip),%ymm2
+vmovdqa		cdecl(PQCLEAN_DILITHIUM4_AVX2_8xqinv)(%rip),%ymm0
+vmovdqa		cdecl(PQCLEAN_DILITHIUM4_AVX2_8xq)(%rip),%ymm1
+vmovdqa		cdecl(PQCLEAN_DILITHIUM4_AVX2_8x256q)(%rip),%ymm2
 
 #load
 vmovdqa		(%rsi),%ymm4
@@ -237,7 +237,7 @@ vpbroadcastd	24(%rdx),%ymm3
 butterfly	4,5,6,7,8,9,10,11,3,3
 
 #consts
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM4_AVX2_8xdiv)(%rip),%ymm3
+vmovdqa		cdecl(PQCLEAN_DILITHIUM4_AVX2_8xdiv)(%rip),%ymm3
 
 vpmuludq        %ymm3,%ymm4,%ymm4
 vpmuludq        %ymm3,%ymm5,%ymm5
@@ -261,7 +261,7 @@ vpsrlq          $32,%ymm6,%ymm6
 vpsrlq          $32,%ymm7,%ymm7
 
 #store
-vmovdqa         cdecl(_PQCLEAN_DILITHIUM4_AVX2_mask)(%rip),%ymm3
+vmovdqa         cdecl(PQCLEAN_DILITHIUM4_AVX2_mask)(%rip),%ymm3
 vpermd          %ymm4,%ymm3,%ymm4
 vpermd          %ymm5,%ymm3,%ymm5
 vpermd          %ymm6,%ymm3,%ymm6
diff --git a/crypto_sign/dilithium4/avx2/ntt.S b/crypto_sign/dilithium4/avx2/ntt.S
index f76fc616..b3f499bc 100644
--- a/crypto_sign/dilithium4/avx2/ntt.S
+++ b/crypto_sign/dilithium4/avx2/ntt.S
@@ -44,9 +44,9 @@ vpsubd		%ymm15,%ymm\rh3,%ymm\rh3
 .global cdecl(PQCLEAN_DILITHIUM4_AVX2_ntt_levels0t2_avx)
 cdecl(PQCLEAN_DILITHIUM4_AVX2_ntt_levels0t2_avx):
 #consts
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM4_AVX2_8xqinv)(%rip),%ymm0
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM4_AVX2_8xq)(%rip),%ymm1
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM4_AVX2_8x2q)(%rip),%ymm2
+vmovdqa		cdecl(PQCLEAN_DILITHIUM4_AVX2_8xqinv)(%rip),%ymm0
+vmovdqa		cdecl(PQCLEAN_DILITHIUM4_AVX2_8xq)(%rip),%ymm1
+vmovdqa		cdecl(PQCLEAN_DILITHIUM4_AVX2_8x2q)(%rip),%ymm2
 
 level0:
 #zetas
@@ -95,9 +95,9 @@ ret
 .global cdecl(PQCLEAN_DILITHIUM4_AVX2_ntt_levels3t8_avx)
 cdecl(PQCLEAN_DILITHIUM4_AVX2_ntt_levels3t8_avx):
 #consts
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM4_AVX2_8xqinv)(%rip),%ymm0
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM4_AVX2_8xq)(%rip),%ymm1
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM4_AVX2_8x2q)(%rip),%ymm2
+vmovdqa		cdecl(PQCLEAN_DILITHIUM4_AVX2_8xqinv)(%rip),%ymm0
+vmovdqa		cdecl(PQCLEAN_DILITHIUM4_AVX2_8xq)(%rip),%ymm1
+vmovdqa		cdecl(PQCLEAN_DILITHIUM4_AVX2_8x2q)(%rip),%ymm2
 
 #load
 vmovdqa		(%rsi),%ymm4
diff --git a/crypto_sign/dilithium4/avx2/nttconsts.c b/crypto_sign/dilithium4/avx2/nttconsts.c
index 7bf71ea7..20aa2120 100644
--- a/crypto_sign/dilithium4/avx2/nttconsts.c
+++ b/crypto_sign/dilithium4/avx2/nttconsts.c
@@ -5,19 +5,19 @@
 #define DIV (((MONT*MONT % Q) * (Q-1) % Q) * ((Q-1) >> 8) % Q)
 
 
-const aligned_uint32x8_t _PQCLEAN_DILITHIUM4_AVX2_8xqinv = {.as_arr = {QINV, QINV, QINV, QINV, QINV, QINV, QINV, QINV}};
-const aligned_uint32x8_t _PQCLEAN_DILITHIUM4_AVX2_8xq = {.as_arr = {Q, Q, Q, Q, Q, Q, Q, Q}};
-const aligned_uint32x8_t _PQCLEAN_DILITHIUM4_AVX2_8x2q = {.as_arr = {2 * Q, 2 * Q, 2 * Q, 2 * Q, 2 * Q, 2 * Q, 2 * Q, 2 * Q}};
-const aligned_uint32x8_t _PQCLEAN_DILITHIUM4_AVX2_8x256q = {.as_arr = {256 * Q, 256 * Q, 256 * Q, 256 * Q, 256 * Q, 256 * Q, 256 * Q,
-                                                                       256 * Q
-                                                                      }
-                                                           };
-const aligned_uint32x8_t _PQCLEAN_DILITHIUM4_AVX2_mask = {.as_arr = {0, 2, 4, 6, 0, 0, 0, 0}};
-const aligned_uint32x8_t _PQCLEAN_DILITHIUM4_AVX2_8x23ones = {.as_arr = {0x7FFFFF, 0x7FFFFF, 0x7FFFFF, 0x7FFFFF, 0x7FFFFF, 0x7FFFFF,
-                                                                         0x7FFFFF, 0x7FFFFF
-                                                                        }
-                                                             };
-const aligned_uint32x8_t _PQCLEAN_DILITHIUM4_AVX2_8xdiv = { .as_arr = {DIV, DIV, DIV, DIV, DIV, DIV, DIV, DIV}};
+const aligned_uint32x8_t PQCLEAN_DILITHIUM4_AVX2_8xqinv = {.as_arr = {QINV, QINV, QINV, QINV, QINV, QINV, QINV, QINV}};
+const aligned_uint32x8_t PQCLEAN_DILITHIUM4_AVX2_8xq = {.as_arr = {Q, Q, Q, Q, Q, Q, Q, Q}};
+const aligned_uint32x8_t PQCLEAN_DILITHIUM4_AVX2_8x2q = {.as_arr = {2 * Q, 2 * Q, 2 * Q, 2 * Q, 2 * Q, 2 * Q, 2 * Q, 2 * Q}};
+const aligned_uint32x8_t PQCLEAN_DILITHIUM4_AVX2_8x256q = {.as_arr = {256 * Q, 256 * Q, 256 * Q, 256 * Q, 256 * Q, 256 * Q, 256 * Q,
+                                                                      256 * Q
+                                                                     }
+                                                          };
+const aligned_uint32x8_t PQCLEAN_DILITHIUM4_AVX2_mask = {.as_arr = {0, 2, 4, 6, 0, 0, 0, 0}};
+const aligned_uint32x8_t PQCLEAN_DILITHIUM4_AVX2_8x23ones = {.as_arr = {0x7FFFFF, 0x7FFFFF, 0x7FFFFF, 0x7FFFFF, 0x7FFFFF, 0x7FFFFF,
+                                                                        0x7FFFFF, 0x7FFFFF
+                                                                       }
+                                                            };
+const aligned_uint32x8_t PQCLEAN_DILITHIUM4_AVX2_8xdiv = { .as_arr = {DIV, DIV, DIV, DIV, DIV, DIV, DIV, DIV}};
 
 #undef QINV
 #undef MONT
diff --git a/crypto_sign/dilithium4/avx2/nttconsts.h b/crypto_sign/dilithium4/avx2/nttconsts.h
index 1904a67b..5a340bbf 100644
--- a/crypto_sign/dilithium4/avx2/nttconsts.h
+++ b/crypto_sign/dilithium4/avx2/nttconsts.h
@@ -12,13 +12,13 @@ typedef ALIGNED_UINT32(8) aligned_uint32x8_t;
 typedef ALIGNED_UINT32(N) aligned_uint32xN_t;
 
 
-extern const aligned_uint32x8_t _PQCLEAN_DILITHIUM4_AVX2_8xqinv;
-extern const aligned_uint32x8_t _PQCLEAN_DILITHIUM4_AVX2_8xq;
-extern const aligned_uint32x8_t _PQCLEAN_DILITHIUM4_AVX2_8x2q;
-extern const aligned_uint32x8_t _PQCLEAN_DILITHIUM4_AVX2_8x256q;
-extern const aligned_uint32x8_t _PQCLEAN_DILITHIUM4_AVX2_mask;
-extern const aligned_uint32x8_t _PQCLEAN_DILITHIUM4_AVX2_8x23ones;
-extern const aligned_uint32x8_t _PQCLEAN_DILITHIUM4_AVX2_8xdiv;
+extern const aligned_uint32x8_t PQCLEAN_DILITHIUM4_AVX2_8xqinv;
+extern const aligned_uint32x8_t PQCLEAN_DILITHIUM4_AVX2_8xq;
+extern const aligned_uint32x8_t PQCLEAN_DILITHIUM4_AVX2_8x2q;
+extern const aligned_uint32x8_t PQCLEAN_DILITHIUM4_AVX2_8x256q;
+extern const aligned_uint32x8_t PQCLEAN_DILITHIUM4_AVX2_mask;
+extern const aligned_uint32x8_t PQCLEAN_DILITHIUM4_AVX2_8x23ones;
+extern const aligned_uint32x8_t PQCLEAN_DILITHIUM4_AVX2_8xdiv;
 
 extern const aligned_uint32xN_t PQCLEAN_DILITHIUM4_AVX2_zetas;
 extern const aligned_uint32xN_t PQCLEAN_DILITHIUM4_AVX2_zetas_inv;
diff --git a/crypto_sign/dilithium4/avx2/pointwise.S b/crypto_sign/dilithium4/avx2/pointwise.S
index 9f61dd58..a9d3ddd3 100644
--- a/crypto_sign/dilithium4/avx2/pointwise.S
+++ b/crypto_sign/dilithium4/avx2/pointwise.S
@@ -4,8 +4,8 @@
 .global cdecl(PQCLEAN_DILITHIUM4_AVX2_pointwise_avx)
 cdecl(PQCLEAN_DILITHIUM4_AVX2_pointwise_avx):
 #consts
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM4_AVX2_8xqinv)(%rip),%ymm0
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM4_AVX2_8xq)(%rip),%ymm1
+vmovdqa		cdecl(PQCLEAN_DILITHIUM4_AVX2_8xqinv)(%rip),%ymm0
+vmovdqa		cdecl(PQCLEAN_DILITHIUM4_AVX2_8xq)(%rip),%ymm1
 
 xor		%eax,%eax
 _looptop1:
@@ -136,8 +136,8 @@ vpaddq		%ymm9,%ymm5,%ymm5
 .global cdecl(PQCLEAN_DILITHIUM4_AVX2_pointwise_acc_avx)
 cdecl(PQCLEAN_DILITHIUM4_AVX2_pointwise_acc_avx):
 #consts
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM4_AVX2_8xqinv)(%rip),%ymm0
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM4_AVX2_8xq)(%rip),%ymm1
+vmovdqa		cdecl(PQCLEAN_DILITHIUM4_AVX2_8xqinv)(%rip),%ymm0
+vmovdqa		cdecl(PQCLEAN_DILITHIUM4_AVX2_8xq)(%rip),%ymm1
 
 xor		%eax,%eax
 _looptop2:
diff --git a/crypto_sign/dilithium4/avx2/poly.c b/crypto_sign/dilithium4/avx2/poly.c
index 6fde3258..f5d28f56 100644
--- a/crypto_sign/dilithium4/avx2/poly.c
+++ b/crypto_sign/dilithium4/avx2/poly.c
@@ -82,7 +82,7 @@ void PQCLEAN_DILITHIUM4_AVX2_poly_add(poly *c, const poly *a, const poly *b)  {
 **************************************************/
 void PQCLEAN_DILITHIUM4_AVX2_poly_sub(poly *c, const poly *a, const poly *b) {
     __m256i vec0, vec1;
-    const __m256i twoq = _mm256_load_si256(_PQCLEAN_DILITHIUM4_AVX2_8x2q.as_vec);
+    const __m256i twoq = _mm256_load_si256(PQCLEAN_DILITHIUM4_AVX2_8x2q.as_vec);
 
     for (size_t i = 0; i < N / 8; i++) {
         vec0 = _mm256_load_si256(&a->coeffs_x8[i]);
diff --git a/crypto_sign/dilithium4/avx2/reduce.S b/crypto_sign/dilithium4/avx2/reduce.S
index c6d226f5..ae1dcdad 100644
--- a/crypto_sign/dilithium4/avx2/reduce.S
+++ b/crypto_sign/dilithium4/avx2/reduce.S
@@ -3,7 +3,7 @@
 .global cdecl(PQCLEAN_DILITHIUM4_AVX2_reduce_avx)
 cdecl(PQCLEAN_DILITHIUM4_AVX2_reduce_avx):
 #consts
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM4_AVX2_8x23ones)(%rip),%ymm0
+vmovdqa		cdecl(PQCLEAN_DILITHIUM4_AVX2_8x23ones)(%rip),%ymm0
 
 xor		%eax,%eax
 _looptop_rdc32:
@@ -51,7 +51,7 @@ ret
 .global cdecl(PQCLEAN_DILITHIUM4_AVX2_csubq_avx)
 cdecl(PQCLEAN_DILITHIUM4_AVX2_csubq_avx):
 #consts
-vmovdqa		cdecl(_PQCLEAN_DILITHIUM4_AVX2_8xq)(%rip),%ymm0
+vmovdqa		cdecl(PQCLEAN_DILITHIUM4_AVX2_8xq)(%rip),%ymm0
 
 xor		%eax,%eax
 _looptop_csubq: