-[ ]*3[ ]+_start:
-[ ]*4[ ]+mov \{sae\}, %eax\{%k1\}
-[ ]*5[ ]+mov \{sae\}, %eax
-[ ]*6[ ]+mov %ebx, %eax\{%k2\}
-[ ]*7[ ]+vaddps %zmm3, %zmm1, %zmm2\{z\}\{%k1\}\{z\}
-[ ]*8[ ]+vaddps %zmm3, %zmm1\{%k3\}, %zmm2\{z\}
-[ ]*9[ ]+vaddps %zmm3, %zmm1\{%k1\}, %zmm2\{%k2\}
-[ ]*10[ ]+
-[ ]*11[ ]+vcvtps2pd \(%eax\), %zmm1\{1to8\}
-[ ]*12[ ]+vcvtps2pd \(%eax\)\{1to16\}, %zmm1
-[ ]*13[ ]+
-[ ]*14[ ]+vcvtps2pd \(%eax\)\{%k1\}, %zmm1
-[ ]*15[ ]+vcvtps2pd \(%eax\)\{z\}, %zmm1
-[ ]*16[ ]+
-[ ]*17[ ]+vgatherqpd \(%rdi,%zmm2,8\),%zmm6
-[ ]*18[ ]+
-[ ]*19[ ]+\.intel_syntax noprefix
-[ ]*20[ ]+mov eax\{k1\}, \{sae\}
-[ ]*21[ ]+mov eax, \{sae\}
-[ ]*22[ ]+mov eax\{k2\}, ebx
-[ ]*23[ ]+vaddps zmm2\{z\}\{k1\}\{z\}, zmm1, zmm3
-[ ]*24[ ]+vaddps zmm2\{z\}, zmm1\{k3\}, zmm3
-[ ]*25[ ]+vaddps zmm2\{k2\}, zmm1\{k1\}, zmm3
-[ ]*26[ ]+
-[ ]*27[ ]+vcvtps2pd zmm1\{1to8\}, \[eax\]
-[ ]*28[ ]+vcvtps2pd zmm1, \[eax\]\{1to16\}
-[ ]*29[ ]+
-[ ]*30[ ]+vcvtps2pd zmm1, \[eax\]\{k1\}
-[ ]*31[ ]+vcvtps2pd zmm1, \[eax\]\{z\}
-[ ]*32[ ]+
-[ ]*33[ ]+vgatherqpd zmm6, ZMMWORD PTR \[rdi\+zmm2\*8\]
+[ ]*3[ ]+\.allow_index_reg
+[ ]*4[ ]+_start:
+[ ]*5[ ]+mov \{sae\}, %eax\{%k1\}
+[ ]*6[ ]+mov \{sae\}, %eax
+[ ]*7[ ]+mov %ebx, %eax\{%k2\}
+[ ]*8[ ]+vaddps %zmm3, %zmm1, %zmm2\{z\}\{%k1\}\{z\}
+[ ]*9[ ]+vaddps %zmm3, %zmm1\{%k3\}, %zmm2\{z\}
+[ ]*10[ ]+vaddps %zmm3, %zmm1\{%k1\}, %zmm2\{%k2\}
+[ ]*11[ ]+
+[ ]*12[ ]+vcvtps2pd \(%eax\), %zmm1\{1to8\}
+[ ]*13[ ]+vcvtps2pd \(%eax\)\{1to16\}, %zmm1
+[ ]*14[ ]+
+[ ]*15[ ]+vcvtps2pd \(%eax\)\{%k1\}, %zmm1
+[ ]*16[ ]+vcvtps2pd \(%eax\)\{z\}, %zmm1
+[ ]*17[ ]+
+[ ]*18[ ]+vgatherqpd \(%rdi,%zmm2,8\),%zmm6
+[ ]*19[ ]+vgatherqpd \(%edi\),%zmm6\{%k1\}
+[ ]*20[ ]+vgatherqpd \(%zmm2\),%zmm6\{%k1\}
+[ ]*21[ ]+vpscatterdd %zmm6,\(%edi\)\{%k1\}
+[ ]*22[ ]+vpscatterdd %zmm6,\(%zmm2\)\{%k1\}
+[ ]*23[ ]+
+[ ]*24[ ]+\.intel_syntax noprefix
+[ ]*25[ ]+mov eax\{k1\}, \{sae\}
+[ ]*26[ ]+mov eax, \{sae\}
+[ ]*27[ ]+mov eax\{k2\}, ebx
+[ ]*28[ ]+vaddps zmm2\{z\}\{k1\}\{z\}, zmm1, zmm3
+[ ]*29[ ]+vaddps zmm2\{z\}, zmm1\{k3\}, zmm3
+[ ]*30[ ]+vaddps zmm2\{k2\}, zmm1\{k1\}, zmm3
+[ ]*31[ ]+
+[ ]*32[ ]+vcvtps2pd zmm1\{1to8\}, \[eax\]
+[ ]*33[ ]+vcvtps2pd zmm1, \[eax\]\{1to16\}