Included appro's updates to AES-NI.

diff --git a/lib/accelerated/intel/asm/appro-aes-x86-64.s b/lib/accelerated/intel/asm/appro-aes-x86-64.s
index 98204d1ba0b5b6b258f1cb2ec2bb93ce87d152eb..f286fb12bdfba3784a223f4788c3b38dc93f9bed 100644 (file)
--- a/lib/accelerated/intel/asm/appro-aes-x86-64.s
+++ b/lib/accelerated/intel/asm/appro-aes-x86-64.s
@@ -42,14 +42,14 @@
 aesni_encrypt:
        movups  (%rdi),%xmm2
        movl    240(%rdx),%eax
-       movaps  (%rdx),%xmm0
-       movaps  16(%rdx),%xmm1
+       movups  (%rdx),%xmm0
+       movups  16(%rdx),%xmm1
        leaq    32(%rdx),%rdx
        xorps   %xmm0,%xmm2
 .Loop_enc1_1:
 .byte  102,15,56,220,209
        decl    %eax
-       movaps  (%rdx),%xmm1
+       movups  (%rdx),%xmm1
        leaq    16(%rdx),%rdx
        jnz     .Loop_enc1_1    
 .byte  102,15,56,221,209
@@ -63,14 +63,14 @@ aesni_encrypt:
 aesni_decrypt:
        movups  (%rdi),%xmm2
        movl    240(%rdx),%eax
-       movaps  (%rdx),%xmm0
-       movaps  16(%rdx),%xmm1
+       movups  (%rdx),%xmm0
+       movups  16(%rdx),%xmm1
        leaq    32(%rdx),%rdx
        xorps   %xmm0,%xmm2
 .Loop_dec1_2:
 .byte  102,15,56,222,209
        decl    %eax
-       movaps  (%rdx),%xmm1
+       movups  (%rdx),%xmm1
        leaq    16(%rdx),%rdx
        jnz     .Loop_dec1_2    
 .byte  102,15,56,223,209
@@ -80,26 +80,26 @@ aesni_decrypt:
 .type  _aesni_encrypt3,@function
 .align 16
 _aesni_encrypt3:
-       movaps  (%rcx),%xmm0
+       movups  (%rcx),%xmm0
        shrl    $1,%eax
-       movaps  16(%rcx),%xmm1
+       movups  16(%rcx),%xmm1
        leaq    32(%rcx),%rcx
        xorps   %xmm0,%xmm2
        xorps   %xmm0,%xmm3
        xorps   %xmm0,%xmm4
-       movaps  (%rcx),%xmm0
+       movups  (%rcx),%xmm0
 
 .Lenc_loop3:
 .byte  102,15,56,220,209
 .byte  102,15,56,220,217
        decl    %eax
 .byte  102,15,56,220,225
-       movaps  16(%rcx),%xmm1
+       movups  16(%rcx),%xmm1
 .byte  102,15,56,220,208
 .byte  102,15,56,220,216
        leaq    32(%rcx),%rcx
 .byte  102,15,56,220,224
-       movaps  (%rcx),%xmm0
+       movups  (%rcx),%xmm0
        jnz     .Lenc_loop3
 
 .byte  102,15,56,220,209
@@ -113,26 +113,26 @@ _aesni_encrypt3:
 .type  _aesni_decrypt3,@function
 .align 16
 _aesni_decrypt3:
-       movaps  (%rcx),%xmm0
+       movups  (%rcx),%xmm0
        shrl    $1,%eax
-       movaps  16(%rcx),%xmm1
+       movups  16(%rcx),%xmm1
        leaq    32(%rcx),%rcx
        xorps   %xmm0,%xmm2
        xorps   %xmm0,%xmm3
        xorps   %xmm0,%xmm4
-       movaps  (%rcx),%xmm0
+       movups  (%rcx),%xmm0
 
 .Ldec_loop3:
 .byte  102,15,56,222,209
 .byte  102,15,56,222,217
        decl    %eax
 .byte  102,15,56,222,225
-       movaps  16(%rcx),%xmm1
+       movups  16(%rcx),%xmm1
 .byte  102,15,56,222,208
 .byte  102,15,56,222,216
        leaq    32(%rcx),%rcx
 .byte  102,15,56,222,224
-       movaps  (%rcx),%xmm0
+       movups  (%rcx),%xmm0
        jnz     .Ldec_loop3
 
 .byte  102,15,56,222,209
@@ -146,15 +146,15 @@ _aesni_decrypt3:
 .type  _aesni_encrypt4,@function
 .align 16
 _aesni_encrypt4:
-       movaps  (%rcx),%xmm0
+       movups  (%rcx),%xmm0
        shrl    $1,%eax
-       movaps  16(%rcx),%xmm1
+       movups  16(%rcx),%xmm1
        leaq    32(%rcx),%rcx
        xorps   %xmm0,%xmm2
        xorps   %xmm0,%xmm3
        xorps   %xmm0,%xmm4
        xorps   %xmm0,%xmm5
-       movaps  (%rcx),%xmm0
+       movups  (%rcx),%xmm0
 
 .Lenc_loop4:
 .byte  102,15,56,220,209
@@ -162,13 +162,13 @@ _aesni_encrypt4:
        decl    %eax
 .byte  102,15,56,220,225
 .byte  102,15,56,220,233
-       movaps  16(%rcx),%xmm1
+       movups  16(%rcx),%xmm1
 .byte  102,15,56,220,208
 .byte  102,15,56,220,216
        leaq    32(%rcx),%rcx
 .byte  102,15,56,220,224
 .byte  102,15,56,220,232
-       movaps  (%rcx),%xmm0
+       movups  (%rcx),%xmm0
        jnz     .Lenc_loop4
 
 .byte  102,15,56,220,209
@@ -184,15 +184,15 @@ _aesni_encrypt4:
 .type  _aesni_decrypt4,@function
 .align 16
 _aesni_decrypt4:
-       movaps  (%rcx),%xmm0
+       movups  (%rcx),%xmm0
        shrl    $1,%eax
-       movaps  16(%rcx),%xmm1
+       movups  16(%rcx),%xmm1
        leaq    32(%rcx),%rcx
        xorps   %xmm0,%xmm2
        xorps   %xmm0,%xmm3
        xorps   %xmm0,%xmm4
        xorps   %xmm0,%xmm5
-       movaps  (%rcx),%xmm0
+       movups  (%rcx),%xmm0
 
 .Ldec_loop4:
 .byte  102,15,56,222,209
@@ -200,13 +200,13 @@ _aesni_decrypt4:
        decl    %eax
 .byte  102,15,56,222,225
 .byte  102,15,56,222,233
-       movaps  16(%rcx),%xmm1
+       movups  16(%rcx),%xmm1
 .byte  102,15,56,222,208
 .byte  102,15,56,222,216
        leaq    32(%rcx),%rcx
 .byte  102,15,56,222,224
 .byte  102,15,56,222,232
-       movaps  (%rcx),%xmm0
+       movups  (%rcx),%xmm0
        jnz     .Ldec_loop4
 
 .byte  102,15,56,222,209
@@ -222,9 +222,9 @@ _aesni_decrypt4:
 .type  _aesni_encrypt6,@function
 .align 16
 _aesni_encrypt6:
-       movaps  (%rcx),%xmm0
+       movups  (%rcx),%xmm0
        shrl    $1,%eax
-       movaps  16(%rcx),%xmm1
+       movups  16(%rcx),%xmm1
        leaq    32(%rcx),%rcx
        xorps   %xmm0,%xmm2
        pxor    %xmm0,%xmm3
@@ -238,7 +238,7 @@ _aesni_encrypt6:
        pxor    %xmm0,%xmm7
        decl    %eax
 .byte  102,15,56,220,241
-       movaps  (%rcx),%xmm0
+       movups  (%rcx),%xmm0
 .byte  102,15,56,220,249
        jmp     .Lenc_loop6_enter
 .align 16
@@ -251,7 +251,7 @@ _aesni_encrypt6:
 .byte  102,15,56,220,241
 .byte  102,15,56,220,249
 .Lenc_loop6_enter:
-       movaps  16(%rcx),%xmm1
+       movups  16(%rcx),%xmm1
 .byte  102,15,56,220,208
 .byte  102,15,56,220,216
        leaq    32(%rcx),%rcx
@@ -259,7 +259,7 @@ _aesni_encrypt6:
 .byte  102,15,56,220,232
 .byte  102,15,56,220,240
 .byte  102,15,56,220,248
-       movaps  (%rcx),%xmm0
+       movups  (%rcx),%xmm0
        jnz     .Lenc_loop6
 
 .byte  102,15,56,220,209
@@ -279,9 +279,9 @@ _aesni_encrypt6:
 .type  _aesni_decrypt6,@function
 .align 16
 _aesni_decrypt6:
-       movaps  (%rcx),%xmm0
+       movups  (%rcx),%xmm0
        shrl    $1,%eax
-       movaps  16(%rcx),%xmm1
+       movups  16(%rcx),%xmm1
        leaq    32(%rcx),%rcx
        xorps   %xmm0,%xmm2
        pxor    %xmm0,%xmm3
@@ -295,7 +295,7 @@ _aesni_decrypt6:
        pxor    %xmm0,%xmm7
        decl    %eax
 .byte  102,15,56,222,241
-       movaps  (%rcx),%xmm0
+       movups  (%rcx),%xmm0
 .byte  102,15,56,222,249
        jmp     .Ldec_loop6_enter
 .align 16
@@ -308,7 +308,7 @@ _aesni_decrypt6:
 .byte  102,15,56,222,241
 .byte  102,15,56,222,249
 .Ldec_loop6_enter:
-       movaps  16(%rcx),%xmm1
+       movups  16(%rcx),%xmm1
 .byte  102,15,56,222,208
 .byte  102,15,56,222,216
        leaq    32(%rcx),%rcx
@@ -316,7 +316,7 @@ _aesni_decrypt6:
 .byte  102,15,56,222,232
 .byte  102,15,56,222,240
 .byte  102,15,56,222,248
-       movaps  (%rcx),%xmm0
+       movups  (%rcx),%xmm0
        jnz     .Ldec_loop6
 
 .byte  102,15,56,222,209
@@ -336,9 +336,9 @@ _aesni_decrypt6:
 .type  _aesni_encrypt8,@function
 .align 16
 _aesni_encrypt8:
-       movaps  (%rcx),%xmm0
+       movups  (%rcx),%xmm0
        shrl    $1,%eax
-       movaps  16(%rcx),%xmm1
+       movups  16(%rcx),%xmm1
        leaq    32(%rcx),%rcx
        xorps   %xmm0,%xmm2
        xorps   %xmm0,%xmm3
@@ -355,10 +355,10 @@ _aesni_encrypt8:
        pxor    %xmm0,%xmm8
 .byte  102,15,56,220,249
        pxor    %xmm0,%xmm9
-       movaps  (%rcx),%xmm0
+       movups  (%rcx),%xmm0
 .byte  102,68,15,56,220,193
 .byte  102,68,15,56,220,201
-       movaps  16(%rcx),%xmm1
+       movups  16(%rcx),%xmm1
        jmp     .Lenc_loop8_enter
 .align 16
 .Lenc_loop8:
@@ -371,7 +371,7 @@ _aesni_encrypt8:
 .byte  102,15,56,220,249
 .byte  102,68,15,56,220,193
 .byte  102,68,15,56,220,201
-       movaps  16(%rcx),%xmm1
+       movups  16(%rcx),%xmm1
 .Lenc_loop8_enter:
 .byte  102,15,56,220,208
 .byte  102,15,56,220,216
@@ -382,7 +382,7 @@ _aesni_encrypt8:
 .byte  102,15,56,220,248
 .byte  102,68,15,56,220,192
 .byte  102,68,15,56,220,200
-       movaps  (%rcx),%xmm0
+       movups  (%rcx),%xmm0
        jnz     .Lenc_loop8
 
 .byte  102,15,56,220,209
@@ -406,9 +406,9 @@ _aesni_encrypt8:
 .type  _aesni_decrypt8,@function
 .align 16
 _aesni_decrypt8:
-       movaps  (%rcx),%xmm0
+       movups  (%rcx),%xmm0
        shrl    $1,%eax
-       movaps  16(%rcx),%xmm1
+       movups  16(%rcx),%xmm1
        leaq    32(%rcx),%rcx
        xorps   %xmm0,%xmm2
        xorps   %xmm0,%xmm3
@@ -425,10 +425,10 @@ _aesni_decrypt8:
        pxor    %xmm0,%xmm8
 .byte  102,15,56,222,249
        pxor    %xmm0,%xmm9
-       movaps  (%rcx),%xmm0
+       movups  (%rcx),%xmm0
 .byte  102,68,15,56,222,193
 .byte  102,68,15,56,222,201
-       movaps  16(%rcx),%xmm1
+       movups  16(%rcx),%xmm1
        jmp     .Ldec_loop8_enter
 .align 16
 .Ldec_loop8:
@@ -441,7 +441,7 @@ _aesni_decrypt8:
 .byte  102,15,56,222,249
 .byte  102,68,15,56,222,193
 .byte  102,68,15,56,222,201
-       movaps  16(%rcx),%xmm1
+       movups  16(%rcx),%xmm1
 .Ldec_loop8_enter:
 .byte  102,15,56,222,208
 .byte  102,15,56,222,216
@@ -452,7 +452,7 @@ _aesni_decrypt8:
 .byte  102,15,56,222,248
 .byte  102,68,15,56,222,192
 .byte  102,68,15,56,222,200
-       movaps  (%rcx),%xmm0
+       movups  (%rcx),%xmm0
        jnz     .Ldec_loop8
 
 .byte  102,15,56,222,209
@@ -481,7 +481,7 @@ aesni_ecb_encrypt:
        jz      .Lecb_ret
 
        movl    240(%rcx),%eax
-       movaps  (%rcx),%xmm0
+       movups  (%rcx),%xmm0
        movq    %rcx,%r11
        movl    %eax,%r10d
        testl   %r8d,%r8d
@@ -572,14 +572,14 @@ aesni_ecb_encrypt:
        jmp     .Lecb_ret
 .align 16
 .Lecb_enc_one:
-       movaps  (%rcx),%xmm0
-       movaps  16(%rcx),%xmm1
+       movups  (%rcx),%xmm0
+       movups  16(%rcx),%xmm1
        leaq    32(%rcx),%rcx
        xorps   %xmm0,%xmm2
 .Loop_enc1_3:
 .byte  102,15,56,220,209
        decl    %eax
-       movaps  (%rcx),%xmm1
+       movups  (%rcx),%xmm1
        leaq    16(%rcx),%rcx
        jnz     .Loop_enc1_3    
 .byte  102,15,56,221,209
@@ -670,7 +670,7 @@ aesni_ecb_encrypt:
 
        call    _aesni_decrypt8
 
-       movaps  (%r11),%xmm0
+       movups  (%r11),%xmm0
        subq    $128,%rdx
        jnc     .Lecb_dec_loop8
 
@@ -705,7 +705,7 @@ aesni_ecb_encrypt:
        movups  80(%rdi),%xmm7
        je      .Lecb_dec_six
        movups  96(%rdi),%xmm8
-       movaps  (%rcx),%xmm0
+       movups  (%rcx),%xmm0
        call    _aesni_decrypt8
        movups  %xmm2,(%rsi)
        movups  %xmm3,16(%rsi)
@@ -717,14 +717,14 @@ aesni_ecb_encrypt:
        jmp     .Lecb_ret
 .align 16
 .Lecb_dec_one:
-       movaps  (%rcx),%xmm0
-       movaps  16(%rcx),%xmm1
+       movups  (%rcx),%xmm0
+       movups  16(%rcx),%xmm1
        leaq    32(%rcx),%rcx
        xorps   %xmm0,%xmm2
 .Loop_dec1_4:
 .byte  102,15,56,222,209
        decl    %eax
-       movaps  (%rcx),%xmm1
+       movups  (%rcx),%xmm1
        leaq    16(%rcx),%rcx
        jnz     .Loop_dec1_4    
 .byte  102,15,56,223,209
@@ -779,54 +779,54 @@ aesni_ecb_encrypt:
 .type  aesni_ccm64_encrypt_blocks,@function
 .align 16
 aesni_ccm64_encrypt_blocks:
+       movl    240(%rcx),%eax
        movdqu  (%r8),%xmm9
-       movdqu  (%r9),%xmm3
-       movdqa  .Lincrement64(%rip),%xmm8
-       movdqa  .Lbswap_mask(%rip),%xmm9
-.byte  102,69,15,56,0,201
+       movdqa  .Lincrement64(%rip),%xmm6
+       movdqa  .Lbswap_mask(%rip),%xmm7
 
-       movl    240(%rcx),%eax
-       movq    %rcx,%r11
-       movl    %eax,%r10d
+       shrl    $1,%eax
+       leaq    0(%rcx),%r11
+       movdqu  (%r9),%xmm3
        movdqa  %xmm9,%xmm2
-
+       movl    %eax,%r10d
+       jmp     .Lccm64_enc_outer
+.align 16
 .Lccm64_enc_outer:
-       movups  (%rdi),%xmm8
-.byte  102,65,15,56,0,209
-       movq    %r11,%rcx
+       movups  (%r11),%xmm0
        movl    %r10d,%eax
+       movups  (%rdi),%xmm8
 
-       movaps  (%rcx),%xmm0
-       shrl    $1,%eax
-       movaps  16(%rcx),%xmm1
-       xorps   %xmm0,%xmm8
-       leaq    32(%rcx),%rcx
        xorps   %xmm0,%xmm2
-       xorps   %xmm3,%xmm8
-       movaps  (%rcx),%xmm0
+       movups  16(%r11),%xmm1
+       xorps   %xmm8,%xmm0
+       leaq    32(%r11),%rcx
+       xorps   %xmm0,%xmm3
+       movups  (%rcx),%xmm0
 
 .Lccm64_enc2_loop:
 .byte  102,15,56,220,209
        decl    %eax
 .byte  102,15,56,220,217
-       movaps  16(%rcx),%xmm1
+       movups  16(%rcx),%xmm1
 .byte  102,15,56,220,208
        leaq    32(%rcx),%rcx
 .byte  102,15,56,220,216
-       movaps  0(%rcx),%xmm0
+       movups  0(%rcx),%xmm0
        jnz     .Lccm64_enc2_loop
+.byte  102,68,15,56,0,207
 .byte  102,15,56,220,209
 .byte  102,15,56,220,217
+       paddq   %xmm6,%xmm9
 .byte  102,15,56,221,208
 .byte  102,15,56,221,216
 
-       paddq   %xmm8,%xmm9
        decq    %rdx
        leaq    16(%rdi),%rdi
        xorps   %xmm2,%xmm8
        movdqa  %xmm9,%xmm2
        movups  %xmm8,(%rsi)
        leaq    16(%rsi),%rsi
+.byte  102,68,15,56,0,207
        jnz     .Lccm64_enc_outer
 
        movups  %xmm3,(%r9)
@@ -836,35 +836,36 @@ aesni_ccm64_encrypt_blocks:
 .type  aesni_ccm64_decrypt_blocks,@function
 .align 16
 aesni_ccm64_decrypt_blocks:
-       movdqu  (%r8),%xmm9
+       movl    240(%rcx),%eax
+       movups  (%r8),%xmm9
        movdqu  (%r9),%xmm3
-       movdqa  .Lincrement64(%rip),%xmm8
-       movdqa  .Lbswap_mask(%rip),%xmm9
+       movdqa  .Lincrement64(%rip),%xmm6
+       movdqa  .Lbswap_mask(%rip),%xmm7
 
-       movl    240(%rcx),%eax
-       movdqa  %xmm9,%xmm2
-.byte  102,69,15,56,0,201
+       movaps  %xmm9,%xmm2
        movl    %eax,%r10d
        movq    %rcx,%r11
-       movaps  (%rcx),%xmm0
-       movaps  16(%rcx),%xmm1
+.byte  102,68,15,56,0,207
+       movups  (%rcx),%xmm0
+       movups  16(%rcx),%xmm1
        leaq    32(%rcx),%rcx
        xorps   %xmm0,%xmm2
 .Loop_enc1_5:
 .byte  102,15,56,220,209
        decl    %eax
-       movaps  (%rcx),%xmm1
+       movups  (%rcx),%xmm1
        leaq    16(%rcx),%rcx
        jnz     .Loop_enc1_5    
 .byte  102,15,56,221,209
-.Lccm64_dec_outer:
-       paddq   %xmm8,%xmm9
        movups  (%rdi),%xmm8
+       paddq   %xmm6,%xmm9
+.byte  102,68,15,56,0,207
+       leaq    16(%rdi),%rdi
+       jmp     .Lccm64_dec_outer
+.align 16
+.Lccm64_dec_outer:
        xorps   %xmm2,%xmm8
        movdqa  %xmm9,%xmm2
-       leaq    16(%rdi),%rdi
-.byte  102,65,15,56,0,209
-       movq    %r11,%rcx
        movl    %r10d,%eax
        movups  %xmm8,(%rsi)
        leaq    16(%rsi),%rsi
@@ -872,41 +873,48 @@ aesni_ccm64_decrypt_blocks:
        subq    $1,%rdx
        jz      .Lccm64_dec_break
 
-       movaps  (%rcx),%xmm0
+       movups  (%r11),%xmm0
        shrl    $1,%eax
-       movaps  16(%rcx),%xmm1
+       movups  16(%r11),%xmm1
        xorps   %xmm0,%xmm8
-       leaq    32(%rcx),%rcx
+       leaq    32(%r11),%rcx
        xorps   %xmm0,%xmm2
        xorps   %xmm8,%xmm3
-       movaps  (%rcx),%xmm0
+       movups  (%rcx),%xmm0
 
 .Lccm64_dec2_loop:
 .byte  102,15,56,220,209
        decl    %eax
 .byte  102,15,56,220,217
-       movaps  16(%rcx),%xmm1
+       movups  16(%rcx),%xmm1
 .byte  102,15,56,220,208
        leaq    32(%rcx),%rcx
 .byte  102,15,56,220,216
-       movaps  0(%rcx),%xmm0
+       movups  0(%rcx),%xmm0
        jnz     .Lccm64_dec2_loop
+       movups  (%rdi),%xmm8
+       paddq   %xmm6,%xmm9
 .byte  102,15,56,220,209
 .byte  102,15,56,220,217
+.byte  102,68,15,56,0,207
+       leaq    16(%rdi),%rdi
 .byte  102,15,56,221,208
+.byte  102,15,56,221,216
        jmp     .Lccm64_dec_outer
 
 .align 16
 .Lccm64_dec_break:
-       movaps  (%rcx),%xmm0
-       movaps  16(%rcx),%xmm1
-       leaq    32(%rcx),%rcx
-       xorps   %xmm0,%xmm3
+
+       movups  (%r11),%xmm0
+       movups  16(%r11),%xmm1
+       xorps   %xmm0,%xmm8
+       leaq    32(%r11),%r11
+       xorps   %xmm8,%xmm3
 .Loop_enc1_6:
 .byte  102,15,56,220,217
        decl    %eax
-       movaps  (%rcx),%xmm1
-       leaq    16(%rcx),%rcx
+       movups  (%r11),%xmm1
+       leaq    16(%r11),%r11
        jnz     .Loop_enc1_6    
 .byte  102,15,56,221,217
        movups  %xmm3,(%r9)
@@ -960,10 +968,10 @@ aesni_ctr32_encrypt_blocks:
 .Lctr32_loop6:
        pshufd  $192,%xmm13,%xmm5
        por     %xmm14,%xmm2
-       movaps  (%r11),%xmm0
+       movups  (%r11),%xmm0
        pshufd  $128,%xmm13,%xmm6
        por     %xmm14,%xmm3
-       movaps  16(%r11),%xmm1
+       movups  16(%r11),%xmm1
        pshufd  $64,%xmm13,%xmm7
        por     %xmm14,%xmm4
        por     %xmm14,%xmm5
@@ -986,7 +994,7 @@ aesni_ctr32_encrypt_blocks:
        pxor    %xmm0,%xmm6
 .byte  102,15,56,220,233
        pxor    %xmm0,%xmm7
-       movaps  (%rcx),%xmm0
+       movups  (%rcx),%xmm0
        decl    %eax
 .byte  102,15,56,220,241
 .byte  102,15,56,220,249
@@ -1001,7 +1009,7 @@ aesni_ctr32_encrypt_blocks:
 .byte  102,15,56,220,241
 .byte  102,15,56,220,249
 .Lctr32_enc_loop6_enter:
-       movaps  16(%rcx),%xmm1
+       movups  16(%rcx),%xmm1
 .byte  102,15,56,220,208
 .byte  102,15,56,220,216
        leaq    32(%rcx),%rcx
@@ -1009,7 +1017,7 @@ aesni_ctr32_encrypt_blocks:
 .byte  102,15,56,220,232
 .byte  102,15,56,220,240
 .byte  102,15,56,220,248
-       movaps  (%rcx),%xmm0
+       movups  (%rcx),%xmm0
        jnz     .Lctr32_enc_loop6
 
 .byte  102,15,56,220,209
@@ -1109,14 +1117,14 @@ aesni_ctr32_encrypt_blocks:
        movups  (%rdi),%xmm8
        movl    240(%rcx),%eax
 .Lctr32_one:
-       movaps  (%rcx),%xmm0
-       movaps  16(%rcx),%xmm1
+       movups  (%rcx),%xmm0
+       movups  16(%rcx),%xmm1
        leaq    32(%rcx),%rcx
        xorps   %xmm0,%xmm2
 .Loop_enc1_7:
 .byte  102,15,56,220,209
        decl    %eax
-       movaps  (%rcx),%xmm1
+       movups  (%rcx),%xmm1
        leaq    16(%rcx),%rcx
        jnz     .Loop_enc1_7    
 .byte  102,15,56,221,209
@@ -1168,14 +1176,14 @@ aesni_xts_encrypt:
        movups  (%r9),%xmm15
        movl    240(%r8),%eax
        movl    240(%rcx),%r10d
-       movaps  (%r8),%xmm0
-       movaps  16(%r8),%xmm1
+       movups  (%r8),%xmm0
+       movups  16(%r8),%xmm1
        leaq    32(%r8),%r8
        xorps   %xmm0,%xmm15
 .Loop_enc1_8:
 .byte  102,68,15,56,220,249
        decl    %eax
-       movaps  (%r8),%xmm1
+       movups  (%r8),%xmm1
        leaq    16(%r8),%r8
        jnz     .Loop_enc1_8    
 .byte  102,68,15,56,221,249
@@ -1242,13 +1250,13 @@ aesni_xts_encrypt:
        movdqu  80(%rdi),%xmm7
        leaq    96(%rdi),%rdi
        pxor    %xmm13,%xmm5
-       movaps  (%r11),%xmm0
+       movups  (%r11),%xmm0
        pxor    %xmm14,%xmm6
        pxor    %xmm15,%xmm7
 
 
 
-       movaps  16(%r11),%xmm1
+       movups  16(%r11),%xmm1
        pxor    %xmm0,%xmm2
        pxor    %xmm0,%xmm3
        movdqa  %xmm10,0(%rsp)
@@ -1264,7 +1272,7 @@ aesni_xts_encrypt:
        movdqa  %xmm13,48(%rsp)
 .byte  102,15,56,220,233
        pxor    %xmm0,%xmm7
-       movaps  (%rcx),%xmm0
+       movups  (%rcx),%xmm0
        decl    %eax
        movdqa  %xmm14,64(%rsp)
 .byte  102,15,56,220,241
@@ -1284,7 +1292,7 @@ aesni_xts_encrypt:
 .byte  102,15,56,220,241
 .byte  102,15,56,220,249
 .Lxts_enc_loop6_enter:
-       movaps  16(%rcx),%xmm1
+       movups  16(%rcx),%xmm1
 .byte  102,15,56,220,208
 .byte  102,15,56,220,216
        leaq    32(%rcx),%rcx
@@ -1292,7 +1300,7 @@ aesni_xts_encrypt:
 .byte  102,15,56,220,232
 .byte  102,15,56,220,240
 .byte  102,15,56,220,248
-       movaps  (%rcx),%xmm0
+       movups  (%rcx),%xmm0
        jnz     .Lxts_enc_loop6
 
        pshufd  $19,%xmm14,%xmm9
@@ -1307,7 +1315,7 @@ aesni_xts_encrypt:
 .byte  102,15,56,220,233
 .byte  102,15,56,220,241
 .byte  102,15,56,220,249
-       movaps  16(%rcx),%xmm1
+       movups  16(%rcx),%xmm1
 
        pshufd  $19,%xmm14,%xmm9
        pxor    %xmm14,%xmm14
@@ -1322,7 +1330,7 @@ aesni_xts_encrypt:
 .byte  102,15,56,220,232
 .byte  102,15,56,220,240
 .byte  102,15,56,220,248
-       movaps  32(%rcx),%xmm0
+       movups  32(%rcx),%xmm0
 
        pshufd  $19,%xmm14,%xmm9
        pxor    %xmm14,%xmm14
@@ -1432,14 +1440,14 @@ aesni_xts_encrypt:
        movups  (%rdi),%xmm2
        leaq    16(%rdi),%rdi
        xorps   %xmm10,%xmm2
-       movaps  (%rcx),%xmm0
-       movaps  16(%rcx),%xmm1
+       movups  (%rcx),%xmm0
+       movups  16(%rcx),%xmm1
        leaq    32(%rcx),%rcx
        xorps   %xmm0,%xmm2
 .Loop_enc1_9:
 .byte  102,15,56,220,209
        decl    %eax
-       movaps  (%rcx),%xmm1
+       movups  (%rcx),%xmm1
        leaq    16(%rcx),%rcx
        jnz     .Loop_enc1_9    
 .byte  102,15,56,221,209
@@ -1537,14 +1545,14 @@ aesni_xts_encrypt:
 
        movups  -16(%rsi),%xmm2
        xorps   %xmm10,%xmm2
-       movaps  (%rcx),%xmm0
-       movaps  16(%rcx),%xmm1
+       movups  (%rcx),%xmm0
+       movups  16(%rcx),%xmm1
        leaq    32(%rcx),%rcx
        xorps   %xmm0,%xmm2
 .Loop_enc1_10:
 .byte  102,15,56,220,209
        decl    %eax
-       movaps  (%rcx),%xmm1
+       movups  (%rcx),%xmm1
        leaq    16(%rcx),%rcx
        jnz     .Loop_enc1_10   
 .byte  102,15,56,221,209
@@ -1564,14 +1572,14 @@ aesni_xts_decrypt:
        movups  (%r9),%xmm15
        movl    240(%r8),%eax
        movl    240(%rcx),%r10d
-       movaps  (%r8),%xmm0
-       movaps  16(%r8),%xmm1
+       movups  (%r8),%xmm0
+       movups  16(%r8),%xmm1
        leaq    32(%r8),%r8
        xorps   %xmm0,%xmm15
 .Loop_enc1_11:
 .byte  102,68,15,56,220,249
        decl    %eax
-       movaps  (%r8),%xmm1
+       movups  (%r8),%xmm1
        leaq    16(%r8),%r8
        jnz     .Loop_enc1_11   
 .byte  102,68,15,56,221,249
@@ -1644,13 +1652,13 @@ aesni_xts_decrypt:
        movdqu  80(%rdi),%xmm7
        leaq    96(%rdi),%rdi
        pxor    %xmm13,%xmm5
-       movaps  (%r11),%xmm0
+       movups  (%r11),%xmm0
        pxor    %xmm14,%xmm6
        pxor    %xmm15,%xmm7
 
 
 
-       movaps  16(%r11),%xmm1
+       movups  16(%r11),%xmm1
        pxor    %xmm0,%xmm2
        pxor    %xmm0,%xmm3
        movdqa  %xmm10,0(%rsp)
@@ -1666,7 +1674,7 @@ aesni_xts_decrypt:
        movdqa  %xmm13,48(%rsp)
 .byte  102,15,56,222,233
        pxor    %xmm0,%xmm7
-       movaps  (%rcx),%xmm0
+       movups  (%rcx),%xmm0
        decl    %eax
        movdqa  %xmm14,64(%rsp)
 .byte  102,15,56,222,241
@@ -1686,7 +1694,7 @@ aesni_xts_decrypt:
 .byte  102,15,56,222,241
 .byte  102,15,56,222,249
 .Lxts_dec_loop6_enter:
-       movaps  16(%rcx),%xmm1
+       movups  16(%rcx),%xmm1
 .byte  102,15,56,222,208
 .byte  102,15,56,222,216
        leaq    32(%rcx),%rcx
@@ -1694,7 +1702,7 @@ aesni_xts_decrypt:
 .byte  102,15,56,222,232
 .byte  102,15,56,222,240
 .byte  102,15,56,222,248
-       movaps  (%rcx),%xmm0
+       movups  (%rcx),%xmm0
        jnz     .Lxts_dec_loop6
 
        pshufd  $19,%xmm14,%xmm9
@@ -1709,7 +1717,7 @@ aesni_xts_decrypt:
 .byte  102,15,56,222,233
 .byte  102,15,56,222,241
 .byte  102,15,56,222,249
-       movaps  16(%rcx),%xmm1
+       movups  16(%rcx),%xmm1
 
        pshufd  $19,%xmm14,%xmm9
        pxor    %xmm14,%xmm14
@@ -1724,7 +1732,7 @@ aesni_xts_decrypt:
 .byte  102,15,56,222,232
 .byte  102,15,56,222,240
 .byte  102,15,56,222,248
-       movaps  32(%rcx),%xmm0
+       movups  32(%rcx),%xmm0
 
        pshufd  $19,%xmm14,%xmm9
        pxor    %xmm14,%xmm14
@@ -1843,14 +1851,14 @@ aesni_xts_decrypt:
        movups  (%rdi),%xmm2
        leaq    16(%rdi),%rdi
        xorps   %xmm10,%xmm2
-       movaps  (%rcx),%xmm0
-       movaps  16(%rcx),%xmm1
+       movups  (%rcx),%xmm0
+       movups  16(%rcx),%xmm1
        leaq    32(%rcx),%rcx
        xorps   %xmm0,%xmm2
 .Loop_dec1_12:
 .byte  102,15,56,222,209
        decl    %eax
-       movaps  (%rcx),%xmm1
+       movups  (%rcx),%xmm1
        leaq    16(%rcx),%rcx
        jnz     .Loop_dec1_12   
 .byte  102,15,56,223,209
@@ -1947,14 +1955,14 @@ aesni_xts_decrypt:
 
        movups  (%rdi),%xmm2
        xorps   %xmm11,%xmm2
-       movaps  (%rcx),%xmm0
-       movaps  16(%rcx),%xmm1
+       movups  (%rcx),%xmm0
+       movups  16(%rcx),%xmm1
        leaq    32(%rcx),%rcx
        xorps   %xmm0,%xmm2
 .Loop_dec1_13:
 .byte  102,15,56,222,209
        decl    %eax
-       movaps  (%rcx),%xmm1
+       movups  (%rcx),%xmm1
        leaq    16(%rcx),%rcx
        jnz     .Loop_dec1_13   
 .byte  102,15,56,223,209
@@ -1977,14 +1985,14 @@ aesni_xts_decrypt:
 
        movups  (%rsi),%xmm2
        xorps   %xmm10,%xmm2
-       movaps  (%rcx),%xmm0
-       movaps  16(%rcx),%xmm1
+       movups  (%rcx),%xmm0
+       movups  16(%rcx),%xmm1
        leaq    32(%rcx),%rcx
        xorps   %xmm0,%xmm2
 .Loop_dec1_14:
 .byte  102,15,56,222,209
        decl    %eax
-       movaps  (%rcx),%xmm1
+       movups  (%rcx),%xmm1
        leaq    16(%rcx),%rcx
        jnz     .Loop_dec1_14   
 .byte  102,15,56,223,209
@@ -2019,15 +2027,15 @@ aesni_cbc_encrypt:
        movups  (%rdi),%xmm3
        leaq    16(%rdi),%rdi
 
-       movaps  (%rcx),%xmm0
-       movaps  16(%rcx),%xmm1
+       movups  (%rcx),%xmm0
+       movups  16(%rcx),%xmm1
        xorps   %xmm0,%xmm3
        leaq    32(%rcx),%rcx
        xorps   %xmm3,%xmm2
 .Loop_enc1_15:
 .byte  102,15,56,220,209
        decl    %eax
-       movaps  (%rcx),%xmm1
+       movups  (%rcx),%xmm1
        leaq    16(%rcx),%rcx
        jnz     .Loop_enc1_15   
 .byte  102,15,56,221,209
@@ -2074,10 +2082,10 @@ aesni_cbc_encrypt:
        movups  %xmm9,(%rsi)
        leaq    16(%rsi),%rsi
 .Lcbc_dec_loop8_enter:
-       movaps  (%rcx),%xmm0
+       movups  (%rcx),%xmm0
        movups  (%rdi),%xmm2
        movups  16(%rdi),%xmm3
-       movaps  16(%rcx),%xmm1
+       movups  16(%rcx),%xmm1
 
        leaq    32(%rcx),%rcx
        movdqu  32(%rdi),%xmm4
@@ -2101,10 +2109,10 @@ aesni_cbc_encrypt:
        pxor    %xmm0,%xmm8
 .byte  102,15,56,222,249
        pxor    %xmm0,%xmm9
-       movaps  (%rcx),%xmm0
+       movups  (%rcx),%xmm0
 .byte  102,68,15,56,222,193
 .byte  102,68,15,56,222,201
-       movaps  16(%rcx),%xmm1
+       movups  16(%rcx),%xmm1
 
        call    .Ldec_loop8_enter
 
@@ -2202,14 +2210,14 @@ aesni_cbc_encrypt:
        jmp     .Lcbc_dec_tail_collected
 .align 16
 .Lcbc_dec_one:
-       movaps  (%rcx),%xmm0
-       movaps  16(%rcx),%xmm1
+       movups  (%rcx),%xmm0
+       movups  16(%rcx),%xmm1
        leaq    32(%rcx),%rcx
        xorps   %xmm0,%xmm2
 .Loop_dec1_16:
 .byte  102,15,56,222,209
        decl    %eax
-       movaps  (%rcx),%xmm1
+       movups  (%rcx),%xmm1
        leaq    16(%rcx),%rcx
        jnz     .Loop_dec1_16   
 .byte  102,15,56,223,209
@@ -2332,28 +2340,28 @@ aesni_set_decrypt_key:
        jnz     .Ldec_key_ret
        leaq    16(%rdx,%rsi,1),%rdi
 
-       movaps  (%rdx),%xmm0
-       movaps  (%rdi),%xmm1
-       movaps  %xmm0,(%rdi)
-       movaps  %xmm1,(%rdx)
+       movups  (%rdx),%xmm0
+       movups  (%rdi),%xmm1
+       movups  %xmm0,(%rdi)
+       movups  %xmm1,(%rdx)
        leaq    16(%rdx),%rdx
        leaq    -16(%rdi),%rdi
 
 .Ldec_key_inverse:
-       movaps  (%rdx),%xmm0
-       movaps  (%rdi),%xmm1
+       movups  (%rdx),%xmm0
+       movups  (%rdi),%xmm1
 .byte  102,15,56,219,192
 .byte  102,15,56,219,201
        leaq    16(%rdx),%rdx
        leaq    -16(%rdi),%rdi
-       movaps  %xmm0,16(%rdi)
-       movaps  %xmm1,-16(%rdx)
+       movups  %xmm0,16(%rdi)
+       movups  %xmm1,-16(%rdx)
        cmpq    %rdx,%rdi
        ja      .Ldec_key_inverse
 
-       movaps  (%rdx),%xmm0
+       movups  (%rdx),%xmm0
 .byte  102,15,56,219,192
-       movaps  %xmm0,(%rdi)
+       movups  %xmm0,(%rdi)
 .Ldec_key_ret:
        addq    $8,%rsp
        .byte   0xf3,0xc3
@@ -2383,7 +2391,7 @@ __aesni_set_encrypt_key:
 
 .L10rounds:
        movl    $9,%esi
-       movaps  %xmm0,(%rdx)
+       movups  %xmm0,(%rdx)
 .byte  102,15,58,223,200,1
        call    .Lkey_expansion_128_cold
 .byte  102,15,58,223,200,2
@@ -2404,7 +2412,7 @@ __aesni_set_encrypt_key:
        call    .Lkey_expansion_128
 .byte  102,15,58,223,200,54
        call    .Lkey_expansion_128
-       movaps  %xmm0,(%rax)
+       movups  %xmm0,(%rax)
        movl    %esi,80(%rax)
        xorl    %eax,%eax
        jmp     .Lenc_key_ret
@@ -2413,7 +2421,7 @@ __aesni_set_encrypt_key:
 .L12rounds:
        movq    16(%rdi),%xmm2
        movl    $11,%esi
-       movaps  %xmm0,(%rdx)
+       movups  %xmm0,(%rdx)
 .byte  102,15,58,223,202,1
        call    .Lkey_expansion_192a_cold
 .byte  102,15,58,223,202,2
@@ -2430,7 +2438,7 @@ __aesni_set_encrypt_key:
        call    .Lkey_expansion_192a
 .byte  102,15,58,223,202,128
        call    .Lkey_expansion_192b
-       movaps  %xmm0,(%rax)
+       movups  %xmm0,(%rax)
        movl    %esi,48(%rax)
        xorq    %rax,%rax
        jmp     .Lenc_key_ret
@@ -2440,8 +2448,8 @@ __aesni_set_encrypt_key:
        movups  16(%rdi),%xmm2
        movl    $13,%esi
        leaq    16(%rax),%rax
-       movaps  %xmm0,(%rdx)
-       movaps  %xmm2,16(%rdx)
+       movups  %xmm0,(%rdx)
+       movups  %xmm2,16(%rdx)
 .byte  102,15,58,223,202,1
        call    .Lkey_expansion_256a_cold
 .byte  102,15,58,223,200,1
@@ -2468,7 +2476,7 @@ __aesni_set_encrypt_key:
        call    .Lkey_expansion_256b
 .byte  102,15,58,223,202,64
        call    .Lkey_expansion_256a
-       movaps  %xmm0,(%rax)
+       movups  %xmm0,(%rax)
        movl    %esi,16(%rax)
        xorq    %rax,%rax
        jmp     .Lenc_key_ret
@@ -2483,7 +2491,7 @@ __aesni_set_encrypt_key:
 
 .align 16
 .Lkey_expansion_128:
-       movaps  %xmm0,(%rax)
+       movups  %xmm0,(%rax)
        leaq    16(%rax),%rax
 .Lkey_expansion_128_cold:
        shufps  $16,%xmm0,%xmm4
@@ -2496,7 +2504,7 @@ __aesni_set_encrypt_key:
 
 .align 16
 .Lkey_expansion_192a:
-       movaps  %xmm0,(%rax)
+       movups  %xmm0,(%rax)
        leaq    16(%rax),%rax
 .Lkey_expansion_192a_cold:
        movaps  %xmm2,%xmm5
@@ -2518,15 +2526,15 @@ __aesni_set_encrypt_key:
 .Lkey_expansion_192b:
        movaps  %xmm0,%xmm3
        shufps  $68,%xmm0,%xmm5
-       movaps  %xmm5,(%rax)
+       movups  %xmm5,(%rax)
        shufps  $78,%xmm2,%xmm3
-       movaps  %xmm3,16(%rax)
+       movups  %xmm3,16(%rax)
        leaq    32(%rax),%rax
        jmp     .Lkey_expansion_192b_warm
 
 .align 16
 .Lkey_expansion_256a:
-       movaps  %xmm2,(%rax)
+       movups  %xmm2,(%rax)
        leaq    16(%rax),%rax
 .Lkey_expansion_256a_cold:
        shufps  $16,%xmm0,%xmm4
@@ -2539,7 +2547,7 @@ __aesni_set_encrypt_key:
 
 .align 16
 .Lkey_expansion_256b:
-       movaps  %xmm0,(%rax)
+       movups  %xmm0,(%rax)
        leaq    16(%rax),%rax
 
        shufps  $16,%xmm2,%xmm4

diff --git a/lib/accelerated/intel/asm/appro-aes-x86.s b/lib/accelerated/intel/asm/appro-aes-x86.s
index a8dc2ac46b4f982ebf73af92b92bc6af2ad480cb..2084749838a647ed903099ef9c39ba2e7e14575f 100644 (file)
--- a/lib/accelerated/intel/asm/appro-aes-x86.s
+++ b/lib/accelerated/intel/asm/appro-aes-x86.s
@@ -47,14 +47,14 @@ aesni_encrypt:
        movups  (%eax),%xmm2
        movl    240(%edx),%ecx
        movl    8(%esp),%eax
-       movaps  (%edx),%xmm0
-       movaps  16(%edx),%xmm1
+       movups  (%edx),%xmm0
+       movups  16(%edx),%xmm1
        leal    32(%edx),%edx
        xorps   %xmm0,%xmm2
 .L000enc1_loop_1:
 .byte  102,15,56,220,209
        decl    %ecx
-       movaps  (%edx),%xmm1
+       movups  (%edx),%xmm1
        leal    16(%edx),%edx
        jnz     .L000enc1_loop_1
 .byte  102,15,56,221,209
@@ -71,14 +71,14 @@ aesni_decrypt:
        movups  (%eax),%xmm2
        movl    240(%edx),%ecx
        movl    8(%esp),%eax
-       movaps  (%edx),%xmm0
-       movaps  16(%edx),%xmm1
+       movups  (%edx),%xmm0
+       movups  16(%edx),%xmm1
        leal    32(%edx),%edx
        xorps   %xmm0,%xmm2
 .L001dec1_loop_2:
 .byte  102,15,56,222,209
        decl    %ecx
-       movaps  (%edx),%xmm1
+       movups  (%edx),%xmm1
        leal    16(%edx),%edx
        jnz     .L001dec1_loop_2
 .byte  102,15,56,223,209
@@ -88,25 +88,25 @@ aesni_decrypt:
 .type  _aesni_encrypt3,@function
 .align 16
 _aesni_encrypt3:
-       movaps  (%edx),%xmm0
+       movups  (%edx),%xmm0
        shrl    $1,%ecx
-       movaps  16(%edx),%xmm1
+       movups  16(%edx),%xmm1
        leal    32(%edx),%edx
        xorps   %xmm0,%xmm2
        pxor    %xmm0,%xmm3
        pxor    %xmm0,%xmm4
-       movaps  (%edx),%xmm0
+       movups  (%edx),%xmm0
 .L002enc3_loop:
 .byte  102,15,56,220,209
 .byte  102,15,56,220,217
        decl    %ecx
 .byte  102,15,56,220,225
-       movaps  16(%edx),%xmm1
+       movups  16(%edx),%xmm1
 .byte  102,15,56,220,208
 .byte  102,15,56,220,216
        leal    32(%edx),%edx
 .byte  102,15,56,220,224
-       movaps  (%edx),%xmm0
+       movups  (%edx),%xmm0
        jnz     .L002enc3_loop
 .byte  102,15,56,220,209
 .byte  102,15,56,220,217
@@ -119,25 +119,25 @@ _aesni_encrypt3:
 .type  _aesni_decrypt3,@function
 .align 16
 _aesni_decrypt3:
-       movaps  (%edx),%xmm0
+       movups  (%edx),%xmm0
        shrl    $1,%ecx
-       movaps  16(%edx),%xmm1
+       movups  16(%edx),%xmm1
        leal    32(%edx),%edx
        xorps   %xmm0,%xmm2
        pxor    %xmm0,%xmm3
        pxor    %xmm0,%xmm4
-       movaps  (%edx),%xmm0
+       movups  (%edx),%xmm0
 .L003dec3_loop:
 .byte  102,15,56,222,209
 .byte  102,15,56,222,217
        decl    %ecx
 .byte  102,15,56,222,225
-       movaps  16(%edx),%xmm1
+       movups  16(%edx),%xmm1
 .byte  102,15,56,222,208
 .byte  102,15,56,222,216
        leal    32(%edx),%edx
 .byte  102,15,56,222,224
-       movaps  (%edx),%xmm0
+       movups  (%edx),%xmm0
        jnz     .L003dec3_loop
 .byte  102,15,56,222,209
 .byte  102,15,56,222,217
@@ -150,28 +150,28 @@ _aesni_decrypt3:
 .type  _aesni_encrypt4,@function
 .align 16
 _aesni_encrypt4:
-       movaps  (%edx),%xmm0
-       movaps  16(%edx),%xmm1
+       movups  (%edx),%xmm0
+       movups  16(%edx),%xmm1
        shrl    $1,%ecx
        leal    32(%edx),%edx
        xorps   %xmm0,%xmm2
        pxor    %xmm0,%xmm3
        pxor    %xmm0,%xmm4
        pxor    %xmm0,%xmm5
-       movaps  (%edx),%xmm0
+       movups  (%edx),%xmm0
 .L004enc4_loop:
 .byte  102,15,56,220,209
 .byte  102,15,56,220,217
        decl    %ecx
 .byte  102,15,56,220,225
 .byte  102,15,56,220,233
-       movaps  16(%edx),%xmm1
+       movups  16(%edx),%xmm1
 .byte  102,15,56,220,208
 .byte  102,15,56,220,216
        leal    32(%edx),%edx
 .byte  102,15,56,220,224
 .byte  102,15,56,220,232
-       movaps  (%edx),%xmm0
+       movups  (%edx),%xmm0
        jnz     .L004enc4_loop
 .byte  102,15,56,220,209
 .byte  102,15,56,220,217
@@ -186,28 +186,28 @@ _aesni_encrypt4:
 .type  _aesni_decrypt4,@function
 .align 16
 _aesni_decrypt4:
-       movaps  (%edx),%xmm0
-       movaps  16(%edx),%xmm1
+       movups  (%edx),%xmm0
+       movups  16(%edx),%xmm1
        shrl    $1,%ecx
        leal    32(%edx),%edx
        xorps   %xmm0,%xmm2
        pxor    %xmm0,%xmm3
        pxor    %xmm0,%xmm4
        pxor    %xmm0,%xmm5
-       movaps  (%edx),%xmm0
+       movups  (%edx),%xmm0
 .L005dec4_loop:
 .byte  102,15,56,222,209
 .byte  102,15,56,222,217
        decl    %ecx
 .byte  102,15,56,222,225
 .byte  102,15,56,222,233
-       movaps  16(%edx),%xmm1
+       movups  16(%edx),%xmm1
 .byte  102,15,56,222,208
 .byte  102,15,56,222,216
        leal    32(%edx),%edx
 .byte  102,15,56,222,224
 .byte  102,15,56,222,232
-       movaps  (%edx),%xmm0
+       movups  (%edx),%xmm0
        jnz     .L005dec4_loop
 .byte  102,15,56,222,209
 .byte  102,15,56,222,217
@@ -222,9 +222,9 @@ _aesni_decrypt4:
 .type  _aesni_encrypt6,@function
 .align 16
 _aesni_encrypt6:
-       movaps  (%edx),%xmm0
+       movups  (%edx),%xmm0
        shrl    $1,%ecx
-       movaps  16(%edx),%xmm1
+       movups  16(%edx),%xmm1
        leal    32(%edx),%edx
        xorps   %xmm0,%xmm2
        pxor    %xmm0,%xmm3
@@ -238,7 +238,7 @@ _aesni_encrypt6:
 .byte  102,15,56,220,233
        pxor    %xmm0,%xmm7
 .byte  102,15,56,220,241
-       movaps  (%edx),%xmm0
+       movups  (%edx),%xmm0
 .byte  102,15,56,220,249
        jmp     .L_aesni_encrypt6_enter
 .align 16
@@ -252,7 +252,7 @@ _aesni_encrypt6:
 .byte  102,15,56,220,249
 .align 16
 .L_aesni_encrypt6_enter:
-       movaps  16(%edx),%xmm1
+       movups  16(%edx),%xmm1
 .byte  102,15,56,220,208
 .byte  102,15,56,220,216
        leal    32(%edx),%edx
@@ -260,7 +260,7 @@ _aesni_encrypt6:
 .byte  102,15,56,220,232
 .byte  102,15,56,220,240
 .byte  102,15,56,220,248
-       movaps  (%edx),%xmm0
+       movups  (%edx),%xmm0
        jnz     .L006enc6_loop
 .byte  102,15,56,220,209
 .byte  102,15,56,220,217
@@ -279,9 +279,9 @@ _aesni_encrypt6:
 .type  _aesni_decrypt6,@function
 .align 16
 _aesni_decrypt6:
-       movaps  (%edx),%xmm0
+       movups  (%edx),%xmm0
        shrl    $1,%ecx
-       movaps  16(%edx),%xmm1
+       movups  16(%edx),%xmm1
        leal    32(%edx),%edx
        xorps   %xmm0,%xmm2
        pxor    %xmm0,%xmm3
@@ -295,7 +295,7 @@ _aesni_decrypt6:
 .byte  102,15,56,222,233
        pxor    %xmm0,%xmm7
 .byte  102,15,56,222,241
-       movaps  (%edx),%xmm0
+       movups  (%edx),%xmm0
 .byte  102,15,56,222,249
        jmp     .L_aesni_decrypt6_enter
 .align 16
@@ -309,7 +309,7 @@ _aesni_decrypt6:
 .byte  102,15,56,222,249
 .align 16
 .L_aesni_decrypt6_enter:
-       movaps  16(%edx),%xmm1
+       movups  16(%edx),%xmm1
 .byte  102,15,56,222,208
 .byte  102,15,56,222,216
        leal    32(%edx),%edx
@@ -317,7 +317,7 @@ _aesni_decrypt6:
 .byte  102,15,56,222,232
 .byte  102,15,56,222,240
 .byte  102,15,56,222,248
-       movaps  (%edx),%xmm0
+       movups  (%edx),%xmm0
        jnz     .L007dec6_loop
 .byte  102,15,56,222,209
 .byte  102,15,56,222,217
@@ -418,14 +418,14 @@ aesni_ecb_encrypt:
        jmp     .L008ecb_ret
 .align 16
 .L013ecb_enc_one:
-       movaps  (%edx),%xmm0
-       movaps  16(%edx),%xmm1
+       movups  (%edx),%xmm0
+       movups  16(%edx),%xmm1
        leal    32(%edx),%edx
        xorps   %xmm0,%xmm2
 .L017enc1_loop_3:
 .byte  102,15,56,220,209
        decl    %ecx
-       movaps  (%edx),%xmm1
+       movups  (%edx),%xmm1
        leal    16(%edx),%edx
        jnz     .L017enc1_loop_3
 .byte  102,15,56,221,209
@@ -521,14 +521,14 @@ aesni_ecb_encrypt:
        jmp     .L008ecb_ret
 .align 16
 .L021ecb_dec_one:
-       movaps  (%edx),%xmm0
-       movaps  16(%edx),%xmm1
+       movups  (%edx),%xmm0
+       movups  16(%edx),%xmm1
        leal    32(%edx),%edx
        xorps   %xmm0,%xmm2
 .L025dec1_loop_4:
 .byte  102,15,56,222,209
        decl    %ecx
-       movaps  (%edx),%xmm1
+       movups  (%edx),%xmm1
        leal    16(%edx),%edx
        jnz     .L025dec1_loop_4
 .byte  102,15,56,223,209
@@ -583,56 +583,55 @@ aesni_ccm64_encrypt_blocks:
        movl    %ebp,48(%esp)
        movdqu  (%ebx),%xmm7
        movdqu  (%ecx),%xmm3
+       movl    240(%edx),%ecx
        movl    $202182159,(%esp)
        movl    $134810123,4(%esp)
        movl    $67438087,8(%esp)
        movl    $66051,12(%esp)
-       movl    $1,%ecx
+       movl    $1,%ebx
        xorl    %ebp,%ebp
-       movl    %ecx,16(%esp)
+       movl    %ebx,16(%esp)
        movl    %ebp,20(%esp)
        movl    %ebp,24(%esp)
        movl    %ebp,28(%esp)
-       movdqa  (%esp),%xmm5
-.byte  102,15,56,0,253
-       movl    240(%edx),%ecx
-       movl    %edx,%ebp
-       movl    %ecx,%ebx
+       shrl    $1,%ecx
+       leal    (%edx),%ebp
        movdqa  %xmm7,%xmm2
+       movl    %ecx,%ebx
+       movdqa  (%esp),%xmm5
 .L026ccm64_enc_outer:
-       movups  (%esi),%xmm6
-.byte  102,15,56,0,213
-       movl    %ebp,%edx
+       movups  (%ebp),%xmm0
        movl    %ebx,%ecx
-       movaps  (%edx),%xmm0
-       shrl    $1,%ecx
-       movaps  16(%edx),%xmm1
-       xorps   %xmm0,%xmm6
-       leal    32(%edx),%edx
+       movups  (%esi),%xmm6
        xorps   %xmm0,%xmm2
-       xorps   %xmm6,%xmm3
-       movaps  (%edx),%xmm0
+       movups  16(%ebp),%xmm1
+       xorps   %xmm6,%xmm0
+       leal    32(%ebp),%edx
+       xorps   %xmm0,%xmm3
+       movups  (%edx),%xmm0
 .L027ccm64_enc2_loop:
 .byte  102,15,56,220,209
        decl    %ecx
 .byte  102,15,56,220,217
-       movaps  16(%edx),%xmm1
+       movups  16(%edx),%xmm1
 .byte  102,15,56,220,208
        leal    32(%edx),%edx
 .byte  102,15,56,220,216
-       movaps  (%edx),%xmm0
+       movups  (%edx),%xmm0
        jnz     .L027ccm64_enc2_loop
+.byte  102,15,56,0,253
 .byte  102,15,56,220,209
 .byte  102,15,56,220,217
+       paddq   16(%esp),%xmm7
 .byte  102,15,56,221,208
 .byte  102,15,56,221,216
-       paddq   16(%esp),%xmm7
        decl    %eax
        leal    16(%esi),%esi
        xorps   %xmm2,%xmm6
        movdqa  %xmm7,%xmm2
        movups  %xmm6,(%edi)
        leal    16(%edi),%edi
+.byte  102,15,56,0,253
        jnz     .L026ccm64_enc_outer
        movl    48(%esp),%esp
        movl    40(%esp),%edi
@@ -664,80 +663,86 @@ aesni_ccm64_decrypt_blocks:
        movl    %ebp,48(%esp)
        movdqu  (%ebx),%xmm7
        movdqu  (%ecx),%xmm3
+       movl    240(%edx),%ecx
        movl    $202182159,(%esp)
        movl    $134810123,4(%esp)
        movl    $67438087,8(%esp)
        movl    $66051,12(%esp)
-       movl    $1,%ecx
+       movl    $1,%ebx
        xorl    %ebp,%ebp
-       movl    %ecx,16(%esp)
+       movl    %ebx,16(%esp)
        movl    %ebp,20(%esp)
        movl    %ebp,24(%esp)
        movl    %ebp,28(%esp)
        movdqa  (%esp),%xmm5
        movdqa  %xmm7,%xmm2
-.byte  102,15,56,0,253
-       movl    240(%edx),%ecx
        movl    %edx,%ebp
        movl    %ecx,%ebx
-       movaps  (%edx),%xmm0
-       movaps  16(%edx),%xmm1
+.byte  102,15,56,0,253
+       movups  (%edx),%xmm0
+       movups  16(%edx),%xmm1
        leal    32(%edx),%edx
        xorps   %xmm0,%xmm2
 .L028enc1_loop_5:
 .byte  102,15,56,220,209
        decl    %ecx
-       movaps  (%edx),%xmm1
+       movups  (%edx),%xmm1
        leal    16(%edx),%edx
        jnz     .L028enc1_loop_5
 .byte  102,15,56,221,209
-.L029ccm64_dec_outer:
-       paddq   16(%esp),%xmm7
        movups  (%esi),%xmm6
+       paddq   16(%esp),%xmm7
+.byte  102,15,56,0,253
+       leal    16(%esi),%esi
+       jmp     .L029ccm64_dec_outer
+.align 16
+.L029ccm64_dec_outer:
        xorps   %xmm2,%xmm6
        movdqa  %xmm7,%xmm2
-       leal    16(%esi),%esi
-.byte  102,15,56,0,213
-       movl    %ebp,%edx
        movl    %ebx,%ecx
        movups  %xmm6,(%edi)
        leal    16(%edi),%edi
        subl    $1,%eax
        jz      .L030ccm64_dec_break
-       movaps  (%edx),%xmm0
+       movups  (%ebp),%xmm0
        shrl    $1,%ecx
-       movaps  16(%edx),%xmm1
+       movups  16(%ebp),%xmm1
        xorps   %xmm0,%xmm6
-       leal    32(%edx),%edx
+       leal    32(%ebp),%edx
        xorps   %xmm0,%xmm2
        xorps   %xmm6,%xmm3
-       movaps  (%edx),%xmm0
+       movups  (%edx),%xmm0
 .L031ccm64_dec2_loop:
 .byte  102,15,56,220,209
        decl    %ecx
 .byte  102,15,56,220,217
-       movaps  16(%edx),%xmm1
+       movups  16(%edx),%xmm1
 .byte  102,15,56,220,208
        leal    32(%edx),%edx
 .byte  102,15,56,220,216
-       movaps  (%edx),%xmm0
+       movups  (%edx),%xmm0
        jnz     .L031ccm64_dec2_loop
+       movups  (%esi),%xmm6
+       paddq   16(%esp),%xmm7
 .byte  102,15,56,220,209
 .byte  102,15,56,220,217
+.byte  102,15,56,0,253
+       leal    16(%esi),%esi
 .byte  102,15,56,221,208
 .byte  102,15,56,221,216
        jmp     .L029ccm64_dec_outer
 .align 16
 .L030ccm64_dec_break:
-       movaps  (%edx),%xmm0
-       movaps  16(%edx),%xmm1
+       movl    %ebp,%edx
+       movups  (%edx),%xmm0
+       movups  16(%edx),%xmm1
        xorps   %xmm0,%xmm6
        leal    32(%edx),%edx
        xorps   %xmm6,%xmm3
 .L032enc1_loop_6:
 .byte  102,15,56,220,217
        decl    %ecx
-       movaps  (%edx),%xmm1
+       movups  (%edx),%xmm1
        leal    16(%edx),%edx
        jnz     .L032enc1_loop_6
 .byte  102,15,56,221,217
@@ -826,8 +831,8 @@ aesni_ctr32_encrypt_blocks:
        por     %xmm1,%xmm5
        por     %xmm1,%xmm6
        por     %xmm1,%xmm7
-       movaps  (%ebp),%xmm0
-       movaps  16(%ebp),%xmm1
+       movups  (%ebp),%xmm0
+       movups  16(%ebp),%xmm1
        leal    32(%ebp),%edx
        decl    %ecx
        pxor    %xmm0,%xmm2
@@ -841,7 +846,7 @@ aesni_ctr32_encrypt_blocks:
 .byte  102,15,56,220,233
        pxor    %xmm0,%xmm7
 .byte  102,15,56,220,241
-       movaps  (%edx),%xmm0
+       movups  (%edx),%xmm0
 .byte  102,15,56,220,249
        call    .L_aesni_encrypt6_enter
        movups  (%esi),%xmm1
@@ -920,14 +925,14 @@ aesni_ctr32_encrypt_blocks:
        movups  (%ebx),%xmm2
        movl    240(%edx),%ecx
 .L037ctr32_one:
-       movaps  (%edx),%xmm0
-       movaps  16(%edx),%xmm1
+       movups  (%edx),%xmm0
+       movups  16(%edx),%xmm1
        leal    32(%edx),%edx
        xorps   %xmm0,%xmm2
 .L041enc1_loop_7:
 .byte  102,15,56,220,209
        decl    %ecx
-       movaps  (%edx),%xmm1
+       movups  (%edx),%xmm1
        leal    16(%edx),%edx
        jnz     .L041enc1_loop_7
 .byte  102,15,56,221,209
@@ -994,14 +999,14 @@ aesni_xts_encrypt:
        movl    40(%esp),%esi
        movl    240(%edx),%ecx
        movups  (%esi),%xmm2
-       movaps  (%edx),%xmm0
-       movaps  16(%edx),%xmm1
+       movups  (%edx),%xmm0
+       movups  16(%edx),%xmm1
        leal    32(%edx),%edx
        xorps   %xmm0,%xmm2
 .L042enc1_loop_8:
 .byte  102,15,56,220,209
        decl    %ecx
-       movaps  (%edx),%xmm1
+       movups  (%edx),%xmm1
        leal    16(%edx),%edx
        jnz     .L042enc1_loop_8
 .byte  102,15,56,221,209
@@ -1064,7 +1069,7 @@ aesni_xts_encrypt:
        pshufd  $19,%xmm0,%xmm7
        movdqa  %xmm1,64(%esp)
        paddq   %xmm1,%xmm1
-       movaps  (%ebp),%xmm0
+       movups  (%ebp),%xmm0
        pand    %xmm3,%xmm7
        movups  (%esi),%xmm2
        pxor    %xmm1,%xmm7
@@ -1082,7 +1087,7 @@ aesni_xts_encrypt:
        pxor    (%esp),%xmm2
        movdqa  %xmm7,80(%esp)
        pxor    %xmm1,%xmm7
-       movaps  16(%ebp),%xmm1
+       movups  16(%ebp),%xmm1
        leal    32(%ebp),%edx
        pxor    16(%esp),%xmm3
 .byte  102,15,56,220,209
@@ -1095,7 +1100,7 @@ aesni_xts_encrypt:
 .byte  102,15,56,220,233
        pxor    %xmm0,%xmm7
 .byte  102,15,56,220,241
-       movaps  (%edx),%xmm0
+       movups  (%edx),%xmm0
 .byte  102,15,56,220,249
        call    .L_aesni_encrypt6_enter
        movdqa  80(%esp),%xmm1
@@ -1196,14 +1201,14 @@ aesni_xts_encrypt:
        movups  (%esi),%xmm2
        leal    16(%esi),%esi
        xorps   %xmm5,%xmm2
-       movaps  (%edx),%xmm0
-       movaps  16(%edx),%xmm1
+       movups  (%edx),%xmm0
+       movups  16(%edx),%xmm1
        leal    32(%edx),%edx
        xorps   %xmm0,%xmm2
 .L051enc1_loop_9:
 .byte  102,15,56,220,209
        decl    %ecx
-       movaps  (%edx),%xmm1
+       movups  (%edx),%xmm1
        leal    16(%edx),%edx
        jnz     .L051enc1_loop_9
 .byte  102,15,56,221,209
@@ -1307,14 +1312,14 @@ aesni_xts_encrypt:
        movl    %ebx,%ecx
        movups  -16(%edi),%xmm2
        xorps   %xmm5,%xmm2
-       movaps  (%edx),%xmm0
-       movaps  16(%edx),%xmm1
+       movups  (%edx),%xmm0
+       movups  16(%edx),%xmm1
        leal    32(%edx),%edx
        xorps   %xmm0,%xmm2
 .L054enc1_loop_10:
 .byte  102,15,56,220,209
        decl    %ecx
-       movaps  (%edx),%xmm1
+       movups  (%edx),%xmm1
        leal    16(%edx),%edx
        jnz     .L054enc1_loop_10
 .byte  102,15,56,221,209
@@ -1341,14 +1346,14 @@ aesni_xts_decrypt:
        movl    40(%esp),%esi
        movl    240(%edx),%ecx
        movups  (%esi),%xmm2
-       movaps  (%edx),%xmm0
-       movaps  16(%edx),%xmm1
+       movups  (%edx),%xmm0
+       movups  16(%edx),%xmm1
        leal    32(%edx),%edx
        xorps   %xmm0,%xmm2
 .L055enc1_loop_11:
 .byte  102,15,56,220,209
        decl    %ecx
-       movaps  (%edx),%xmm1
+       movups  (%edx),%xmm1
        leal    16(%edx),%edx
        jnz     .L055enc1_loop_11
 .byte  102,15,56,221,209
@@ -1416,7 +1421,7 @@ aesni_xts_decrypt:
        pshufd  $19,%xmm0,%xmm7
        movdqa  %xmm1,64(%esp)
        paddq   %xmm1,%xmm1
-       movaps  (%ebp),%xmm0
+       movups  (%ebp),%xmm0
        pand    %xmm3,%xmm7
        movups  (%esi),%xmm2
        pxor    %xmm1,%xmm7
@@ -1434,7 +1439,7 @@ aesni_xts_decrypt:
        pxor    (%esp),%xmm2
        movdqa  %xmm7,80(%esp)
        pxor    %xmm1,%xmm7
-       movaps  16(%ebp),%xmm1
+       movups  16(%ebp),%xmm1
        leal    32(%ebp),%edx
        pxor    16(%esp),%xmm3
 .byte  102,15,56,222,209
@@ -1447,7 +1452,7 @@ aesni_xts_decrypt:
 .byte  102,15,56,222,233
        pxor    %xmm0,%xmm7
 .byte  102,15,56,222,241
-       movaps  (%edx),%xmm0
+       movups  (%edx),%xmm0
 .byte  102,15,56,222,249
        call    .L_aesni_decrypt6_enter
        movdqa  80(%esp),%xmm1
@@ -1548,14 +1553,14 @@ aesni_xts_decrypt:
        movups  (%esi),%xmm2
        leal    16(%esi),%esi
        xorps   %xmm5,%xmm2
-       movaps  (%edx),%xmm0
-       movaps  16(%edx),%xmm1
+       movups  (%edx),%xmm0
+       movups  16(%edx),%xmm1
        leal    32(%edx),%edx
        xorps   %xmm0,%xmm2
 .L064dec1_loop_12:
 .byte  102,15,56,222,209
        decl    %ecx
-       movaps  (%edx),%xmm1
+       movups  (%edx),%xmm1
        leal    16(%edx),%edx
        jnz     .L064dec1_loop_12
 .byte  102,15,56,223,209
@@ -1656,14 +1661,14 @@ aesni_xts_decrypt:
        movl    %ebx,%ecx
        movups  (%esi),%xmm2
        xorps   %xmm5,%xmm2
-       movaps  (%edx),%xmm0
-       movaps  16(%edx),%xmm1
+       movups  (%edx),%xmm0
+       movups  16(%edx),%xmm1
        leal    32(%edx),%edx
        xorps   %xmm0,%xmm2
 .L067dec1_loop_13:
 .byte  102,15,56,222,209
        decl    %ecx
-       movaps  (%edx),%xmm1
+       movups  (%edx),%xmm1
        leal    16(%edx),%edx
        jnz     .L067dec1_loop_13
 .byte  102,15,56,223,209
@@ -1683,14 +1688,14 @@ aesni_xts_decrypt:
        movl    %ebx,%ecx
        movups  (%edi),%xmm2
        xorps   %xmm6,%xmm2
-       movaps  (%edx),%xmm0
-       movaps  16(%edx),%xmm1
+       movups  (%edx),%xmm0
+       movups  16(%edx),%xmm1
        leal    32(%edx),%edx
        xorps   %xmm0,%xmm2
 .L069dec1_loop_14:
 .byte  102,15,56,222,209
        decl    %ecx
-       movaps  (%edx),%xmm1
+       movups  (%edx),%xmm1
        leal    16(%edx),%edx
        jnz     .L069dec1_loop_14
 .byte  102,15,56,223,209
@@ -1740,15 +1745,15 @@ aesni_cbc_encrypt:
 .L073cbc_enc_loop:
        movups  (%esi),%xmm7
        leal    16(%esi),%esi
-       movaps  (%edx),%xmm0
-       movaps  16(%edx),%xmm1
+       movups  (%edx),%xmm0
+       movups  16(%edx),%xmm1
        xorps   %xmm0,%xmm7
        leal    32(%edx),%edx
        xorps   %xmm7,%xmm2
 .L074enc1_loop_15:
 .byte  102,15,56,220,209
        decl    %ecx
-       movaps  (%edx),%xmm1
+       movups  (%edx),%xmm1
        leal    16(%edx),%edx
        jnz     .L074enc1_loop_15
 .byte  102,15,56,221,209
@@ -1863,14 +1868,14 @@ aesni_cbc_encrypt:
        jmp     .L079cbc_dec_tail_collected
 .align 16
 .L080cbc_dec_one:
-       movaps  (%edx),%xmm0
-       movaps  16(%edx),%xmm1
+       movups  (%edx),%xmm0
+       movups  16(%edx),%xmm1
        leal    32(%edx),%edx
        xorps   %xmm0,%xmm2
 .L084dec1_loop_16:
 .byte  102,15,56,222,209
        decl    %ecx
-       movaps  (%edx),%xmm1
+       movups  (%edx),%xmm1
        leal    16(%edx),%edx
        jnz     .L084dec1_loop_16
 .byte  102,15,56,223,209
@@ -1961,7 +1966,7 @@ _aesni_set_encrypt_key:
 .align 16
 .L09010rounds:
        movl    $9,%ecx
-       movaps  %xmm0,-16(%edx)
+       movups  %xmm0,-16(%edx)
 .byte  102,15,58,223,200,1
        call    .L091key_128_cold
 .byte  102,15,58,223,200,2
@@ -1982,13 +1987,13 @@ _aesni_set_encrypt_key:
        call    .L092key_128
 .byte  102,15,58,223,200,54
        call    .L092key_128
-       movaps  %xmm0,(%edx)
+       movups  %xmm0,(%edx)
        movl    %ecx,80(%edx)
        xorl    %eax,%eax
        ret
 .align 16
 .L092key_128:
-       movaps  %xmm0,(%edx)
+       movups  %xmm0,(%edx)
        leal    16(%edx),%edx
 .L091key_128_cold:
        shufps  $16,%xmm0,%xmm4
@@ -2002,7 +2007,7 @@ _aesni_set_encrypt_key:
 .L08812rounds:
        movq    16(%eax),%xmm2
        movl    $11,%ecx
-       movaps  %xmm0,-16(%edx)
+       movups  %xmm0,-16(%edx)
 .byte  102,15,58,223,202,1
        call    .L093key_192a_cold
 .byte  102,15,58,223,202,2
@@ -2019,13 +2024,13 @@ _aesni_set_encrypt_key:
        call    .L095key_192a
 .byte  102,15,58,223,202,128
        call    .L094key_192b
-       movaps  %xmm0,(%edx)
+       movups  %xmm0,(%edx)
        movl    %ecx,48(%edx)
        xorl    %eax,%eax
        ret
 .align 16
 .L095key_192a:
-       movaps  %xmm0,(%edx)
+       movups  %xmm0,(%edx)
        leal    16(%edx),%edx
 .align 16
 .L093key_192a_cold:
@@ -2047,9 +2052,9 @@ _aesni_set_encrypt_key:
 .L094key_192b:
        movaps  %xmm0,%xmm3
        shufps  $68,%xmm0,%xmm5
-       movaps  %xmm5,(%edx)
+       movups  %xmm5,(%edx)
        shufps  $78,%xmm2,%xmm3
-       movaps  %xmm3,16(%edx)
+       movups  %xmm3,16(%edx)
        leal    32(%edx),%edx
        jmp     .L096key_192b_warm
 .align 16
@@ -2057,8 +2062,8 @@ _aesni_set_encrypt_key:
        movups  16(%eax),%xmm2
        movl    $13,%ecx
        leal    16(%edx),%edx
-       movaps  %xmm0,-32(%edx)
-       movaps  %xmm2,-16(%edx)
+       movups  %xmm0,-32(%edx)
+       movups  %xmm2,-16(%edx)
 .byte  102,15,58,223,202,1
        call    .L097key_256a_cold
 .byte  102,15,58,223,200,1
@@ -2085,13 +2090,13 @@ _aesni_set_encrypt_key:
        call    .L098key_256b
 .byte  102,15,58,223,202,64
        call    .L099key_256a
-       movaps  %xmm0,(%edx)
+       movups  %xmm0,(%edx)
        movl    %ecx,16(%edx)
        xorl    %eax,%eax
        ret
 .align 16
 .L099key_256a:
-       movaps  %xmm2,(%edx)
+       movups  %xmm2,(%edx)
        leal    16(%edx),%edx
 .L097key_256a_cold:
        shufps  $16,%xmm0,%xmm4
@@ -2103,7 +2108,7 @@ _aesni_set_encrypt_key:
        ret
 .align 16
 .L098key_256b:
-       movaps  %xmm0,(%edx)
+       movups  %xmm0,(%edx)
        leal    16(%edx),%edx
        shufps  $16,%xmm2,%xmm4
        xorps   %xmm4,%xmm2
@@ -2146,26 +2151,26 @@ aesni_set_decrypt_key:
        testl   %eax,%eax
        jnz     .L100dec_key_ret
        leal    16(%edx,%ecx,1),%eax
-       movaps  (%edx),%xmm0
-       movaps  (%eax),%xmm1
-       movaps  %xmm0,(%eax)
-       movaps  %xmm1,(%edx)
+       movups  (%edx),%xmm0
+       movups  (%eax),%xmm1
+       movups  %xmm0,(%eax)
+       movups  %xmm1,(%edx)
        leal    16(%edx),%edx
        leal    -16(%eax),%eax
 .L101dec_key_inverse:
-       movaps  (%edx),%xmm0
-       movaps  (%eax),%xmm1
+       movups  (%edx),%xmm0
+       movups  (%eax),%xmm1
 .byte  102,15,56,219,192
 .byte  102,15,56,219,201
        leal    16(%edx),%edx
        leal    -16(%eax),%eax
-       movaps  %xmm0,16(%eax)
-       movaps  %xmm1,-16(%edx)
+       movups  %xmm0,16(%eax)
+       movups  %xmm1,-16(%edx)
        cmpl    %edx,%eax
        ja      .L101dec_key_inverse
-       movaps  (%edx),%xmm0
+       movups  (%edx),%xmm0
 .byte  102,15,56,219,192
-       movaps  %xmm0,(%edx)
+       movups  %xmm0,(%edx)
        xorl    %eax,%eax
 .L100dec_key_ret:
        ret