--- test-without-peel.s 2015-10-29 10:44:34.940278732 +0100
+++ test.s  2015-10-29 10:46:27.017510990 +0100
@@ -26,35 +26,61 @@
    cmp r1, #0
-   ble .L17
-   subs    r2, r1, #4
-   push    {r4, r5}
+   ble .L23
+   sbfx    r3, r0, #2, #1
+   push    {r4, r5, r6, r7}
+   and r3, r3, #3
+   mov r5, r0
+   cmp r3, r1
+   it  cs
+   movcs   r3, r1
+   cmp r1, #6
+   it  le
+   movle   r3, r1
+   bgt .L26
+.L3:
+   mov r4, r3
+   movs    r2, #0
+.L5:
+   adds    r2, r2, #1
+   vstmia.32   r5!, {s0}
+   cmp r2, r3
+   bne .L5
+   cmp r1, r3
+   beq .L1
+.L4:
+   subs    r6, r1, r3
+   subs    r5, r1, #1
+   subs    r2, r6, #4
+   subs    r5, r5, r3
    lsrs    r2, r2, #2
-   subs    r4, r1, #1
-   cmp r4, #2
+   cmp r5, #2
    add r2, r2, #1
-   lsl r3, r2, #2
-   bls .L9
+   lsl r7, r2, #2
+   bls .L7
+   add r5, r0, r3, lsl #2
    vdup.32 q8, d0[0]
-   mov r5, r0
    movs    r4, #0
-.L4:
+.L8:
    adds    r4, r4, #1
-   vst1.32 {q8}, [r5]
-   cmp r2, r4
+   vst1.64 {d16-d17}, [r5:64]
+   cmp r4, r2
    add r5, r5, #16
-   bhi .L4
-   cmp r1, r3
+   bcc .L8
+   cmp r6, r7
+   add r4, r3, r7
    beq .L1
-.L3:
-   add r0, r0, r3, lsl #2
-.L6:
-   adds    r3, r3, #1
+.L7:
+   add r0, r0, r4, lsl #2
+.L10:
+   adds    r4, r4, #1
    vstmia.32   r0!, {s0}
-   cmp r1, r3
-   bgt .L6
+   cmp r1, r4
+   bgt .L10
 .L1:
-   pop {r4, r5}
-.L17:
+   pop {r4, r5, r6, r7}
+.L23:
    bx  lr
-.L9:
-   movs    r3, #0
+.L26:
+   mov r4, r3
+   cmp r3, #0
+   beq .L4
    b   .L3