Duplicate assembly when filling small arrays

[Andrew Kravchuk]

Hello! Consider the following example (opcodes and byte numbers edited out for clarity):

(disassemble #'(lambda () (declare (optimize (speed 3) (safety 0) (debug 0)))
  (make-array 6 :element-type 'fixnum :initial-element -1)))
; disassembly for (LAMBDA ())
; Size: 124 bytes. Origin: #x5398F226 ; (LAMBDA ())
; MOV [R13+40], R13 ; thread.pseudo-atomic-bits
; MOV RCX, [R13+112] ; thread.mixed-tlab
; LEA RAX, [RCX+64]
; CMP RAX, [R13+120]
; JA L2
; MOV [R13+112], RAX ; thread.mixed-tlab
; L0: MOV BYTE PTR [RCX], -55
; MOV BYTE PTR [RCX+8], 12
; OR CL, 15
; XOR [R13+40], R13 ; thread.pseudo-atomic-bits
; JEQ L1
; INT3 9 ; pending interrupt trap
; L1: MOV RDX, -2
; MOV [RCX+1], RDX
; MOV RDX, -2
; MOV [RCX+9], RDX
; MOV RDX, -2
; MOV [RCX+17], RDX
; MOV RDX, -2
; MOV [RCX+25], RDX
; MOV RDX, -2
; MOV [RCX+33], RDX
; MOV RDX, -2
; MOV [RCX+41], RDX
; MOV RDX, RCX
; LEAVE
; CLC
; RET
; L2: PUSH 64
; CALL [#x50000578] ; #x52A004F0: ALLOC-TRAMP
; POP RCX
; JMP L0
NIL

Have a closer look at code starting at L1 label. I appreciate compiler's effort on unrolling the loop to fill that small array with -1s, but wouldn't it make sense to not to put it into RDX every time?..

My SBCL version is 2.3.7 on Linux.