gather-downcase.lisp

(defun simd-downcase (vector start end)
  ;; (declare (type index start end)
  ;;          (optimize speed (safety 0)))
  (let ((pages #.(map   '(vector (unsigned-byte 32))
                        (lambda (x)
                          (ash x 6))
                        sb-impl::+character-case-pages+))
        (cases sb-impl::+character-cases+))
    (declare (optimize sb-c::preserve-single-use-debug-variables))
    (with-pinned-objects (vector pages)
      (inline-vop (((32-bit-array* sap-reg t) (vector-sap vector))
                   ((pages sap-reg t) (vector-sap pages))
                   ((cases sap-reg t) (vector-sap cases))
                   ((start any-reg) start)
                   ((end any-reg) end)
                   ((left))
                   ((32-bit-array sap-reg t))
                   ((bytes int-avx2-reg))
                   ((page-index))
                   ((page-index-mask))
                   ((search))
                   ((mask)))
          ()

        (inst lea 32-bit-array (ea 32-bit-array* start 2))
        (inst mov left 63)
        (inst vmovd page-index-mask left)
        (inst vpbroadcastd page-index-mask page-index-mask)
        
        
        LOOP
        (inst vmovdqu bytes (ea 32-bit-array))
        (inst vpsrld-imm page-index bytes 6)
        (inst vpcmpeqd mask mask mask)

        (inst vpgatherdd search (ea pages page-index 4) mask)
        (inst vpand bytes bytes page-index-mask)

        (inst vpaddd bytes bytes search)
        (inst vpslld-imm bytes bytes 1)
        (inst vpcmpeqd mask mask mask)

        (inst vpgatherdd search (ea cases bytes 4) mask)

        (inst vmovdqu (ea 32-bit-array) search)
        (inst vzeroupper)
        ))
    vector))