sse-seq.lisp

   1 (in-package :sb-vm)
   2
   3 (defmacro vect-ea (base &optional idx (width :dword))
   4   (let ((disp
   5          (if (and idx (numberp idx))
   6              `(+ (- (* VECTOR-DATA-OFFSET N-WORD-BYTES) OTHER-POINTER-LOWTAG) ,idx)
   7              `(- (* VECTOR-DATA-OFFSET N-WORD-BYTES) OTHER-POINTER-LOWTAG))))
   8
   9 ;;    (format t "ea ~A ~A ~A~%" base idx (and idx (symbolp idx)))
  10     (if (and idx (symbolp idx))
  11         `(make-ea ,width :base ,base :index ,idx :disp ,disp)
  12         `(make-ea ,width :base ,base :disp ,disp))))
  13
  14
  15 (DEFINE-VOP (%sse-seq=)
  16     (:POLICY :FAST-SAFE)
  17   (:ARGS (seq1 :SCS (DESCRIPTOR-REG))
  18          (seq2 :SCS (DESCRIPTOR-REG)))
  19   (:ARG-TYPES simple-array-unsigned-byte-8 simple-array-unsigned-byte-8 )
  20
  21   (:results (RESULT :SCS (DESCRIPTOR-REG)))
  22
  23   (:result-types fixnum)
  24
  25   (:TEMPORARY (:SC XMM-REG) X0)
  26   (:TEMPORARY (:SC XMM-REG) X1)
  27   (:TEMPORARY (:SC XMM-REG) X2)
  28   (:TEMPORARY (:SC XMM-REG) X3)
  29   (:TEMPORARY (:SC XMM-REG) X4)
  30   (:TEMPORARY (:SC XMM-REG) X5)
  31
  32 ;;  (:TEMPORARY (:SC unsigned-reg :offset edx-offset) edx)
  33
  34   (:TEMPORARY (:SC unsigned-reg :offset ebx-offset) index)
  35   (:TEMPORARY (:SC unsigned-reg :offset ecx-offset) length)
  36
  37   (:GENERATOR 10
  38
  39     (let ((top (gen-label))
  40 ;;        (top2 (gen-label))
  41           (length-ok (gen-label))
  42           (fail (gen-label))
  43           (the-end (gen-label))
  44           (end (gen-label)))
  45
  46               (loadw index seq1 vector-length-slot other-pointer-lowtag)
  47               (loadw length seq2 vector-length-slot other-pointer-lowtag)
  48
  49               ;; same length ?
  50               (inst cmp index length)
  51               (inst jmp :eq length-ok)
  52
  53               ;; not same length, fail
  54               (inst mov result -1)
  55               (inst jmp end)
  56
  57               (emit-label length-ok)
  58
  59               ;; un-fixnumize length
  60               (inst shr length 2)
  61
  62               ;; calc number of 256bit blocks
  63               (inst shr length (floor (log (/ 256 8) 2)))
  64
  65               ;; init indices
  66               (inst xor index index)
  67
  68               ;; zero eq-regs
  69               (inst pxor x4 x4)
  70               (inst pxor x5 x5)
  71
  72               (emit-label top)
  73
  74               ;; load first blocks
  75               (inst movdqu x0 (vect-ea seq1 index :xmmword))
  76               (inst movdqu x1 (vect-ea seq2 index :xmmword))
  77
  78               ;; load second blocks
  79               (inst movdqu x2
  80                     (make-ea :xmmword :base seq1 :index index
  81                              :disp (+ (- (* VECTOR-DATA-OFFSET N-WORD-BYTES) OTHER-POINTER-LOWTAG) 16)))
  82               (inst movdqu x3
  83                     (make-ea :xmmword :base seq2 :index index
  84                              :disp (+ (- (* VECTOR-DATA-OFFSET N-WORD-BYTES) OTHER-POINTER-LOWTAG) 16)))
  85
  86
  87               ;; xor first/second blocks (i.e. if equal, xor will be zero)
  88               (inst pxor x0 x1)
  89               (inst pxor x2 x3)
  90
  91               ;; add index
  92               (inst add index 32)
  93
  94               ;; or bits to eq-regs (if not eq, some bits will be nonzero)
  95               (inst por x4 x0)
  96               (inst por x5 x2)
  97
  98               ;; loop
  99               (inst dec length)
 100               (inst jmp :nz top)
 101
 102
 103               ;; all 256bit blocks done
 104
 105               ;; or each 32bit word from x4 to x5
 106               (inst por x4 x5)
 107               (inst movdqa x0 x4)
 108
 109               (inst psrldq-ib x4 4)  ;; this is number of bytes, not bits
 110               (inst por x0 x4)
 111
 112               (inst psrldq-ib x4 4)
 113               (inst por x0 x4)
 114
 115               (inst psrldq-ib x4 4)
 116               (inst por x0 x4)
 117
 118               ;; now low 32bits of x0 will be non-zero if seq's not equal
 119
 120               (inst movd result x0)
 121
 122               ;; end
 123               (emit-label end)
 124
 125               (inst test result result)
 126               (inst jmp :nz fail)
 127
 128               (inst mov result (fixnumize 0))
 129               (inst jmp the-end)
 130
 131               (emit-label fail)
 132               (inst mov result (fixnumize 1))
 133
 134               (emit-label the-end)
 135
 136               )))