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 eax-offset :to (:result 0)) temp1)
  33   (:TEMPORARY (:SC unsigned-reg :offset edx-offset) temp2)
  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 jmp fail)
  55
  56               (emit-label length-ok)
  57
  58               ;; un-fixnumize length
  59               (inst shr length 2)
  60
  61               ;; calc number of 256bit blocks
  62               (inst shr length (floor (log (/ 256 8) 2)))
  63
  64               ;; init indices
  65               (inst xor index index)
  66
  67               ;; zero eq-regs
  68 ;;            (inst pxor x4 x4)
  69 ;;            (inst pxor x5 x5)
  70
  71               (emit-label top)
  72
  73               ;; load first blocks
  74               (inst movdqu x0 (vect-ea seq1 index :xmmword))
  75               (inst movdqu x1 (vect-ea seq2 index :xmmword))
  76
  77               (inst pxor x4 x4)
  78               (inst pxor x5 x5)
  79
  80               ;; load second blocks
  81               (inst movdqu x2
  82                     (make-ea :xmmword :base seq1 :index index
  83                              :disp (+ (- (* VECTOR-DATA-OFFSET N-WORD-BYTES) OTHER-POINTER-LOWTAG) 16)))
  84               (inst movdqu x3
  85                     (make-ea :xmmword :base seq2 :index index
  86                              :disp (+ (- (* VECTOR-DATA-OFFSET N-WORD-BYTES) OTHER-POINTER-LOWTAG) 16)))
  87
  88
  89               ;; xor first/second blocks (i.e. if equal, xor will be zero)
  90               (inst pxor x0 x1)
  91               (inst pxor x2 x3)
  92
  93               ;; add index
  94               (inst add index 32)
  95
  96               ;; check for non-equality
  97               (inst pcmpeqd x4 x0)
  98               (inst pcmpeqd x5 x2)
  99
 100               (inst pmovmskb temp1 x4)
 101               (inst pmovmskb temp2 x5)
 102
 103               (inst cmp temp1 #x0000FFFF)
 104               (inst jmp :ne fail)
 105
 106               (inst cmp temp2 #x0000FFFF)
 107               (inst jmp :ne fail)
 108
 109               ;; loop
 110               (inst dec length)
 111               (inst jmp :nz top)
 112
 113
 114               ;; all 256bit blocks done
 115
 116
 117               ;; check remaining bytes
 118               (loadw length seq1 vector-length-slot other-pointer-lowtag)
 119               (inst shr length 2)
 120               (inst and length (1- (/ 256 8)))
 121
 122               ;; no bytes left ?
 123               (inst test length length)
 124               (inst jmp :z end)
 125
 126               (inst xor temp1 temp1)
 127               (inst xor temp2 temp2)
 128
 129               (emit-label top2)
 130
 131               ;; test bytes
 132               (inst movzx temp1 (vect-ea seq1 index :byte))
 133               (inst movzx temp2 (vect-ea seq2 index :byte))
 134               (inst xor temp1 temp2)
 135               (inst inc index)
 136
 137               ;; if not zero, fail
 138               (inst test temp1 temp1)
 139               (inst jmp :nz fail)
 140
 141               ;; loop
 142               (inst dec length)
 143               (inst jmp :nz top2)
 144
 145               ;; end
 146               (emit-label end)
 147
 148               (inst mov result (fixnumize 0))
 149               (inst jmp the-end)
 150
 151               ;; fail
 152               (emit-label fail)
 153               (inst mov result (fixnumize 1))
 154
 155               ;; the-end
 156               (emit-label the-end)
 157
 158               )))