sim/testsuite/sim/aarch64/mla.s

   1 # mach: aarch64
   2
   3 # Check the vector multiply add instruction: mla.
   4
   5 .include "testutils.inc"
   6
   7 input:
   8         .word 0x04030201
   9         .word 0x08070605
  10         .word 0x0c0b0a09
  11         .word 0x100f0e0d
  12 m8b:
  13         .word 0x110a0502
  14         .word 0x4132251a
  15 m16b:
  16         .word 0x110a0502
  17         .word 0x4132251a
  18         .word 0x917a6552
  19         .word 0x01e2c5aa
  20 m4h:
  21         .word 0x180a0402
  22         .word 0x70323c1a
  23 m8h:
  24         .word 0x180a0402
  25         .word 0x70323c1a
  26         .word 0x087ab452
  27         .word 0xe0e26caa
  28 m2s:
  29         .word 0x140a0402
  30         .word 0xa46a3c1a
  31 m4s:
  32         .word 0x140a0402
  33         .word 0xa46a3c1a
  34         .word 0xb52ab452
  35         .word 0x464b6caa
  36
  37         start
  38         adrp x0, input
  39         ldr q0, [x0, #:lo12:input]
  40
  41         movi v1.8b, #1
  42         mla v1.8b, v0.8b, v0.8b
  43         mov x1, v1.d[0]
  44         adrp x3, m8b
  45         ldr x4, [x3, #:lo12:m8b]
  46         cmp x1, x4
  47         bne .Lfailure
  48
  49         movi v1.16b, #1
  50         mla v1.16b, v0.16b, v0.16b
  51         mov x1, v1.d[0]
  52         mov x2, v1.d[1]
  53         adrp x3, m16b
  54         ldr x4, [x3, #:lo12:m16b]
  55         cmp x1, x4
  56         bne .Lfailure
  57         ldr x5, [x3, #:lo12:m16b+8]
  58         cmp x2, x5
  59         bne .Lfailure
  60
  61         movi v1.4h, #1
  62         mla v1.4h, v0.4h, v0.4h
  63         mov x1, v1.d[0]
  64         adrp x3, m4h
  65         ldr x4, [x3, #:lo12:m4h]
  66         cmp x1, x4
  67         bne .Lfailure
  68
  69         movi v1.8h, #1
  70         mla v1.8h, v0.8h, v0.8h
  71         mov x1, v1.d[0]
  72         mov x2, v1.d[1]
  73         adrp x3, m8h
  74         ldr x4, [x3, #:lo12:m8h]
  75         cmp x1, x4
  76         bne .Lfailure
  77         ldr x5, [x3, #:lo12:m8h+8]
  78         cmp x2, x5
  79         bne .Lfailure
  80
  81         movi v1.2s, #1
  82         mla v1.2s, v0.2s, v0.2s
  83         mov x1, v1.d[0]
  84         adrp x3, m2s
  85         ldr x4, [x3, #:lo12:m2s]
  86         cmp x1, x4
  87         bne .Lfailure
  88
  89         movi v1.4s, #1
  90         mla v1.4s, v0.4s, v0.4s
  91         mov x1, v1.d[0]
  92         mov x2, v1.d[1]
  93         adrp x3, m4s
  94         ldr x4, [x3, #:lo12:m4s]
  95         cmp x1, x4
  96         bne .Lfailure
  97         ldr x5, [x3, #:lo12:m4s+8]
  98         cmp x2, x5
  99         bne .Lfailure
 100
 101         pass
 102 .Lfailure:
 103         fail