RFR (S): JDK-8191328: Avoid unnecessary overhead in CRC32C

[Andrew Haley]

On 15/11/17 15:38, Alan Bateman wrote:
> Moving the nativeOrder out of the loop make sense but I'm curious about 
> the context for improving this implementation.

I wonder about lifting ByteOrder.nativeOrder().  Maybe it fails to
inline because the method is too large: if that happens, we really
lose.  I'm not seeing that, though: it seems to be inlined just fine,
and has no effect.

In any case, this patch doesn't help anything on my test hardware.

Before:
CRC.crc    avgt    5  1050.374 ? 1.533  ns/op

After:
CRC.crc    avgt    5  1088.962 ? 0.212  ns/op

And here, for the sake of anyone interested, is the AArch64 code,
before:

 ;; B10: #	B10 B11 <- B9 B10 	Loop: B10-B10 inner main of N120 Freq: 127.001

  0x000003ffa4bb5900: ldr	x2, [x11,w3,sxtw]  ;*invokevirtual getLong {reexecute=0 rethrow=0 return_oop=0}
                                                ; - java.util.zip.CRC32C::updateBytes at 147 (line 237)

  0x000003ffa4bb5904: eor	w4, w2, w1      ;*ixor {reexecute=0 rethrow=0 return_oop=0}
                                                ; - java.util.zip.CRC32C::updateBytes at 193 (line 246)

  0x000003ffa4bb5908: lsr	x2, x2, #32
  0x000003ffa4bb590c: ubfiz	x5, x4, #2, #8
  0x000003ffa4bb5910: mov	w19, w2         ;*l2i {reexecute=0 rethrow=0 return_oop=0}
                                                ; - java.util.zip.CRC32C::updateBytes at 171 (line 240)

  0x000003ffa4bb5914: ubfx	w2, w4, #16, #8
  0x000003ffa4bb5918: ubfx	w1, w4, #8, #8
  0x000003ffa4bb591c: ldr	w6, [x15,x5]
  0x000003ffa4bb5920: lsr	w5, w4, #24
  0x000003ffa4bb5924: ldr	w4, [x16,w2,sxtw #2]
  0x000003ffa4bb5928: ldr	w1, [x14,w1,sxtw #2]
  0x000003ffa4bb592c: ldr	w2, [x17,w5,sxtw #2]
  0x000003ffa4bb5930: eor	w5, w1, w6
  0x000003ffa4bb5934: ubfiz	x23, x19, #2, #8
  0x000003ffa4bb5938: ubfx	w6, w19, #8, #8
  0x000003ffa4bb593c: eor	w4, w5, w4
  0x000003ffa4bb5940: ubfx	w5, w19, #16, #8
  0x000003ffa4bb5944: ldr	w1, [x18,x23]
  0x000003ffa4bb5948: eor	w2, w4, w2
  0x000003ffa4bb594c: ldr	w6, [x0,w6,sxtw #2]
  0x000003ffa4bb5950: lsr	w4, w19, #24
  0x000003ffa4bb5954: ldr	w19, [x13,w5,sxtw #2]
  0x000003ffa4bb5958: eor	w2, w2, w1
  0x000003ffa4bb595c: add	w5, w3, #0x8
  0x000003ffa4bb5960: ldr	w1, [x12,w4,sxtw #2]
  0x000003ffa4bb5964: eor	w29, w2, w6
  0x000003ffa4bb5968: eor	w4, w29, w19

and after:

 ;; B10: #	B10 B11 <- B9 B10 	Loop: B10-B10 inner main of N121 Freq: 127.001

  0x000003ff8cb98200: ldr	x10, [x21,w3,sxtw]
                                                ;*invokevirtual getLong {reexecute=0 rethrow=0 return_oop=0}
                                                ; - java.util.zip.CRC32C::updateBytes at 359 (line 264)

  0x000003ff8cb98204: eor	w1, w10, w1     ;*ixor {reexecute=0 rethrow=0 return_oop=0}
                                                ; - java.util.zip.CRC32C::updateBytes at 380 (line 268)

  0x000003ff8cb98208: ubfx	w2, w1, #8, #8
  0x000003ff8cb9820c: lsr	x10, x10, #32
  0x000003ff8cb98210: ubfx	w5, w1, #16, #8
  0x000003ff8cb98214: mov	w10, w10        ;*l2i {reexecute=0 rethrow=0 return_oop=0}
                                                ; - java.util.zip.CRC32C::updateBytes at 374 (line 266)

  0x000003ff8cb98218: ldr	w0, [x15,w2,sxtw #2]
  0x000003ff8cb9821c: lsr	w6, w1, #24
  0x000003ff8cb98220: ubfiz	x4, x1, #2, #8
  0x000003ff8cb98224: ubfx	w1, w10, #8, #8
  0x000003ff8cb98228: ldr	w2, [x17,w5,sxtw #2]
  0x000003ff8cb9822c: ubfiz	x20, x10, #2, #8
  0x000003ff8cb98230: lsr	w5, w10, #24
  0x000003ff8cb98234: ubfx	w19, w10, #16, #8
  0x000003ff8cb98238: ldr	w10, [x13,w1,sxtw #2]
  0x000003ff8cb9823c: ldr	w6, [x18,w6,sxtw #2]
  0x000003ff8cb98240: ldr	w4, [x16,x4]
  0x000003ff8cb98244: ldr	w1, [x12,w19,sxtw #2]
  0x000003ff8cb98248: ldr	w7, [x14,x20]
  0x000003ff8cb9824c: eor	w6, w2, w6
  0x000003ff8cb98250: eor	w0, w0, w4
  0x000003ff8cb98254: ldr	w19, [x11,w5,sxtw #2]
  0x000003ff8cb98258: add	w20, w3, #0x8
  0x000003ff8cb9825c: eor	w29, w0, w6
  0x000003ff8cb98260: eor	w10, w10, w7
  0x000003ff8cb98264: eor	w0, w1, w19
  0x000003ff8cb98268: eor	w1, w10, w29

-- 
Andrew Haley
Java Platform Lead Engineer
Red Hat UK Ltd. <https://www.redhat.com>
EAC8 43EB D3EF DB98 CC77 2FAD A5CD 6035 332F A671