SIMD auto-vectorization in hotspot

[Krystal Mok]

Hi,

Apparently, the source code to read is opto/superword.[hpp|cpp] [1][2]. The
flag to trace it is -XX:+TraceSuperWord.
There's a paper mentioned in [1], which is the origin of the algorithm used.

An example to see superword in action is a simple array copy loop:

// byte[] src = ...
// byte[] dest = ...
// assert(src.length == dest.length)
for (int i = 0; i < src.length; i++) {
  dest[i] = src[i];
}

This loop will be vectorized by superword and then unrolled a bit, turning
the actual copy into something like: (on 32-bit x86)

movq   0xc(%ecx,%edi,4),%xmm0
movq   %xmm0,0xc(%edx,%edi,4)

HTH,
- Kris

[1]:
http://hg.openjdk.java.net/hsx/hotspot-main/hotspot/file/tip/src/share/vm/opto/superword.hpp
[2]:
http://hg.openjdk.java.net/hsx/hotspot-main/hotspot/file/tip/src/share/vm/opto/superword.cpp<http://hg.openjdk.java.net/hsx/hotspot-main/hotspot/file/tip/src/share/vm/opto/superword.hpp>


On Tue, Jan 17, 2012 at 1:45 PM, Vitaly Davidovich <vitalyd at gmail.com>wrote:

> Hi guys,
>
> Does the -XX:+UseSuperWord (I see it's set to true on 6u23) optimization
> enable auto-vectorization of certain loops? If so, are there any switches
> that can be turned on (in product VM) to see if vectorization is being
> applied or not? In general, would someone be able to describe the current
> state of loop vectorization (i.e. implementation limits, restrictions,
> etc)? Which ISAs are supported (i.e. SSE, AVX)? Any other info on this
> subject would be greatly appreciated ...
>
> Thanks very much,
>
> Vitaly
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.openjdk.java.net/pipermail/hotspot-compiler-dev/attachments/20120117/db942185/attachment.html