RFR: 8221404: Remove double alignment of RegMasks in Matcher
claes.redestad at oracle.com
Mon Mar 25 18:16:11 UTC 2019
Withdrawing with intent to rework this a bit.
- A single RegMask on my Intel ivy bridge workstation is 22 4-byte mask
words - 88 bytes - plus the 8 bytes for _lwm and _hwm, so 96 bytes. As
double-aligning only shifts the object at most 4 bytes we'll still look
at RegMasks that span 2 cache lines.
- Quick performance runs show neutral results.
- I'm having second thoughts about the utility of _lwm since it sticks
to 0 for many (most?) masks, and my intent with removing the double-
alignment was to make the RegMask layout more malleable to
experimentation with layout without memory waste. I have some data that
suggest we could improve a bit by putting watermarks and the AllStack
bit closer to the first masks.
On 2019-03-25 18:20, Vladimir Kozlov wrote:
> Intel's Skylake has 64 bytes L1 cache line.
> Keeping first 64 register masks bits in one cache line should help. I
> don't think we should do this change.
> On 3/25/19 6:44 AM, Claes Redestad wrote:
>> RegMask are allocated double-aligned, which doesn't seem to have any
>> real effect on any of our supported platforms. Simplify.
>> Bug: https://bugs.openjdk.java.net/browse/JDK-8221404
>> Webrev: http://cr.openjdk.java.net/~redestad/8221404/open.00
>> Testing: tier1-3 (together with JDK-8221343)
More information about the hotspot-compiler-dev