RFR: 8218675: Reduce verification overhead in ClassFileParser
David Holmes
david.holmes at oracle.com
Mon Feb 11 11:25:49 UTC 2019
Sorry didn't see this before responding to your previous mail.
On 11/02/2019 9:09 pm, Claes Redestad wrote:
> On 2019-02-11 10:07, Claes Redestad wrote:
>>
>>
>> On 2019-02-11 09:44, David Holmes wrote:
>>>
>>>> All of the modified UTF8-encoded bytes UTF8::next would skip over will
>>>> have values >= 128 (assuming legal UTF-8, which UTF8::next already
>>>> assumes), so they can't match the special characters we're looking
>>>> for anyway, so the switch will never match any of them. Thus the new
>>>> method
>>>> is semantically equivalent, but measurably faster.
>>>
>>> New code compares p1, p2, p3 ... to each of the special characters to
>>> see that it doesn't match. Or are you saying the switch somehow
>>> generates a set of range checks rather than comparisons??
>>
>> No, I'm saying that any byte value that would have been skipped over by
>> UTF8::next will be skipped over by the switch, since they are all 128
>> and higher. It's impossible that, say, a ';' hides in one of the bytes
>> of a codepoint.
>>
>> UTF8 was designed to allow scanning for ASCII characters byte-by-byte
>> like this, and modified UTF-8 is no different except for '\0' (which we
>> don't care about here).
>
> If you were just questioning the performance results, GCC does compile
> the switch prelude in this method to a quick range check:
>
> sub $0x2e,%ecx // 2e is '.', the lowest character
> cmp $0x2d,%cl // 2e + 2d is '[', the highest character
> ja ...
Okay that's what I was wondering. But seems very compiler specific - are
they all this clever or does the new code only work better on some
platforms?
David
> The most common range of characters, a-z, is outside of this range: for
> all these we do one subtraction and a compare + branch in the common
> case, compared to the baseline which did at least two branches per byte.
>
> Specifically, this optimization amounts to ~100k fewer instructions on
> startup even with -Xverify:none, a large chunk of which were cmp + jump.
>
> /Claes
More information about the hotspot-runtime-dev
mailing list