Measure and potentially fix build time regression

[sdeleuze]

My measurements on webmvc-tomcat sample

• 0.8.2 - 20.2.0 : 64.8s
• 0.8.2 - 20.3.0 : 60.9s
• 0.8.3 - 20.2.0 : 74.7s
• 0.8.3 - 20.3.0 : 71.8s

So there is a regression between 0.8.2 and 0.8.3 we should identify and remove.

Originally raised via https://spring.io/blog/2020/11/23/spring-native-for-graalvm-0-8-3-available-now#comment-5163901805.

[sdeleuze]

@aclement I compared both latest 0.9.0-SNAPSHOT and 0.8.2 and I confirm there is a significant regression visible on commandinerunner and webmvc-tomcat samples. Notice this is not specific to bti, the regression happened between 0.8.2 and 0.8.3 and leads to bigger native-image so longer build. Comparing what is included in the image should allow us to identify and maybe fix this regression.

[sdeleuze]

I think I have found the root cause and a fix, seems to be related to lack of code removal at build time, I will share more details later.

[sdeleuze]

Based on my tests, it seems leveraging the new NativeDetector utility seems to reduce image size and build time by 15%, to this issue should be fixed when #432 will be fixed.

[sdeleuze]

Ok I get more evidences, previous finding as just due to oracle/graal#3163 so not relevant.

I did a diff of the classes shipped in the native-image between 0.8.2 (known to compile fast) and master. The difference seems to be due to reflection entries bellow present only in master:
AbstractErrorController
AutoConfigurationImportSelector methods
BasicErrorController methods
CacheType
Compression
DataSize
DateTokenConverter
DefaultMessageCodesResolver$Format
DeserializationFeature
DocumentBuilderFactoryImpl
Encoding
ErrorProperties
GenericWebApplicationContext methods
Http2
HttpStatus
IntegerTokenConverter
JsonAutoDetect
JsonGenerator methods
JsonParser
Jsp
MapperFeature
MediaType
PropertyAccessor
SerializationFeature
Session
SpringApplication
Ssl
Yaml -> see #484

Some could be legit to fix compat, some could be just footprint regression due to hint updates. See attached diff for more details.
diff-0.8.2.txt

[aclement]

Some work done. Some of the new types that are appearing in the diff now are due us being better/more-compatible - in particular they relate to the changes in handling for configuration properties (using getters return types). So what we are doing is more correct but we are paying a heavy price for properties that will never be bound. I've enhanced that property code again, it is smarter now, which helps. I also wired that code up to the build-time-property-checking flags for the first time (defaults to off but you can turn it on) - previously that was only wired to the @ConditionalOnProperty checks.

I've stopped using J11 for this kind of analysis, the unpredictable numbers drive me crazy. So comparing J8 below.

Comparing 0.8.2 and 0.9.0: (first are 0.8.2, second are 0.9.0 and third are 0.9.0 with build time property checking turned on):

commandlinerunner
Build Time:  47.1s > 45.8s > 46.5s
RSS Memory: 23.1 > 21.8M > 21.6M
Image Size: 20.7M > 20.9M > 20.9M

webmvc-tomcat
Build Time: 70.2s > 76.7s > 76.6s
RSS Memory: 52.0M > 50.1M > 48.2M
Image Size: 43.1M > 44.8M > 43.5M

Some of this included hint tweaks so I need to see a clean samples run on the CI to verify nothing is really broken. I'll do a bit more on it this week whilst these changes bed in.