RFR: 8270459: Conflict inlining decisions by C1/C2 with the same CompileCommand [v2]

[Vladimir Kozlov]

On Thu, 15 Jul 2021 06:56:44 GMT, Jie Fu <jiefu at openjdk.org> wrote:

>> Hi all,
>> 
>> Conflict inlining decisions can be observed by C1/C2 with the same CompileCommand.
>> And I came across this situation while I was analyzing a performance issue using the jmh, which made me quite confused.
>> 
>> For the example in the JBS.
>> 
>>      28    1 %     3       TestInline::main @ 11 (43 bytes)
>>                               @ 23   TestInline::caller (4 bytes)   disallowed by CompileCommand
>>      29    2       3       TestInline::main (43 bytes)
>>                               @ 23   TestInline::caller (4 bytes)   disallowed by CompileCommand
>>      31    3 %     4       TestInline::main @ 11 (43 bytes)
>>      33    1 %     3       TestInline::main @ 11 (43 bytes)   made not entrant
>>                               @ 23   TestInline::caller (4 bytes)   force inline by CompileCommand
>> 
>> 
>> This can be easily reproduced by specifying both the `inline` and `dontinline` compile commands for a method.
>> 
>>    -XX:CompileCommand=inline,TestInline::caller \
>>    -XX:CompileCommand=dontinline,TestInline::caller \
>> 
>> 
>> The fix will ignore the inline/not-inline directives if both the `inline` and `dontinline` compile commands are specified for a method.
>> 
>> Effect of the patch:
>> 
>>      27    1 %     3       TestInline::main @ 11 (43 bytes)
>>                               @ 23   TestInline::caller (4 bytes)   inline
>>      28    2       3       TestInline::main (43 bytes)
>>                               @ 23   TestInline::caller (4 bytes)   inline
>>      28    3 %     4       TestInline::main @ 11 (43 bytes)
>>      30    1 %     3       TestInline::main @ 11 (43 bytes)   made not entrant
>>                               @ 23   TestInline::caller (4 bytes)   inline (hot)
>>      31    3 %     4       TestInline::main @ 11 (43 bytes)   made not entrant
>> 
>> 
>> Testing:
>>   - tier 1~3 on Linux/x64, no regression
>> 
>> Thanks.
>> Best regards,
>> Jie
>
> Jie Fu has updated the pull request incrementally with two additional commits since the last revision:
> 
>  - The jtreg test requires a compiler
>  - The last specified inlining option should overwrite all previous

src/hotspot/share/compiler/compilerOracle.cpp line 355:

> 353:       // Find the last one for that method and return the predicate accordingly
> 354:       enum CompileCommand last_one = CompileCommand::Unknown;
> 355:       TypedMethodOptionMatcher* current = option_list;

Add comment that `option_list` lists options in reverse order. So the first option we find is the last which was specified.

-------------

PR: https://git.openjdk.java.net/jdk/pull/4780