RFR(L): 8028468: Add inlining information into ciReplay

[Vladimir Kozlov]

On 1/7/14 4:49 AM, Roland Westrelin wrote:
> One more question:
>
> When can the code below cause an early return from process_compile()? Can that really happen?

When replaying inlining it will skip OSR compilation if you dumped 
inlining for normal compilation before (entry_bci != _entry_bci). And 
reverse.

Dumping inlining from C2 and run application with Tiered. We need to 
skip C1 compilations (when we have replay inlining in C1) (comp_level != 
_comp_level).

You can specify several DumpInline commands for different methods and 
one output inline data file so it will contains several lines of 
inlining data for different methods. We need to find only related data 
for current compilation.

thanks,
Vladimir

>
>   493       if (entry_bci != _entry_bci || comp_level != _comp_level) {
>   494         return;
>   495       }
>   496       const char* iklass_name  = _imethod->method_holder()->name()->as_utf8();
>   497       const char* imethod_name = _imethod->name()->as_utf8();
>   498       const char* isignature   = _imethod->signature()->as_utf8();
>   499       const char* klass_name   = method->method_holder()->name()->as_utf8();
>   500       const char* method_name  = method->name()->as_utf8();
>   501       const char* signature    = method->signature()->as_utf8();
>   502       if (strcmp(iklass_name,  klass_name)  != 0 ||
>   503           strcmp(imethod_name, method_name) != 0 ||
>   504           strcmp(isignature,   signature)   != 0) {
>   505         return;
>   506       }
>   507     }
>
> Roland.
>
> On Jan 7, 2014, at 11:39 AM, Roland Westrelin <roland.westrelin at oracle.com> wrote:
>
>>> http://cr.openjdk.java.net/~kvn/8028468/webrev/
>>
>> Should the agent/doc/cireplay.html be updated? Is there another doc on how to use the replay support somewhere (wiki)?
>>
>> ciReplay.cpp
>> typos:
>> // Use pointer because we may need to retirn inline records
>>
>> // Replay Inlinig
>>
>> vmError.cpp
>> typo:
>> // Do not overwite file during replay
>>
>> compile.hpp
>> Shouldn’t _replay_inline_data be private with an accessor?
>>
>> 1183   // Dump inlining replay data to the stream.
>> 1184   void dump_inline_data(outputStream* out);
>> 1185   void* _replay_inline_data;
>>
>> ciReplay.cpp
>> Why was this changed? Why was it there in the first place?
>> 1052   // Make sure we don't run with background compilation
>> 1053   //  BackgroundCompilation = false;
>>
>> Existing fields in class CompileReplay that don’t start with _ make the code confusing:
>>
>>   char* bufptr;
>>   char* buffer;
>>   int   buffer_length;
>>   int   buffer_end;
>>   int   line_no;
>>
>> etc. Maybe it would be worthwhile to fix that as part of this change?
>>
>> Can buffer be a growableArray? If not factor this code and other similar code in a method?
>> 437       if (pos + 1 >= buffer_length) {
>> 438         int newl = buffer_length * 2;
>> 439         char* newb = NEW_RESOURCE_ARRAY(char, newl);
>> 440         memcpy(newb, buffer, pos);
>> 441         buffer = newb;
>> 442         buffer_length = newl;
>> 443       }
>>
>> Same for:
>> 445         // null terminate it, reset the pointer and process the line
>> 446         buffer[pos] = '\0';
>> 447         buffer_end = pos++;
>> 448         bufptr = buffer;
>>
>> Shouldn’t the fields below be private?
>> 422   ciKlass* _iklass;
>> 423   Method* _imethod;
>> 424   int _entry_bci;
>> 425   int _comp_level;
>>
>> I think
>> ciInlineRecord* find_ciInlineRecord(Method* method, int bci, int inline_depth)
>> should be implemented by calling:
>> static ciInlineRecord* find_ciInlineRecord(GrowableArray<ciInlineRecord*>*  records, Method* method, int bci, int inline_depth)
>> to avoid duplication of code.
>>
>> Roland.
>