[DYNAREC] Implement perf map

[xiangzhai]

Hi,

To find out DynaRec jit "hot" code:

perf record -e cpu-cycles box64 scimark4-x64

scimark: https://math.nist.gov/scimark2/download_c.html

Before no perf map no Symbol:

Samples: 133K of event 'cpu-cycles:u', Event count (approx.): 66380844254
Overhead  Command       Shared Object          Symbol
   2.38%  scimark4-x64  perf-19878.map         [.] 0x000000fff42c3888
   2.35%  scimark4-x64  perf-19878.map         [.] 0x000000fff42c388c
   1.18%  scimark4-x64  perf-19878.map         [.] 0x000000fff42c62ec
   1.14%  scimark4-x64  perf-19878.map         [.] 0x000000fff42c62fc
   0.96%  scimark4-x64  perf-19878.map         [.] 0x000000fff42c62ac
   0.94%  scimark4-x64  perf-19878.map         [.] 0x000000fff42c0c7c
   0.88%  scimark4-x64  perf-19878.map         [.] 0x000000fff42c0c6c
   0.82%  scimark4-x64  perf-19878.map         [.] 0x000000fff42c0b7c
   0.80%  scimark4-x64  perf-19878.map         [.] 0x000000fff42c0b74
   0.78%  scimark4-x64  perf-19878.map         [.] 0x000000fff42c0cdc
   0.76%  scimark4-x64  perf-19878.map         [.] 0x000000fff42c60ec
   0.73%  scimark4-x64  perf-19878.map         [.] 0x000000fff42c0c74
   0.68%  scimark4-x64  perf-19878.map         [.] 0x000000fff42c0b78
   0.62%  scimark4-x64  perf-19878.map         [.] 0x000000fff42bcec0
   0.62%  scimark4-x64  perf-19878.map         [.] 0x000000fff42bceb0
   0.57%  scimark4-x64  perf-19878.map         [.] 0x000000fff42c0c78
   0.57%  scimark4-x64  perf-19878.map         [.] 0x000000fff42c0c80
   0.56%  scimark4-x64  perf-19878.map         [.] 0x000000fff42c6114
   0.54%  scimark4-x64  perf-19878.map         [.] 0x000000fff42bcef0
   0.52%  scimark4-x64  perf-19878.map         [.] 0x000000fff42c0bf4
   0.52%  scimark4-x64  perf-19878.map         [.] 0x000000fff42c625c
   0.50%  scimark4-x64  perf-19878.map         [.] 0x000000fff42c65c4
...

After generated perf map just use x86 INST NAME as the Symbol name:

Samples: 133K of event 'cpu-cycles:u', Event count (approx.): 66395424841
Overhead  Command       Shared Object          Symbol
   2.39%  scimark4-x64  perf-20589.map         [.] 0x000000fff306388c
   2.31%  scimark4-x64  perf-20589.map         [.] 0x000000fff3063888
   1.14%  scimark4-x64  perf-20589.map         [.] 0x000000fff30662ec
   1.13%  scimark4-x64  perf-20589.map         [.] 0x000000fff30662fc
   0.95%  scimark4-x64  perf-20589.map         [.] ADD Ed, Ib
   0.93%  scimark4-x64  perf-20589.map         [.] 0x000000fff3060c7c
   0.89%  scimark4-x64  perf-20589.map         [.] 0x000000fff3060c6c
   0.86%  scimark4-x64  perf-20589.map         [.] RET
   0.82%  scimark4-x64  perf-20589.map         [.] 0x000000fff3060b7c
   0.81%  scimark4-x64  perf-20589.map         [.] MOVSD Ex, Gx
   0.78%  scimark4-x64  perf-20589.map         [.] JZ ib
   0.75%  scimark4-x64  perf-20589.map         [.] 0x000000fff3060b74
   0.75%  scimark4-x64  perf-20589.map         [.] 0x000000fff3060c74
   0.67%  scimark4-x64  perf-20589.map         [.] ADDSD Gx, Ex
   0.62%  scimark4-x64  perf-20589.map         [.] SUBSD Gx, Ex
   0.61%  scimark4-x64  perf-20589.map         [.] UNPCKHPD Gx, Ex
   0.60%  scimark4-x64  perf-20589.map         [.] MULSD Gx, Ex
   0.57%  scimark4-x64  perf-20589.map         [.] JZ ib
   0.56%  scimark4-x64  perf-20589.map         [.] ADD Ed, Gd
   0.55%  scimark4-x64  perf-20589.map         [.] 0x000000fff306625c
   0.53%  scimark4-x64  perf-20589.map         [.] ADDSD Gx, Ex
...

Please review my patch and give some suggestion.

Thanks,
Leslie Zhai

[ksco]

Thank you! While it's cool, in my opinion, I don't think showing the x86 instruction name is of much use to us. Perf maps are mainly used for perfing guest programs to make it easier to find hot code, so it will be useful for users to find hot code in Java or Python programs with this implemented in JVM or Python VM, but it is not helpful for JVM developers.

But in any case, I will wait for @ptitSeb to review this.

[ptitSeb]

I would prefer the perf-map to represent the name of the x64 function (if any), or the x64 address for anonymous map (that could be later named after the file name map when this will be tracked properly), and then use a simple hex offset for each instruction instead of the opcode name that is kind of anonymous at tis level.

If you think you can easily do the change, I can wait. I you prefer this pr to be merge and look at a "2" level with function name, I can merge this quickly (but indeed, I don't see this function really usefull in it's current state)

[xiangzhai]

I would prefer the perf-map to represent the name of the x64 function

Yup, JVM use the function alike (method or jvmci) name: https://github.com/openjdk/jdk/blob/master/src/hotspot/share/code/codeCache.cpp#L1833
I will change the Symbol to function name.

Thanks,
Leslie Zhai

[xiangzhai]

Use function name as the Symbol:

Samples: 133K of event 'cpu-cycles:u', Event count (approx.): 66363113293
Overhead  Command       Shared Object          Symbol
   2.42%  scimark4-x64  perf-27184.map         [.] 0x000000fff4bdb88c
   2.27%  scimark4-x64  perf-27184.map         [.] 0x000000fff4bdb888
   1.18%  scimark4-x64  perf-27184.map         [.] 0x000000fff4bde2ec
   1.14%  scimark4-x64  perf-27184.map         [.] 0x000000fff4bde2fc
   0.87%  scimark4-x64  perf-27184.map         [.] SparseCompRow_matmult
   0.84%  scimark4-x64  perf-27184.map         [.] SOR_execute
   0.83%  scimark4-x64  perf-27184.map         [.] Random_nextDouble
   0.81%  scimark4-x64  perf-27184.map         [.] SOR_execute
   0.80%  scimark4-x64  perf-27184.map         [.] SparseCompRow_matmult
   0.77%  scimark4-x64  perf-27184.map         [.] SOR_execute
   0.69%  scimark4-x64  perf-27184.map         [.] SOR_execute
   0.68%  scimark4-x64  perf-27184.map         [.] 0x000000fff4bd8c6c
   0.65%  scimark4-x64  perf-27184.map         [.] 0x000000fff4bd8cd8
   0.64%  scimark4-x64  perf-27184.map         [.] 0x000000fff4bd8b74
   0.64%  scimark4-x64  perf-27184.map         [.] FFT_transform_internal
   0.62%  scimark4-x64  perf-27184.map         [.] 0x000000fff4bd8c74
   0.62%  scimark4-x64  perf-27184.map         [.] 0x000000fff4bd8c7c
   0.61%  scimark4-x64  perf-27184.map         [.] FFT_transform_internal
   0.61%  scimark4-x64  perf-27184.map         [.] 0x000000fff4bd8b7c
   0.60%  scimark4-x64  perf-27184.map         [.] SOR_execute
   0.60%  scimark4-x64  perf-27184.map         [.] SparseCompRow_matmult
   0.57%  scimark4-x64  perf-27184.map         [.] FFT_transform_internal
   0.51%  scimark4-x64  perf-27184.map         [.] SOR_execute
   0.50%  scimark4-x64  perf-27184.map         [.] 0x000000fff4bde5e4

Please review my patch again.

Thanks,
Leslie Zhai

[ksco]

The PR looks good to me now.

[ptitSeb]

LGTM too.

[xiangzhai]

Thanks 🍻