compiler cache stats now display information on a per lang+compiler basis

[robertmaynard]

So with the proposed expansion of clang as a CUDA compiler ( #1892 ) I have noticed that the stats tracking for sccache binned the clang++ cuda compilations as a C/C++ invocation due to the presumption that the clang compiler meant C/C++.

So this PR breaks down stats to be tracked on a way finer granulatiry so that people can see exactly what compinations of language + compiler are causing hits or misses.

[robertmaynard]

Previous sccache stat output would be:

Compile requests                     18
Compile requests executed            18
Cache hits                            9
Cache hits (C/C++)                    6
Cache hits (CUDA)                     3
Cache misses                          9
Cache misses (C/C++)                  6
Cache misses (CUDA)                   3

With this PR it would look like:

Compile requests                     18
Compile requests executed            18
Cache hits                            9
Cache hits (c [clang])                1
Cache hits (c [gcc])                  1
Cache hits (c [nvcc])                 1
Cache hits (c++ [clang])              1
Cache hits (c++ [gcc])                2
Cache hits (c++ [nvcc])               1
Cache hits (cuda [clang])             1
Cache hits (cuda [nvcc])              1
Cache misses                          9
Cache misses (c [clang])              1
Cache misses (c [gcc])                1
Cache misses (c [nvcc])               1
Cache misses (c++ [clang])            1
Cache misses (c++ [gcc])              2
Cache misses (c++ [nvcc])             1
Cache misses (cuda [clang])           1
Cache misses (cuda [nvcc])            1

This is an obvious worse case as I have used 3 different compilers, but I wanted to show the level of detail we now get

[sylvestre]

this will break projects parsing the output or json.
maybe we should keep the original information and add your break down ?

[robertmaynard]

this will break projects parsing the output or json. maybe we should keep the original information and add your break down ?

That was what I was wondering about.

Would it be better to guard the detailed stats behind a different command line switch ( sccache -s -v )? That would make it completly opt-in.

[sylvestre]

yeah, that works too :)

[robertmaynard]

I am going to close this PR in preference to more finer grained PR's that build up to more verbose stat tracking.

The first of those PRs is: #1895