Summary:
With flambda (`-O3`), compilation time is ~5x slower, but the backend is ~25% faster!
To mitigate the atrocious compilation times, introduce a new `opt` build mode in the jbuilder files.
- build in "opt" mode by default from the toplevel (so that install scripts and external users get the fastest infer by default), in "default" mode by default from infer/src (since the latter is only called directly by infer devs, for faster builds)
- `make byte` is as fast as before in any mode
- `make test` will build "opt" by default, which is very slow. Solution for testing (or building the models) locally: `make BUILD_MODE=default test`.
- You can even change the default locally with `export BUILD_MODE=default`.
The benchmarks are to be taken with a sizable pinch of salt because I ran them only once and other stuff could be running in the background. That said, the perf win is consistent across all projects, with 15-20% win in wallclock time and around 25% win in total CPU time, ~9% win in sys time, and ~25% fewer minor allocations, and ~5-10% fewer overall allocations. This is only for the backend; the capture is by and large unaffected (either the same or a tad faster within noise range).
Here are the results running on OpenSSL 1.0.2d on osx (12 cores, 32G RAM)
=== base
infer binary: 26193088 bytes
compile time: 40s
capture:
```lang=text
real 1m7.513s
user 3m11.437s
sys 0m55.236s
```
analysis:
```lang=text
real 5m41.580s
user 61m37.855s
sys 1m12.870s
```
Memory profile:
```lang=json
{
...
"minor_gb": 0.1534719169139862,
"promoted_gb": 0.0038930922746658325,
"major_gb": 0.4546157643198967,
"allocated_gb": 0.6041945889592171,
"minor_collections": 78,
"major_collections": 23,
"compactions": 7,
"top_heap_gb": 0.07388687133789062,
"stack_kb": 0.3984375,
"minor_heap_kb": 8192.0,
...
}
```
=== flambda with stock options (no `-Oclassic`, just the same flags as base)
Exactly the same as base.
=== flambda `-O3`
infer binary: 56870376 bytes (2.17x bigger)
compile time: 191s (4.78x slower)
capture is the same as base:
```lang=text
real 1m9.203s
user 3m12.242s
sys 0m58.905s
```
analysis is ~20% wallclock time faster, ~25% CPU time faster:
```lang=text
real 4m32.656s
user 46m43.987s
sys 1m2.424s
```
memory usage is a bit lower too:
```lang=json
{
...
"minor_gb": 0.11583046615123749, // 75% of previous
"promoted_gb": 0.00363825261592865, // 93% of previous
"major_gb": 0.45415670424699783, // about same
"allocated_gb": 0.5663489177823067, // 94% of previous
"minor_collections": 73,
"major_collections": 22,
"compactions": 7,
"top_heap_gb": 0.07165145874023438,
"stack_kb": 0.3359375,
"minor_heap_kb": 8192.0,
...
}
```
=== flambda `-O2`
Not nearly as exciting as `-O3`, but the compilation cost is still quite high:
infer: 37826856 bytes
compilation of infer: 100s
Capture and analysis timings are mostly the same as base.
Reviewed By: jberdine
Differential Revision: D4867979
fbshipit-source-id: 99230b7
Jules Villard
f8d7c81045