H3 library is not inlined

[Komzpa]

Hi,

Investingating performance issues I found that H3 actively resists inlining.

objdump -D h3.so:

Simple math like *pi/180 is not getting inlined.

After some fight I got some pieces but haven't got to destination.

CMake supposedly enables inline like this:

cmake_minimum_required(VERSION 3.9)
cmake_policy(SET CMP0069 NEW)
set(CMAKE_INTERPROCEDURAL_OPTIMIZATION TRUE)

These guys helped a bit inside the h3-pg itself but haven't got me to let linktime optimizations to cross library boundary.

CFLAGS += -flto -O2
LDFLAGS += -flto -Wl,-Bsymbolic-functions -Wl,-z,relro

Any ideas help.

[zachasme]

Thanks for the PR, I'll try looking into it. Could it be related to the way h3 is built?

cmake \
  -DCMAKE_C_FLAGS=-fPIC \
  -DBUILD_TESTING=OFF \
  -DENABLE_COVERAGE=OFF \
  -DENABLE_DOCS=OFF \
  -DENABLE_FORMAT=OFF \
  -DENABLE_LINTING=OFF \
  ..

Is there a Release flag I'm missing or similar?

Thanks for opening the issue on h3 core as well.

[Komzpa]

Per uber/h3#326 (comment) flag is -DCMAKE_BUILD_TYPE=Release, thanks @isaacbrodsky.

[Komzpa]

I've spent another couple of hours with it and give up for today.
Test is simple: after you build everything right, this should show empty. In non-inlined build it shows four callsites for square and the function def.
objdump -D h3.so | grep _square

[AbelVM]

Per uber/h3#326 (comment) flag is -DCMAKE_BUILD_TYPE=Release, thanks @isaacbrodsky.

Why is this flag not actually used in the latest v3.6.2 release? It looks harmless, and the 2x/3x improvements shown at uber/h3#326 are huge

[zachasme]

Release v3.6.3 includes -DCMAKE_BUILD_TYPE=Release which should improve performance.

Still objdump -D h3.so | grep _square shows three callsites, so I'll keep this issue open until we figure out how to inline h3. Hopefully something good will come from uber/h3#326 but apart from that I have no ideas.

@Komzpa do you have any ideas for benchmarking the built extension? I would like to be able to compare changes like these (after being built into extension and not just h3 core directly).

[Komzpa]

@zachasme pgbench is an industry standard to benchmark query execution times in postgres.

[zachasme]

We should also keep an eye on uber/h3#360.

[zachasme]

Solved by #75