Unable to build ceres-solver for iOS with c/c++ flag `-fembed-bitcode`

[AndreasSchacherbauer]

Hi,
I tried to build ceres for iOS with the bitcode option enabled but I got the following error: clang: error: -mllvm is not supported with -fembed-bitcode.
I had to remove the -mllvm cxx flag in the main CMakeLists.txt file and was then able to specify -fembed-bitcode in the iOS.cmake toolchain file line 250-255.

For what is the -mllvm option needed or is there another way to resolve my bitcode build error?

[alexsmac]

The -mllvm option is required to pass the -inline-threshold=600 option through Clang to the LLVM pipeline. As per the comment in the CMakeLists, historically we have found that increasing the inlining threshold on Clang made a big performance improvement overall as otherwise Eigen was being hobbled as various functions were not being inlined.

As it seems from the Clang/LLVM code that -mllvm and -fembed-bitcode are incompatible options, I don't think that there is a workaround here other than removing the -mllvm -inline-threshold=600, however that would impact performance.

[sandwichmaker]

@alexsmac it maybe worth looking at the -inline-threshold again. Its been a while since we evaluated the need for it.

[sandwichmaker]

A quick informal test on my laptop does not indicate that this is particularly useful.
@alexsmac do you remember what prompted this inline threshold to begin with?

[sandwichmaker]

So this seems to be the CL that added this

3d6eceb

so I need to check as to what happens to the performance of ITERATIVE_SCHUR with and without this.

[alexsmac]

I think it was something @keir added to improve Jet performance, without it I think autodiff cost functions were being clobbered when evaluating Jets with larger parameter sizes.

[sandwichmaker]

Turns out this has to do with the Schur based solvers rather than Jets. And there is a performance differential, but I need to spend some time to see if we can do this via compiler annotations instead of compiler flags.