Why use ARCHIVE_START and ARCHIVE_END in Fluid

[wangkuiyi]

In paddle/fluid/inference/*, there are two uses of ARCHIVE_START and ARCHIVE_END:

1. Paddle/paddle/fluid/inference/CMakeLists.txt

   Line 14 in 0821ee7

   DEPS ARCHIVE_START ${GLOB_OP_LIB} ${FLUID_CORE_MODULES} ARCHIVE_END)

2. Paddle/paddle/fluid/inference/tests/book/CMakeLists.txt

   Line 20 in 09b4a1a

   DEPS ARCHIVE_START paddle_fluid ARCHIVE_END

ARCHIVE_START and ARCHIVE_END are not part of cmake/generic.cmake, thus is not supposed to be used in Fluid. Why do we need them in paddle/fluid/inference?

[wangkuiyi]

I am suffering from this problem: #9071 (comment)

[Xreki]

ARCHIVE_START and ARCHIVE_END is defined in https://github.com/PaddlePaddle/Paddle/blob/develop/cmake/util.cmake#L8. It is a couple of flags needed by the linker ld. The usage is different for gcc and clang.

• For gcc, you should use gcc -shared -o libab.so -Wl,--whole-archive liba.a -Wl,--no-whole-archive libb.a

For the effect of -Wl,--whole-archive, you can refer to https://blog.csdn.net/Aquester/article/details/49800237. Without this linking flags, user codes for inference will fail by the error: operators xxx is not registered.

It is also used in caffe2:

https://github.com/caffe2/caffe2/blob/0dd3284525079f3870df92f61bed3b94eb45ff53/cmake/public/utils.cmake#L12-L24

[wangkuiyi]

I know the purpose of -Wl,--whole-archive. I am afraid that it would prevent us from building a binary (maybe a shared object file) which contains only the operators in use?

Anyway, I see that we have the USE_OP mechanism with Fluid:

Paddle/paddle/fluid/operators/dropout_op_test.cc

Line 33 in 767f453

USE_OP(dropout);

I suppose that USE_OP would make us no longer need -Wl,--whole-archive, would it?

[Xreki]

I think you are right, I'll try.