the new optimizations should be moved into a separated dispatch-able source that only holds targets "baseline vsx4" e.g. loops_modulo.dispatch.c.src to avoid the unnecessary increase of binary size on x86 for the un-affected targets (sse41 avx2 avx512f avx512_skx) within "loops_arithmetic.dispatch.c.src ".

Hi @seiko2plus, thank you for reviewing my code.

Below are two points I would like to discuss with you:

1. one of the operations I optimized in this PR was floor_divide (signed integer types -- array // array). This operation, in particular, has some similarities compared to divmod, which allowed me to reuse the same code to implement/optimize both operations for VSX4. I think that if I move divmod to a new dispatchable source file, I will have to duplicate a considerable part of the code there.

2. the optimizations I applied within the dispatchable source file loops_arithmetic.dispatch.c.src are inside an ifdef macro (#if defined(NPY_HAVE_VSX4)), which means that the compiler's preprocessor will remove this code before generating the binary for the other targets.

I would like to know if even considering the points above I should still move the code to another dispatchable source file. If you still think I should apply the suggested changes, I will do it.

Hi, I would like to request a review of my code (cc: @mattip , @seiko2plus )

For this PR, specifically, I would like to know if what I said above makes sense, if not, let me know so I will apply the changes that were asked.

The errors in the CI seem to be something related to timeout. The PRs
of the first page were closed and opened (in sequence) last week.

Thanks!

Hi, I analyzed the size of the binaries generated for x86, and as stated by @seiko2plus, the contributions of this PR are indeed generating some increase in the binaries since the code moved from loops.c.src is being replicated for each of the SIMD extensions defined as targets within loops_arithmetic.dispatch.c.src (see below).

**BEFORE --> AFTER**

loops_arithmetic.dispatch.avx2.o
16K -->  26K

loops_arithmetic.dispatch.avx512f.o
17K --> 27K

loops_arithmetic.dispatch.avx512_skx.o
16K --> 26K

loops_arithmetic.dispatch.sse41.o
16K --> 26K

loops.o
1055KB --> 1046KB

I will move the operations fmod, remainder and divmod to a new dispatchable souce file as requested.

Hi @seiko2plus, I modified the code as you asked.

With the new changes, the binaries generated for the other architectures now have the same size with/without this PR.

See below the same table I shared in my previous comment updated:

**BEFORE --> AFTER**

loops_arithmetic.dispatch.avx2.o
16K -->  16K

loops_arithmetic.dispatch.avx512f.o
17K --> 17K

loops_arithmetic.dispatch.avx512_skx.o
16K --> 16K

loops_arithmetic.dispatch.sse41.o
16K --> 16K

loops.o
1055KB --> 1046KB

loops_modulo.dispatch.o
0 --> 11K

Thank you! 👍

Thanks @rafaelcfsousa

Thank you @rafaelcfsousa, @mattip. sorry for delayed response