File-specific compilation flags

[mratsim]

As mentioned a long time ago on IRC and here #11202 (comment) file-specific compilation flags would be useful.

Currently I have to maintain a complex cfg and/or nimble files with variations for Windows/Mac/Linux + making sure that the compiler doesn't compile a SSE2 file with AVX2 otherwise the SSE2 platform will not get SSE2 SIMD.

https://github.com/numforge/laser/blob/3c3c6f526d877cb8fdb7c8bf6e0f0867259e6a42/nim.cfg#L18-L37
https://github.com/mratsim/Arraymancer/blob/master/arraymancer.nimble#L40-L81

This would be useful for:

• restricting CUDA/OpenCL compilation to specific files as the compilers sometime break on things that GCC/Clang support. And it also requires stripping all compilation flags of Nim and the users and me maintaining parity with the official Nim.cfg
• allow me to mix SSE2 and AVX in the same codebase. I can use the GCC Function attribute __attribute(target('avx')) as a workaround with codegendecl but it also forces me to rewrite all Nim macros like N_INLINE, N_CLOSURE, ... see codegenDecl on functions overrides all attributes #10682
• avoid people being confused when import or copy my code: gemm_strided: error: always_inline function '_mm256_setzero_pd' requires target feature 'xsave' mratsim/laser#22
• having file-level fast-math when the loss of precision is acceptable and the normal behaviour otherwise.

I'm not alone in needing this:

• see @zacharycarter and @genotrance for GLSL: https://irclogs.nim-lang.org/06-11-2018.html#16:38:40
• @slomp: https://irclogs.nim-lang.org/09-10-2018.html and https://irclogs.nim-lang.org/26-02-2019.html#15:27:04

@Araq I couldn't find the last discussion (April) where you said that it was planned for the future, so tracking here so that it's not forgotten.

[slomp]

Related: idea for push/pop pragmas to work with passC pragmas et al.:
https://irclogs.nim-lang.org/09-10-2018.html#21:19:24

[timotheecour]

one example use case among many: it would allow compiling different nim files with different modes, eg: nim c, nim cpp, nim objc and linking all them together. In particular it would simplify #12653 (comment) by compiling the objective C code inside nimextras_objc (as an emit statement inside a nim program) as nim objc so there wouldn't be a separate compilation step needed for that nimextras_objc.mm.

[slomp]

it would allow compiling different nim files with different modes, eg: nim c, nim cpp, nim objc and linking all them together.

Hi @timotheecour
Would you mind elaborating a bit more on that?
Suppose I have a set of .c files, a set of .cpp files, and a set of .mm files, and that I have some nim files (say, cfiles.nim, cppfiles.nim and mmfiles.nim) that call the appropriate compile/pass pragmas. How do you tell nim to generate objects out of each of them and then subsequently link those objects together into an executable?
( I know this Issue is closed, but I'm really curious :) )