Re: compilation flags

[Jay Carlson <nop@...>]

On 2018-12-11, at 11:58 AM, Luiz Henrique de Figueiredo <lhf@tecgraf.puc-rio.br> wrote:

> The Makefile in the Lua tarball contains targets for specific
> platforms. It also is committed to gcc. Therefore, we could use better
> compilation flags, at least for some platforms. In particular, adding
> -march=native seems desirable and could boost performance at no cost.
> What other compilation flags could be used, especially for linux and
> macosx?

Only if the person compiling *does not intend to distribute*. I can go hunting for options for non-distributors if you think that's useful.

For people who are going to distribute, I would be very, very reluctant to implicitly tie the results of compilation to the nature of the machine being built on. If my build machine is an i7-4770K:

>      'haswell'
>           Intel Haswell CPU with 64-bit extensions, MOVBE, MMX, SSE,
>           SSE2, SSE3, SSSE3, SSE4.1, SSE4.2, POPCNT, AVX, AVX2, AES,
>           PCLMUL, FSGSBASE, RDRND, FMA, BMI, BMI2 and F16C instruction
>           set support.

but some of the machines my users have Intel Atom Zs:

>      'bonnell'
>           Intel Bonnell CPU with 64-bit extensions, MOVBE, MMX, SSE,
>           SSE2, SSE3 and SSSE3 instruction set support.

...I could get illegal instruction traps, and tear my hair out...until I think to try doing my builds on those Atom machines.

You could use -mtune=native safely, but in principle you get the same problem: hidden dependency on build machine. This drives build engineers crazy.

-- 
Jay