A proposal for a modern build-system

From: Hassan Sajjad (hassan.sajjad069_at_[hidden])
Date: 2025-04-16 11:34:35

Hi.

I compiled 25 Boost libraries with C++20 header-units, including examples
and Tests of a few.
This resulted in a 1.44x speed-up if we exclude the scanning. Scanning was
as slow as the header-unit build.

I have a solution for that. I am sending this paper (Making some final
edits)
https://htmlpreview.github.io/?https://github.com/HassanSajjad-302/iso-papers/blob/main/generated/my-paper.html
,
which allows for a build-system to use a scanning-less approach to build
header-units.This method requires only minimal modifications to the
compiler and enables memory-mapped file sharing, which significantly
reduces filesystem overhead and further boosts speed.

My proposal for a scanning-less approach will allow HMake to have another
amazing feature, monolithic Big header-units. BigHu will allow you to map
all the headers from a directory to a single header-unit. This header-unit
includes all these header-files and consumes them as header-files. Now,
whenever the compiler requests a header-unit from this directory, the
build-system instead sends this consolidated BigHu. This will not only
improve the clean build speed, because of fewer process launches. But also
massively boosts the rebuild speed as few header-units need to be read
instead of lots of files. No source-code editing will be needed to use this
feature. And you will have the option to either build a library as BigHu
(if you don't mean to edit it) or as small header-units (if you are working
on it).

HMake will soon support compile_commands.json, which I guess can work fine
with header-units. Thus you can benefit from both, faster compilation and
working intelliscence.

The API shown in hmake.cpp is a work in progress. HMake build-system
benefits are numerous, even if we leave out the header-units support. This
makes it a perfect replacement for B2. I believe we can complete the switch
in the next 2 months. And if some compiler supports my paper, we get >3x
faster compilation as well.

Best,
Hassan Sajjad