Hacker News new | ask | show | jobs
by tpoacher 765 days ago
Very nice.

Whenever I see comments about people trying to reinvent make, my conclusion is that make "as a tool" is inevitably far superior, but people are mostly familiar with a small subset of its functionality, which makes it appear clunky. That, and the syntax is very flexible, which means it can easily be abused to write ugly/convoluted makefiles.

If I had one thing I'd want to improve about make, it wouldn't be the tool itself, but its documentation, and a set of reasonable defaults/templates, like you have here.

The documentation is amazing in one sense, in that it's very comprehensive, but very clunky and hard to master in another, in that it's organised in a very high-level manner that doesn't allow it to be used as a quick reference / point of truth, if you don't know what you're looking for. I found myself making anki notes for it, as it's not organised in the kind of way that would simply allow me to refer to the docs for your usecase and get things done. You really either know some functionality exists and can look it up or you don't, and unless you're someone who uses make in expert mode on a daily basis, it's hard to know all the specialised components and be able to combine them all together seamlessly. Hence my anki notes, hahah.

But make really is an amazing tool. I wish people focused on improving make workflows instead of reinventing the wheel all the time with other tools (which are then forced on the user).

2 comments

> Whenever I see comments about people trying to reinvent make, my conclusion is that make "as a tool" is inevitably far superior, but people are mostly familiar with a small subset of its functionality, which makes it appear clunky.

If you're looking for a task runner, not a build tool, then make is not "far superior" in any sense and there are much better alternatives, the most prominent probably being "just". Even something as basic as accepting parameters to a task (make fetch <arg1> <arg2>) is awkward because make doesn't understand this syntax and interprets the arguments as targets. The only way to make it work is either through named arguments or empty target hacks and in both cases you need to write the validation logic yourself. In just it's simply:

  fetch PACKAGE VER:
    @echo fetching {{PACKAGE}} at version {{VER}}

  $ just fetch foo 
  error: Recipe `fetch` got 1 argument but takes 2
  usage:
    just fetch PACKAGE VER

  $ just fetch foo 1.2.3
  fetching foo at version 1.2.3

https://just.systems/man/en/chapter_1.html
Just is fine, don't get me wrong. But Make targets can take arguements too. They just have to be named, or (if you stretch the definition a bit) wildcard targets.

AFAIK, Just doesn't have any way to mark "this task is complete, so don't re-run it". It also doesn't have file-based dependendicies, which are pretty common for any kind of programming.

Why would I want to lose all of Make's flexibility and power in exchange for slightly prettier UI syntax?

> AFAIK, Just doesn't have any way to mark "this task is complete, so don't re-run it". It also doesn't have file-based dependendicies, which are pretty common for any kind of programming.

That's fair but also not what I'm looking for in a task runner, that's why I explicitly said "not a build tool", which to me this falls under (although I realize everybody has a different definition of these terms and where they draw the line if at all).

Because it makes the common things easier.

We make these tradeoffs all the time in software engineering. Raw machine code is the most flexible and powerful programming language, and yet most of us here aren't using it today. Make isn't machine code and Just isn't Python, but there's an analogy: you give up something by using the latter in exchange for making it easier to do everyday things.

Just doesn't replace Make. For the thousand everyday tasks where Make's power isn't needed, it's likely a better choice for most people.

I think this post is well intentioned but misses the point entirely.

    GOOS=darwin GOARCH=arm64 go build -ldflags="-X 'main.Version=v1.2.3'"
Executing this with variables for os/arch/version is not straightforward in a makefile. You might argue that it's a flaw with the `go build` command, but replace go with cargo/dotnet/maven and you have the same problem.

> AFAIK, Just doesn't have any way to mark "this task is complete, so don't re-run it"

What does a makefile look like for "install this list of dependencies with pip/homebrew/apt, and don't run it twice?"

> It also doesn't have file-based dependendicies, which are pretty common for any kind of programming

C and C++ compilers at this stage are the only places that I work with file based dependencies. Unless you explicitly declare all the headers as dependencies in a makefile, make won't rebuild a cpp file if you change a header. IME ninja (plus cmake to generate it) has entirely superseded make in this space. Tools and languages comes with it's own build tool (see above - go/cargo/dotnet/etc) or state tracking (docker) that break make's assumption of file based dependencies.

> Why would I want to lose all of Make's flexibility and power in exchange for slightly prettier UI syntax?

Because every make file I've ever used in recent memory has been 30% workarounds to avoid file based dependencies, and work around subtle footguns that make has, rather than actually doing what I want it to do - execute a build command.

> What does a makefile look like for "install this list of dependencies with pip/homebrew/apt, and don't run it twice?"

It looks like this:

    default: your_task
    
    VENV_DIR := venv
    
    $(VENV_DIR)/bin/python: requirements.txt
        python -m venv $(VENV_DIR)
        . $(VENV_DIR)/bin/activate && pip install -r requirements.txt
    
    your_task: $(VENV_DIR)/bin/python
        do_your_thing
    
    clean:
        rm -rf $(VENV_DIR)

> C and C++ compilers at this stage are the only places that I work with file based dependencies. Unless you explicitly declare all the headers as dependencies in a makefile, make won't rebuild a cpp file if you change a header. IME ninja (plus cmake to generate it) has entirely superseded make in this space. Tools and languages comes with it's own build tool (see above - go/cargo/dotnet/etc) or state tracking (docker) that break make's assumption of file based dependencies.

In our Python example above, you have a file-based dependency between requirements.txt and your virtual environment. And then you have a file-based dependency on your virtual-environment's Python for whatever task you're trying to run.

> Because every make file I've ever used in recent memory has been 30% workarounds to avoid file based dependencies, and work around subtle footguns that make has, rather than actually doing what I want it to do - execute a build command.

Why would you want to avoid expressing your dependencies? File dependencies exist, and are the simplest / most correct model for 75% of all build automation. When somebody has to sit down and read your Makefile, isn't it nice to have a way to express "these are the files that actually matter for task X"?

I think this post is well intentioned but misses the point entirely.

    GOOS=darwin GOARCH=arm64 go build -ldflags="-X 'main.Version=v1.2.3'"
Executing this with variables for os/arch/version is not straightforward in a makefile. You might argue that it's a flaw with the `go build` command, but replace go with cargo/dotnet/maven and you have the same problem.

What makes that hard? I think the following works correctly:

    $ cat Makefile
    goos := FOO
    goarch := BAR
    mainversion := BAZ

    .PHONY: build
    build: ; @GOOS=$(goos) GOARCH=$(goarch) go build -ldflags="-X 'main.Version=v$(mainversion)'"*

    $ make -n
    GOOS=FOO GOARCH=BAR go build -ldflags="-X 'main.Version=vBAZ'"
> I think this post is well intentioned but misses the point entirely. > > GOOS=darwin GOARCH=arm64 go build -ldflags="-X 'main.Version=v1.2.3'" > > Executing this with variables for os/arch/version is not straightforward in a makefile. You might argue that it's a flaw with the `go build` command, but replace go with cargo/dotnet/maven and you have the same problem.

This is pretty easy in Make, especially since GOOS and GOARCH are environment variables. One way to do it could look like this:

    export GOOS ?= darwin
    export GOARCH ?= arm64 
    VERSION ?= 0.0.0+$(shell git describe --dirty --always)

    build:
        $(if $(VERSION),,$(error VERSION was not set))
        go build -ldflags="-X 'main.Version=v$(VERSION)'"
you could run it as:

    $ make build (uses darwin/arm64/0.0.0+<git ref>)
    $ make build GOOS=linux VERSION=1.0.0 (uses linux/arm64/1.0.0)
    $ GOOS=templeos GOARCH=sparc make build VERSION=1.1.0 (uses templeos/sparc/1.1.0)

    ---- 
    fetch PACKAGE VER:
        @echo fetching {{PACKAGE}} at version {{VER}}
    ---- 
    $ just fetch foo 1.2.3
Here is the Make equivalent, with equivalent error-checking

    ---- 
    fetch:
        $(if $(PACKAGE),,$(error PACKAGE was not set))
        $(if $(VERSION),,$(error VERSION was not set))
        @echo fetching $(PACKAGE) at $(VERSION)
    ----
    $ make fetch PACKAGE=foo VERSION=1.2.3
I agree that the `just` version is a little neater / tidier. But it's not like Make can't do the same thing. And Make is a much more powerful tool.
I like the idea of Just (and the many tools like it), I only wish it had come along a few decades earlier. Make has the advantage of being a ubiquitous de facto standard. It is probably already installed anywhere you want to deploy your repo, or can be very easily installed with one command.

As a personal preference, I actually do tend to avoid Make as a task runner because a quick shell script is almost as easy to write and far more flexible. I am pretty familiar with Make but have still painted myself into a corner with it more times than I'd like to admit.

You are not alone in this. I have found myself stubbornly writing a complex Makefile when I should have just written a simple shell script and be done with it.
Any tool that treats spaces and tabs differently is not sane.
Agreed, I don't like that about Make at all.

But also it's never something I have to worry about. Pretty much every syntax-aware text editor knows how to handle Makefiles, and knows that recipe steps start with a tab (unless you change Make's default). So for me, that falls in the "ugly, but not really a usability issue" bucket. And Make's power/expressiveness is worth accepting a few "I'd have done it differently" syntax things.

It always seems to bite when I've got nothing but an SSH link and a default vim install to hand.
lol that's fair enough. This is a tangent and unrelated to Make - but I also find myself in that situation pretty frequently, since I do a lot of embedded systems and some CI/CD work.

If you find yourself in that situation frequently, I'd recommend one of:

    1. Use sshfs to mount your remote target locally on your dev machine, or:
    2. Use vscode's remote-ssh mode (if the agent can run on your target).
Vscode's remote-ssh mode is really nice. It's good enough that it single-handedly made me switch from my previous editor, and I wish that I knew more tools like it.