We should move to a D-based build system to build dmd/druntime/phobos/d-programming-language/tools.
Reasons, in case it's not obvious:
1) DRY: makefile is full of repetitions
2) cross platform: different makefiles for different architectures are needed (even win32 vs win64!)
3) safety: makefile happily ignores that $VAR variables are not set, and can create havoc; likewise tools/update.sh is a bash script and may not stop on 1st error.
Even a bad D-based build system will be better than a makefile, but it's not hard to write a reasonable one at least for the particular task of building dmd/druntime/phobos/d-programming-language/tools. It'll lower the entry point for people to contribute to dmd/phobos by making rebuilding trivial.
Here's how it'd work:
1) a mixed D-based/makefile build that'll work 'from scratch' via bootstrapping:
1a) dmd: first uses a makefile to build dmd+druntime
1b) rdmd built from a D file using dmd
1c) phobos, d-programming-language, tools built from a D file using rdmd
2) a pure D-based build that'll work using an existing dmd binary (eg distributed binary)
that's the one most will use.