Optimise build of Cabal library in Stage0 & StageH

[snowleopard]

The current state of affairs is very inefficient:

• We build Stage0 Cabal for ghc-pkg -- results go to _build/stage0/libraries/Cabal.
• We also build parts of Cabal for Stage0 ghc-cabal -- results go to _build/stage0/utils/ghc-cabal. We may at some point get rid of ghc-cabal, but there is still a lot of work to be done Get rid of ghc-cabal and package-data.mk #18.
• We also build parts of Cabal for Hadrian -- results go to _build/hadrian (we can call this StageH). See Switch to in-tree Cabal library #228 on switching to using in-tree Cabal library.

So we end up building big chunks of Cabal three times in Stage0 + StageH! Presumably, we could only do it once when we use the same GHC both for Stage0 and for StageH (i.e. for compiling Hadrian), which is likely to be the most common case.

[snowleopard]

Why does Hadrian currently require Cabal? We need to parse .cabal files for two reasons:

1. To compute boot package constraints:

Cabal == 1.25.0.0
binary == 0.8.3.0
ghc-boot == 8.1
hoopl == 3.10.2.1
hpc == 0.6.0.3
template-haskell == 2.11.0.0
transformers == 0.5.2.0

Make build system does this using the following fragile code:

BOOT_PKG_CONSTRAINTS := \
    $(foreach d,$(PACKAGES_STAGE0),\
        $(foreach p,$(basename $(notdir $(wildcard libraries/$d/*.cabal))),\
            --constraint "$p == $(shell grep -i "^Version:" libraries/$d/$p.cabal | sed "s/[^0-9.]//g")"))

2. To determine the order in which we need to configure packages (by extracting their dependencies from .cabal files). Make build system sequentialised things with the following hack:

# -------------------------------------------
# Note [Dependencies between package-data.mk files].

# We cannot run ghc-cabal to configure a package until we have
# configured and registered all of its dependencies.  So the following
# hack forces all the configure steps to happen in exactly the following order:
#
#  $(PACKAGES_STAGE1) ghc(stage2) $(PACKAGES_STAGE2)
#
# Ideally we should use the correct dependencies here to allow more
# parallelism, but we don't know the dependencies until we've
# generated the package-data.mk files.
define fixed_pkg_dep
libraries/$1/$2/package-data.mk : $$(fixed_pkg_prev)
fixed_pkg_prev:=libraries/$1/$2/package-data.mk
endef

ifneq "$(BINDIST)" "YES"
fixed_pkg_prev=
$(foreach pkg,$(PACKAGES_STAGE1),$(eval $(call fixed_pkg_dep,$(pkg),dist-install)))

[snowleopard]

We no longer build ghc-cabal, so some duplication is gone. I think we can close this issue although there is still some scope for optimisation by sharing the build artefacts of Hadrian and Stage0 Cabal library.