haskell-cabal.eclass: unleash full parallelism of parallel ghc

I've explored scalability of 'ghc --make -j' a bit in
    https://ghc.haskell.org/trac/ghc/ticket/9221

Some takeaways:
- never specify -j<N> with N > CPU. garbage collector threads
  waste kernel time running sched_yield()
- GHC allocates A Lot: large nursery decreases GC interruptions.
  We fix it with '-A256M'
- for large nursery enabling work-stealing makes GC finish faster
  on each collection cycle.
  We fix it with -qb0

While at it move HCFLAGS setup after parallel defaults.
That allows user to override defaults with own HCFLAGS.

Signed-off-by: Sergei Trofimovich <slyfox@gentoo.org>

1 files changed, 8 insertions, 12 deletions

diff --git a/eclass/haskell-cabal.eclass b/eclass/haskell-cabal.eclass
index c1a13300abb0..94ef7f35ad75 100644
--- a/eclass/haskell-cabal.eclass
+++ b/eclass/haskell-cabal.eclass
@@ -372,24 +372,20 @@ cabal-configure() {
 		cabalconf+=($(cabal-constraint "ghc"))
 	fi
 
+	# parallel on all available cores
+	if ghc-supports-parallel-make; then
+		# It should have been just -j$(makeopts_jobs)
+		# but GHC does not yet have nice defaults:
+		#    https://ghc.haskell.org/trac/ghc/ticket/9221#comment:57
+		cabalconf+=(--ghc-options="-j$(makeopts_jobs) +RTS -A256M -qb0 -RTS")
+	fi
+
 	local option
 	for option in ${HCFLAGS}
 	do
 		cabalconf+=(--ghc-option="$option")
 	done
 
-	# parallel on all available cores
-	if ghc-supports-parallel-make; then
-		local max_jobs=$(makeopts_jobs)
-
-		# limit to very small value, as parallelism
-		# helps slightly, but makes things severely worse
-		# when amount of threads is Very Large.
-		[[ ${max_jobs} -gt 4 ]] && max_jobs=4
-
-		cabalconf+=(--ghc-option=-j"$max_jobs")
-	fi
-
 	# Building GHCi libs on ppc64 causes "TOC overflow".
 	if use ppc64; then
 		cabalconf+=(--disable-library-for-ghci)