diff options
| author | Matthew Pickering <matthewtpickering@gmail.com> | 2019-08-17 18:35:09 +0100 |
|---|---|---|
| committer | Matthew Pickering <matthewtpickering@gmail.com> | 2019-08-17 18:35:09 +0100 |
| commit | 5cff43d301b0773e2deeedb0a397da5e47cb89e1 (patch) | |
| tree | 074d42d6a9cd8e82703b298d7f55b1078416dcd6 /ghc/Main.hs | |
| parent | 47e162374051ed3e874ed7916cc811df288cbd95 (diff) | |
| download | haskell-wip/faststring-allocs.tar.gz | |
FastString: Reimplement in terms of pointer equalitywip/faststring-allocs
GHC allocates a lot of FastStrings when compiling so reducing the
overhead of a FastString will result in a large decrease in allocations.
Therefore we remove all the unecessary fields from the FastString.
1. Unique: Computed from the pointer
2. Equality: Computed directly by comparing the pointer
3. Length: Only used in 3 places, so the overhead is not worth it.
4. HasZEncoding: Caching the zencoding for a string should be localised
to code generation rather than any consumer having to pay a large
allocation penalty.
Comparing pointer equality is safe as the string table guarantees that
each string is only allocated once. As the memory is pinned, we don't
have to worry about GC moving around the locations of the strings.
Thanks to Jamie (Hanue) Willis for suggesting the idea of using pointer
equality.
Diffstat (limited to 'ghc/Main.hs')
| -rw-r--r-- | ghc/Main.hs | 4 |
1 files changed, 0 insertions, 4 deletions
diff --git a/ghc/Main.hs b/ghc/Main.hs index 614b45f277..3b32c2de99 100644 --- a/ghc/Main.hs +++ b/ghc/Main.hs @@ -819,7 +819,6 @@ dumpFastStringStats dflags = do bucketsPerSegment = map length segments entriesPerBucket = map length buckets entries = sum entriesPerBucket - hasZ = sum $ map (length . filter hasZEncoding) buckets msg = text "FastString stats:" $$ nest 4 (vcat [ text "segments: " <+> int (length segments) , text "buckets: " <+> int (sum bucketsPerSegment) @@ -827,7 +826,6 @@ dumpFastStringStats dflags = do , text "largest segment: " <+> int (maximum bucketsPerSegment) , text "smallest segment: " <+> int (minimum bucketsPerSegment) , text "longest bucket: " <+> int (maximum entriesPerBucket) - , text "has z-encoding: " <+> (hasZ `pcntOf` entries) ]) -- we usually get more "has z-encoding" than "z-encoded", because -- when we z-encode a string it might hash to the exact same string, @@ -835,8 +833,6 @@ dumpFastStringStats dflags = do -- Z-encoding is different from the original string are counted in -- the "z-encoded" total. putMsg dflags msg - where - x `pcntOf` y = int ((x * 100) `quot` y) Outputable.<> char '%' showPackages, dumpPackages, dumpPackagesSimple :: DynFlags -> IO () showPackages dflags = putStrLn (showSDoc dflags (pprPackages dflags)) |