Dear GHC developers,
I think I have found an optimisation problem in GHC
pre-4.07-20000613. I ran across that problem when compiling nhc98
(version 1.0pre19) without and then with -O.
The follwoing program is a cut down version of a code excerpt from
nhc98 (taken from nhc98-1.0pre19/src/compiler98/EmitState.hs).
I tried this on
- i386/Linux (Debian 2.2) with GHC 4.06 and pre-4.07-20000613, gcc 2.95.2
- sparc/Solaris with GHC 4.06, gcc 2.7.2.3
The results were the same in all three configurations.
module Main where
data Label = Define String | Use String deriving Show
testfun :: [Label] -> String
testfun labs =
let isUse (Use _) = True
isUse _ = False
isDefine (Define _) = True
isDefine _ = False
defines = filter isDefine labs
uses = filter isUse labs
externs = filter (\use-> notElemBy useAfterDef use defines) uses
notElemBy :: (a->b->Bool) -> a -> [b] -> Bool
notElemBy ok x = all (not . ok x)
useAfterDef (Use sym) (Define sym') = (sym==sym')
useAfterDef u d = error ("useAfterDef (" ++ show u ++ ") (" ++
show d ++ ")\n")
in
concat (map show externs)
lablist1 = [Define "a"]
lablist2 = [Use "a",
Define "b",
Define "a"]
lablist3 = [Use "b",
Define "c",
Define "a",
Use "c",
Define "b",
Use "a"]
main :: IO ()
main = print (testfun lablist1)
The function `useAfterDef' is always called (due to the filters) as
"useAfterDef (Use _) (Define _)", so the `error ...' clause should
never be reached, right?
When I compile without -O, the program runs correctly, but with -O (or
-O2), the `... = error ...' clause is reached:
$ ghc -o main Main.hs && ./main
""
$ ghc -O -o main Main.hs && ./main
Fail: useAfterDef (Define "a") (Define "a")
With `lablist2' instead of `lablist1' the result is
Fail: useAfterDef (Define "b") (Define "b")
and with `lablist3' the program produces
Fail: useAfterDef (Define "c") (Define "c")
When I try to show the value of `uses' in the call to error, i.e.
... = error (show uses)
the problem disappers and the program runs correctly with -O.
Armin
