compiling truncate :: Float -> Int with -O2 panics ghc

[conal]

bash-3.2$ cat Bug.hs
{-# OPTIONS -fglasgow-exts -O2 #-}
module Bug (tst) where
tst :: Float -> Bool
tst x = truncate x > (0::Int)

bash-3.2$ ghc Bug.hs
ghc.exe: panic! (the 'impossible' happened)
  (GHC version 6.8.1.20071102 for i386-unknown-mingw32):
	narrowTo
Please report this as a GHC bug:  http://www.haskell.org/ghc/reportabug

Same results reported on linux.

[int-e]

proposed patch

[int-e]

The code contains an expression of the form (comparison (conversion expr) int-literal). In cmmMachOpFold, the CMM optimizer tries to narrow the literal down to the type of expr, to see if it fits and the conversion can be eliminated. However this narrowing fails if expr has a floating point type. And that's good, because the optimization is not valid in that case.

The patch changes cmmMachOpFold to not even attempt narrowing to floating point types.

It fixes the example above.

Affected architectures are x86 and x86_64 as it seems.

[igloo]

Thanks for the report, and for the patch! Let's try to look at this for 6.8.2.

[simonmar]

mine

[simonmar]

applied.

Thu Nov 22 09:55:13 GMT 2007  Bertram Felgenhauer <int-e@gmx.de>
  * FIX #1916: don't try to convert float constants to int in CMM optimizer

[igloo]

Merged