[GHC] #2209: MagicHash extraction is wrong on x86_64 with -fasm -O2

[GHC]

#2209: MagicHash extraction is wrong on x86_64 with -fasm -O2
----------------------+-----------------------------------------------------
 Reporter:  quark     |          Owner:  simonmar      
     Type:  bug       |         Status:  new           
 Priority:  normal    |      Milestone:  6.8.3         
Component:  Compiler  |        Version:  6.8.2         
 Severity:  major     |     Resolution:                
 Keywords:            |     Difficulty:  Unknown       
 Testcase:            |   Architecture:  x86_64 (amd64)
       Os:  Linux     |  
----------------------+-----------------------------------------------------
Changes (by simonmar):

  * owner:  => simonmar

Comment:

 Replying to [comment:2 igloo]:

 > We do claim that this ("Casting an unboxed type to another unboxed type
 of the same size") works in the `unsafeCoerce#` docs, so we ought to fix
 it!

 I'm tempted to alter the docs here.  The problem is that the Cmm generated
 for this program has a type error, which you can see by compiling with
 `-dcmm-lint`.  This breaks some assumptions that the native code generator
 is making, with the result that it ends up being confused about floating
 point vs. scalar register assignments.

 You can achieve the desired effect using castSTUArray, incedentally (this
 is the way we do it in GHC).

 To make the direct coercion work, we'd have to make the `unsafeCoerce#`
 really compile into an operation at the Cmm level.  Which is possible, but
 we'd have to add new primitives for the operation (the current float->int
 conversions do truncation, not bit-for-bit conversion), and add some code
 to CoreToStg to generate the operations.

-- 
Ticket URL: <http://hackage.haskell.org/trac/ghc/ticket/2209#comment:3>
GHC <http://www.haskell.org/ghc/>
The Glasgow Haskell Compiler