GHC panic when compiling unsafeCoerce

The GHC compiler crashes on the following program.

module Test where

import Unsafe.Coerce
import Graphics.Rendering.OpenGL (color, Color3, GLdouble)

data D4 = D4 Double Double Double Double

crash :: D4 -> IO ()
crash c = color (invalidCast c)

invalidCast :: D4 -> Color3 GLdouble
invalidCast = unsafeCoerce

With the following output:

[1 of 1] Compiling Test             ( Test.hs, Test.o )
ghc: panic! (the 'impossible' happened)
  (GHC version 7.6.1 for x86_64-apple-darwin):
	cgLookupPanic (probably invalid Core; try -dcore-lint)
    ww_s28U{v} [lid]
    static binds for:
    local binds for:

I've made several attempts to make the input program smaller, but all modifications will result in successful compilation.

The cast is invalid but should produce an runtime error, not a compile time error.

Changes I made that somehow prevent this bug from appearing:

• Changing the crash functions into point free style color . invalidCast.
• Decreasing the amount of doubles in the D4 datatype.
• Inlining the exact definitions from the OpenGL package.
• Using the print function instead of the color function.

(btw, I know I shouldn't use unsafeCoerce)

Trac metadata

Trac field	Value
Version	7.6.1
Type	Bug
TypeOfFailure	OtherFailure
Priority	normal
Resolution	Unresolved
Component	Compiler
Test case
Differential revisions
BlockedBy
Related
Blocking
CC
Operating system
Architecture

changed weight to 5

added Tbug Trac import labels

Can you give a test case that doesn't depend on OpenGL? Eg make a stub module that has the relevant bits of OpenGL in it?

I can't reproduce on Linux, GHC 7.6.3, Open#2 (closed).6 or Open#2 (closed).8. What version of OpenGL do you use? Also what happens when you compile with -dcore-lint?

I can reproduce a perhaps related issue with GHC 7.7 and OpenGL-2.9.0.0,

$ ghc -O Test.hs
[1 of 1] Compiling Test             ( Test.hs, Test.o )
ghc: panic! (the 'impossible' happened)
  (GHC version 7.7.20130919 for x86_64-unknown-linux):
	StgCmmEnv: variable not found
    ww_s3rs{v} [lid]
    local binds for:
    main:Test.crash{v r1pI} [gid]
    main:Test.invalidCast{v r1pJ} [gid]
    main:Test.crash1{v r3ss} [gid]
    main:Test.$wa{v r3st} [gid]
    ww{v s3tT} [lid]
    ww1{v s3tU} [lid]
    ww2{v s3tV} [lid]
    sat_s3tX{v} [lid]
    sat_s3tY{v} [lid]
    sat_s3tZ{v} [lid]

Great, I can now reproduce. I looked into internals of OpenGL and OpenGLRaw and here's a version which does not have the dependency. Compile with -O to get the panic, both 7.6 and HEAD. (edit: made it smaller)

module Test where

import Unsafe.Coerce
import Foreign.C.Types
import System.IO.Unsafe

data D4 = D4 CInt CInt CInt
data Color3 = Color3 CInt CInt

crash :: D4 -> IO ()
crash = color . unsafeCoerce

color :: Color3 -> IO ()
color (Color3 r g) = f (unsafePerformIO undefined) r g

foreign import ccall f :: CInt -> CInt -> CInt -> IO ()

Great test case.

You'll see that it unsafely converts a data value with three fields to one with four. I grant that this shouldn't make GHC fail, but it's very likely to lead to a runtime crash. If this is what OpenGL is really doing, I am highly dubious!

Simon

OpenGL doesn't have the invalid cast; the original reporter wrote it him/herself. So it's a problem that GHC panics on a rather artificial program.

mentioned in commit a1b69324

mentioned in commit bd612773

closed

Thanks. I've robustified this.

Simon

Trac metadata

Trac field	Value
Resolution	Unresolved → ResolvedFixed
Test case	→ stranal/should_compile/T8037

added Pnormal label

marked this issue as related to #22039 (closed)