allocateRegsAndSpill: Cannot read from uninitialized register

[erikd]

Building git HEAD on linux-powerpc64 and I get:

ghc-stage1: panic! (the 'impossible' happened)
  (GHC version 7.7.20121227 for powerpc-unknown-linux):
        allocateRegsAndSpill: Cannot read from uninitialized register
    %vI_na5b

[erikd]

From compiler/nativeGen/RegAlloc/Linear/Main.hs:

    -- NOTE: if the input to the NCG contains some
    -- unreachable blocks with junk code, this panic
    -- might be triggered.  Make sure you only feed
    -- sensible code into the NCG.  In CmmPipeline we
    -- call removeUnreachableBlocks at the end for this
    -- reason.

[simonpj]

Simon M, could this be anything to do with the new codegen path? I'll asssign it to you for now.

Simon

[simonmar]

So the background is that this panic used to be commented out, but I enabled it recently because it makes debugging much easier (with the panic commented out, the compiler will happily generate completely bogus code that will crash at runtime). However, it adds another invariant, as mentioned in the comment: all register must be written before they are read (not an unreasonable assumption, I'm sure you'll agree) I presume this invariant is being violated somewhere.

To debug it you'll need to use -ddump-cmm and find out where the bogus code comes from.

[simonmar]

More info needed to make progress on this one, the repro case only happens on PPC/Linux.

[PHO]

Isn't this related to #7442?

[simonmar]

I don't immediately see the connection to #7442?

[PHO]

Replying to simonmar:

I don't immediately see the connection to #7442?

Just a thought. Even though PPC/Linux doesn't have the problem described in the note [inconsistent-pic-reg], PIC.initialisePicBase_ppc still semms to be assuming only the first block is an entry point.

[erikd]

I added -ddump-cmm to the command line of the compile that was failing and saved the output to a file. Unfortunately there is no sign on the uninitialized register %vI_na5b in the dumped file.

[simonmar]

This could be the same as #7574, I described the fix that was needed on that ticket and @thoughtpolice said he would look into it.