Problem with CodeExtractor.cpp

[llvmbot]

Bugzilla Link	1004
Resolution	INVALID
Resolved on	Feb 22, 2010 12:45
Version	trunk
OS	Linux
Reporter	LLVM Bugzilla Contributor

Extended Description

I have this code:

loopexit.6:
... // something
no_exit.7:
... // something
loopexit.7: ; preds = %no_exit.7, %loopexit.6
%n.3.1 = phi int [ 0, %loopexit.6 ], [ %n.2, %no_exit.7 ]
...

I want extract code region, which contains loopexit.6 and no_exit.7.
CodeExtractor creates new function and puts those basic blocks to this function.
CodeExtractor creates codeReplacer block containing call-instr to new function.
So CodeExtractor modifies phi-node instruction in loopexit.7 :
%n.3.1 = phi int [ 0, %codeRepl1 ], [ %n.2.reload, %codeRepl1 ]
And verifier say 'Error' :

PHINode should have one entry for each predecessor of its parent basic block!
%n.3.1 = phi int [ 0, %codeRepl1 ], [ %n.2.reload, %codeRepl1 ]
; [#uses=1]
Broken module found, compilation aborted!

[lattner]

Can you provide some information on how to reproduce this? The code extractor requires critical edges to
be broken, are they?

If so, this case shouldn't occur.

-Chris

[llvmbot]

bug reproduced information
bug reproduced information

[llvmbot]

bug reproduced information
There is in the attachment.
*example1.bc (example program)
*one.cpp/h (bug demonstrated code (using only on example1.bc))
*bad_modification.txt (resulted bc in text format)

Compile one.cpp and run on example1.bc. You will see error message:
<--
PHINode should have one entry for each predecessor of its parent basic block!
%res.0.1 = phi int [ 0, %codeRepl ], [ %tmp.30.reload, %codeRepl ]
; [#uses=2]
Broken module found, compilation aborted! -->

I think that CodeExtractor doesn't correctly modify phi instruction.

[nlewycky]

Right, so you're taking code which hasn't been run through break-crit-edges,
then directly calling ExtractCodeRegion like so:

Function* func = ExtractCodeRegion((DominatorSet)0, region, false);

The trouble is that there's no PassManager to recognize the need to run BCE
first and make the code valid. The bad phi node is caused by a critical edge
between label %loopexit.0 and label %loopexit.1.

Because your code is special-case anyways, I don't think you should bother with
a PassManager. Instead just use BCE directly -- something like this:

FunctionPass *BCE = createBreakCriticalEdgesPass();
BCE->runOnFunction(*mainfunc);

before you run the code extractor ought to fix it up for you.

[lattner]

The problem is that the code extractor expects critical edges to be broken and they are not.

-Chris