llvm dev - Jan 2012 - [LLVMdev] ValueMapper question: no type mapping for GlobalValue?

=?ISO-8859-1?Q?Rafael_=C1vila_de_Esp=EDndola?= wrote:
> On 11/01/12 07:50 AM, Michael Muller wrote:
> > 
> > Hi all, I was looking at the ValueMapper code this morning and I
notice that
> > it doesn't do type mapping for GlobalValues.  Is this correct?
> > 
> > I ask because I am seeing a case where I'm failing type assertions
from
> > ModuleLinker::linkGlobalInits() when an array initializer references
an
> > external global.  It looks like the external global is being mapped
directly
> > and this causes an element type check to fail in ConstantArray::get().
> > Curiously, this only happens when that external global is also
referenced from
> > within a function in the same module.
> > 
> 
> Do you have a testcase?
> 
> I tried to reproduce this with
> 
> -----------
> a = global i64 42, align 8
> -----------
> 
> and
> 
> ----------
> @a = external global i32
> @b = constant [1 x i32*] [i32* @a], align 8
> 
> define i32 @f() nounwind uwtable readonly optsize {
> entry:
>   %0 = load i32* @a, align 4
>   ret i32 %0
> }
> ----------
> 
> but llvm-link correctly produced
> 
> 
> @b = constant [1 x i32*] [i32* bitcast (i64* @a to i32*)], align 8
> @a = global i64 42, align 8
> 
> define i32 @f() nounwind uwtable readonly optsize {
> entry:
>   %0 = load i32* bitcast (i64* @a to i32*), align 4
>   ret i32 %0
> }
I can reproduce this consistently, but only under an extremely large pile of
code :-)  I haven't tried to strip it down to minimal a test case yet, but I
will.  One salient difference with the code you've provided is that in my
code, @a is a struct type.  However, changing the example to use a minimal
structure doesn't trigger the assertion failure.

But I'm still curious: why doesn't the value mapper do type mapping on
global
variables?
> 
> 
> Cheers,
> Rafael
> 
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
> 

============================================================================michaelMuller
= mmuller at enduden.com | http://www.mindhog.net/~mmuller
-----------------------------------------------------------------------------
The natural progress of things is for liberty to yield and government to
gain control. - Thomas Jefferson
=============================================================================

> I can reproduce this consistently, but only under an extremely large pile
of
> code :-)  I haven't tried to strip it down to minimal a test case yet,
but I
> will.  One salient difference with the code you've provided is that in
my
> code, @a is a struct type.  However, changing the example to use a minimal
> structure doesn't trigger the assertion failure.
Is the example something you can share even in unreduced for?
> But I'm still curious: why doesn't the value mapper do type mapping
on global
> variables?
I am not sure. I would probably have to look at a failing example to
know what is going on.

Cheers,
Rafael

=?ISO-8859-1?Q?Rafael_=C1vila_de_Esp=EDndola?= wrote:
> > I can reproduce this consistently, but only under an extremely large
pile of
> > code :-)  I haven't tried to strip it down to minimal a test case
yet, but I
> > will.  One salient difference with the code you've provided is
that in my
> > code, @a is a struct type.  However, changing the example to use a
minimal
> > structure doesn't trigger the assertion failure.
> 
> Is the example something you can share even in unreduced for?
As it turns out, the root cause was that I had a GlobalVariable that wasn't
being copied to an extern in the new module under some circumstances, but was
being used directly.  I would have expected the verifier pass to catch this,
but it looks like we're probably not using the verifier except when running
non-optimized + debug, which I never tried.

I'll recreate this situation just to verify that the verifier pass
identifies
the problem, if it doesn't I'll proceed from there but I'm currently
assuming
the problem is entirely ours :-)

If you think the behavior of the ValueMapper is incorrect in this situation,
I'd prefer to furnish you with a minimal test case - the unreduced form
involves checking out and building a specific revision of the Crack compiler
and running one of the tests in "native" (AOT compile) mode.

Thank you for looking into this :-)
> 
> > But I'm still curious: why doesn't the value mapper do type
mapping on global
> > variables?
> 
> I am not sure. I would probably have to look at a failing example to
> know what is going on.
> 
> Cheers,
> Rafael
> 

============================================================================michaelMuller
= mmuller at enduden.com | http://www.mindhog.net/~mmuller
-----------------------------------------------------------------------------
We are the music-makers, and we are the dreamers of dreams
 - Arthur O'Shaughnessy
=============================================================================