[LLVMdev] canonicalizing add in constant expressions

[Jay Foad]

> On Fri, 2 May 2008, Jay Foad wrote:
> > I wrote some code that fell over when it encountered this
> ConstantExpr:
> >  i32 add (i32 24, i32 ptrtoint ([20 x %"struct.stlpmtx_std::string"]*
> > @_ZN12BulletMLNode11name2stringE to i32)))
> >
> > because it expected the LHS of the add to be another ConstantExpr.
> (This is
> > the first time that bit of code has fallen over, despite having
> compiled
> > thousands of test cases and many megabytes of C and C++ source, so
> constant
> > expressions like this must at least be pretty rare.)
> >
> > Is there a rule, or a convention, that the plain old constant should
> go on
> > the RHS of an add like this? If so, I'll try to track down what
> created this
> > constant expression. Otherwise I won't bother!
> 
> The compiler shouldn't crash, VMCore or ConstantFolding in this case
> (which I assume is what is crashing) should do the canonicalization.

No, it was an LLVM pass that I wrote myself that was crashing. I can't see
any code in lib/VMCore/ that tries to canonicalize the order of operands of
an add.

Thanks,
Jay.