[llvm-commits] CVS: llvm/lib/Target/X86/X86Subtarget.cpp X86Subtarget.h X86TargetMachine.cpp

[Reid Spencer]

On Wed, 2006-12-20 at 00:25 -0800, Evan Cheng wrote:
> Ok. I see PPC does exactly the same thing. But this seems like a  
> temporary solution. Could we add some kind of JIT relocation model  
> which are target specific? Or any method that would allow llc to  
> codegen (by specifying some command line option?) in the exact same  
> way as the JIT would?

Having debugged a number of JIT failures now, I'll throw in my $0.02
worth that we *must* retain this functionality. Sometimes the
only/best/easiest way to find a JIT bug is by generating code with llc
and looking at it. There needs to be a way to ensure/force JIT and LLC
to produce identical code (as far as possible).

Reid.

> 
> Evan
> On Dec 19, 2006, at 9:59 PM, Chris Lattner wrote:
> 
> > On Dec 19, 2006, at 9:46 PM, Evan Cheng wrote:
> >> Let me clarify. It's important for llc and lli to behave exactly the
> >> same up to the final code emission pass. With your patch, the two
> >> would behave differently and that means we can no longer use llc to
> >> reproduce jit codegen bug.
> >
> > PPC64 has a similar problem.  It wants to do codegen in PIC mode,  
> > but it doesn't want the JIT to make dyld stubs.  As such, it does  
> > 'JIT PIC' mode, which is like PIC but doesn't do the extra load to  
> > go through dyld stubs.
> >
> > An alternative approach would be to have the JIT synthesize dyld  
> > stubs, but that takes work to implement and yields slower programs.
> >
> > -Chris
> >
> >
> 
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits