Eli,
Thanks. I don't think it's a bug in the linker, really… it's
required to treat (stripped) opaque types as different, since it really has no
way to know if the caller and the callee really are referring to the same thing…
though it then seems to turn right around and do exactly that by inserting the
bitcast. ;-) (That does come from the linker.)
That the linker 'works differently' if the modules are not stripped
is arguably a 'bug', since even if the opaque types happen to have the
same name, they still may not be referring to the same type.
That said, I think what the linker is doing with unstripped modules is
reasonable… Using an opaque structure's name to "match" types is
really the only option. And they are, in fact, in this case, referring to the
same type.
I would go a bit further, though, and say that IF the linker is (as it seems
to be) using opaque structure names to match types, then the "strip"
pass ought to leave opaque structure names intact. In this particular case, the
names aren't useless… Or perhaps ALL (stripped) opaque types should merely
be smashed together to avoid the troublesome bitcast.
I'll look in to instcombine, but if you've got any further words of
advice on how to use it, they would be greatly appreciated.
Thanks,
Richard
On Nov 9, 2012, at 3:33 PM, Eli Friedman <eli.friedman at gmail.com>
wrote:
> On Fri, Nov 9, 2012 at 3:25 PM, Relph, Richard <Richard.Relph at
amd.com> wrote:
>> I've got a call instruction:
>> call void bitcast (void (%4 addrspace(1)*, <2 x i32>, <4 x
float>)* @_Z12write_imagefPU3AS110_image2d_tDv2_iDv4_f to void (%9
addrspace(1)*, <2 x i32>, <4 x float>)*)(%9 addrspace(1)* %dstimg,
<2 x i32> %28, <4 x float> %26) nounwind
>>
>> %4 and %9 are both (stripped) opaque structs.
>>
>> InlineFunction() does not inline this because
call->getCalledFunction() returns 0. call->getCalledValue() returns a cast
ConstantExpr. The called Function * is available at (leaving out necessary casts
and checks) call->getCalledValue()->getOperand(0). I've tried using
call->setCalledFunction(), but that isn't successful either, I'm
guessing because the types don't match.
>>
>> If I don't strip the module, the linker produces an ordinary call
to the target function, with no bitcast because both formal arg and passed arg
have the same opaque struct name, and InlineFunction() works great.
>
> This sounds like a bug in the IR linker (which is normally pretty good
> at unifying struct types).
>
>> Is this a "bug" in InlineFunction? Or is there something I am
expected to do to the call before calling InlineFunction()? (If yes, what? ;-)
>
> instcombine knows how to fix this.
>
> -Eli
>
On Fri, Nov 9, 2012 at 3:58 PM, Relph, Richard <Richard.Relph at amd.com> wrote:> Eli, > Thanks. I don't think it's a bug in the linker, really… it's required to treat (stripped) opaque types as different, since it really has no way to know if the caller and the callee really are referring to the same thing… though it then seems to turn right around and do exactly that by inserting the bitcast. ;-) (That does come from the linker.) > That the linker 'works differently' if the modules are not stripped is arguably a 'bug', since even if the opaque types happen to have the same name, they still may not be referring to the same type. > That said, I think what the linker is doing with unstripped modules is reasonable… Using an opaque structure's name to "match" types is really the only option. And they are, in fact, in this case, referring to the same type. > I would go a bit further, though, and say that IF the linker is (as it seems to be) using opaque structure names to match types, then the "strip" pass ought to leave opaque structure names intact. In this particular case, the names aren't useless… Or perhaps ALL (stripped) opaque types should merely be smashed together to avoid the troublesome bitcast.The name isn't a fundamental property of an LLVM struct type; we preserve it for readability, but not at the expense of correctness. I think the linker uses the names as a hint to merge struct types, but it also should be taking the structure of globals and instructions which use the struct type into account.> I'll look in to instcombine, but if you've got any further words of advice on how to use it, they would be greatly appreciated.It's a pass; you run it. -Eli
> I'll look in to instcombine, but if you've got any further words of advice on how to use it, they would be greatly appreciated.It's a pass; you run it. No joy running the instcombine pass between the link and my inlining code... still a call to a bitcast.
Spoke too soon... running the inst combine pass does seem to resolve the linker/inlining issue. But the resulting code fails to run correctly. I was expecting the same result that I get when running with unstripped modules, that is passing test runs. When it still failed, I guessed the failure mode to be the same, but it isn't. Not sure WHAT to make of this... ________________________________________ From: Eli Friedman [eli.friedman at gmail.com] Sent: Friday, November 09, 2012 4:08 PM To: Relph, Richard Cc: LLVMdev at cs.uiuc.edu Subject: Re: [LLVMdev] Inlining bitcast functions... On Fri, Nov 9, 2012 at 3:58 PM, Relph, Richard <Richard.Relph at amd.com> wrote:> Eli, > Thanks. I don't think it's a bug in the linker, really… it's required to treat (stripped) opaque types as different, since it really has no way to know if the caller and the callee really are referring to the same thing… though it then seems to turn right around and do exactly that by inserting the bitcast. ;-) (That does come from the linker.) > That the linker 'works differently' if the modules are not stripped is arguably a 'bug', since even if the opaque types happen to have the same name, they still may not be referring to the same type. > That said, I think what the linker is doing with unstripped modules is reasonable… Using an opaque structure's name to "match" types is really the only option. And they are, in fact, in this case, referring to the same type. > I would go a bit further, though, and say that IF the linker is (as it seems to be) using opaque structure names to match types, then the "strip" pass ought to leave opaque structure names intact. In this particular case, the names aren't useless… Or perhaps ALL (stripped) opaque types should merely be smashed together to avoid the troublesome bitcast.The name isn't a fundamental property of an LLVM struct type; we preserve it for readability, but not at the expense of correctness. I think the linker uses the names as a hint to merge struct types, but it also should be taking the structure of globals and instructions which use the struct type into account.> I'll look in to instcombine, but if you've got any further words of advice on how to use it, they would be greatly appreciated.It's a pass; you run it. -Eli