llvm dev - Sep 2014 - [LLVMdev] Machine code sinking pass

Hi,

I ran into MachineVerifier "Virtual register killed in block, but needed
live out."

It was MachineSinking:PerformTrivialForwardCoalescing() that coalesced a COPY
inside a single-block loop, but left the
kill-flag and then MachineVerifier complains that a register in vregsRequired is
killed in MBB.

In the example, %vreg520 is replaced by %vreg368 by
PerformTrivialForwardCoalescing(), without clearing the kill flag:

BB#13: derived from LLVM BB %CF250
    Predecessors according to CFG: BB#12 BB#13 BB#14
       ...
        %vreg520<def> = COPY %vreg368
        %vreg568<def,tied1> = cmp %vreg341<tied0>,
%vreg520<kill>
        brr_cond <BB#13>
        brr_uncond <BB#14>
    Successors according to CFG: BB#13, BB#14

Into

BB#13: derived from LLVM BB %CF250
    Predecessors according to CFG: BB#12 BB#13 BB#14
        ...
        %vreg568<def,tied1> = cmp %vreg341<tied0>,
%vreg368<kill>
        brr_cond <BB#13>
        brr_uncond <BB#14>

=>

*** Bad machine code: Virtual register killed in block, but needed live out. ***
- function:    autogen_SD15028
- basic block: BB#13 CF250 (0x1c75890)
Virtual register %vreg368 is used after the block.

There is only one use of %vreg368 in the function.

One thing that strikes me is that one might want to clear the kill flag for a
use inside a loop of a register defined prior to
the loop?

Best regards,

Jonas Paulsson

-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20140902/1d4198e6/attachment.html>

Hi Jonas,

This looks like a bug in Machine code sinking pass.

Please file a PR on llvm.org/bugs with a reduced test case to keep track of the
issue.

Thanks,
-Quentin

On Sep 2, 2014, at 5:57 AM, Jonas Paulsson <jonas.paulsson at
ericsson.com> wrote:
> Hi,
>  
> I ran into MachineVerifier ”Virtual register killed in block, but needed
live out.”
>  
> It was MachineSinking:PerformTrivialForwardCoalescing() that coalesced a
COPY inside a single-block loop, but left the
> kill-flag and then MachineVerifier complains that a register in
vregsRequired is killed in MBB.
>  
> In the example, %vreg520 is replaced by %vreg368 by
PerformTrivialForwardCoalescing(), without clearing the kill flag:
>  
> BB#13: derived from LLVM BB %CF250
>     Predecessors according to CFG: BB#12 BB#13 BB#14
>        …
>         %vreg520<def> = COPY %vreg368
>         %vreg568<def,tied1> = cmp %vreg341<tied0>,
%vreg520<kill>
>         brr_cond <BB#13>
>         brr_uncond <BB#14>
>     Successors according to CFG: BB#13, BB#14
>  
> Into
>  
> BB#13: derived from LLVM BB %CF250
>     Predecessors according to CFG: BB#12 BB#13 BB#14
>         …
>         %vreg568<def,tied1> = cmp %vreg341<tied0>,
%vreg368<kill>
>         brr_cond <BB#13>
>         brr_uncond <BB#14>
>  
> =>
>  
> *** Bad machine code: Virtual register killed in block, but needed live
out. ***
> - function:    autogen_SD15028
> - basic block: BB#13 CF250 (0x1c75890)
> Virtual register %vreg368 is used after the block.
>  
> There is only one use of %vreg368 in the function.
>  
> One thing that strikes me is that one might want to clear the kill flag for
a use inside a loop of a register defined prior to
> the loop?
>  
> Best regards,
>  
> Jonas Paulsson
>  
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20140902/19625725/attachment.html>