llvm dev - Nov 2016 - getMinimalPhysRegClass : Couldn't find the register class

I've implemented a function pass, and some associated hacks to llvm and
clang, and am hitting an error in a DAG pass when my transformed IR is
processed. That transformed bytecode attempts to insert a call around a load of
a 'long double'. The error is:


3. Running pass 'Function Pass Manager' on module 'big.c'.


4. Running pass 'X86 DAG->DAG Instruction Selection' on function
'@ld'



The assert is after a failure to lookup a BestRC pointer in this function:




(gdb) b 141 if (BestRC == 0)


Breakpoint 3 at 0x7ffff6755436: file
/home/pjoot/llvm/lib/CodeGen/TargetRegisterInfo.cpp, line 141.


(gdb) c


Continuing.


Breakpoint 3, llvm::TargetRegisterInfo::getMinimalPhysRegClass (this=0x5cc4f8,
reg=44, VT=...) at /home/pjoot/llvm/lib/CodeGen/TargetRegisterInfo.cpp:141


141  assert(BestRC && "Couldn't find the register class");


(gdb) p reg


$4 = 44


(gdb) p c


$5 = (const llvm::TargetRegisterClass *) 0x0


(gdb) p VT


$6 = {SimpleTy = llvm::MVT::SimpleValueType::i32}


The bytecode that my pass is producing is:


; Function Attrs: nounwind uwtable


define x86_fp80 @ld(x86_fp80* nocapture readonly %d) local_unnamed_addr #0 {


entry:


%0 = load x86_fp80, x86_fp80* %d, align 16, !tbaa !1


%swap16 = tail call x86_fp80 @__lz_bswapld(x86_fp80 %0) #1


ret x86_fp80 %swap16


}

where the untransformed bytecode was:


define x86_fp80 @ld(x86_fp80* nocapture readonly %d) local_unnamed_addr #0 {


entry:


%0 = load x86_fp80, x86_fp80* %d, align 16, !tbaa !1


ret x86_fp80 %0


}

What about this transformed bytecode does the later DAG->DAG pass dislike?
What is the valid range of registers in this function (i.e. is the reg=45 value
that I see in this function valid)?



--
Peeter
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20161109/b8ef5d96/attachment-0001.html>

Could you post the output from -debug-only=isel?

The transformed IR works for me.

-Krzysztof

On 11/9/2016 12:58 PM, Peeter Joot via llvm-dev wrote:
> I've implemented a function pass, and some associated hacks to llvm and
> clang, and am hitting an error in a DAG pass when my transformed IR is
> processed.  That transformed bytecode attempts to insert a call around a
> load of a 'long double'.  The error is:
>
>     3.      Running pass 'Function Pass Manager' on module
'big.c'.
>
>     4.      Running pass 'X86 DAG->DAG Instruction Selection' on
> function '@ld'
>
> The assert is after a failure to lookup a BestRC pointer in this function:
>
>
>
>     (gdb) b 141 if (BestRC == 0)
>
>     Breakpoint 3 at 0x7ffff6755436: file
> /home/pjoot/llvm/lib/CodeGen/TargetRegisterInfo.cpp, line 141.
>
>     (gdb) c
>
>     Continuing.
>
>     Breakpoint 3, llvm::TargetRegisterInfo::getMinimalPhysRegClass
> (this=0x5cc4f8, reg=44, VT=...) at
> /home/pjoot/llvm/lib/CodeGen/TargetRegisterInfo.cpp:141
>
>     141       assert(BestRC && "Couldn't find the register
class");
>
>     (gdb) p reg
>
>     $4 = 44
>
>     (gdb) p c
>
>     $5 = (const llvm::TargetRegisterClass *) 0x0
>
>     (gdb) p VT
>
>     $6 = {SimpleTy = llvm::MVT::SimpleValueType::i32}
>
>
> The bytecode that my pass is producing is:
>
>     ; Function Attrs: nounwind uwtable
>
>     define x86_fp80 @ld(x86_fp80* nocapture readonly %d)
> local_unnamed_addr #0 {
>
>     entry:
>
>       %0 = load x86_fp80, x86_fp80* %d, align 16, !tbaa !1
>
>       %swap16 = tail call x86_fp80 @__lz_bswapld(x86_fp80 %0) #1
>
>       ret x86_fp80 %swap16
>
>     }
>
> where the untransformed bytecode was:
>
>     define x86_fp80 @ld(x86_fp80* nocapture readonly %d)
> local_unnamed_addr #0 {
>
>     entry:
>
>       %0 = load x86_fp80, x86_fp80* %d, align 16, !tbaa !1
>
>       ret x86_fp80 %0
>
>     }
>
> What about this transformed bytecode does the later DAG->DAG pass
> dislike?  What is the valid range of registers in this function (i.e. is
> the reg=45 value that I see in this function valid)?
>
> --
> Peeter
>
>
>
>
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>
-- 
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, 
hosted by The Linux Foundation