llvm dev - May 2019 - Spilling multiple individual instructions not supported in [Target]InstrInfo::storeRegToStackSlot()? (error: <<Assertion `itr != mi2iMap.end() && "Instruction not found in maps."' failed.>>)

Hello.
     Could you please tell me: is it not allowed to create 2 or more individual
(that is,
not bundled) instructions in ConnexInstrInfo::storeRegToStackSlot()? Such a
thing was
possible with an older version, LLVM 3.9 from Jul 2016, but with a very recent
version of
LLVM from Mar 2019 (SVN revision 357222) it seems it's not possible anymore.

     For example, when I add 1 NOP and 1 STORE instruction (I can even create
more than 2
instructions) at iterator I in ConnexInstrInfo::storeRegToStackSlot() I get the
following
error:
       <<Assertion `itr != mi2iMap.end() && "Instruction not
found in maps."' failed.>>
     But if I create only 1 instruction then everything is OK.

      I found some info from 2017 that supports my reasoning at 
https://lists.llvm.org/pipermail/llvm-dev/2017-February/110281.html where they
write:
       - "This does make the assumption that storeRegtoStackSlot only
inserts a single
instruction. From what I can see, that is true for all in-tree backends."
       - "modify storeRegToStackSlot to return the number of MachineInstr
inserted and use
that value when marking instructions as FrameSetup. This is more invasive, as it
will
affect all in tree and out-of-tree backends."


     Please let me know what is the exact situation now in LLVM: can't we
create more than
1 individual instruction in 1 call of [Target]InstrInfo::storeRegToStackSlot()?

   Thank you,
     Alex

On 05/03/2019 06:13 PM, Alex Susu via llvm-dev wrote:
>   Hello.
>     Could you please tell me: is it not allowed to create 2 or more
individual (that is, not bundled) instructions in
ConnexInstrInfo::storeRegToStackSlot()? Such a thing was possible with an older
version, LLVM 3.9 from Jul 2016, but with a very recent version of LLVM from Mar
2019 (SVN revision 357222) it seems it's not possible anymore.
> 
>     For example, when I add 1 NOP and 1 STORE instruction (I can even
create more than 2 instructions) at iterator I in
ConnexInstrInfo::storeRegToStackSlot() I get the following error:
>       <<Assertion `itr != mi2iMap.end() && "Instruction
not found in maps."' failed.>>
>     But if I create only 1 instruction then everything is OK.
> 
>      I found some info from 2017 that supports my reasoning at
https://lists.llvm.org/pipermail/llvm-dev/2017-February/110281.html where they
write:
>       - "This does make the assumption that storeRegtoStackSlot only
inserts a single
> instruction. From what I can see, that is true for all in-tree
backends."
>       - "modify storeRegToStackSlot to return the number of
MachineInstr inserted and use that value when marking instructions as
FrameSetup. This is more invasive, as it will affect all in tree and out-of-tree
backends."
> 
> 
>     Please let me know what is the exact situation now in LLVM: can't
we create more than 1 individual instruction in 1 call of
[Target]InstrInfo::storeRegToStackSlot()?
> 
I think this is still the case that you are only allowed to add 1 instruction.
The AMDGPU backend[1] works around this by emitting a single pseudo
instruction during register allocation and then expanding it to multiple
instructions in a later pass.

-Tom

[1]
https://github.com/llvm/llvm-project/blob/master/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp#L915
>   Thank you,
>     Alex
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev