llvm dev - Oct 2021 - Help: A Question about AArch64 Register Class

Hello, 

I am modifying the new committed pass "aarch64-mi-peephole-opt", but
get unexpected errors.

My code:

  const TargetRegisterClass *RC 
      (RegSize == 32) ? &AArch64::GPR32spRegClass :
&AArch64::GPR64spRegClass;

  Register DstReg = MI.getOperand(0).getReg();

  Register SrcReg = MI.getOperand(1).getReg();

  Register TmpReg = MRI->createVirtualRegister(RC);

  BuildMI(*MBB, MI, DL, TII->get(AArch64::SUBXri), TmpReg)

      .addReg(SrcReg)

      .addImm(Imm)

      .addImm(12);

 

But when I tried to compile a simple function with

"llc a.ll -mtriple=aarch64 -verify-machineinstrs", I got errors like

 

*** Bad machine code: Illegal virtual register for instruction ***

- function:    add_two_parts_imm_i64

- basic block: %bb.0  (0x561f681b0410)

- instruction: %3:gpr64common = SUBXri %0:gpr64, 2730, 12

- operand 1:   %0:gpr64

Expected a GPR64sp register, but got a GPR64 register

LLVM ERROR: Found 1 machine code errors.

 

But if I changed to Opcode from SUBXri to ANDXri, then the errors were
eliminated.

 

What is difference between ANDXri and SUBXri? (also ADDXri has the same errors)

What should I do to avoid the errors with Opcode SUBXri/ADDXri ?

 

Ben Shi

 

 
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20211003/f2e3f26e/attachment.html>

Hi Ben,

As the error message mentioned, the first operand of SUBXri needs to be GPR64sp
register class. The ANDXri one is GPR64.
In order to align the register classes, you can use constrainRegClass with the
SrcReg and it will constraint the register class to GPR64common which is common
between GPR64sp and GPR64.

JinGu Kang

From: llvm-dev <llvm-dev-bounces at lists.llvm.org> On Behalf Of Ben Shi
via llvm-dev
Sent: 03 October 2021 15:46
To: llvm-dev <llvm-dev at lists.llvm.org>
Subject: [llvm-dev] Help: A Question about AArch64 Register Class


Hello,
I am modifying the new committed pass "aarch64-mi-peephole-opt", but
get unexpected errors.
My code:
  const TargetRegisterClass *RC       (RegSize == 32) ?
&AArch64::GPR32spRegClass : &AArch64::GPR64spRegClass;
  Register DstReg = MI.getOperand(0).getReg();
  Register SrcReg = MI.getOperand(1).getReg();
  Register TmpReg = MRI->createVirtualRegister(RC);
  BuildMI(*MBB, MI, DL, TII->get(AArch64::SUBXri), TmpReg)
      .addReg(SrcReg)
      .addImm(Imm)
      .addImm(12);

But when I tried to compile a simple function with
"llc a.ll -mtriple=aarch64 -verify-machineinstrs", I got errors like

*** Bad machine code: Illegal virtual register for instruction ***
- function:    add_two_parts_imm_i64
- basic block: %bb.0  (0x561f681b0410)
- instruction: %3:gpr64common = SUBXri %0:gpr64, 2730, 12
- operand 1:   %0:gpr64
Expected a GPR64sp register, but got a GPR64 register
LLVM ERROR: Found 1 machine code errors.

But if I changed to Opcode from SUBXri to ANDXri, then the errors were
eliminated.

What is difference between ANDXri and SUBXri? (also ADDXri has the same errors)
What should I do to avoid the errors with Opcode SUBXri/ADDXri ?

Ben Shi


-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20211004/b4d0fc60/attachment.html>