[AMDGPU] LiveIntervals is inaccurate after TwoAddressInstructionPass

[paperchalice]

There are some test failures in pull request #98632 when LiveIntervals analysis happens before TwoAddressInstructionPass, because SIInstrInfo::convertToThreeAddress doesn't handle live intervals after folding immediate values:

llvm-project/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp

Lines 3949 to 3961 in ae63db7

	const auto killDef = [&]() -> void {
	const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
	// The only user is the instruction which will be killed.
	Register DefReg = DefMI->getOperand(0).getReg();
	if (!MRI.hasOneNonDBGUse(DefReg))
	return;
	// We cannot just remove the DefMI here, calling pass will crash.
	DefMI->setDesc(get(AMDGPU::IMPLICIT_DEF));
	for (unsigned I = DefMI->getNumOperands() - 1; I != 0; --I)
	DefMI->removeOperand(I);
	if (LV)
	LV->getVarInfo(DefReg).AliveBlocks.clear();
	};

The live interval of DefReg is no longer valid.

---

In #66775, -early-live-intervals will not work when using -run-pass, the Liveintervals LIS in TwoAddressInstructionPass is nullptr when I debugging this part. A possible way to test -early-live-intervals is using:

llc -march=amdgcn -mcpu=gfx1010 %s -run-pass liveintervals -run-pass twoaddressinstruction -verify-machineinstrs -o -

but llc will crash.

[llvmbot]

@llvm/issue-subscribers-backend-amdgpu

Author: None (paperchalice)

There are some test failures in pull request #98632 when `LiveIntervals` analysis happens before `TwoAddressInstructionPass`, because `SIInstrInfo::convertToThreeAddress` doesn't handle live intervals after folding immediate values: https://github.com/llvm/llvm-project/blob/ae63db78828621feca713016e33f34717da626a5/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp#L3949-L3961 The live interval of `DefReg` is no longer valid.