R600/SI: Fix a bug with handling of INSERT_SUBREG in SIFixSGPRCopies

This prevents a future commit from regressing the load-i1.ll test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208887 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
Tom Stellard 2014-05-15 14:41:55 +00:00
parent c1de569ce8
commit 704fe35d88
2 changed files with 28 additions and 7 deletions

View File

@ -259,15 +259,18 @@ bool SIFixSGPRCopies::runOnMachineFunction(MachineFunction &MF) {
break;
}
case AMDGPU::INSERT_SUBREG: {
const TargetRegisterClass *DstRC, *SrcRC;
const TargetRegisterClass *DstRC, *Src0RC, *Src1RC;
DstRC = MRI.getRegClass(MI.getOperand(0).getReg());
SrcRC = MRI.getRegClass(MI.getOperand(1).getReg());
if (!TRI->isSGPRClass(DstRC) || !TRI->hasVGPRs(SrcRC))
break;
Src0RC = MRI.getRegClass(MI.getOperand(1).getReg());
Src1RC = MRI.getRegClass(MI.getOperand(2).getReg());
if (TRI->isSGPRClass(DstRC) &&
(TRI->hasVGPRs(Src0RC) || TRI->hasVGPRs(Src1RC))) {
DEBUG(dbgs() << " Fixing INSERT_SUBREG:\n");
DEBUG(MI.print(dbgs()));
TII->moveToVALU(MI);
}
break;
}
}
}
}

View File

@ -691,6 +691,7 @@ bool SIInstrInfo::canReadVGPR(const MachineInstr &MI, unsigned OpNo) const {
case AMDGPU::COPY:
case AMDGPU::REG_SEQUENCE:
case AMDGPU::PHI:
case AMDGPU::INSERT_SUBREG:
return RI.hasVGPRs(getOpRegClass(MI, 0));
default:
return RI.hasVGPRs(getOpRegClass(MI, OpNo));
@ -924,6 +925,23 @@ void SIInstrInfo::legalizeOperands(MachineInstr *MI) const {
}
}
// Legalize INSERT_SUBREG
// src0 must have the same register class as dst
if (MI->getOpcode() == AMDGPU::INSERT_SUBREG) {
unsigned Dst = MI->getOperand(0).getReg();
unsigned Src0 = MI->getOperand(1).getReg();
const TargetRegisterClass *DstRC = MRI.getRegClass(Dst);
const TargetRegisterClass *Src0RC = MRI.getRegClass(Src0);
if (DstRC != Src0RC) {
MachineBasicBlock &MBB = *MI->getParent();
unsigned NewSrc0 = MRI.createVirtualRegister(DstRC);
BuildMI(MBB, MI, MI->getDebugLoc(), get(AMDGPU::COPY), NewSrc0)
.addReg(Src0);
MI->getOperand(1).setReg(NewSrc0);
}
return;
}
// Legalize MUBUF* instructions
// FIXME: If we start using the non-addr64 instructions for compute, we
// may need to legalize them here.