unsigned SrcReg = DefMI->getOperand(1).getReg();
if (!TargetRegisterInfo::isVirtualRegister(SrcReg))
continue;
- if (DefMI->getOperand(0).getSubReg() || DefMI->getOperand(1).getSubReg())
+ if (DefMI->getOperand(0).getSubReg())
continue;
- if (!MRI->constrainRegClass(SrcReg, MRI->getRegClass(Reg)))
+ unsigned SrcSubReg = DefMI->getOperand(1).getSubReg();
+ const TargetRegisterClass *RC = MRI->getRegClass(Reg);
+ if (SrcSubReg)
+ RC = TRI->getMatchingSuperRegClass(MRI->getRegClass(SrcReg), RC,
+ SrcSubReg);
+ if (!MRI->constrainRegClass(SrcReg, RC))
continue;
DEBUG(dbgs() << "Coalescing: " << *DefMI);
DEBUG(dbgs() << "*** to: " << *MI);
- MO.setReg(SrcReg);
+ MO.substVirtReg(SrcReg, SrcSubReg, *TRI);
MRI->clearKillFlags(SrcReg);
DefMI->eraseFromParent();
++NumCoalesces;
unsigned LastCopiedReg = 0;
SlotIndex LastCopyIdx;
unsigned RegB = 0;
+ unsigned SubRegB = 0;
for (unsigned tpi = 0, tpe = TiedPairs.size(); tpi != tpe; ++tpi) {
unsigned SrcIdx = TiedPairs[tpi].first;
unsigned DstIdx = TiedPairs[tpi].second;
// Grab RegB from the instruction because it may have changed if the
// instruction was commuted.
RegB = MI->getOperand(SrcIdx).getReg();
+ SubRegB = MI->getOperand(SrcIdx).getSubReg();
if (RegA == RegB) {
// The register is tied to multiple destinations (or else we would
#endif
// Emit a copy.
- BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
- TII->get(TargetOpcode::COPY), RegA).addReg(RegB);
+ MachineInstrBuilder MIB = BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
+ TII->get(TargetOpcode::COPY), RegA);
+ // If this operand is folding a truncation, the truncation now moves to the
+ // copy so that the register classes remain valid for the operands.
+ MIB.addReg(RegB, 0, SubRegB);
+ const TargetRegisterClass *RC = MRI->getRegClass(RegB);
+ if (SubRegB) {
+ if (TargetRegisterInfo::isVirtualRegister(RegA)) {
+ assert(TRI->getMatchingSuperRegClass(MRI->getRegClass(RegB),
+ MRI->getRegClass(RegA), SubRegB) &&
+ "tied subregister must be a truncation");
+ // The superreg class will not be used to constrain the subreg class.
+ RC = 0;
+ }
+ else {
+ assert(TRI->getMatchingSuperReg(RegA, SubRegB, MRI->getRegClass(RegB))
+ && "tied subregister must be a truncation");
+ }
+ }
// Update DistanceMap.
MachineBasicBlock::iterator PrevMI = MI;
}
}
- DEBUG(dbgs() << "\t\tprepend:\t" << *PrevMI);
+ DEBUG(dbgs() << "\t\tprepend:\t" << *MIB);
MachineOperand &MO = MI->getOperand(SrcIdx);
assert(MO.isReg() && MO.getReg() == RegB && MO.isUse() &&
// Make sure regA is a legal regclass for the SrcIdx operand.
if (TargetRegisterInfo::isVirtualRegister(RegA) &&
TargetRegisterInfo::isVirtualRegister(RegB))
- MRI->constrainRegClass(RegA, MRI->getRegClass(RegB));
-
+ MRI->constrainRegClass(RegA, RC);
MO.setReg(RegA);
+ MO.setSubReg(0);
// Propagate SrcRegMap.
SrcRegMap[RegA] = RegB;
// Replace other (un-tied) uses of regB with LastCopiedReg.
for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
MachineOperand &MO = MI->getOperand(i);
- if (MO.isReg() && MO.getReg() == RegB && MO.isUse()) {
+ if (MO.isReg() && MO.getReg() == RegB && MO.getSubReg() == SubRegB &&
+ MO.isUse()) {
if (MO.isKill()) {
MO.setIsKill(false);
RemovedKillFlag = true;
}
MO.setReg(LastCopiedReg);
+ MO.setSubReg(0);
}
}
}
--- /dev/null
+; RUN: llc < %s -mtriple=x86_64-darwin -mcpu=generic | FileCheck %s
+; rdar:15661073 simple example of redundant adds
+;
+; MachineCSE should coalesce trivial subregister copies.
+;
+; The extra movl+addl should be removed during MachineCSE.
+; CHECK-LABEL: redundantadd
+; CHECK: cmpq
+; CHECK: movq
+; CHECK-NOT: movl
+; CHECK: addl
+; CHECK-NOT: addl
+; CHECK: ret
+
+define i64 @redundantadd(i64* %a0, i64* %a1) {
+entry:
+ %tmp8 = load i64* %a0, align 8
+ %tmp12 = load i64* %a1, align 8
+ %tmp13 = icmp ult i64 %tmp12, -281474976710656
+ br i1 %tmp13, label %exit1, label %body
+
+exit1:
+ unreachable
+
+body:
+ %tmp14 = trunc i64 %tmp8 to i32
+ %tmp15 = trunc i64 %tmp12 to i32
+ %tmp16 = tail call { i32, i1 } @llvm.sadd.with.overflow.i32(i32 %tmp14, i32 %tmp15)
+ %tmp17 = extractvalue { i32, i1 } %tmp16, 1
+ br i1 %tmp17, label %exit2, label %return
+
+exit2:
+ unreachable
+
+return:
+ %tmp18 = add i64 %tmp12, %tmp8
+ %tmp19 = and i64 %tmp18, 4294967295
+ %tmp20 = or i64 %tmp19, -281474976710656
+ ret i64 %tmp20
+}
+
+declare { i32, i1 } @llvm.sadd.with.overflow.i32(i32, i32)