Transfer implicit operands in NEONMoveFixPass.

Later passes /are/ using this information when running the register
scavenger.

This fixes the second problem in PR10520.

llvm-svn: 136440
This commit is contained in:
Jakob Stoklund Olesen 2011-07-29 00:27:35 +00:00
parent 9c3badceba
commit b28ee4115d
2 changed files with 41 additions and 10 deletions
llvm
lib/Target/ARM
test/CodeGen/ARM

View File

@ -40,6 +40,8 @@ namespace {
typedef DenseMap<unsigned, const MachineInstr*> RegMap;
bool InsertMoves(MachineBasicBlock &MBB);
void TransferImpOps(MachineInstr &Old, MachineInstr &New);
};
char NEONMoveFixPass::ID = 0;
}
@ -49,6 +51,16 @@ static bool inNEONDomain(unsigned Domain, bool isA8) {
(isA8 && (Domain & ARMII::DomainNEONA8));
}
/// Transfer implicit kill and def operands from Old to New.
void NEONMoveFixPass::TransferImpOps(MachineInstr &Old, MachineInstr &New) {
for (unsigned i = 0, e = Old.getNumOperands(); i != e; ++i) {
MachineOperand &MO = Old.getOperand(i);
if (!MO.isReg() || !MO.isImplicit())
continue;
New.addOperand(MO);
}
}
bool NEONMoveFixPass::InsertMoves(MachineBasicBlock &MBB) {
RegMap Defs;
bool Modified = false;
@ -82,17 +94,15 @@ bool NEONMoveFixPass::InsertMoves(MachineBasicBlock &MBB) {
DEBUG({errs() << "vmov convert: "; MI->dump();});
// It's safe to ignore imp-defs / imp-uses here, since:
// - We're running late, no intelligent condegen passes should be run
// afterwards
// - The imp-defs / imp-uses are superregs only, we don't care about
// them.
// We need to preserve imp-defs / imp-uses here. Following passes may
// use the register scavenger to update liveness.
MachineInstr *NewMI =
AddDefaultPred(BuildMI(MBB, *MI, MI->getDebugLoc(),
TII->get(ARM::VORRd), DestReg)
.addReg(SrcReg).addReg(SrcReg));
TransferImpOps(*MI, *NewMI);
MBB.erase(MI);
MachineBasicBlock::iterator I = prior(NextMII);
MI = &*I;
MI = NewMI;
DEBUG({errs() << " into: "; MI->dump();});

View File

@ -48,3 +48,24 @@ bb:
store <4 x float> %tmp20, <4 x float>* undef, align 16
ret void
}
; PR10520, second bug. NEONMoveFixPass needs to preserve implicit operands.
define arm_aapcs_vfpcc void @pr10520_2() nounwind align 2 {
bb:
%tmp76 = shufflevector <2 x i64> zeroinitializer, <2 x i64> zeroinitializer, <1 x i32> <i32 1>
%tmp77 = bitcast <1 x i64> %tmp76 to <2 x float>
%tmp78 = shufflevector <2 x float> %tmp77, <2 x float> %tmp77, <4 x i32> zeroinitializer
%tmp81 = fmul <4 x float> undef, %tmp78
%tmp82 = fadd <4 x float> %tmp81, undef
%tmp85 = fadd <4 x float> %tmp82, undef
%tmp86 = bitcast <4 x float> %tmp85 to i128
%tmp136 = bitcast i128 %tmp86 to <4 x float>
%tmp137 = bitcast <4 x float> %tmp136 to i128
%tmp138 = bitcast i128 %tmp137 to <4 x float>
%tmp139 = bitcast <4 x float> %tmp138 to i128
%tmp152 = bitcast i128 %tmp139 to <4 x float>
%tmp153 = bitcast <4 x float> %tmp152 to i128
%tmp154 = bitcast i128 %tmp153 to <4 x float>
store <4 x float> %tmp154, <4 x float>* undef, align 16
ret void
}