forked from OSchip/llvm-project
Improve sub-register def handling in ProcessImplicitDefs.
This boils down to using MachineOperand::readsReg() more. This fixes PR11829 where a use ended up after the first def when lowering REG_SEQUENCE instructions involving IMPLICIT_DEFs. llvm-svn: 148996
This commit is contained in:
parent
9562f39e2f
commit
4864a81aa3
|
@ -50,10 +50,10 @@ ProcessImplicitDefs::CanTurnIntoImplicitDef(MachineInstr *MI,
|
|||
SmallSet<unsigned, 8> &ImpDefRegs) {
|
||||
switch(OpIdx) {
|
||||
case 1:
|
||||
return MI->isCopy() && (MI->getOperand(0).getSubReg() == 0 ||
|
||||
return MI->isCopy() && (!MI->getOperand(0).readsReg() ||
|
||||
ImpDefRegs.count(MI->getOperand(0).getReg()));
|
||||
case 2:
|
||||
return MI->isSubregToReg() && (MI->getOperand(0).getSubReg() == 0 ||
|
||||
return MI->isSubregToReg() && (!MI->getOperand(0).readsReg() ||
|
||||
ImpDefRegs.count(MI->getOperand(0).getReg()));
|
||||
default: return false;
|
||||
}
|
||||
|
@ -66,7 +66,7 @@ static bool isUndefCopy(MachineInstr *MI, unsigned Reg,
|
|||
MachineOperand &MO1 = MI->getOperand(1);
|
||||
if (MO1.getReg() != Reg)
|
||||
return false;
|
||||
if (!MO0.getSubReg() || ImpDefRegs.count(MO0.getReg()))
|
||||
if (!MO0.readsReg() || ImpDefRegs.count(MO0.getReg()))
|
||||
return true;
|
||||
return false;
|
||||
}
|
||||
|
@ -105,7 +105,9 @@ bool ProcessImplicitDefs::runOnMachineFunction(MachineFunction &fn) {
|
|||
MachineInstr *MI = &*I;
|
||||
++I;
|
||||
if (MI->isImplicitDef()) {
|
||||
if (MI->getOperand(0).getSubReg())
|
||||
ImpDefMIs.push_back(MI);
|
||||
// Is this a sub-register read-modify-write?
|
||||
if (MI->getOperand(0).readsReg())
|
||||
continue;
|
||||
unsigned Reg = MI->getOperand(0).getReg();
|
||||
ImpDefRegs.insert(Reg);
|
||||
|
@ -113,12 +115,11 @@ bool ProcessImplicitDefs::runOnMachineFunction(MachineFunction &fn) {
|
|||
for (const unsigned *SS = TRI->getSubRegisters(Reg); *SS; ++SS)
|
||||
ImpDefRegs.insert(*SS);
|
||||
}
|
||||
ImpDefMIs.push_back(MI);
|
||||
continue;
|
||||
}
|
||||
|
||||
// Eliminate %reg1032:sub<def> = COPY undef.
|
||||
if (MI->isCopy() && MI->getOperand(0).getSubReg()) {
|
||||
if (MI->isCopy() && MI->getOperand(0).readsReg()) {
|
||||
MachineOperand &MO = MI->getOperand(1);
|
||||
if (MO.isUndef() || ImpDefRegs.count(MO.getReg())) {
|
||||
if (MO.isKill()) {
|
||||
|
@ -140,7 +141,7 @@ bool ProcessImplicitDefs::runOnMachineFunction(MachineFunction &fn) {
|
|||
bool ChangedToImpDef = false;
|
||||
for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
|
||||
MachineOperand& MO = MI->getOperand(i);
|
||||
if (!MO.isReg() || (MO.isDef() && !MO.getSubReg()) || MO.isUndef())
|
||||
if (!MO.isReg() || !MO.readsReg())
|
||||
continue;
|
||||
unsigned Reg = MO.getReg();
|
||||
if (!Reg)
|
||||
|
@ -172,10 +173,10 @@ bool ProcessImplicitDefs::runOnMachineFunction(MachineFunction &fn) {
|
|||
continue;
|
||||
}
|
||||
if (MO.isKill() || MI->isRegTiedToDefOperand(i)) {
|
||||
// Make sure other uses of
|
||||
// Make sure other reads of Reg are also marked <undef>.
|
||||
for (unsigned j = i+1; j != e; ++j) {
|
||||
MachineOperand &MOJ = MI->getOperand(j);
|
||||
if (MOJ.isReg() && MOJ.isUse() && MOJ.getReg() == Reg)
|
||||
if (MOJ.isReg() && MOJ.getReg() == Reg && MOJ.readsReg())
|
||||
MOJ.setIsUndef();
|
||||
}
|
||||
ImpDefRegs.erase(Reg);
|
||||
|
|
|
@ -1,5 +1,6 @@
|
|||
; RUN: llc < %s -mcpu=cortex-a8 -verify-machineinstrs -verify-coalescing
|
||||
; PR11841
|
||||
; PR11829
|
||||
target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:64:128-a0:0:64-n32-S64"
|
||||
target triple = "armv7-none-linux-eabi"
|
||||
|
||||
|
@ -37,7 +38,30 @@ bb:
|
|||
ret void
|
||||
}
|
||||
|
||||
define arm_aapcs_vfpcc void @foo2() nounwind uwtable {
|
||||
entry:
|
||||
br i1 undef, label %for.end, label %cond.end295
|
||||
|
||||
cond.end295: ; preds = %entry
|
||||
%shuffle.i39.i.i1035 = shufflevector <2 x i64> undef, <2 x i64> undef, <1 x i32> zeroinitializer
|
||||
%shuffle.i38.i.i1036 = shufflevector <2 x i64> zeroinitializer, <2 x i64> undef, <1 x i32> zeroinitializer
|
||||
%shuffle.i37.i.i1037 = shufflevector <1 x i64> %shuffle.i39.i.i1035, <1 x i64> %shuffle.i38.i.i1036, <2 x i32> <i32 0, i32 1>
|
||||
%0 = bitcast <2 x i64> %shuffle.i37.i.i1037 to <4 x float>
|
||||
%1 = bitcast <4 x float> undef to <2 x i64>
|
||||
%shuffle.i36.i.i = shufflevector <2 x i64> %1, <2 x i64> undef, <1 x i32> zeroinitializer
|
||||
%shuffle.i35.i.i = shufflevector <2 x i64> undef, <2 x i64> undef, <1 x i32> zeroinitializer
|
||||
%shuffle.i34.i.i = shufflevector <1 x i64> %shuffle.i36.i.i, <1 x i64> %shuffle.i35.i.i, <2 x i32> <i32 0, i32 1>
|
||||
%2 = bitcast <2 x i64> %shuffle.i34.i.i to <4 x float>
|
||||
tail call void @llvm.arm.neon.vst1.v4f32(i8* undef, <4 x float> %0, i32 4) nounwind
|
||||
tail call void @llvm.arm.neon.vst1.v4f32(i8* undef, <4 x float> %2, i32 4) nounwind
|
||||
unreachable
|
||||
|
||||
for.end: ; preds = %entry
|
||||
ret void
|
||||
}
|
||||
|
||||
declare arm_aapcs_vfpcc void @bar(i8*, float, float, float)
|
||||
declare void @llvm.arm.neon.vst1.v4f32(i8*, <4 x float>, i32) nounwind
|
||||
|
||||
!0 = metadata !{metadata !"omnipotent char", metadata !1}
|
||||
!1 = metadata !{metadata !"Simple C/C++ TBAA", null}
|
||||
|
|
Loading…
Reference in New Issue