aboutsummaryrefslogtreecommitdiff
path: root/lib/Target/ARM
diff options
context:
space:
mode:
Diffstat (limited to 'lib/Target/ARM')
-rw-r--r--lib/Target/ARM/A15SDOptimizer.cpp12
-rw-r--r--lib/Target/ARM/ARMBaseInstrInfo.cpp2
-rw-r--r--lib/Target/ARM/ARMExpandPseudoInsts.cpp4
-rw-r--r--lib/Target/ARM/ARMISelDAGToDAG.cpp75
-rw-r--r--lib/Target/ARM/ARMInstrNEON.td8
5 files changed, 87 insertions, 14 deletions
diff --git a/lib/Target/ARM/A15SDOptimizer.cpp b/lib/Target/ARM/A15SDOptimizer.cpp
index ff585b41a2aa..3e805a2b68bc 100644
--- a/lib/Target/ARM/A15SDOptimizer.cpp
+++ b/lib/Target/ARM/A15SDOptimizer.cpp
@@ -418,7 +418,8 @@ SmallVector<unsigned, 8> A15SDOptimizer::getReadDPRs(MachineInstr *MI) {
if (!MO.isReg() || !MO.isUse())
continue;
if (!usesRegClass(MO, &ARM::DPRRegClass) &&
- !usesRegClass(MO, &ARM::QPRRegClass))
+ !usesRegClass(MO, &ARM::QPRRegClass) &&
+ !usesRegClass(MO, &ARM::DPairRegClass)) // Treat DPair as QPR
continue;
Defs.push_back(MO.getReg());
@@ -538,7 +539,10 @@ A15SDOptimizer::optimizeAllLanesPattern(MachineInstr *MI, unsigned Reg) {
InsertPt++;
unsigned Out;
- if (MRI->getRegClass(Reg)->hasSuperClassEq(&ARM::QPRRegClass)) {
+ // DPair has the same length as QPR and also has two DPRs as subreg.
+ // Treat DPair as QPR.
+ if (MRI->getRegClass(Reg)->hasSuperClassEq(&ARM::QPRRegClass) ||
+ MRI->getRegClass(Reg)->hasSuperClassEq(&ARM::DPairRegClass)) {
unsigned DSub0 = createExtractSubreg(MBB, InsertPt, DL, Reg,
ARM::dsub_0, &ARM::DPRRegClass);
unsigned DSub1 = createExtractSubreg(MBB, InsertPt, DL, Reg,
@@ -571,7 +575,9 @@ A15SDOptimizer::optimizeAllLanesPattern(MachineInstr *MI, unsigned Reg) {
default: llvm_unreachable("Unknown preferred lane!");
}
- bool UsesQPR = usesRegClass(MI->getOperand(0), &ARM::QPRRegClass);
+ // Treat DPair as QPR
+ bool UsesQPR = usesRegClass(MI->getOperand(0), &ARM::QPRRegClass) ||
+ usesRegClass(MI->getOperand(0), &ARM::DPairRegClass);
Out = createImplicitDef(MBB, InsertPt, DL);
Out = createInsertSubreg(MBB, InsertPt, DL, Out, PrefLane, Reg);
diff --git a/lib/Target/ARM/ARMBaseInstrInfo.cpp b/lib/Target/ARM/ARMBaseInstrInfo.cpp
index f835a4e5b5fe..658af8380d8d 100644
--- a/lib/Target/ARM/ARMBaseInstrInfo.cpp
+++ b/lib/Target/ARM/ARMBaseInstrInfo.cpp
@@ -3684,6 +3684,7 @@ ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
case ARM::VLD3d16Pseudo:
case ARM::VLD3d32Pseudo:
case ARM::VLD1d64TPseudo:
+ case ARM::VLD1d64TPseudoWB_fixed:
case ARM::VLD3d8Pseudo_UPD:
case ARM::VLD3d16Pseudo_UPD:
case ARM::VLD3d32Pseudo_UPD:
@@ -3700,6 +3701,7 @@ ARMBaseInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
case ARM::VLD4d16Pseudo:
case ARM::VLD4d32Pseudo:
case ARM::VLD1d64QPseudo:
+ case ARM::VLD1d64QPseudoWB_fixed:
case ARM::VLD4d8Pseudo_UPD:
case ARM::VLD4d16Pseudo_UPD:
case ARM::VLD4d32Pseudo_UPD:
diff --git a/lib/Target/ARM/ARMExpandPseudoInsts.cpp b/lib/Target/ARM/ARMExpandPseudoInsts.cpp
index e6f7f86c5587..3e62b649e0dc 100644
--- a/lib/Target/ARM/ARMExpandPseudoInsts.cpp
+++ b/lib/Target/ARM/ARMExpandPseudoInsts.cpp
@@ -136,7 +136,9 @@ static const NEONLdStTableEntry NEONLdStTable[] = {
{ ARM::VLD1LNq8Pseudo_UPD, ARM::VLD1LNd8_UPD, true, true, true, EvenDblSpc, 1, 8 ,true},
{ ARM::VLD1d64QPseudo, ARM::VLD1d64Q, true, false, false, SingleSpc, 4, 1 ,false},
+{ ARM::VLD1d64QPseudoWB_fixed, ARM::VLD1d64Qwb_fixed, true, true, false, SingleSpc, 4, 1 ,false},
{ ARM::VLD1d64TPseudo, ARM::VLD1d64T, true, false, false, SingleSpc, 3, 1 ,false},
+{ ARM::VLD1d64TPseudoWB_fixed, ARM::VLD1d64Twb_fixed, true, true, false, SingleSpc, 3, 1 ,false},
{ ARM::VLD2LNd16Pseudo, ARM::VLD2LNd16, true, false, false, SingleSpc, 2, 4 ,true},
{ ARM::VLD2LNd16Pseudo_UPD, ARM::VLD2LNd16_UPD, true, true, true, SingleSpc, 2, 4 ,true},
@@ -1071,6 +1073,7 @@ bool ARMExpandPseudo::ExpandMI(MachineBasicBlock &MBB,
case ARM::VLD3d16Pseudo:
case ARM::VLD3d32Pseudo:
case ARM::VLD1d64TPseudo:
+ case ARM::VLD1d64TPseudoWB_fixed:
case ARM::VLD3d8Pseudo_UPD:
case ARM::VLD3d16Pseudo_UPD:
case ARM::VLD3d32Pseudo_UPD:
@@ -1087,6 +1090,7 @@ bool ARMExpandPseudo::ExpandMI(MachineBasicBlock &MBB,
case ARM::VLD4d16Pseudo:
case ARM::VLD4d32Pseudo:
case ARM::VLD1d64QPseudo:
+ case ARM::VLD1d64QPseudoWB_fixed:
case ARM::VLD4d8Pseudo_UPD:
case ARM::VLD4d16Pseudo_UPD:
case ARM::VLD4d32Pseudo_UPD:
diff --git a/lib/Target/ARM/ARMISelDAGToDAG.cpp b/lib/Target/ARM/ARMISelDAGToDAG.cpp
index 87d15226947a..6d9b18877f72 100644
--- a/lib/Target/ARM/ARMISelDAGToDAG.cpp
+++ b/lib/Target/ARM/ARMISelDAGToDAG.cpp
@@ -1673,9 +1673,61 @@ SDValue ARMDAGToDAGISel::GetVLDSTAlign(SDValue Align, unsigned NumVecs,
return CurDAG->getTargetConstant(Alignment, MVT::i32);
}
+static bool isVLDfixed(unsigned Opc)
+{
+ switch (Opc) {
+ default: return false;
+ case ARM::VLD1d8wb_fixed : return true;
+ case ARM::VLD1d16wb_fixed : return true;
+ case ARM::VLD1d64Qwb_fixed : return true;
+ case ARM::VLD1d32wb_fixed : return true;
+ case ARM::VLD1d64wb_fixed : return true;
+ case ARM::VLD1d64TPseudoWB_fixed : return true;
+ case ARM::VLD1d64QPseudoWB_fixed : return true;
+ case ARM::VLD1q8wb_fixed : return true;
+ case ARM::VLD1q16wb_fixed : return true;
+ case ARM::VLD1q32wb_fixed : return true;
+ case ARM::VLD1q64wb_fixed : return true;
+ case ARM::VLD2d8wb_fixed : return true;
+ case ARM::VLD2d16wb_fixed : return true;
+ case ARM::VLD2d32wb_fixed : return true;
+ case ARM::VLD2q8PseudoWB_fixed : return true;
+ case ARM::VLD2q16PseudoWB_fixed : return true;
+ case ARM::VLD2q32PseudoWB_fixed : return true;
+ case ARM::VLD2DUPd8wb_fixed : return true;
+ case ARM::VLD2DUPd16wb_fixed : return true;
+ case ARM::VLD2DUPd32wb_fixed : return true;
+ }
+}
+
+static bool isVSTfixed(unsigned Opc)
+{
+ switch (Opc) {
+ default: return false;
+ case ARM::VST1d8wb_fixed : return true;
+ case ARM::VST1d16wb_fixed : return true;
+ case ARM::VST1d32wb_fixed : return true;
+ case ARM::VST1d64wb_fixed : return true;
+ case ARM::VST1q8wb_fixed : return true;
+ case ARM::VST1q16wb_fixed : return true;
+ case ARM::VST1q32wb_fixed : return true;
+ case ARM::VST1q64wb_fixed : return true;
+ case ARM::VST1d64TPseudoWB_fixed : return true;
+ case ARM::VST1d64QPseudoWB_fixed : return true;
+ case ARM::VST2d8wb_fixed : return true;
+ case ARM::VST2d16wb_fixed : return true;
+ case ARM::VST2d32wb_fixed : return true;
+ case ARM::VST2q8PseudoWB_fixed : return true;
+ case ARM::VST2q16PseudoWB_fixed : return true;
+ case ARM::VST2q32PseudoWB_fixed : return true;
+ }
+}
+
// Get the register stride update opcode of a VLD/VST instruction that
// is otherwise equivalent to the given fixed stride updating instruction.
static unsigned getVLDSTRegisterUpdateOpcode(unsigned Opc) {
+ assert((isVLDfixed(Opc) || isVSTfixed(Opc))
+ && "Incorrect fixed stride updating instruction.");
switch (Opc) {
default: break;
case ARM::VLD1d8wb_fixed: return ARM::VLD1d8wb_register;
@@ -1686,6 +1738,10 @@ static unsigned getVLDSTRegisterUpdateOpcode(unsigned Opc) {
case ARM::VLD1q16wb_fixed: return ARM::VLD1q16wb_register;
case ARM::VLD1q32wb_fixed: return ARM::VLD1q32wb_register;
case ARM::VLD1q64wb_fixed: return ARM::VLD1q64wb_register;
+ case ARM::VLD1d64Twb_fixed: return ARM::VLD1d64Twb_register;
+ case ARM::VLD1d64Qwb_fixed: return ARM::VLD1d64Qwb_register;
+ case ARM::VLD1d64TPseudoWB_fixed: return ARM::VLD1d64TPseudoWB_register;
+ case ARM::VLD1d64QPseudoWB_fixed: return ARM::VLD1d64QPseudoWB_register;
case ARM::VST1d8wb_fixed: return ARM::VST1d8wb_register;
case ARM::VST1d16wb_fixed: return ARM::VST1d16wb_register;
@@ -1785,11 +1841,11 @@ SDNode *ARMDAGToDAGISel::SelectVLD(SDNode *N, bool isUpdating, unsigned NumVecs,
SDValue Inc = N->getOperand(AddrOpIdx + 1);
// FIXME: VLD1/VLD2 fixed increment doesn't need Reg0. Remove the reg0
// case entirely when the rest are updated to that form, too.
- if ((NumVecs == 1 || NumVecs == 2) && !isa<ConstantSDNode>(Inc.getNode()))
+ if ((NumVecs <= 2) && !isa<ConstantSDNode>(Inc.getNode()))
Opc = getVLDSTRegisterUpdateOpcode(Opc);
- // We use a VLD1 for v1i64 even if the pseudo says vld2/3/4, so
+ // FIXME: We use a VLD1 for v1i64 even if the pseudo says vld2/3/4, so
// check for that explicitly too. Horribly hacky, but temporary.
- if ((NumVecs != 1 && NumVecs != 2 && Opc != ARM::VLD1q64wb_fixed) ||
+ if ((NumVecs > 2 && !isVLDfixed(Opc)) ||
!isa<ConstantSDNode>(Inc.getNode()))
Ops.push_back(isa<ConstantSDNode>(Inc.getNode()) ? Reg0 : Inc);
}
@@ -1937,11 +1993,12 @@ SDNode *ARMDAGToDAGISel::SelectVST(SDNode *N, bool isUpdating, unsigned NumVecs,
// case entirely when the rest are updated to that form, too.
if (NumVecs <= 2 && !isa<ConstantSDNode>(Inc.getNode()))
Opc = getVLDSTRegisterUpdateOpcode(Opc);
- // We use a VST1 for v1i64 even if the pseudo says vld2/3/4, so
+ // FIXME: We use a VST1 for v1i64 even if the pseudo says vld2/3/4, so
// check for that explicitly too. Horribly hacky, but temporary.
- if ((NumVecs > 2 && Opc != ARM::VST1q64wb_fixed) ||
- !isa<ConstantSDNode>(Inc.getNode()))
- Ops.push_back(isa<ConstantSDNode>(Inc.getNode()) ? Reg0 : Inc);
+ if (!isa<ConstantSDNode>(Inc.getNode()))
+ Ops.push_back(Inc);
+ else if (NumVecs > 2 && !isVSTfixed(Opc))
+ Ops.push_back(Reg0);
}
Ops.push_back(SrcReg);
Ops.push_back(Pred);
@@ -2834,7 +2891,7 @@ SDNode *ARMDAGToDAGISel::Select(SDNode *N) {
static const uint16_t DOpcodes[] = { ARM::VLD3d8Pseudo_UPD,
ARM::VLD3d16Pseudo_UPD,
ARM::VLD3d32Pseudo_UPD,
- ARM::VLD1q64wb_fixed};
+ ARM::VLD1d64TPseudoWB_fixed};
static const uint16_t QOpcodes0[] = { ARM::VLD3q8Pseudo_UPD,
ARM::VLD3q16Pseudo_UPD,
ARM::VLD3q32Pseudo_UPD };
@@ -2848,7 +2905,7 @@ SDNode *ARMDAGToDAGISel::Select(SDNode *N) {
static const uint16_t DOpcodes[] = { ARM::VLD4d8Pseudo_UPD,
ARM::VLD4d16Pseudo_UPD,
ARM::VLD4d32Pseudo_UPD,
- ARM::VLD1q64wb_fixed};
+ ARM::VLD1d64QPseudoWB_fixed};
static const uint16_t QOpcodes0[] = { ARM::VLD4q8Pseudo_UPD,
ARM::VLD4q16Pseudo_UPD,
ARM::VLD4q32Pseudo_UPD };
diff --git a/lib/Target/ARM/ARMInstrNEON.td b/lib/Target/ARM/ARMInstrNEON.td
index 43bd4c21dc39..0b05c08ed948 100644
--- a/lib/Target/ARM/ARMInstrNEON.td
+++ b/lib/Target/ARM/ARMInstrNEON.td
@@ -730,6 +730,8 @@ defm VLD1d32Twb : VLD1D3WB<{1,0,0,?}, "32">;
defm VLD1d64Twb : VLD1D3WB<{1,1,0,?}, "64">;
def VLD1d64TPseudo : VLDQQPseudo<IIC_VLD1x3>;
+def VLD1d64TPseudoWB_fixed : VLDQQWBfixedPseudo<IIC_VLD1x3>;
+def VLD1d64TPseudoWB_register : VLDQQWBregisterPseudo<IIC_VLD1x3>;
// ...with 4 registers
class VLD1D4<bits<4> op7_4, string Dt>
@@ -769,6 +771,8 @@ defm VLD1d32Qwb : VLD1D4WB<{1,0,?,?}, "32">;
defm VLD1d64Qwb : VLD1D4WB<{1,1,?,?}, "64">;
def VLD1d64QPseudo : VLDQQPseudo<IIC_VLD1x4>;
+def VLD1d64QPseudoWB_fixed : VLDQQWBfixedPseudo<IIC_VLD1x4>;
+def VLD1d64QPseudoWB_register : VLDQQWBregisterPseudo<IIC_VLD1x4>;
// VLD2 : Vector Load (multiple 2-element structures)
class VLD2<bits<4> op11_8, bits<4> op7_4, string Dt, RegisterOperand VdTy,
@@ -1671,7 +1675,7 @@ defm VST1d32Twb : VST1D3WB<{1,0,0,?}, "32">;
defm VST1d64Twb : VST1D3WB<{1,1,0,?}, "64">;
def VST1d64TPseudo : VSTQQPseudo<IIC_VST1x3>;
-def VST1d64TPseudoWB_fixed : VSTQQWBPseudo<IIC_VST1x3u>;
+def VST1d64TPseudoWB_fixed : VSTQQWBfixedPseudo<IIC_VST1x3u>;
def VST1d64TPseudoWB_register : VSTQQWBPseudo<IIC_VST1x3u>;
// ...with 4 registers
@@ -1714,7 +1718,7 @@ defm VST1d32Qwb : VST1D4WB<{1,0,?,?}, "32">;
defm VST1d64Qwb : VST1D4WB<{1,1,?,?}, "64">;
def VST1d64QPseudo : VSTQQPseudo<IIC_VST1x4>;
-def VST1d64QPseudoWB_fixed : VSTQQWBPseudo<IIC_VST1x4u>;
+def VST1d64QPseudoWB_fixed : VSTQQWBfixedPseudo<IIC_VST1x4u>;
def VST1d64QPseudoWB_register : VSTQQWBPseudo<IIC_VST1x4u>;
// VST2 : Vector Store (multiple 2-element structures)
generated by cgit v1.2.3 (git 2.25.1) at 2025-04-15 03:52:25 +0000