[Hexagon] Add support for Hexagon v67t microarchitecture (tiny core)
diff --git a/llvm/lib/Target/Hexagon/AsmParser/HexagonAsmParser.cpp b/llvm/lib/Target/Hexagon/AsmParser/HexagonAsmParser.cpp
index 108303b..2f12ea6 100644
--- a/llvm/lib/Target/Hexagon/AsmParser/HexagonAsmParser.cpp
+++ b/llvm/lib/Target/Hexagon/AsmParser/HexagonAsmParser.cpp
@@ -469,13 +469,16 @@
LLVM_DEBUG(dbgs() << "--\n");
MCB.setLoc(IDLoc);
+
// Check the bundle for errors.
const MCRegisterInfo *RI = getContext().getRegisterInfo();
- HexagonMCChecker Check(getContext(), MII, getSTI(), MCB, *RI);
+ MCSubtargetInfo const &STI = getSTI();
- bool CheckOk = HexagonMCInstrInfo::canonicalizePacket(MII, getSTI(),
- getContext(), MCB,
- &Check);
+ MCInst OrigBundle = MCB;
+ HexagonMCChecker Check(getContext(), MII, STI, MCB, *RI, true);
+
+ bool CheckOk = HexagonMCInstrInfo::canonicalizePacket(
+ MII, STI, getContext(), MCB, &Check, true);
if (CheckOk) {
if (HexagonMCInstrInfo::bundleSize(MCB) == 0) {
@@ -484,15 +487,12 @@
// Empty packets are valid yet aren't emitted
return false;
}
- Out.EmitInstruction(MCB, getSTI());
- } else {
- // If compounding and duplexing didn't reduce the size below
- // 4 or less we have a packet that is too big.
- if (HexagonMCInstrInfo::bundleSize(MCB) > HEXAGON_PACKET_SIZE) {
- Error(IDLoc, "invalid instruction packet: out of slots");
- }
+
+ assert(HexagonMCInstrInfo::isBundle(MCB));
+
+ Out.EmitInstruction(MCB, STI);
+ } else
return true; // Error
- }
return false; // No error
}
@@ -520,6 +520,8 @@
HexagonMCInstrInfo::setMemReorderDisabled(MCB);
else
return getParser().Error(IDLoc, MemNoShuffMsg);
+ } else if (Option.compare_lower("mem_no_order") == 0) {
+ // Nothing.
} else
return getParser().Error(IDLoc, llvm::Twine("'") + Option +
"' is not a valid bundle option");
diff --git a/llvm/lib/Target/Hexagon/Disassembler/HexagonDisassembler.cpp b/llvm/lib/Target/Hexagon/Disassembler/HexagonDisassembler.cpp
index 7a90d58..d71409d 100644
--- a/llvm/lib/Target/Hexagon/Disassembler/HexagonDisassembler.cpp
+++ b/llvm/lib/Target/Hexagon/Disassembler/HexagonDisassembler.cpp
@@ -185,7 +185,10 @@
return Result;
if (Size > HEXAGON_MAX_PACKET_SIZE)
return MCDisassembler::Fail;
- HexagonMCChecker Checker(getContext(), *MCII, STI, MI,
+
+ const auto ArchSTI = Hexagon_MC::getArchSubtarget(&STI);
+ const auto STI_ = (ArchSTI != nullptr) ? *ArchSTI : STI;
+ HexagonMCChecker Checker(getContext(), *MCII, STI_, MI,
*getContext().getRegisterInfo(), false);
if (!Checker.check())
return MCDisassembler::Fail;
diff --git a/llvm/lib/Target/Hexagon/Hexagon.td b/llvm/lib/Target/Hexagon/Hexagon.td
index a8f93a5..04c8b23 100644
--- a/llvm/lib/Target/Hexagon/Hexagon.td
+++ b/llvm/lib/Target/Hexagon/Hexagon.td
@@ -23,6 +23,9 @@
// Hexagon Architectures
include "HexagonDepArch.td"
+def ProcTinyCore: SubtargetFeature<"tinycore", "HexagonProcFamily",
+ "TinyCore", "Hexagon Tiny Core">;
+
// Hexagon ISA Extensions
def ExtensionZReg: SubtargetFeature<"zreg", "UseZRegOps", "true",
"Hexagon ZReg extension instructions">;
@@ -114,6 +117,11 @@
def HasMemNoShuf : Predicate<"HST->hasMemNoShuf()">,
AssemblerPredicate<"FeatureMemNoShuf">;
def UseUnsafeMath : Predicate<"HST->useUnsafeMath()">;
+def NotOptTinyCore : Predicate<"!HST->isTinyCore() ||"
+ "MF->getFunction().hasOptSize()"> {
+ let RecomputePerFunction = 1;
+}
+def UseSmallData : Predicate<"HST->useSmallData()">;
def Hvx64: HwMode<"+hvx-length64b">;
def Hvx128: HwMode<"+hvx-length128b">;
@@ -386,6 +394,14 @@
[ArchV5, ArchV55, ArchV60, ArchV62, ArchV65, ArchV66, ArchV67,
FeatureCompound, FeatureDuplex, FeatureMemNoShuf, FeatureMemops,
FeatureNVJ, FeatureNVS, FeaturePackets, FeatureSmallData]>;
+// Need to update the correct features for tiny core.
+// Disable NewValueJumps since the packetizer is unable to handle a packet with
+// a new value jump and another SLOT0 instruction.
+def : Proc<"hexagonv67t", HexagonModelV67T,
+ [ArchV5, ArchV55, ArchV60, ArchV62, ArchV65, ArchV66, ArchV67,
+ ProcTinyCore, ExtensionAudio,
+ FeatureCompound, FeatureMemNoShuf, FeatureMemops,
+ FeatureNVS, FeaturePackets, FeatureSmallData]>;
//===----------------------------------------------------------------------===//
// Declare the target which we are implementing
diff --git a/llvm/lib/Target/Hexagon/HexagonBitSimplify.cpp b/llvm/lib/Target/Hexagon/HexagonBitSimplify.cpp
index 799b85e..49edb0d 100644
--- a/llvm/lib/Target/Hexagon/HexagonBitSimplify.cpp
+++ b/llvm/lib/Target/Hexagon/HexagonBitSimplify.cpp
@@ -1433,10 +1433,16 @@
.addImm(int32_t(Lo));
return Reg;
}
+ MachineFunction *MF = B.getParent();
+ auto &HST = MF->getSubtarget<HexagonSubtarget>();
- BuildMI(B, At, DL, HII.get(Hexagon::CONST64), Reg)
- .addImm(C);
- return Reg;
+ // Disable CONST64 for tiny core since it takes a LD resource.
+ if (!HST.isTinyCore() ||
+ MF->getFunction().hasOptSize()) {
+ BuildMI(B, At, DL, HII.get(Hexagon::CONST64), Reg)
+ .addImm(C);
+ return Reg;
+ }
}
if (RC == &Hexagon::PredRegsRegClass) {
diff --git a/llvm/lib/Target/Hexagon/HexagonConstPropagation.cpp b/llvm/lib/Target/Hexagon/HexagonConstPropagation.cpp
index 5b61d10..5821e72 100644
--- a/llvm/lib/Target/Hexagon/HexagonConstPropagation.cpp
+++ b/llvm/lib/Target/Hexagon/HexagonConstPropagation.cpp
@@ -2836,6 +2836,9 @@
if (MI.isCopy())
return false;
+ MachineFunction *MF = MI.getParent()->getParent();
+ auto &HST = MF->getSubtarget<HexagonSubtarget>();
+
// Collect all virtual register-def operands.
SmallVector<unsigned,2> DefRegs;
for (const MachineOperand &MO : MI.operands()) {
@@ -2923,11 +2926,13 @@
NewMI = BuildMI(B, At, DL, *NewD, NewR)
.addImm(Hi)
.addImm(Lo);
- } else {
+ } else if (MF->getFunction().hasOptSize() || !HST.isTinyCore()) {
+ // Disable CONST64 for tiny core since it takes a LD resource.
NewD = &HII.get(Hexagon::CONST64);
NewMI = BuildMI(B, At, DL, *NewD, NewR)
.addImm(V);
- }
+ } else
+ return false;
}
}
(void)NewMI;
diff --git a/llvm/lib/Target/Hexagon/HexagonCopyToCombine.cpp b/llvm/lib/Target/Hexagon/HexagonCopyToCombine.cpp
index 394a329..ae45d8d 100644
--- a/llvm/lib/Target/Hexagon/HexagonCopyToCombine.cpp
+++ b/llvm/lib/Target/Hexagon/HexagonCopyToCombine.cpp
@@ -212,7 +212,7 @@
// There is a combine of two constant extended values into CONST64,
// provided both constants are true immediates.
if (isGreaterThanNBitTFRI<16>(HighRegInst) &&
- isGreaterThanNBitTFRI<16>(LowRegInst))
+ isGreaterThanNBitTFRI<16>(LowRegInst) && !IsConst64Disabled)
return (HighRegInst.getOperand(1).isImm() &&
LowRegInst.getOperand(1).isImm());
@@ -279,11 +279,11 @@
// A reverse_iterator instantiated like below starts before I2, and I1
// respectively.
// Look at instructions I in between I2 and (excluding) I1.
- MachineBasicBlock::reverse_iterator I(I2),
- End = --(MachineBasicBlock::reverse_iterator(I1));
+ MachineBasicBlock::reverse_iterator I = ++I2.getIterator().getReverse();
+ MachineBasicBlock::reverse_iterator End = I1.getIterator().getReverse();
// At 03 we got better results (dhrystone!) by being more conservative.
if (!ShouldCombineAggressively)
- End = MachineBasicBlock::reverse_iterator(I1);
+ End = ++I1.getIterator().getReverse();
// If I2 kills its operand and we move I2 over an instruction that also
// uses I2's use reg we need to modify that (first) instruction to now kill
// this reg.
@@ -477,6 +477,10 @@
ShouldCombineAggressively =
MF.getTarget().getOptLevel() <= CodeGenOpt::Default;
+ // Disable CONST64 for tiny core since it takes a LD resource.
+ if (!OptForSize && ST->isTinyCore())
+ IsConst64Disabled = true;
+
// Traverse basic blocks.
for (MachineFunction::iterator BI = MF.begin(), BE = MF.end(); BI != BE;
++BI) {
diff --git a/llvm/lib/Target/Hexagon/HexagonDepArch.h b/llvm/lib/Target/Hexagon/HexagonDepArch.h
index ad879e7..45b4cf0 100644
--- a/llvm/lib/Target/Hexagon/HexagonDepArch.h
+++ b/llvm/lib/Target/Hexagon/HexagonDepArch.h
@@ -26,10 +26,10 @@
static constexpr StringLiteral ArchValsTextArray[] = { "v5", "v55", "v60", "v62", "v65", "v66", "v67" };
static constexpr ArrayRef<StringLiteral> ArchValsText(ArchValsTextArray);
-static constexpr StringLiteral CpuValsTextArray[] = { "hexagonv5", "hexagonv55", "hexagonv60", "hexagonv62", "hexagonv65", "hexagonv66", "hexagonv67" };
+static constexpr StringLiteral CpuValsTextArray[] = { "hexagonv5", "hexagonv55", "hexagonv60", "hexagonv62", "hexagonv65", "hexagonv66", "hexagonv67", "hexagonv67t" };
static constexpr ArrayRef<StringLiteral> CpuValsText(CpuValsTextArray);
-static constexpr StringLiteral CpuNickTextArray[] = { "v5", "v55", "v60", "v62", "v65", "v66", "v67" };
+static constexpr StringLiteral CpuNickTextArray[] = { "v5", "v55", "v60", "v62", "v65", "v66", "v67", "v67t" };
static constexpr ArrayRef<StringLiteral> CpuNickText(CpuNickTextArray);
static const std::map<std::string, ArchEnum> CpuTable{
@@ -41,6 +41,7 @@
{"hexagonv65", Hexagon::ArchEnum::V65},
{"hexagonv66", Hexagon::ArchEnum::V66},
{"hexagonv67", Hexagon::ArchEnum::V67},
+ {"hexagonv67t", Hexagon::ArchEnum::V67},
};
} // namespace Hexagon
} // namespace llvm;
diff --git a/llvm/lib/Target/Hexagon/HexagonDepIICScalar.td b/llvm/lib/Target/Hexagon/HexagonDepIICScalar.td
index 9d9f70f..fecccb2 100644
--- a/llvm/lib/Target/Hexagon/HexagonDepIICScalar.td
+++ b/llvm/lib/Target/Hexagon/HexagonDepIICScalar.td
@@ -6253,3 +6253,831 @@
[Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>
];
}
+
+class DepScalarItinV67T {
+ list<InstrItinData> DepScalarItinV67T_list = [
+ InstrItinData <tc_011e0e9d, /*tc_st*/
+ [InstrStage<1, [SLOT0]>], [2, 1, 2, 3],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_01d44cb2, /*tc_2*/
+ [InstrStage<1, [SLOT2, SLOT3]>], [4, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_01e1be3b, /*tc_3x*/
+ [InstrStage<1, [SLOT3]>], [4, 2, 1, 1],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_02fe1c65, /*tc_4x*/
+ [InstrStage<1, [SLOT3]>], [5, 1, 1],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_0655b949, /*tc_st*/
+ [InstrStage<1, [SLOT0]>], [2, 3],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_075c8dd8, /*tc_ld*/
+ [InstrStage<1, [SLOT0]>], [4, 3, 1, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_0a195f2c, /*tc_4x*/
+ [InstrStage<1, [SLOT3]>], [5, 2, 1, 1],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_0a6c20ae, /*tc_st*/
+ [InstrStage<1, [SLOT0]>], [2, 1, 1, 2, 3],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_0ba0d5da, /*tc_3stall*/
+ [InstrStage<1, [SLOT2]>], [1],
+ [Hex_FWD]>,
+
+ InstrItinData <tc_0dfac0a7, /*tc_2*/
+ [InstrStage<1, [SLOT2, SLOT3]>], [4, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_0fac1eb8, /*tc_st*/
+ [InstrStage<1, [SLOT0]>], [3, 2, 3],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_1044324a, /*tc_3stall*/
+ [InstrStage<1, [SLOT3]>], [1, 1],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_10b884b7, /*tc_3stall*/
+ [InstrStage<1, [SLOT2]>], [],
+ []>,
+
+ InstrItinData <tc_112d30d6, /*tc_1*/
+ [InstrStage<1, [SLOT0, SLOT2, SLOT3]>], [2],
+ [Hex_FWD]>,
+
+ InstrItinData <tc_1242dc2a, /*tc_ld*/
+ [InstrStage<1, [SLOT0]>], [2],
+ [Hex_FWD]>,
+
+ InstrItinData <tc_1248597c, /*tc_3x*/
+ [InstrStage<1, [SLOT3]>], [2, 2],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_14ab4f41, /*tc_newvjump*/
+ [InstrStage<1, [SLOT0]>], [3, 3, 1],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_151bf368, /*tc_1*/
+ [InstrStage<1, [SLOT2, SLOT3]>], [3, 2],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_158aa3f7, /*tc_st*/
+ [InstrStage<1, [SLOT0]>], [1, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_197dce51, /*tc_3x*/
+ [InstrStage<1, [SLOT3]>], [4, 2, 1, 1],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_1981450d, /*tc_newvjump*/
+ [InstrStage<1, [SLOT0]>], [3],
+ [Hex_FWD]>,
+
+ InstrItinData <tc_1b8138fc, /*tc_3stall*/
+ [InstrStage<1, [SLOT0]>], [4, 1],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_1c2c7a4a, /*tc_1*/
+ [InstrStage<1, [SLOT0, SLOT2, SLOT3]>], [3, 2, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_1c7522a8, /*tc_ld*/
+ [InstrStage<1, [SLOT0]>], [4, 3, 2, 1, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_1d41f8b7, /*tc_1*/
+ [InstrStage<1, [SLOT2, SLOT3]>], [3, 4, 2, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_1e7875f0, /*tc_2early*/
+ [InstrStage<1, [SLOT2]>], [1, 1],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_1fcb8495, /*tc_2*/
+ [InstrStage<1, [SLOT3]>], [4, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_1fe4ab69, /*tc_st*/
+ [InstrStage<1, [SLOT0]>], [2, 1, 1, 2, 3],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_20131976, /*tc_2*/
+ [InstrStage<1, [SLOT2, SLOT3]>], [4, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_2237d952, /*tc_ld*/
+ [InstrStage<1, [SLOT0]>], [1, 2],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_234f8560, /*tc_ld*/
+ [InstrStage<1, [SLOT0]>], [4, 1, 1],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_23708a21, /*tc_1*/
+ [InstrStage<1, [SLOT0, SLOT2, SLOT3]>], [],
+ []>,
+
+ InstrItinData <tc_24e109c7, /*tc_newvjump*/
+ [InstrStage<1, [SLOT0]>], [3, 3, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_24f426ab, /*tc_1*/
+ [InstrStage<1, [SLOT0, SLOT2, SLOT3]>], [2, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_27106296, /*tc_3x*/
+ [InstrStage<1, [SLOT3]>], [4, 1, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_280f7fe1, /*tc_st*/
+ [InstrStage<1, [SLOT0]>], [1, 1, 2, 3],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_28e55c6f, /*tc_3x*/
+ [InstrStage<1, [SLOT3]>], [1, 1],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_2c13e7f5, /*tc_2*/
+ [InstrStage<1, [SLOT3]>], [4, 2, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_2c3e17fc, /*tc_3x*/
+ [InstrStage<1, [SLOT3]>], [1],
+ [Hex_FWD]>,
+
+ InstrItinData <tc_2f573607, /*tc_1*/
+ [InstrStage<1, [SLOT2]>], [2, 2],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_2f669c77, /*tc_3stall*/
+ [InstrStage<1, [SLOT0]>], [1, 1],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_362b0be2, /*tc_3*/
+ [InstrStage<1, [SLOT2]>], [1],
+ [Hex_FWD]>,
+
+ InstrItinData <tc_38382228, /*tc_3x*/
+ [InstrStage<1, [SLOT3]>], [4, 1, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_388f9897, /*tc_1*/
+ [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3]>], [3, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_38e0bae9, /*tc_3x*/
+ [InstrStage<1, [SLOT3]>], [4, 4, 2, 1, 1],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_3d14a17b, /*tc_1*/
+ [InstrStage<1, [SLOT0]>], [3, 2],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_3edca78f, /*tc_2*/
+ [InstrStage<1, [SLOT3]>], [4, 2],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_3fbf1042, /*tc_1*/
+ [InstrStage<1, [SLOT0]>], [3],
+ [Hex_FWD]>,
+
+ InstrItinData <tc_407e96f9, /*tc_1*/
+ [InstrStage<1, [SLOT2, SLOT3]>], [3, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_40d64c94, /*tc_newvjump*/
+ [InstrStage<1, [SLOT0]>], [3, 1],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_4222e6bf, /*tc_ld*/
+ [InstrStage<1, [SLOT0]>], [4, 1, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_42ff66ba, /*tc_1*/
+ [InstrStage<1, [SLOT2]>], [2, 2],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_442395f3, /*tc_2latepred*/
+ [InstrStage<1, [SLOT0, SLOT2, SLOT3]>], [4, 3, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_449acf79, /*tc_latepredstaia*/
+ [InstrStage<1, [SLOT0]>], [4, 3, 1, 2, 1],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_44d5a428, /*tc_st*/
+ [InstrStage<1, [SLOT0]>], [1, 2],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_44fffc58, /*tc_3*/
+ [InstrStage<1, [SLOT2, SLOT3]>], [2],
+ [Hex_FWD]>,
+
+ InstrItinData <tc_45791fb8, /*tc_ld*/
+ [InstrStage<1, [SLOT0]>], [4, 2, 1, 1, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_45f9d1be, /*tc_2early*/
+ [InstrStage<1, [SLOT2]>], [2],
+ [Hex_FWD]>,
+
+ InstrItinData <tc_49fdfd4b, /*tc_3stall*/
+ [InstrStage<1, [SLOT3]>], [4, 1],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_4a55d03c, /*tc_1*/
+ [InstrStage<1, [SLOT2, SLOT3]>], [3, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_4abdbdc6, /*tc_3x*/
+ [InstrStage<1, [SLOT3]>], [2, 2],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_4ac61d92, /*tc_2latepred*/
+ [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3]>], [4, 3, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_4c1520ae, /*tc_3x*/
+ [InstrStage<1, [SLOT3]>], [2, 1],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_503ce0f3, /*tc_3x*/
+ [InstrStage<1, [SLOT3]>], [4, 2, 2, 1],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_53c851ab, /*tc_3stall*/
+ [InstrStage<1, [SLOT2]>], [4, 1, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_5502c366, /*tc_1*/
+ [InstrStage<1, [SLOT2, SLOT3]>], [3, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_55255f2b, /*tc_3stall*/
+ [InstrStage<1, [SLOT3]>], [],
+ []>,
+
+ InstrItinData <tc_556f6577, /*tc_3x*/
+ [InstrStage<1, [SLOT3]>], [4, 1, 1],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_55a9a350, /*tc_st*/
+ [InstrStage<1, [SLOT0]>], [1, 2, 2, 3],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_55b33fda, /*tc_1*/
+ [InstrStage<1, [SLOT2, SLOT3]>], [3, 2],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_56a124a7, /*tc_1*/
+ [InstrStage<1, [SLOT0, SLOT2, SLOT3]>], [2, 2],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_57a55b54, /*tc_1*/
+ [InstrStage<1, [SLOT3]>], [2, 2],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_5944960d, /*tc_ld*/
+ [InstrStage<1, [SLOT0]>], [1, 1, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_59a7822c, /*tc_1*/
+ [InstrStage<1, [SLOT0]>], [2, 2],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_5a4b5e58, /*tc_3x*/
+ [InstrStage<1, [SLOT3]>], [4, 1, 1],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_5b347363, /*tc_1*/
+ [InstrStage<1, [SLOT0]>], [3, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_5ceb2f9e, /*tc_ld*/
+ [InstrStage<1, [SLOT0]>], [4, 3, 1, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_5d636bc7, /*tc_3stall*/
+ [InstrStage<1, [SLOT0]>], [4, 1],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_5da50c4b, /*tc_1*/
+ [InstrStage<1, [SLOT2, SLOT3]>], [3, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_5deb5e47, /*tc_st*/
+ [InstrStage<1, [SLOT0]>], [1, 2, 3],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_5e4cf0e8, /*tc_2*/
+ [InstrStage<1, [SLOT3]>], [4, 2, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_5f2afaf7, /*tc_latepredldaia*/
+ [InstrStage<1, [SLOT0]>], [4, 4, 3, 1, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_60e324ff, /*tc_1*/
+ [InstrStage<1, [SLOT2]>], [2],
+ [Hex_FWD]>,
+
+ InstrItinData <tc_63567288, /*tc_2latepred*/
+ [InstrStage<1, [SLOT0]>], [4],
+ [Hex_FWD]>,
+
+ InstrItinData <tc_64b00d8a, /*tc_ld*/
+ [InstrStage<1, [SLOT0]>], [4, 1],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_651cbe02, /*tc_1*/
+ [InstrStage<1, [SLOT2, SLOT3]>], [3, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_65279839, /*tc_2*/
+ [InstrStage<1, [SLOT2, SLOT3]>], [4, 2],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_65cbd974, /*tc_st*/
+ [InstrStage<1, [SLOT0]>], [3, 1, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_69bfb303, /*tc_3*/
+ [InstrStage<1, [SLOT2, SLOT3]>], [2, 2],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_6ae3426b, /*tc_3x*/
+ [InstrStage<1, [SLOT3]>], [4, 1],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_6d861a95, /*tc_3x*/
+ [InstrStage<1, [SLOT3]>], [2, 1],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_6e20402a, /*tc_st*/
+ [InstrStage<1, [SLOT0]>], [2, 3],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_6f42bc60, /*tc_3stall*/
+ [InstrStage<1, [SLOT0]>], [4, 1, 1],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_6fb32599, /*tc_3stall*/
+ [InstrStage<1, [SLOT3]>], [1],
+ [Hex_FWD]>,
+
+ InstrItinData <tc_6fc5dbea, /*tc_1*/
+ [InstrStage<1, [SLOT2, SLOT3]>], [3, 2, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_711c805f, /*tc_1*/
+ [InstrStage<1, [SLOT0, SLOT2, SLOT3]>], [2, 2],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_713b66bf, /*tc_1*/
+ [InstrStage<1, [SLOT0, SLOT2, SLOT3]>], [3, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_7401744f, /*tc_2*/
+ [InstrStage<1, [SLOT3]>], [4, 4, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_7476d766, /*tc_3stall*/
+ [InstrStage<1, [SLOT3]>], [4, 2],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_74a42bda, /*tc_ld*/
+ [InstrStage<1, [SLOT0, SLOT1]>], [3, 1, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_76bb5435, /*tc_ld*/
+ [InstrStage<1, [SLOT0]>], [4, 3, 2, 1, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_77f94a5e, /*tc_st*/
+ [InstrStage<1, [SLOT0]>], [],
+ []>,
+
+ InstrItinData <tc_788b1d09, /*tc_3x*/
+ [InstrStage<1, [SLOT3]>], [4, 1, 1, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_7b9187d3, /*tc_newvjump*/
+ [InstrStage<1, [SLOT0]>], [3, 2],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_7c31e19a, /*tc_st*/
+ [InstrStage<1, [SLOT0]>], [1, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_7c6d32e4, /*tc_ld*/
+ [InstrStage<1, [SLOT0]>], [4, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_7dc63b5c, /*tc_3x*/
+ [InstrStage<1, [SLOT3]>], [4, 1],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_7dcd9d89, /*tc_st*/
+ [InstrStage<1, [SLOT0]>], [1, 3],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_7f7f45f5, /*tc_4x*/
+ [InstrStage<1, [SLOT3]>], [5, 5, 1],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_7f8ae742, /*tc_3x*/
+ [InstrStage<1, [SLOT3]>], [4, 2, 1, 1],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_8035e91f, /*tc_st*/
+ [InstrStage<1, [SLOT0]>], [2, 1, 2, 3],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_822c3c68, /*tc_ld*/
+ [InstrStage<1, [SLOT0]>], [4, 3, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_829d8a86, /*tc_st*/
+ [InstrStage<1, [SLOT0]>], [3, 1, 1, 2, 3],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_838c4d7a, /*tc_st*/
+ [InstrStage<1, [SLOT0, SLOT1]>], [1, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_84a7500d, /*tc_2*/
+ [InstrStage<1, [SLOT0, SLOT2, SLOT3]>], [4, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_86173609, /*tc_2latepred*/
+ [InstrStage<1, [SLOT0, SLOT2, SLOT3]>], [4, 3, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_887d1bb7, /*tc_st*/
+ [InstrStage<1, [SLOT0]>], [1, 2, 2, 3],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_8a6d0d94, /*tc_ld*/
+ [InstrStage<1, [SLOT0]>], [4, 2],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_8a825db2, /*tc_2*/
+ [InstrStage<1, [SLOT2, SLOT3]>], [4, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_8b5bd4f5, /*tc_2*/
+ [InstrStage<1, [SLOT0, SLOT2, SLOT3]>], [4, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_8e82e8ca, /*tc_st*/
+ [InstrStage<1, [SLOT0]>], [3, 1, 1, 2, 3],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_9124c04f, /*tc_1*/
+ [InstrStage<1, [SLOT0, SLOT1, SLOT2, SLOT3]>], [3, 2],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_9165014d, /*tc_3stall*/
+ [InstrStage<1, [SLOT2]>], [4, 1],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_92240447, /*tc_st*/
+ [InstrStage<1, [SLOT0]>], [3, 1, 2, 3],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_934753bb, /*tc_ld*/
+ [InstrStage<1, [SLOT0]>], [3, 1, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_937dd41c, /*tc_ld*/
+ [InstrStage<1, [SLOT0]>], [],
+ []>,
+
+ InstrItinData <tc_9406230a, /*tc_3x*/
+ [InstrStage<1, [SLOT3]>], [2, 1],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_95a33176, /*tc_2*/
+ [InstrStage<1, [SLOT0, SLOT2, SLOT3]>], [4, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_96ef76ef, /*tc_st*/
+ [InstrStage<1, [SLOT0]>], [1, 1, 2, 3],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_975a4e54, /*tc_newvjump*/
+ [InstrStage<1, [SLOT0]>], [3, 3, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_9783714b, /*tc_4x*/
+ [InstrStage<1, [SLOT3]>], [5, 1],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_988416e3, /*tc_st*/
+ [InstrStage<1, [SLOT0]>], [3],
+ [Hex_FWD]>,
+
+ InstrItinData <tc_9b34f5e0, /*tc_3stall*/
+ [InstrStage<1, [SLOT2]>], [],
+ []>,
+
+ InstrItinData <tc_9b3c0462, /*tc_2*/
+ [InstrStage<1, [SLOT2, SLOT3]>], [4, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_9bcfb2ee, /*tc_st*/
+ [InstrStage<1, [SLOT0]>], [1, 2, 3],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_9c52f549, /*tc_1*/
+ [InstrStage<1, [SLOT0, SLOT2, SLOT3]>], [3, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_9e27f2f9, /*tc_1*/
+ [InstrStage<1, [SLOT0, SLOT2, SLOT3]>], [2, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_9e72dc89, /*tc_4x*/
+ [InstrStage<1, [SLOT3]>], [5, 2, 1, 1],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_9edb7c77, /*tc_4x*/
+ [InstrStage<1, [SLOT3]>], [5, 2, 1, 1, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_9edefe01, /*tc_st*/
+ [InstrStage<1, [SLOT0]>], [3, 2, 1, 2, 3],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_9f6cd987, /*tc_1*/
+ [InstrStage<1, [SLOT2, SLOT3]>], [3, 2],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_a08b630b, /*tc_2*/
+ [InstrStage<1, [SLOT3]>], [4, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_a1297125, /*tc_1*/
+ [InstrStage<1, [SLOT2, SLOT3]>], [3, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_a154b476, /*tc_3x*/
+ [InstrStage<1, [SLOT3]>], [4, 2, 1, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_a2b365d2, /*tc_st*/
+ [InstrStage<1, [SLOT0]>], [3, 1, 2, 3],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_a3070909, /*tc_3stall*/
+ [InstrStage<1, [SLOT0]>], [1, 1],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_a32e03e7, /*tc_ld*/
+ [InstrStage<1, [SLOT0]>], [4, 2, 1, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_a38c45dc, /*tc_3x*/
+ [InstrStage<1, [SLOT3]>], [4, 2, 1, 1, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_a4e22bbd, /*tc_2*/
+ [InstrStage<1, [SLOT3]>], [4, 2, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_a4ee89db, /*tc_2early*/
+ [InstrStage<1, [SLOT0]>], [],
+ []>,
+
+ InstrItinData <tc_a7a13fac, /*tc_1*/
+ [InstrStage<1, [SLOT2, SLOT3]>], [3, 2, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_a7bdb22c, /*tc_2*/
+ [InstrStage<1, [SLOT3]>], [4, 2],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_a9edeffa, /*tc_st*/
+ [InstrStage<1, [SLOT0, SLOT1]>], [1, 2, 3],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_abfd9a6d, /*tc_ld*/
+ [InstrStage<1, [SLOT0]>], [4, 1, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_ac65613f, /*tc_ld*/
+ [InstrStage<1, [SLOT0]>], [4, 3, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_addc37a8, /*tc_st*/
+ [InstrStage<1, [SLOT0]>], [3, 1, 2, 2, 3],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_ae5babd7, /*tc_st*/
+ [InstrStage<1, [SLOT0]>], [1, 2, 3],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_aee6250c, /*tc_ld*/
+ [InstrStage<1, [SLOT0, SLOT1]>], [4, 1],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_b1ae5f67, /*tc_st*/
+ [InstrStage<1, [SLOT0]>], [1],
+ [Hex_FWD]>,
+
+ InstrItinData <tc_b34eb232, /*tc_3stall*/
+ [InstrStage<1, [SLOT0]>], [],
+ []>,
+
+ InstrItinData <tc_b4dc7630, /*tc_st*/
+ [InstrStage<1, [SLOT0]>], [3, 1, 2, 2, 3],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_b570493d, /*tc_3stall*/
+ [InstrStage<1, [SLOT3]>], [4, 1],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_b7c4062a, /*tc_ld*/
+ [InstrStage<1, [SLOT0]>], [4, 3, 1, 1, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_b837298f, /*tc_1*/
+ [InstrStage<1, [SLOT0, SLOT2, SLOT3]>], [],
+ []>,
+
+ InstrItinData <tc_ba9255a6, /*tc_st*/
+ [InstrStage<1, [SLOT0]>], [2, 2, 3],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_bb07f2c5, /*tc_st*/
+ [InstrStage<1, [SLOT0]>], [3, 2, 3],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_bb831a7c, /*tc_2*/
+ [InstrStage<1, [SLOT3]>], [4, 2, 2, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_bf2ffc0f, /*tc_ld*/
+ [InstrStage<1, [SLOT0]>], [4, 1, 1, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_c20701f0, /*tc_2*/
+ [InstrStage<1, [SLOT2, SLOT3]>], [4, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_c21d7447, /*tc_3x*/
+ [InstrStage<1, [SLOT3]>], [4, 1, 1],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_c57d9f39, /*tc_1*/
+ [InstrStage<1, [SLOT0, SLOT2, SLOT3]>], [3, 2],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_c818ff7f, /*tc_newvjump*/
+ [InstrStage<1, [SLOT0]>], [],
+ []>,
+
+ InstrItinData <tc_ce59038e, /*tc_st*/
+ [InstrStage<1, [SLOT0]>], [3, 2, 1, 2, 3],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_cfa0e29b, /*tc_st*/
+ [InstrStage<1, [SLOT0]>], [2, 2, 3],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_d03278fd, /*tc_st*/
+ [InstrStage<1, [SLOT0]>], [2, 1, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_d33e5eee, /*tc_1*/
+ [InstrStage<1, [SLOT0, SLOT2, SLOT3]>], [3, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_d3632d88, /*tc_2*/
+ [InstrStage<1, [SLOT3]>], [4, 2],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_d45ba9cd, /*tc_ld*/
+ [InstrStage<1, [SLOT0]>], [1],
+ [Hex_FWD]>,
+
+ InstrItinData <tc_d47648a2, /*tc_3stall*/
+ [InstrStage<1, [SLOT2]>], [1, 1],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_d57d649c, /*tc_3stall*/
+ [InstrStage<1, [SLOT2]>], [2],
+ [Hex_FWD]>,
+
+ InstrItinData <tc_d61dfdc3, /*tc_2*/
+ [InstrStage<1, [SLOT2, SLOT3]>], [4, 2],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_d68dca5c, /*tc_3stall*/
+ [InstrStage<1, [SLOT3]>], [4, 1, 1],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_d7718fbe, /*tc_3x*/
+ [InstrStage<1, [SLOT3]>], [1],
+ [Hex_FWD]>,
+
+ InstrItinData <tc_db596beb, /*tc_3x*/
+ [InstrStage<1, [SLOT3]>], [4, 1, 1],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_db96aa6b, /*tc_st*/
+ [InstrStage<1, [SLOT0]>], [1],
+ [Hex_FWD]>,
+
+ InstrItinData <tc_dc51281d, /*tc_3*/
+ [InstrStage<1, [SLOT2]>], [2, 1],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_decdde8a, /*tc_1*/
+ [InstrStage<1, [SLOT0, SLOT2, SLOT3]>], [2],
+ [Hex_FWD]>,
+
+ InstrItinData <tc_df4536ae, /*tc_3stall*/
+ [InstrStage<1, [SLOT3]>], [4, 1, 1],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_df5d53f9, /*tc_newvjump*/
+ [InstrStage<1, [SLOT0]>], [3, 2, 1],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_e3d699e3, /*tc_2*/
+ [InstrStage<1, [SLOT2, SLOT3]>], [4, 2],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_e9170fb7, /*tc_ld*/
+ [InstrStage<1, [SLOT0]>], [4, 1],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_ed03645c, /*tc_1*/
+ [InstrStage<1, [SLOT2]>], [3, 2],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_eed07714, /*tc_ld*/
+ [InstrStage<1, [SLOT0, SLOT1]>], [4, 1, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_eeda4109, /*tc_1*/
+ [InstrStage<1, [SLOT0, SLOT2, SLOT3]>], [3, 2],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_ef921005, /*tc_1*/
+ [InstrStage<1, [SLOT2, SLOT3]>], [3, 2],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_f098b237, /*tc_2*/
+ [InstrStage<1, [SLOT2, SLOT3]>], [4, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_f0cdeccf, /*tc_3x*/
+ [InstrStage<1, [SLOT3]>], [4, 1, 1, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_f0e8e832, /*tc_4x*/
+ [InstrStage<1, [SLOT3]>], [5, 1, 1],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_f34c1c21, /*tc_2*/
+ [InstrStage<1, [SLOT2, SLOT3]>], [4, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_f38f92e1, /*tc_newvjump*/
+ [InstrStage<1, [SLOT0]>], [2],
+ [Hex_FWD]>,
+
+ InstrItinData <tc_f529831b, /*tc_latepredstaia*/
+ [InstrStage<1, [SLOT0]>], [4, 3, 1, 2, 3],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_f6e2aff9, /*tc_newvjump*/
+ [InstrStage<1, [SLOT0]>], [3, 2, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_f7569068, /*tc_4x*/
+ [InstrStage<1, [SLOT3]>], [5, 5, 1, 1],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_f999c66e, /*tc_1*/
+ [InstrStage<1, [SLOT0, SLOT2, SLOT3]>], [2, 2],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_fae9dfa5, /*tc_3x*/
+ [InstrStage<1, [SLOT3]>], [4, 2],
+ [Hex_FWD, Hex_FWD]>,
+
+ InstrItinData <tc_fedb7e19, /*tc_ld*/
+ [InstrStage<1, [SLOT0]>], [4, 2, 1, 2],
+ [Hex_FWD, Hex_FWD, Hex_FWD, Hex_FWD]>
+ ];
+}
diff --git a/llvm/lib/Target/Hexagon/HexagonInstrInfo.cpp b/llvm/lib/Target/Hexagon/HexagonInstrInfo.cpp
index 9c424ba..1fa3285 100644
--- a/llvm/lib/Target/Hexagon/HexagonInstrInfo.cpp
+++ b/llvm/lib/Target/Hexagon/HexagonInstrInfo.cpp
@@ -118,6 +118,12 @@
: HexagonGenInstrInfo(Hexagon::ADJCALLSTACKDOWN, Hexagon::ADJCALLSTACKUP),
Subtarget(ST) {}
+namespace llvm {
+namespace HexagonFUnits {
+ bool isSlot0Only(unsigned units);
+}
+}
+
static bool isIntRegForSubInst(unsigned Reg) {
return (Reg >= Hexagon::R0 && Reg <= Hexagon::R7) ||
(Reg >= Hexagon::R16 && Reg <= Hexagon::R23);
@@ -3403,6 +3409,64 @@
: Hexagon::J4_cmpeqi_tp1_jump_nt;
}
+// Returns -1 if there is no opcode found.
+int HexagonInstrInfo::getDuplexOpcode(const MachineInstr &MI,
+ bool ForBigCore) const {
+ // Static table to switch the opcodes across Tiny Core and Big Core.
+ // dup_ opcodes are Big core opcodes.
+ // NOTE: There are special instructions that need to handled later.
+ // L4_return* instructions, they will only occupy SLOT0 (on big core too).
+ // PS_jmpret - This pseudo translates to J2_jumpr which occupies only SLOT2.
+ // The compiler need to base the root instruction to L6_return_map_to_raw
+ // which can go any slot.
+ static const std::map<unsigned, unsigned> DupMap = {
+ {Hexagon::A2_add, Hexagon::dup_A2_add},
+ {Hexagon::A2_addi, Hexagon::dup_A2_addi},
+ {Hexagon::A2_andir, Hexagon::dup_A2_andir},
+ {Hexagon::A2_combineii, Hexagon::dup_A2_combineii},
+ {Hexagon::A2_sxtb, Hexagon::dup_A2_sxtb},
+ {Hexagon::A2_sxth, Hexagon::dup_A2_sxth},
+ {Hexagon::A2_tfr, Hexagon::dup_A2_tfr},
+ {Hexagon::A2_tfrsi, Hexagon::dup_A2_tfrsi},
+ {Hexagon::A2_zxtb, Hexagon::dup_A2_zxtb},
+ {Hexagon::A2_zxth, Hexagon::dup_A2_zxth},
+ {Hexagon::A4_combineii, Hexagon::dup_A4_combineii},
+ {Hexagon::A4_combineir, Hexagon::dup_A4_combineir},
+ {Hexagon::A4_combineri, Hexagon::dup_A4_combineri},
+ {Hexagon::C2_cmoveif, Hexagon::dup_C2_cmoveif},
+ {Hexagon::C2_cmoveit, Hexagon::dup_C2_cmoveit},
+ {Hexagon::C2_cmovenewif, Hexagon::dup_C2_cmovenewif},
+ {Hexagon::C2_cmovenewit, Hexagon::dup_C2_cmovenewit},
+ {Hexagon::C2_cmpeqi, Hexagon::dup_C2_cmpeqi},
+ {Hexagon::L2_deallocframe, Hexagon::dup_L2_deallocframe},
+ {Hexagon::L2_loadrb_io, Hexagon::dup_L2_loadrb_io},
+ {Hexagon::L2_loadrd_io, Hexagon::dup_L2_loadrd_io},
+ {Hexagon::L2_loadrh_io, Hexagon::dup_L2_loadrh_io},
+ {Hexagon::L2_loadri_io, Hexagon::dup_L2_loadri_io},
+ {Hexagon::L2_loadrub_io, Hexagon::dup_L2_loadrub_io},
+ {Hexagon::L2_loadruh_io, Hexagon::dup_L2_loadruh_io},
+ {Hexagon::S2_allocframe, Hexagon::dup_S2_allocframe},
+ {Hexagon::S2_storerb_io, Hexagon::dup_S2_storerb_io},
+ {Hexagon::S2_storerd_io, Hexagon::dup_S2_storerd_io},
+ {Hexagon::S2_storerh_io, Hexagon::dup_S2_storerh_io},
+ {Hexagon::S2_storeri_io, Hexagon::dup_S2_storeri_io},
+ {Hexagon::S4_storeirb_io, Hexagon::dup_S4_storeirb_io},
+ {Hexagon::S4_storeiri_io, Hexagon::dup_S4_storeiri_io},
+ };
+ unsigned OpNum = MI.getOpcode();
+ // Conversion to Big core.
+ if (ForBigCore) {
+ auto Iter = DupMap.find(OpNum);
+ if (Iter != DupMap.end())
+ return Iter->second;
+ } else { // Conversion to Tiny core.
+ for (auto Iter = DupMap.begin(), End = DupMap.end(); Iter != End; ++Iter)
+ if (Iter->second == OpNum)
+ return Iter->first;
+ }
+ return -1;
+}
+
int HexagonInstrInfo::getCondOpcode(int Opc, bool invertPredicate) const {
enum Hexagon::PredSense inPredSense;
inPredSense = invertPredicate ? Hexagon::PredSense_false :
@@ -3735,6 +3799,7 @@
// Rd = memw(Rs+#u4:2)
// Rd = memub(Rs+#u4:0)
case Hexagon::L2_loadri_io:
+ case Hexagon::dup_L2_loadri_io:
DstReg = MI.getOperand(0).getReg();
SrcReg = MI.getOperand(1).getReg();
// Special case this one from Group L2.
@@ -3753,6 +3818,7 @@
}
break;
case Hexagon::L2_loadrub_io:
+ case Hexagon::dup_L2_loadrub_io:
// Rd = memub(Rs+#u4:0)
DstReg = MI.getOperand(0).getReg();
SrcReg = MI.getOperand(1).getReg();
@@ -3772,6 +3838,8 @@
// [if ([!]p0[.new])] jumpr r31
case Hexagon::L2_loadrh_io:
case Hexagon::L2_loadruh_io:
+ case Hexagon::dup_L2_loadrh_io:
+ case Hexagon::dup_L2_loadruh_io:
// Rd = memh/memuh(Rs+#u3:1)
DstReg = MI.getOperand(0).getReg();
SrcReg = MI.getOperand(1).getReg();
@@ -3781,6 +3849,7 @@
return HexagonII::HSIG_L2;
break;
case Hexagon::L2_loadrb_io:
+ case Hexagon::dup_L2_loadrb_io:
// Rd = memb(Rs+#u3:0)
DstReg = MI.getOperand(0).getReg();
SrcReg = MI.getOperand(1).getReg();
@@ -3790,6 +3859,7 @@
return HexagonII::HSIG_L2;
break;
case Hexagon::L2_loadrd_io:
+ case Hexagon::dup_L2_loadrd_io:
// Rdd = memd(r29+#u5:3)
DstReg = MI.getOperand(0).getReg();
SrcReg = MI.getOperand(1).getReg();
@@ -3806,6 +3876,7 @@
case Hexagon::RESTORE_DEALLOC_RET_JMP_V4_PIC:
case Hexagon::L4_return:
case Hexagon::L2_deallocframe:
+ case Hexagon::dup_L2_deallocframe:
return HexagonII::HSIG_L2;
case Hexagon::EH_RETURN_JMPR:
case Hexagon::PS_jmpret:
@@ -3825,6 +3896,7 @@
case Hexagon::SL2_jumpr31_t:
case Hexagon::SL2_jumpr31_f:
case Hexagon::SL2_jumpr31_tnew:
+ case Hexagon::SL2_jumpr31_fnew:
DstReg = MI.getOperand(1).getReg();
SrcReg = MI.getOperand(0).getReg();
// [if ([!]p0[.new])] jumpr r31
@@ -3850,6 +3922,7 @@
// memw(Rs+#u4:2) = Rt
// memb(Rs+#u4:0) = Rt
case Hexagon::S2_storeri_io:
+ case Hexagon::dup_S2_storeri_io:
// Special case this one from Group S2.
// memw(r29+#u5:2) = Rt
Src1Reg = MI.getOperand(0).getReg();
@@ -3866,6 +3939,7 @@
return HexagonII::HSIG_S1;
break;
case Hexagon::S2_storerb_io:
+ case Hexagon::dup_S2_storerb_io:
// memb(Rs+#u4:0) = Rt
Src1Reg = MI.getOperand(0).getReg();
Src2Reg = MI.getOperand(2).getReg();
@@ -3883,6 +3957,7 @@
// memb(Rs+#u4) = #U1
// allocframe(#u5:3)
case Hexagon::S2_storerh_io:
+ case Hexagon::dup_S2_storerh_io:
// memh(Rs+#u3:1) = Rt
Src1Reg = MI.getOperand(0).getReg();
Src2Reg = MI.getOperand(2).getReg();
@@ -3892,6 +3967,7 @@
return HexagonII::HSIG_S1;
break;
case Hexagon::S2_storerd_io:
+ case Hexagon::dup_S2_storerd_io:
// memd(r29+#s6:3) = Rtt
Src1Reg = MI.getOperand(0).getReg();
Src2Reg = MI.getOperand(2).getReg();
@@ -3902,6 +3978,7 @@
return HexagonII::HSIG_S2;
break;
case Hexagon::S4_storeiri_io:
+ case Hexagon::dup_S4_storeiri_io:
// memw(Rs+#u4:2) = #U1
Src1Reg = MI.getOperand(0).getReg();
if (isIntRegForSubInst(Src1Reg) && MI.getOperand(1).isImm() &&
@@ -3910,6 +3987,7 @@
return HexagonII::HSIG_S2;
break;
case Hexagon::S4_storeirb_io:
+ case Hexagon::dup_S4_storeirb_io:
// memb(Rs+#u4) = #U1
Src1Reg = MI.getOperand(0).getReg();
if (isIntRegForSubInst(Src1Reg) &&
@@ -3918,6 +3996,7 @@
return HexagonII::HSIG_S2;
break;
case Hexagon::S2_allocframe:
+ case Hexagon::dup_S2_allocframe:
if (MI.getOperand(2).isImm() &&
isShiftedUInt<5,3>(MI.getOperand(2).getImm()))
return HexagonII::HSIG_S1;
@@ -3941,6 +4020,7 @@
// Rd = sxth/sxtb/zxtb/zxth(Rs)
// Rd = and(Rs,#1)
case Hexagon::A2_addi:
+ case Hexagon::dup_A2_addi:
DstReg = MI.getOperand(0).getReg();
SrcReg = MI.getOperand(1).getReg();
if (isIntRegForSubInst(DstReg)) {
@@ -3962,6 +4042,7 @@
}
break;
case Hexagon::A2_add:
+ case Hexagon::dup_A2_add:
// Rx = add(Rx,Rs)
DstReg = MI.getOperand(0).getReg();
Src1Reg = MI.getOperand(1).getReg();
@@ -3971,6 +4052,7 @@
return HexagonII::HSIG_A;
break;
case Hexagon::A2_andir:
+ case Hexagon::dup_A2_andir:
// Same as zxtb.
// Rd16=and(Rs16,#255)
// Rd16=and(Rs16,#1)
@@ -3983,6 +4065,7 @@
return HexagonII::HSIG_A;
break;
case Hexagon::A2_tfr:
+ case Hexagon::dup_A2_tfr:
// Rd = Rs
DstReg = MI.getOperand(0).getReg();
SrcReg = MI.getOperand(1).getReg();
@@ -3990,6 +4073,7 @@
return HexagonII::HSIG_A;
break;
case Hexagon::A2_tfrsi:
+ case Hexagon::dup_A2_tfrsi:
// Rd = #u6
// Do not test for #u6 size since the const is getting extended
// regardless and compound could be formed.
@@ -4002,6 +4086,10 @@
case Hexagon::C2_cmovenewit:
case Hexagon::C2_cmoveif:
case Hexagon::C2_cmovenewif:
+ case Hexagon::dup_C2_cmoveit:
+ case Hexagon::dup_C2_cmovenewit:
+ case Hexagon::dup_C2_cmoveif:
+ case Hexagon::dup_C2_cmovenewif:
// if ([!]P0[.new]) Rd = #0
// Actual form:
// %r16 = C2_cmovenewit internal %p0, 0, implicit undef %r16;
@@ -4013,6 +4101,7 @@
return HexagonII::HSIG_A;
break;
case Hexagon::C2_cmpeqi:
+ case Hexagon::dup_C2_cmpeqi:
// P0 = cmp.eq(Rs,#u2)
DstReg = MI.getOperand(0).getReg();
SrcReg = MI.getOperand(1).getReg();
@@ -4023,6 +4112,8 @@
break;
case Hexagon::A2_combineii:
case Hexagon::A4_combineii:
+ case Hexagon::dup_A2_combineii:
+ case Hexagon::dup_A4_combineii:
// Rdd = combine(#u2,#U2)
DstReg = MI.getOperand(0).getReg();
if (isDblRegForSubInst(DstReg, HRI) &&
@@ -4035,6 +4126,8 @@
return HexagonII::HSIG_A;
break;
case Hexagon::A4_combineri:
+ case Hexagon::dup_A4_combineri:
+ // Rdd = combine(Rs,#0)
// Rdd = combine(Rs,#0)
DstReg = MI.getOperand(0).getReg();
SrcReg = MI.getOperand(1).getReg();
@@ -4044,6 +4137,7 @@
return HexagonII::HSIG_A;
break;
case Hexagon::A4_combineir:
+ case Hexagon::dup_A4_combineir:
// Rdd = combine(#0,Rs)
DstReg = MI.getOperand(0).getReg();
SrcReg = MI.getOperand(2).getReg();
@@ -4056,6 +4150,10 @@
case Hexagon::A2_sxth:
case Hexagon::A2_zxtb:
case Hexagon::A2_zxth:
+ case Hexagon::dup_A2_sxtb:
+ case Hexagon::dup_A2_sxth:
+ case Hexagon::dup_A2_zxtb:
+ case Hexagon::dup_A2_zxth:
// Rd = sxth/sxtb/zxtb/zxth(Rs)
DstReg = MI.getOperand(0).getReg();
SrcReg = MI.getOperand(1).getReg();
@@ -4199,6 +4297,61 @@
addrMode == HexagonII::BaseLongOffset);
}
+bool HexagonInstrInfo::isPureSlot0(const MachineInstr &MI) const {
+ // Workaround for the Global Scheduler. Sometimes, it creates
+ // A4_ext as a Pseudo instruction and calls this function to see if
+ // it can be added to an existing bundle. Since the instruction doesn't
+ // belong to any BB yet, we can't use getUnits API.
+ if (MI.getOpcode() == Hexagon::A4_ext)
+ return false;
+
+ unsigned FuncUnits = getUnits(MI);
+ return HexagonFUnits::isSlot0Only(FuncUnits);
+}
+
+bool HexagonInstrInfo::isRestrictNoSlot1Store(const MachineInstr &MI) const {
+ const uint64_t F = MI.getDesc().TSFlags;
+ return ((F >> HexagonII::RestrictNoSlot1StorePos) &
+ HexagonII::RestrictNoSlot1StoreMask);
+}
+
+void HexagonInstrInfo::changeDuplexOpcode(MachineBasicBlock::instr_iterator MII,
+ bool ToBigInstrs) const {
+ int Opcode = -1;
+ if (ToBigInstrs) { // To BigCore Instr.
+ // Check if the instruction can form a Duplex.
+ if (getDuplexCandidateGroup(*MII))
+ // Get the opcode marked "dup_*" tag.
+ Opcode = getDuplexOpcode(*MII, ToBigInstrs);
+ } else // To TinyCore Instr.
+ Opcode = getDuplexOpcode(*MII, ToBigInstrs);
+
+ // Change the opcode of the instruction.
+ if (Opcode >= 0)
+ MII->setDesc(get(Opcode));
+}
+
+// This function is used to translate instructions to facilitate generating
+// Duplexes on TinyCore.
+void HexagonInstrInfo::translateInstrsForDup(MachineFunction &MF,
+ bool ToBigInstrs) const {
+ for (auto &MB : MF)
+ for (MachineBasicBlock::instr_iterator Instr = MB.instr_begin(),
+ End = MB.instr_end();
+ Instr != End; ++Instr)
+ changeDuplexOpcode(Instr, ToBigInstrs);
+}
+
+// This is a specialized form of above function.
+void HexagonInstrInfo::translateInstrsForDup(
+ MachineBasicBlock::instr_iterator MII, bool ToBigInstrs) const {
+ MachineBasicBlock *MBB = MII->getParent();
+ while ((MII != MBB->instr_end()) && MII->isInsideBundle()) {
+ changeDuplexOpcode(MII, ToBigInstrs);
+ ++MII;
+ }
+}
+
unsigned HexagonInstrInfo::getMemAccessSize(const MachineInstr &MI) const {
using namespace HexagonII;
diff --git a/llvm/lib/Target/Hexagon/HexagonInstrInfo.h b/llvm/lib/Target/Hexagon/HexagonInstrInfo.h
index f54cf6a..e529d86 100644
--- a/llvm/lib/Target/Hexagon/HexagonInstrInfo.h
+++ b/llvm/lib/Target/Hexagon/HexagonInstrInfo.h
@@ -341,10 +341,10 @@
MachineBasicBlock *TargetBB,
SmallPtrSet<MachineBasicBlock *, 8> &Visited) const;
- bool isBaseImmOffset(const MachineInstr &MI) const;
bool isAbsoluteSet(const MachineInstr &MI) const;
bool isAccumulator(const MachineInstr &MI) const;
bool isAddrModeWithOffset(const MachineInstr &MI) const;
+ bool isBaseImmOffset(const MachineInstr &MI) const;
bool isComplex(const MachineInstr &MI) const;
bool isCompoundBranchInstr(const MachineInstr &MI) const;
bool isConstExtended(const MachineInstr &MI) const;
@@ -387,6 +387,8 @@
bool isPredicated(unsigned Opcode) const;
bool isPredicateLate(unsigned Opcode) const;
bool isPredictedTaken(unsigned Opcode) const;
+ bool isPureSlot0(const MachineInstr &MI) const;
+ bool isRestrictNoSlot1Store(const MachineInstr &MI) const;
bool isSaveCalleeSavedRegsCall(const MachineInstr &MI) const;
bool isSignExtendingLoad(const MachineInstr &MI) const;
bool isSolo(const MachineInstr &MI) const;
@@ -435,6 +437,7 @@
getCompoundCandidateGroup(const MachineInstr &MI) const;
unsigned getCompoundOpcode(const MachineInstr &GA,
const MachineInstr &GB) const;
+ int getDuplexOpcode(const MachineInstr &MI, bool ForBigCore = true) const;
int getCondOpcode(int Opc, bool sense) const;
int getDotCurOp(const MachineInstr &MI) const;
int getNonDotCurOp(const MachineInstr &MI) const;
@@ -480,6 +483,17 @@
void setBundleNoShuf(MachineBasicBlock::instr_iterator MIB) const;
bool getBundleNoShuf(const MachineInstr &MIB) const;
+
+ // When TinyCore with Duplexes is enabled, this function is used to translate
+ // tiny-instructions to big-instructions and vice versa to get the slot
+ // consumption.
+ void changeDuplexOpcode(MachineBasicBlock::instr_iterator MII,
+ bool ToBigInstrs) const;
+ void translateInstrsForDup(MachineFunction &MF,
+ bool ToBigInstrs = true) const;
+ void translateInstrsForDup(MachineBasicBlock::instr_iterator MII,
+ bool ToBigInstrs) const;
+
// Addressing mode relations.
short changeAddrMode_abs_io(short Opc) const;
short changeAddrMode_io_abs(short Opc) const;
diff --git a/llvm/lib/Target/Hexagon/HexagonPatterns.td b/llvm/lib/Target/Hexagon/HexagonPatterns.td
index 6bbde51..2bfd8d7 100644
--- a/llvm/lib/Target/Hexagon/HexagonPatterns.td
+++ b/llvm/lib/Target/Hexagon/HexagonPatterns.td
@@ -362,6 +362,16 @@
// --(1) Immediate -------------------------------------------------------
//
+def Imm64Lo: SDNodeXForm<imm, [{
+ return CurDAG->getTargetConstant(int32_t (N->getSExtValue()),
+ SDLoc(N), MVT::i32);
+}]>;
+def Imm64Hi: SDNodeXForm<imm, [{
+ return CurDAG->getTargetConstant(int32_t (N->getSExtValue()>>32),
+ SDLoc(N), MVT::i32);
+}]>;
+
+
def SDTHexagonCONST32
: SDTypeProfile<1, 1, [SDTCisVT<0, i32>, SDTCisVT<1, i32>, SDTCisPtrTy<0>]>;
@@ -389,7 +399,10 @@
def: Pat<(i1 0), (PS_false)>;
def: Pat<(i1 1), (PS_true)>;
-def: Pat<(i64 imm:$v), (CONST64 imm:$v)>;
+def: Pat<(i64 imm:$v), (CONST64 imm:$v)>,
+ Requires<[UseSmallData,NotOptTinyCore]>;
+def: Pat<(i64 imm:$v),
+ (Combinew (A2_tfrsi (Imm64Hi $v)), (A2_tfrsi (Imm64Lo $v)))>;
def ftoi : SDNodeXForm<fpimm, [{
APInt I = N->getValueAPF().bitcastToAPInt();
diff --git a/llvm/lib/Target/Hexagon/HexagonSchedule.td b/llvm/lib/Target/Hexagon/HexagonSchedule.td
index 4060dd3..5efd02a 100644
--- a/llvm/lib/Target/Hexagon/HexagonSchedule.td
+++ b/llvm/lib/Target/Hexagon/HexagonSchedule.td
@@ -67,3 +67,4 @@
include "HexagonScheduleV65.td"
include "HexagonScheduleV66.td"
include "HexagonScheduleV67.td"
+include "HexagonScheduleV67T.td"
diff --git a/llvm/lib/Target/Hexagon/HexagonScheduleV67T.td b/llvm/lib/Target/Hexagon/HexagonScheduleV67T.td
new file mode 100644
index 0000000..f2bcb1e
--- /dev/null
+++ b/llvm/lib/Target/Hexagon/HexagonScheduleV67T.td
@@ -0,0 +1,61 @@
+//=- HexagonScheduleV67T.td - Hexagon V67 Tiny Core Scheduling Definitions --=//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+class HexagonV67TPseudoItin {
+ list<InstrItinData> V67TPseudoItin_list = [
+ InstrItinData<PSEUDO, [InstrStage<1, [SLOT0, SLOT2, SLOT3]>], [2, 1, 1],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+ InstrItinData<PSEUDOM, [InstrStage<1, [SLOT2, SLOT3], 0>,
+ InstrStage<1, [SLOT2, SLOT3]>],
+ [2, 1, 1],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+ InstrItinData<DUPLEX, [InstrStage<1, [SLOT0]>],
+ [2, 1, 1]>,
+ InstrItinData<tc_ENDLOOP, [InstrStage<1, [SLOT_ENDLOOP]>], [2]>
+ ];
+}
+
+// V67TItin_list and HVXItin contain some old itineraries
+// still used by a handful of instructions. Hopefully, we will be able to
+// get rid of them soon.
+def HexagonV67TItinList : DepScalarItinV67T,
+ DepHVXItinV67, HVXItin, HexagonV67TPseudoItin {
+ list<InstrItinData> V67TItin_list = [
+ InstrItinData<LD_tc_ld_SLOT01, [InstrStage<1, [SLOT0]>],
+ [3, 1, 1],
+ [Hex_FWD, Hex_FWD, Hex_FWD]>,
+ InstrItinData<ST_tc_st_SLOT01, [InstrStage<1, [SLOT0]>],
+ [1, 1, 3, 3],
+ [Hex_FWD, Hex_FWD]>
+ ];
+
+ list<InstrItinData> ItinList =
+ !listconcat(DepScalarItinV67T_list,
+ DepHVXItinV67_list, V67TItin_list,
+ HVXItin_list, V67TPseudoItin_list);
+}
+
+def HexagonItinerariesV67T :
+ ProcessorItineraries<[SLOT0, SLOT1, SLOT2, SLOT3, SLOT_ENDLOOP,
+ CVI_ST, CVI_XLANE, CVI_SHIFT, CVI_MPY0, CVI_MPY1,
+ CVI_LD, CVI_XLSHF, CVI_MPY01, CVI_ALL,
+ CVI_ALL_NOMEM, CVI_ZW],
+ [Hex_FWD, HVX_FWD],
+ HexagonV67TItinList.ItinList>;
+
+
+def HexagonModelV67T : SchedMachineModel {
+ let IssueWidth = 3;
+ let Itineraries = HexagonItinerariesV67T;
+ let LoadLatency = 1;
+ let CompleteModel = 0;
+}
+
+//===----------------------------------------------------------------------===//
+// Hexagon V67 Tiny Core Resource Definitions -
+//===----------------------------------------------------------------------===//
diff --git a/llvm/lib/Target/Hexagon/HexagonSubtarget.cpp b/llvm/lib/Target/Hexagon/HexagonSubtarget.cpp
index 5461687..adeb3e0 100644
--- a/llvm/lib/Target/Hexagon/HexagonSubtarget.cpp
+++ b/llvm/lib/Target/Hexagon/HexagonSubtarget.cpp
@@ -83,6 +83,7 @@
InstrInfo(initializeSubtargetDependencies(CPU, FS)),
RegInfo(getHwMode()), TLInfo(TM, *this),
InstrItins(getInstrItineraryForCPU(CPUString)) {
+ Hexagon_MC::addArchSubtarget(this, FS);
// Beware of the default constructor of InstrItineraryData: it will
// reset all members to 0.
assert(InstrItins.Itineraries != nullptr && "InstrItins not initialized");
@@ -109,6 +110,13 @@
if (OverrideLongCalls.getPosition())
UseLongCalls = OverrideLongCalls;
+ if (isTinyCore()) {
+ // Tiny core has a single thread, so back-to-back scheduling is enabled by
+ // default.
+ if (!EnableBSBSched.getPosition())
+ UseBSBScheduling = false;
+ }
+
FeatureBitset Features = getFeatureBits();
if (HexagonDisableDuplex)
setFeatureBits(Features.reset(Hexagon::FeatureDuplex));
diff --git a/llvm/lib/Target/Hexagon/HexagonSubtarget.h b/llvm/lib/Target/Hexagon/HexagonSubtarget.h
index a2ec665..2c6d489 100644
--- a/llvm/lib/Target/Hexagon/HexagonSubtarget.h
+++ b/llvm/lib/Target/Hexagon/HexagonSubtarget.h
@@ -87,8 +87,14 @@
};
private:
+ enum HexagonProcFamilyEnum { Others, TinyCore };
+
std::string CPUString;
Triple TargetTriple;
+
+ // The following objects can use the TargetTriple, so they must be
+ // declared after it.
+ HexagonProcFamilyEnum HexagonProcFamily = Others;
HexagonInstrInfo InstrInfo;
HexagonRegisterInfo RegInfo;
HexagonTargetLowering TLInfo;
@@ -185,6 +191,9 @@
bool useUnsafeMath() const { return UseUnsafeMath; }
bool useZRegOps() const { return UseZRegOps; }
+ bool isTinyCore() const { return HexagonProcFamily == TinyCore; }
+ bool isTinyCoreWithDuplex() const { return isTinyCore() && EnableDuplex; }
+
bool useHVXOps() const {
return HexagonHVXVersion > Hexagon::ArchEnum::NoArch;
}
diff --git a/llvm/lib/Target/Hexagon/HexagonVLIWPacketizer.cpp b/llvm/lib/Target/Hexagon/HexagonVLIWPacketizer.cpp
index 36d71c4..b9c7be7 100644
--- a/llvm/lib/Target/Hexagon/HexagonVLIWPacketizer.cpp
+++ b/llvm/lib/Target/Hexagon/HexagonVLIWPacketizer.cpp
@@ -242,6 +242,10 @@
}
}
+ // TinyCore with Duplexes: Translate to big-instructions.
+ if (HST.isTinyCoreWithDuplex())
+ HII->translateInstrsForDup(MF, true);
+
// Loop over all of the basic blocks.
for (auto &MB : MF) {
auto Begin = MB.begin(), End = MB.end();
@@ -267,6 +271,10 @@
}
}
+ // TinyCore with Duplexes: Translate to tiny-instructions.
+ if (HST.isTinyCoreWithDuplex())
+ HII->translateInstrsForDup(MF, false);
+
Packetizer.unpacketizeSoloInstrs(MF);
return true;
}
@@ -1802,6 +1810,8 @@
setmemShufDisabled(false);
}
+ PacketHasDuplex = false;
+ PacketHasSLOT0OnlyInsn = false;
ResourceTracker->clearResources();
LLVM_DEBUG(dbgs() << "End packet\n");
}
@@ -1809,7 +1819,64 @@
bool HexagonPacketizerList::shouldAddToPacket(const MachineInstr &MI) {
if (Minimal)
return false;
- return !producesStall(MI);
+
+ // Constrainst for not packetizing this MI with existing instructions in a
+ // packet.
+ // MI is a store instruction.
+ // CurrentPacketMIs has a SLOT0 only instruction with constraint
+ // A_RESTRICT_NOSLOT1_STORE/isRestrictNoSlot1Store.
+ if (MI.mayStore() && isPureSlot0InsnWithNoSlot1Store(MI))
+ return false;
+
+ if (producesStall(MI))
+ return false;
+
+ // If TinyCore with Duplexes is enabled, check if this MI can form a Duplex
+ // with any other instruction in the existing packet.
+ auto &HST = MI.getParent()->getParent()->getSubtarget<HexagonSubtarget>();
+ // Constraint 1: Only one duplex allowed per packet.
+ // Constraint 2: Consider duplex checks only if there is atleast one
+ // instruction in a packet.
+ // Constraint 3: If one of the existing instructions in the packet has a
+ // SLOT0 only instruction that can not be duplexed, do not attempt to form
+ // duplexes. (TODO: This will invalidate the L4_return* instructions to form a
+ // duplex)
+ if (HST.isTinyCoreWithDuplex() && CurrentPacketMIs.size() > 0 &&
+ !PacketHasDuplex) {
+ // Check for SLOT0 only non-duplexable instruction in packet.
+ for (auto &MJ : CurrentPacketMIs)
+ PacketHasSLOT0OnlyInsn |= HII->isPureSlot0(*MJ);
+ // Get the Big Core Opcode (dup_*).
+ int Opcode = HII->getDuplexOpcode(MI, false);
+ if (Opcode >= 0) {
+ // We now have an instruction that can be duplexed.
+ for (auto &MJ : CurrentPacketMIs) {
+ if (HII->isDuplexPair(MI, *MJ) && !PacketHasSLOT0OnlyInsn) {
+ PacketHasDuplex = true;
+ return true;
+ }
+ }
+ // If it can not be duplexed, check if there is a valid transition in DFA
+ // with the original opcode.
+ MachineInstr &MIRef = const_cast<MachineInstr &>(MI);
+ MIRef.setDesc(HII->get(Opcode));
+ return ResourceTracker->canReserveResources(MIRef);
+ }
+ }
+
+ return true;
+}
+
+bool HexagonPacketizerList::isPureSlot0InsnWithNoSlot1Store(
+ const MachineInstr &MI) {
+ bool noSlot1Store = false;
+ bool isSlot0Only = false;
+ for (auto J : CurrentPacketMIs) {
+ noSlot1Store |= HII->isRestrictNoSlot1Store(*J);
+ isSlot0Only |= HII->isPureSlot0(*J);
+ }
+
+ return (noSlot1Store && isSlot0Only);
}
// V60 forward scheduling.
diff --git a/llvm/lib/Target/Hexagon/HexagonVLIWPacketizer.h b/llvm/lib/Target/Hexagon/HexagonVLIWPacketizer.h
index 943b9ac..27a4722 100644
--- a/llvm/lib/Target/Hexagon/HexagonVLIWPacketizer.h
+++ b/llvm/lib/Target/Hexagon/HexagonVLIWPacketizer.h
@@ -57,6 +57,13 @@
// instruction from the previous packet.
bool PacketStalls = false;
+ // Set to true if the packet has a duplex pair of sub-instructions.
+ bool PacketHasDuplex = false;
+
+ // Set to true if the packet has a instruction that can only be executed
+ // in SLOT0.
+ bool PacketHasSLOT0OnlyInsn = false;
+
protected:
/// A handle to the branch probability pass.
const MachineBranchProbabilityInfo *MBPI;
@@ -149,6 +156,7 @@
bool hasRegMaskDependence(const MachineInstr &I, const MachineInstr &J);
bool hasDualStoreDependence(const MachineInstr &I, const MachineInstr &J);
bool producesStall(const MachineInstr &MI);
+ bool isPureSlot0InsnWithNoSlot1Store(const MachineInstr &MI);
};
} // end namespace llvm
diff --git a/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonAsmBackend.cpp b/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonAsmBackend.cpp
index 8f1e5c1..ef7db393 100644
--- a/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonAsmBackend.cpp
+++ b/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonAsmBackend.cpp
@@ -43,6 +43,7 @@
std::unique_ptr <MCInstrInfo> MCII;
std::unique_ptr <MCInst *> RelaxTarget;
MCInst * Extender;
+ unsigned MaxPacketSize;
void ReplaceInstruction(MCCodeEmitter &E, MCRelaxableFragment &RF,
MCInst &HMB) const {
@@ -62,7 +63,8 @@
StringRef CPU)
: MCAsmBackend(support::little), OSABI(OSABI), CPU(CPU), relaxedCnt(0),
MCII(T.createMCInstrInfo()), RelaxTarget(new MCInst *),
- Extender(nullptr) {}
+ Extender(nullptr), MaxPacketSize(HexagonMCInstrInfo::packetSize(CPU))
+ {}
std::unique_ptr<MCObjectTargetWriter>
createObjectTargetWriter() const override {
@@ -685,7 +687,7 @@
ParseIn = 0x00004000, // In packet parse-bits.
ParseEnd = 0x0000c000; // End of packet parse-bits.
- while(Count % HEXAGON_INSTR_SIZE) {
+ while (Count % HEXAGON_INSTR_SIZE) {
LLVM_DEBUG(dbgs() << "Alignment not a multiple of the instruction size:"
<< Count % HEXAGON_INSTR_SIZE << "/"
<< HEXAGON_INSTR_SIZE << "\n");
@@ -693,11 +695,11 @@
OS << '\0';
}
- while(Count) {
+ while (Count) {
Count -= HEXAGON_INSTR_SIZE;
// Close the packet whenever a multiple of the maximum packet size remains
- uint32_t ParseBits = (Count % (HEXAGON_PACKET_SIZE * HEXAGON_INSTR_SIZE))?
- ParseIn: ParseEnd;
+ uint32_t ParseBits = (Count % (MaxPacketSize * HEXAGON_INSTR_SIZE)) ?
+ ParseIn : ParseEnd;
support::endian::write<uint32_t>(OS, Nopcode | ParseBits, Endian);
}
return true;
@@ -728,7 +730,8 @@
MCContext &Context = Asm.getContext();
auto &RF = cast<MCRelaxableFragment>(*K);
auto &Inst = const_cast<MCInst &>(RF.getInst());
- while (Size > 0 && HexagonMCInstrInfo::bundleSize(Inst) < 4) {
+ while (Size > 0 &&
+ HexagonMCInstrInfo::bundleSize(Inst) < MaxPacketSize) {
MCInst *Nop = new (Context) MCInst;
Nop->setOpcode(Hexagon::A2_nop);
Inst.addOperand(MCOperand::createInst(Nop));
diff --git a/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCDuplexInfo.cpp b/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCDuplexInfo.cpp
index 3cbb860..5154a0a 100644
--- a/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCDuplexInfo.cpp
+++ b/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCDuplexInfo.cpp
@@ -10,12 +10,11 @@
//
//===----------------------------------------------------------------------===//
+#include "HexagonMCExpr.h"
#include "MCTargetDesc/HexagonBaseInfo.h"
#include "MCTargetDesc/HexagonMCInstrInfo.h"
#include "MCTargetDesc/HexagonMCTargetDesc.h"
#include "llvm/ADT/SmallVector.h"
-#include "llvm/MC/MCExpr.h"
-#include "llvm/MC/MCInst.h"
#include "llvm/MC/MCSubtargetInfo.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
@@ -296,8 +295,7 @@
DstReg = MCI.getOperand(1).getReg();
SrcReg = MCI.getOperand(0).getReg();
// [if ([!]p0[.new])] jumpr r31
- if ((HexagonMCInstrInfo::isPredReg(SrcReg) && (Hexagon::P0 == SrcReg)) &&
- (Hexagon::R31 == DstReg)) {
+ if ((Hexagon::P0 == SrcReg) && (Hexagon::R31 == DstReg)) {
return HexagonII::HSIG_L2;
}
break;
diff --git a/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCInstrInfo.cpp b/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCInstrInfo.cpp
index 8eb32eb..4f8a432 100644
--- a/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCInstrInfo.cpp
+++ b/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCInstrInfo.cpp
@@ -17,6 +17,7 @@
#include "MCTargetDesc/HexagonMCShuffler.h"
#include "MCTargetDesc/HexagonMCTargetDesc.h"
#include "llvm/ADT/SmallVector.h"
+#include "llvm/ADT/StringSwitch.h"
#include "llvm/MC/MCContext.h"
#include "llvm/MC/MCExpr.h"
#include "llvm/MC/MCInst.h"
@@ -119,10 +120,10 @@
return (1);
}
-bool HexagonMCInstrInfo::canonicalizePacket(MCInstrInfo const &MCII,
- MCSubtargetInfo const &STI,
- MCContext &Context, MCInst &MCB,
- HexagonMCChecker *Check) {
+namespace {
+bool canonicalizePacketImpl(MCInstrInfo const &MCII, MCSubtargetInfo const &STI,
+ MCContext &Context, MCInst &MCB,
+ HexagonMCChecker *Check) {
// Check the bundle for errors.
bool CheckOk = Check ? Check->check(false) : true;
if (!CheckOk)
@@ -132,9 +133,9 @@
if (!HexagonDisableCompound)
HexagonMCInstrInfo::tryCompound(MCII, STI, Context, MCB);
HexagonMCShuffle(Context, false, MCII, STI, MCB);
+
// Examine the packet and convert pairs of instructions to duplex
// instructions when possible.
- MCInst InstBundlePreDuplex = MCInst(MCB);
if (STI.getFeatureBits() [Hexagon::FeatureDuplex]) {
SmallVector<DuplexCandidate, 8> possibleDuplexes;
possibleDuplexes =
@@ -146,8 +147,11 @@
HexagonMCInstrInfo::padEndloop(MCB, Context);
// If compounding and duplexing didn't reduce the size below
// 4 or less we have a packet that is too big.
- if (HexagonMCInstrInfo::bundleSize(MCB) > HEXAGON_PACKET_SIZE)
+ if (HexagonMCInstrInfo::bundleSize(MCB) > HEXAGON_PACKET_SIZE) {
+ if (Check)
+ Check->reportError("invalid instruction packet: out of slots");
return false;
+ }
// Check the bundle for errors.
CheckOk = Check ? Check->check(true) : true;
if (!CheckOk)
@@ -155,6 +159,27 @@
HexagonMCShuffle(Context, true, MCII, STI, MCB);
return true;
}
+} // namespace
+
+bool HexagonMCInstrInfo::canonicalizePacket(MCInstrInfo const &MCII,
+ MCSubtargetInfo const &STI,
+ MCContext &Context, MCInst &MCB,
+ HexagonMCChecker *Check,
+ bool AttemptCompatibility) {
+ auto ArchSTI = Hexagon_MC::getArchSubtarget(&STI);
+ if (!AttemptCompatibility || ArchSTI == nullptr)
+ return canonicalizePacketImpl(MCII, STI, Context, MCB, Check);
+
+ const MCRegisterInfo *RI = Context.getRegisterInfo();
+ HexagonMCChecker DefaultCheck(Context, MCII, STI, MCB, *RI, false);
+ HexagonMCChecker *BaseCheck = (Check == nullptr) ? &DefaultCheck : Check;
+ HexagonMCChecker PerfCheck(*BaseCheck, STI, false);
+ if (canonicalizePacketImpl(MCII, STI, Context, MCB, &PerfCheck))
+ return true;
+
+ HexagonMCChecker ArchCheck(*BaseCheck, *ArchSTI, true);
+ return canonicalizePacketImpl(MCII, *ArchSTI, Context, MCB, &ArchCheck);
+}
MCInst HexagonMCInstrInfo::deriveExtender(MCInstrInfo const &MCII,
MCInst const &Inst,
@@ -493,7 +518,7 @@
MCInst const &HexagonMCInstrInfo::instruction(MCInst const &MCB, size_t Index) {
assert(isBundle(MCB));
- assert(Index < HEXAGON_PACKET_SIZE);
+ assert(Index < HEXAGON_PRESHUFFLE_PACKET_SIZE);
return *MCB.getOperand(bundleInstructionsOffset + Index).getInst();
}
@@ -633,6 +658,12 @@
return ((F >> HexagonII::NewValuePos) & HexagonII::NewValueMask);
}
+bool HexagonMCInstrInfo::isNewValueStore(MCInstrInfo const &MCII,
+ MCInst const &MCI) {
+ const uint64_t F = HexagonMCInstrInfo::getDesc(MCII, MCI).TSFlags;
+ return (F >> HexagonII::NVStorePos) & HexagonII::NVStoreMask;
+}
+
/// Return whether the operand is extendable.
bool HexagonMCInstrInfo::isOpExtendable(MCInstrInfo const &MCII,
MCInst const &MCI, unsigned short O) {
@@ -675,8 +706,17 @@
!((F >> HexagonII::PredicatedFalsePos) & HexagonII::PredicatedFalseMask));
}
-bool HexagonMCInstrInfo::isPredReg(unsigned Reg) {
- return (Reg >= Hexagon::P0 && Reg <= Hexagon::P3_0);
+bool HexagonMCInstrInfo::isPredReg(MCRegisterInfo const &MRI, unsigned Reg) {
+ auto &PredRegClass = MRI.getRegClass(Hexagon::PredRegsRegClassID);
+ return PredRegClass.contains(Reg);
+}
+
+bool HexagonMCInstrInfo::isPredRegister(MCInstrInfo const &MCII,
+ MCInst const &Inst, unsigned I) {
+ MCInstrDesc const &Desc = HexagonMCInstrInfo::getDesc(MCII, Inst);
+
+ return Inst.getOperand(I).isReg() &&
+ Desc.OpInfo[I].RegClass == Hexagon::PredRegsRegClassID;
}
/// Return whether the insn can be packaged only with A and X-type insns.
@@ -773,10 +813,8 @@
}
bool HexagonMCInstrInfo::isVector(MCInstrInfo const &MCII, MCInst const &MCI) {
- if ((getType(MCII, MCI) <= HexagonII::TypeCVI_LAST) &&
- (getType(MCII, MCI) >= HexagonII::TypeCVI_FIRST))
- return true;
- return false;
+ const uint64_t F = HexagonMCInstrInfo::getDesc(MCII, MCI).TSFlags;
+ return (F >> HexagonII::isCVIPos) & HexagonII::isCVIMask;
}
int64_t HexagonMCInstrInfo::minConstant(MCInst const &MCI, size_t Index) {
@@ -822,6 +860,18 @@
return HExpr->s27_2_reloc();
}
+unsigned HexagonMCInstrInfo::packetSizeSlots(MCSubtargetInfo const &STI) {
+ const bool IsTiny = STI.getFeatureBits()[Hexagon::ProcTinyCore];
+
+ return IsTiny ? (HEXAGON_PACKET_SIZE - 1) : HEXAGON_PACKET_SIZE;
+}
+
+unsigned HexagonMCInstrInfo::packetSize(StringRef CPU) {
+ return llvm::StringSwitch<unsigned>(CPU)
+ .Case("hexagonv67t", 3)
+ .Default(4);
+}
+
void HexagonMCInstrInfo::padEndloop(MCInst &MCB, MCContext &Context) {
MCInst Nop;
Nop.setOpcode(Hexagon::A2_nop);
@@ -856,6 +906,33 @@
return (F >> HexagonII::HasTmpDstPos) & HexagonII::HasTmpDstMask;
}
+bool HexagonMCInstrInfo::requiresSlot(MCSubtargetInfo const &STI,
+ MCInst const &MCI) {
+ const unsigned OpCode = MCI.getOpcode();
+ const bool IsTiny = STI.getFeatureBits() [Hexagon::ProcTinyCore];
+ const bool NoSlotReqd = Hexagon::A4_ext == OpCode ||
+ (IsTiny && Hexagon::A2_nop == OpCode) ||
+ (IsTiny && Hexagon::J4_hintjumpr == OpCode);
+
+ return !NoSlotReqd;
+}
+
+unsigned HexagonMCInstrInfo::slotsConsumed(MCInstrInfo const &MCII,
+ MCSubtargetInfo const &STI,
+ MCInst const &MCI) {
+ unsigned slotsUsed = 0;
+ for (auto HMI : bundleInstructions(MCI)) {
+ MCInst const &MCI = *HMI.getInst();
+ if (!requiresSlot(STI, MCI))
+ continue;
+ if (isDuplex(MCII, MCI))
+ slotsUsed += 2;
+ else
+ ++slotsUsed;
+ }
+ return slotsUsed;
+}
+
void HexagonMCInstrInfo::replaceDuplex(MCContext &Context, MCInst &MCB,
DuplexCandidate Candidate) {
assert(Candidate.packetIndexI < MCB.size());
diff --git a/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCInstrInfo.h b/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCInstrInfo.h
index c92cf34..70022aa 100644
--- a/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCInstrInfo.h
+++ b/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCInstrInfo.h
@@ -28,6 +28,7 @@
class MCExpr;
class MCInstrDesc;
class MCInstrInfo;
+class MCRegisterInfo;
class MCSubtargetInfo;
class DuplexCandidate {
@@ -91,7 +92,8 @@
// Put the packet in to canonical form, compound, duplex, pad, and shuffle
bool canonicalizePacket(MCInstrInfo const &MCII, MCSubtargetInfo const &STI,
MCContext &Context, MCInst &MCB,
- HexagonMCChecker *Checker);
+ HexagonMCChecker *Checker,
+ bool AttemptCompatibility = false);
// Create a duplex instruction given the two subinsts
MCInst *deriveDuplex(MCContext &Context, unsigned iClass, MCInst const &inst0,
@@ -257,6 +259,8 @@
// Return whether the insn is a new-value consumer.
bool isNewValue(MCInstrInfo const &MCII, MCInst const &MCI);
+/// Return true if the operand is a new-value store insn.
+bool isNewValueStore(MCInstrInfo const &MCII, MCInst const &MCI);
bool isOpExtendable(MCInstrInfo const &MCII, MCInst const &MCI, unsigned short);
// Can these two instructions be duplexed
@@ -275,8 +279,11 @@
// Return whether the predicate sense is true
bool isPredicatedTrue(MCInstrInfo const &MCII, MCInst const &MCI);
-// Is this a predicate register
-bool isPredReg(unsigned Reg);
+// Return true if this is a scalar predicate register.
+bool isPredReg(MCRegisterInfo const &MRI, unsigned Reg);
+
+// Returns true if the Ith operand is a predicate register.
+bool isPredRegister(MCInstrInfo const &MCII, MCInst const &Inst, unsigned I);
// Return whether the insn is a prefix.
bool isPrefix(MCInstrInfo const &MCII, MCInst const &MCI);
@@ -295,6 +302,21 @@
bool mustExtend(MCExpr const &Expr);
bool mustNotExtend(MCExpr const &Expr);
+// Returns true if this instruction requires a slot to execute.
+bool requiresSlot(MCSubtargetInfo const &STI, MCInst const &MCI);
+
+unsigned packetSize(StringRef CPU);
+
+// Returns the maximum number of slots available in the given
+// subtarget's packets.
+unsigned packetSizeSlots(MCSubtargetInfo const &STI);
+
+// Returns the number of slots consumed by this packet, considering duplexed
+// and compound instructions.
+unsigned slotsConsumed(MCInstrInfo const &MCII, MCSubtargetInfo const &STI,
+ MCInst const &MCI);
+
+
// Pad the bundle with nops to satisfy endloop requirements
void padEndloop(MCInst &MCI, MCContext &Context);
class PredicateInfo {
diff --git a/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCTargetDesc.cpp b/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCTargetDesc.cpp
index 9078be3..5d701a1 100644
--- a/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCTargetDesc.cpp
+++ b/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCTargetDesc.cpp
@@ -75,6 +75,8 @@
cl::init(false));
cl::opt<bool> MV67("mv67", cl::Hidden, cl::desc("Build for Hexagon V67"),
cl::init(false));
+cl::opt<bool> MV67T("mv67t", cl::Hidden, cl::desc("Build for Hexagon V67T"),
+ cl::init(false));
cl::opt<Hexagon::ArchEnum>
EnableHVX("mhvx",
@@ -113,14 +115,20 @@
return "hexagonv66";
if (MV67)
return "hexagonv67";
+ if (MV67T)
+ return "hexagonv67t";
return "";
}
StringRef Hexagon_MC::selectHexagonCPU(StringRef CPU) {
StringRef ArchV = HexagonGetArchVariant();
if (!ArchV.empty() && !CPU.empty()) {
- if (ArchV != CPU)
- report_fatal_error("conflicting architectures specified.");
+ // Tiny cores have a "t" suffix that is discarded when creating a secondary
+ // non-tiny subtarget. See: addArchSubtarget
+ std::pair<StringRef,StringRef> ArchP = ArchV.split('t');
+ std::pair<StringRef,StringRef> CPUP = CPU.split('t');
+ if (!ArchP.first.equals(CPUP.first))
+ report_fatal_error("conflicting architectures specified.");
return CPU;
}
if (ArchV.empty()) {
@@ -175,6 +183,14 @@
}
+namespace llvm {
+namespace HexagonFUnits {
+bool isSlot0Only(unsigned units) {
+ return HexagonItinerariesV62FU::SLOT0 == units;
+}
+} // namespace HexagonFUnits
+} // namespace llvm
+
namespace {
class HexagonTargetAsmStreamer : public HexagonTargetStreamer {
@@ -353,7 +369,8 @@
.Case("hexagonv62", "+hvxv62")
.Case("hexagonv65", "+hvxv65")
.Case("hexagonv66", "+hvxv66")
- .Case("hexagonv67", "+hvxv67"));
+ .Case("hexagonv67", "+hvxv67")
+ .Case("hexagonv67t", "+hvxv67"));
break;
}
case Hexagon::ArchEnum::NoArch:
@@ -376,6 +393,18 @@
Result.second = selectHexagonFS(Result.first, FS);
return Result;
}
+std::mutex ArchSubtargetMutex;
+std::unordered_map<std::string, std::unique_ptr<MCSubtargetInfo const>>
+ ArchSubtarget;
+} // namespace
+
+MCSubtargetInfo const *
+Hexagon_MC::getArchSubtarget(MCSubtargetInfo const *STI) {
+ std::lock_guard<std::mutex> Lock(ArchSubtargetMutex);
+ auto Existing = ArchSubtarget.find(std::string(STI->getCPU()));
+ if (Existing == ArchSubtarget.end())
+ return nullptr;
+ return Existing->second.get();
}
FeatureBitset Hexagon_MC::completeHVXFeatures(const FeatureBitset &S) {
@@ -440,6 +469,8 @@
StringRef ArchFS = Features.second;
MCSubtargetInfo *X = createHexagonMCSubtargetInfoImpl(TT, CPUName, ArchFS);
+ if (X != nullptr && (CPUName == "hexagonv67t"))
+ addArchSubtarget(X, ArchFS);
if (CPU.equals("help"))
exit(0);
@@ -470,6 +501,19 @@
return X;
}
+void Hexagon_MC::addArchSubtarget(MCSubtargetInfo const *STI,
+ StringRef FS) {
+ assert(STI != nullptr);
+ if (STI->getCPU().contains("t")) {
+ auto ArchSTI = createHexagonMCSubtargetInfo(
+ STI->getTargetTriple(),
+ STI->getCPU().substr(0, STI->getCPU().size() - 1), FS);
+ std::lock_guard<std::mutex> Lock(ArchSubtargetMutex);
+ ArchSubtarget[std::string(STI->getCPU())] =
+ std::unique_ptr<MCSubtargetInfo const>(ArchSTI);
+ }
+}
+
unsigned Hexagon_MC::GetELFFlags(const MCSubtargetInfo &STI) {
static std::map<StringRef,unsigned> ElfFlags = {
{"hexagonv5", ELF::EF_HEXAGON_MACH_V5},
@@ -479,6 +523,7 @@
{"hexagonv65", ELF::EF_HEXAGON_MACH_V65},
{"hexagonv66", ELF::EF_HEXAGON_MACH_V66},
{"hexagonv67", ELF::EF_HEXAGON_MACH_V67},
+ {"hexagonv67t", ELF::EF_HEXAGON_MACH_V67T},
};
auto F = ElfFlags.find(STI.getCPU());
diff --git a/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCTargetDesc.h b/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCTargetDesc.h
index 4066a43..6cc6f51 100644
--- a/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCTargetDesc.h
+++ b/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonMCTargetDesc.h
@@ -78,6 +78,9 @@
/// etc. do not need to go through TargetRegistry.
MCSubtargetInfo *createHexagonMCSubtargetInfo(const Triple &TT, StringRef CPU,
StringRef FS);
+ MCSubtargetInfo const *getArchSubtarget(MCSubtargetInfo const *STI);
+ void addArchSubtarget(MCSubtargetInfo const *STI,
+ StringRef FS);
unsigned GetELFFlags(const MCSubtargetInfo &STI);
}
diff --git a/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonShuffler.cpp b/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonShuffler.cpp
index f20c58a..2788b86 100644
--- a/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonShuffler.cpp
+++ b/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonShuffler.cpp
@@ -184,10 +184,6 @@
Packet.push_back(PI);
}
-const static struct {
- unsigned first;
- unsigned second;
-} jumpSlots[] = {{8, 4}, {8, 2}, {8, 1}, {4, 2}, {4, 1}, {2, 1}};
static const unsigned Slot0Mask = 1 << 0;
static const unsigned Slot1Mask = 1 << 1;
@@ -256,6 +252,8 @@
restrictSlot1AOK(Summary);
restrictNoSlot1Store(Summary);
+ permitNonSlot();
+
// These restrictions can modify the slot masks in the instructions
// in the Packet member, but they can also detect constraint failures
// which are fatal.
@@ -279,20 +277,22 @@
return;
}
+ const static std::pair<unsigned, unsigned> jumpSlots[] = {
+ {8, 4}, {8, 2}, {8, 1}, {4, 2}, {4, 1}, {2, 1}};
// try all possible choices
- for (unsigned int i = 0; i < array_lengthof(jumpSlots); ++i) {
+ for (std::pair<unsigned, unsigned> jumpSlot : jumpSlots) {
// validate first jump with this slot rule
- if (!(jumpSlots[i].first & Summary.branchInsts[0]->Core.getUnits()))
+ if (!(jumpSlot.first & Summary.branchInsts[0]->Core.getUnits()))
continue;
// validate second jump with this slot rule
- if (!(jumpSlots[i].second & Summary.branchInsts[1]->Core.getUnits()))
+ if (!(jumpSlot.second & Summary.branchInsts[1]->Core.getUnits()))
continue;
// both valid for this configuration, set new slot rules
const HexagonPacket PacketSave = Packet;
- Summary.branchInsts[0]->Core.setUnits(jumpSlots[i].first);
- Summary.branchInsts[1]->Core.setUnits(jumpSlots[i].second);
+ Summary.branchInsts[0]->Core.setUnits(jumpSlot.first);
+ Summary.branchInsts[1]->Core.setUnits(jumpSlot.second);
const bool HasShuffledPacket = tryAuction(Summary).hasValue();
if (HasShuffledPacket)
@@ -306,6 +306,15 @@
reportError("invalid instruction packet: out of slots");
}
+
+void HexagonShuffler::permitNonSlot() {
+ for (HexagonInstr &ISJ : insts()) {
+ const bool RequiresSlot = HexagonMCInstrInfo::requiresSlot(STI, *ISJ.ID);
+ if (!RequiresSlot)
+ ISJ.Core.setAllUnits();
+ }
+}
+
bool HexagonShuffler::ValidResourceUsage(HexagonPacketSummary const &Summary) {
Optional<HexagonPacket> ShuffledPacket = tryAuction(Summary);
@@ -394,8 +403,16 @@
// A single store must use slot #0.
if (HexagonMCInstrInfo::getDesc(MCII, ID).mayStore()) {
if (!Summary.store0) {
- if (Summary.stores == 1 &&
- (Summary.loads == 0 || !isMemReorderDisabled()))
+ const bool PacketHasNoOnlySlot0 =
+ llvm::none_of(insts(), [&](HexagonInstr const &I) {
+ return I.Core.getUnits() == Slot0Mask &&
+ I.ID->getOpcode() != ID.getOpcode();
+ });
+ const bool SafeToMoveToSlot0 =
+ (Summary.loads == 0) ||
+ (!isMemReorderDisabled() && PacketHasNoOnlySlot0);
+
+ if (Summary.stores == 1 && SafeToMoveToSlot0)
// Pin the store to slot #0 only if isMemReorderDisabled() == false
ISJ->Core.setUnits(ISJ->Core.getUnits() & slotSingleStore);
else if (Summary.stores >= 1) {
@@ -416,12 +433,6 @@
return false;
}
}
-
- if (!ISJ->Core.getUnits()) {
- // Error if insn may not be executed in any slot.
- reportError("invalid instruction packet: out of slots");
- return false;
- }
}
return true;
diff --git a/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonShuffler.h b/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonShuffler.h
index 5f563ce..f7f646c 100644
--- a/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonShuffler.h
+++ b/llvm/lib/Target/Hexagon/MCTargetDesc/HexagonShuffler.h
@@ -19,6 +19,7 @@
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/SmallVector.h"
+#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/StringRef.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/Support/SMLoc.h"
@@ -47,6 +48,9 @@
setWeight(s);
}
+ void setAllUnits() {
+ setUnits(((1u << HEXAGON_PACKET_SIZE) - 1));
+ }
unsigned setWeight(unsigned s);
unsigned getUnits() const { return (Slots); }
@@ -178,12 +182,13 @@
void restrictSlot1AOK(HexagonPacketSummary const &Summary);
void restrictNoSlot1Store(HexagonPacketSummary const &Summary);
void restrictNoSlot1();
-
- Optional<HexagonPacket> tryAuction(HexagonPacketSummary const &Summary) const;
-
bool restrictStoreLoadOrder(HexagonPacketSummary const &Summary);
void restrictBranchOrder(HexagonPacketSummary const &Summary);
void restrictPreferSlot3(HexagonPacketSummary const &Summary);
+ void permitNonSlot();
+
+ Optional<HexagonPacket> tryAuction(HexagonPacketSummary const &Summary) const;
+
HexagonPacketSummary GetPacketSummary();
bool ValidPacketMemoryOps(HexagonPacketSummary const &Summary) const;
bool ValidResourceUsage(HexagonPacketSummary const &Summary);
@@ -227,11 +232,10 @@
using InstPredicate = bool (*)(MCInstrInfo const &, MCInst const &);
bool HasInstWith(InstPredicate Pred) const {
- return llvm::any_of(make_range(cbegin(), cend()),
- [&](HexagonInstr const &I) {
- MCInst const &Inst = I.getDesc();
- return (*Pred)(MCII, Inst);
- });
+ return llvm::any_of(insts(), [&](HexagonInstr const &I) {
+ MCInst const &Inst = I.getDesc();
+ return (*Pred)(MCII, Inst);
+ });
}
// Add insn handle to the bundle .