Difference between revisions of "Extending NaplesPU for 64-bit support"
(→Instruction Lowering) |
(→Instruction Lowering) |
||
| Line 115: | Line 115: | ||
// Create the destination register | // Create the destination register | ||
| − | unsigned DstReg = MI->getOperand(0).getReg(); | + | unsigned DstReg = MI->getOperand(0).getReg(); |
| − | |||
int64_t ImmOp = MI->getOperand(1).getImm(); | int64_t ImmOp = MI->getOperand(1).getImm(); | ||
| Line 142: | Line 141: | ||
} | } | ||
</syntaxhighlight> | </syntaxhighlight> | ||
| + | As it is explained, the load operation for 64-bit immediate value is realized by requiring a ''NuPlus64GPR'' register and writing the split parts on the sub-registers. | ||
Revision as of 15:05, 14 May 2019
nu+ toolchain can be extended to support 64-bit operations. A git branch with full 64-bit support is provided. Consequently, if it is necessary to compile the toolchain supporting this extension, a checkout on llvm-7-64b branch is required.
Changes are related to both frontend and backend.
Contents
nu+ Frontend Modifications
nu+ frontend abstracts target informations through the TargetInfo class, extending it in the NuPlusTargetInfo implementation.
Since 64-bit operations require to support double-integer and double-floating-point formats, the following changes and additions are required in the target definition:
class LLVM_LIBRARY_VISIBILITY NuPlusTargetInfo : public TargetInfo {
...
public:
NuPlusTargetInfo(const llvm::Triple &Triple, const TargetOptions &Opts)
: TargetInfo(Triple) {
...
resetDataLayout("e-m:e-p:32:32-i64:64:64-i32:32:32-f32:32:32-f64:64:64");
LongDoubleWidth = 64;
LongDoubleAlign = 64;
DoubleWidth = 64;
DoubleAlign = 64;
LongWidth = 64;
LongAlign = 64;
LongLongWidth = 64;
LongLongAlign = 64;
}nu+ Backend Modifications
This section describes the backend modification to be applied for 64-bit support.
Registers Definition
The 64-bit support for registers is based on the "Sub-Reg" behaviour. Since nu+ registers are 32-bit wide, a 64-bit variable is stored split in two parts:
- The higher 32-bit are placed in the S[i] register;
- The lower 32-bit are placed in the S[i+1] register.
The following class is declared in NuPlusRegisterInfo.td.
class NuPlus64GPRReg<bits<16> Enc, string n, list<Register> subregs>
: NuPlusRegWithSubRegs<Enc, n, subregs> {
let SubRegIndices = [sub_even, sub_odd];
let CoveredBySubRegs = 1;
}The register instantiation is realized as follows:
foreach i = 0-28 in {
def S#!shl(i, 1)#_S#!add(!shl(i, 1), 1) : NuPlus64GPRReg<!shl(i, 1), "s"#!shl(i, 1)#_64,
[!cast<NuPlusGPRReg>("S"#!shl(i, 1)),
!cast<NuPlusGPRReg>("S"#!add(!shl(i, 1), 1))]>;
}Using the newly defined 64-bit support, it is possible to manage vector registers partitioned in eight cells, in which each one is 64-bit wide.
def VR512L : RegisterClass<"NuPlus", [v8i64, v8f64, v8i8, v8i16, v8i32], 512, (sequence "V%u", 0, 63)>;Calling Convention
Calling conventions for nu+ are modified supporting the 64-bit registers. As a result, the 32-bit Calling Convention is extended by adding the following lines:
CCIfType<[v8i8, v8i16, v8i32], CCPromoteToType<v8i64>>,
CCIfNotVarArg<CCIfType<[i64, f64], CCAssignToReg<[S0_S1, S2_S3, S4_S5, S6_S7, S8_S9, S10_S11, S12_S13, S14_S15]>>>,
CCIfNotVarArg<CCIfType<[v16i32, v16f32, v8i64, v8f64], CCAssignToReg<[V0, V1, V2, V3, V4, V5, V6, V7]>>>,The lines above describes how parameters are passed through registers. Also the stack assignment is described below:
CCIfType<[i64, f64], CCAssignToStack<8, 8>>,
CCIfType<[v16i32, v16f32, v8i64, v8f64], CCAssignToStack<64, 64>>The return value calling convention is extended follows:
CCIfType<[i64, f64], CCAssignToReg<[S0_S1, S2_S3, S4_S5, S6_S7, S8_S9, S10_S11]>>,
CCIfType<[v8i8, v8i16, v8i32], CCPromoteToType<v8i64>>,
CCIfType<[v16i32, v16f32, v8i64, v8f64], CCAssignToReg<[V0, V1, V2, V3, V4, V5]>>ISA Support
Referring to 32-bit ISA Support the 64-bit support on instructions is realized by extending the nu+ instruction hierarchy, adding a bit that acts as a mark.
As an example, in the following code, the bit <FR_TwoOp_Unmasked_64> class is derived from the <FR_TwoOp_Unmasked>, setting the sixth parameter as <1>. Recalling what described in the 32-bit ISA Support, it describes if the instruction refers to a 64-bit behaviour or not.
class FR_TwoOp_Unmasked_64<dag outs, dag ins, string asmstr, list<dag> pattern, bits<6> opcode, Fmt fmt2, Fmt fmt1, Fmt fmt0>
: FR_TwoOp_Unmasked<outs, ins, asmstr, pattern, opcode, 1, fmt2, fmt1, fmt0> {}Instruction Lowering
Since new instructions are defined, a lowering behaviour on them is often required. As a result, NuPlusISelLowering is extended in some points. For instance, since LoadI64 is a pseudo-instruction, it must be expanded as follows:
case NuPlus::LoadI64:
return EmitLoadI64(&MI, BB);
...
MachineBasicBlock *
NuPlusTargetLowering::EmitLoadI64(MachineInstr *MI,
MachineBasicBlock *BB) const {
DebugLoc DL = MI->getDebugLoc();
const TargetInstrInfo *TII = Subtarget.getInstrInfo();
MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
// Create the destination register
unsigned DstReg = MI->getOperand(0).getReg();
int64_t ImmOp = MI->getOperand(1).getImm();
unsigned Immediate = ((ImmOp >> 32) & 0xffffffff);
BuildMI(*BB, MI, DL, TII->get(NuPlus::MOVEIHSI))
.addReg(DstReg, RegState::Define, NuPlus::sub_odd)
.addImm(((Immediate >> 16) & 0xFFFF));
BuildMI(*BB, MI, DL, TII->get(NuPlus::MOVEILSI))
.addReg(DstReg, 0, NuPlus::sub_odd)
.addImm((Immediate & 0xFFFF));
Immediate = (ImmOp & 0xffffffff);
BuildMI(*BB, MI, DL, TII->get(NuPlus::MOVEIHSI))
.addReg(DstReg, 0, NuPlus::sub_even)
.addImm(((Immediate >> 16) & 0xFFFF));
BuildMI(*BB, MI, DL, TII->get(NuPlus::MOVEILSI))
.addReg(DstReg, 0, NuPlus::sub_even)
.addImm((Immediate & 0xFFFF));
MI->eraseFromParent();
return BB;
}As it is explained, the load operation for 64-bit immediate value is realized by requiring a NuPlus64GPR register and writing the split parts on the sub-registers.
