1
0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2024-10-19 11:02:59 +02:00
llvm-mirror/test/TableGen/GlobalISelEmitter.td

1178 lines
66 KiB
TableGen
Raw Normal View History

// RUN: llvm-tblgen -gen-global-isel -I %p/../../include -I %p/Common -optimize-match-table=false %s -o %T/non-optimized.cpp
// RUN: llvm-tblgen -gen-global-isel -I %p/../../include -I %p/Common -optimize-match-table=true %s -o %T/optimized.cpp
// RUN: llvm-tblgen -gen-global-isel -I %p/../../include -I %p/Common %s -o %T/default.cpp
// RUN: FileCheck %s --check-prefixes=CHECK,R19C,R19N -input-file=%T/non-optimized.cpp
// RUN: FileCheck %s --check-prefixes=CHECK,R19C,R19O -input-file=%T/optimized.cpp
// RUN: FileCheck %s --check-prefixes=CHECK,R21C,R21N -input-file=%T/non-optimized.cpp
// RUN: FileCheck %s --check-prefixes=CHECK,R21C,R21O -input-file=%T/optimized.cpp
// RUN: FileCheck %s --check-prefixes=CHECK,R20C,R20N -input-file=%T/non-optimized.cpp
// RUN: FileCheck %s --check-prefixes=CHECK,R20C,R20O -input-file=%T/optimized.cpp
// RUN: FileCheck %s --check-prefixes=CHECK,R00C,R00N -input-file=%T/non-optimized.cpp
// RUN: FileCheck %s --check-prefixes=CHECK,R00C,R00O -input-file=%T/optimized.cpp
// RUN: FileCheck %s --check-prefixes=CHECK,R01C,R01N -input-file=%T/non-optimized.cpp
// RUN: FileCheck %s --check-prefixes=CHECK,R01C,R01O -input-file=%T/optimized.cpp
// RUN: FileCheck %s --check-prefixes=CHECK,R02C,R02N,NOOPT -input-file=%T/non-optimized.cpp
// RUN: FileCheck %s --check-prefixes=CHECK,R02C,R02O -input-file=%T/optimized.cpp
// RUN: diff %T/default.cpp %T/optimized.cpp
include "llvm/Target/Target.td"
include "GlobalISelEmitterCommon.td"
//===- Define the necessary boilerplate for our test target. --------------===//
let TargetPrefix = "mytarget" in {
def int_mytarget_nop : Intrinsic<[llvm_i32_ty], [llvm_i32_ty], [IntrNoMem]>;
}
def complex : Operand<i32>, ComplexPattern<i32, 2, "SelectComplexPattern", []> {
let MIOperandInfo = (ops i32imm, i32imm);
}
def gi_complex :
[globalisel][tablegen] Revise API for ComplexPattern operands to improve flexibility. Summary: Some targets need to be able to do more complex rendering than just adding an operand or two to an instruction. For example, it may need to insert an instruction to extract a subreg first, or it may need to perform an operation on the operand. In SelectionDAG, targets would create SDNode's to achieve the desired effect during the complex pattern predicate. This worked because SelectionDAG had a form of garbage collection that would take care of SDNode's that were created but not used due to a later predicate rejecting a match. This doesn't translate well to GlobalISel and the churn was wasteful. The API changes in this patch enable GlobalISel to accomplish the same thing without the waste. The API is now: InstructionSelector::OptionalComplexRendererFn selectArithImmed(MachineOperand &Root) const; where Root is the root of the match. The return value can be omitted to indicate that the predicate failed to match, or a function with the signature ComplexRendererFn can be returned. For example: return OptionalComplexRendererFn( [=](MachineInstrBuilder &MIB) { MIB.addImm(Immed).addImm(ShVal); }); adds two immediate operands to the rendered instruction. Immed and ShVal are captured from the predicate function. As an added bonus, this also reduces the amount of information we need to provide to GIComplexOperandMatcher. Depends on D31418 Reviewers: aditya_nandakumar, t.p.northover, qcolombet, rovka, ab, javed.absar Reviewed By: ab Subscribers: dberris, kristof.beyls, igorb, llvm-commits Differential Revision: https://reviews.llvm.org/D31761 llvm-svn: 301079
2017-04-22 17:11:04 +02:00
GIComplexOperandMatcher<s32, "selectComplexPattern">,
GIComplexPatternEquiv<complex>;
def complex_rr : Operand<i32>, ComplexPattern<i32, 2, "SelectComplexPatternRR", []> {
let MIOperandInfo = (ops GPR32, GPR32);
}
def gi_complex_rr :
GIComplexOperandMatcher<s32, "selectComplexPatternRR">,
GIComplexPatternEquiv<complex_rr>;
def cimm8_xform : SDNodeXForm<imm, [{
uint64_t Val = N->getZExtValue() << 1;
return CurDAG->getTargetConstant(Val, SDLoc(N), MVT::i64);
}]>;
def cimm8 : Operand<i32>, ImmLeaf<i32, [{return isInt<8>(Imm);}], cimm8_xform>;
def gi_cimm8 : GICustomOperandRenderer<"renderImm">,
GISDNodeXFormEquiv<cimm8_xform>;
def gi_cimm9 : GICustomOperandRenderer<"renderImm">;
def m1 : OperandWithDefaultOps <i32, (ops (i32 -1))>;
def Z : OperandWithDefaultOps <i32, (ops R0)>;
def m1Z : OperandWithDefaultOps <i32, (ops (i32 -1), R0)>;
def HasA : Predicate<"Subtarget->hasA()">;
def HasB : Predicate<"Subtarget->hasB()">;
def HasC : Predicate<"Subtarget->hasC()"> { let RecomputePerFunction = 1; }
//===- Test the function boilerplate. -------------------------------------===//
// CHECK: const unsigned MAX_SUBTARGET_PREDICATES = 3;
// CHECK: using PredicateBitset = llvm::PredicateBitsetImpl<MAX_SUBTARGET_PREDICATES>;
// CHECK-LABEL: #ifdef GET_GLOBALISEL_TEMPORARIES_DECL
// CHECK-NEXT: mutable MatcherState State;
// CHECK-NEXT: typedef ComplexRendererFns(MyTargetInstructionSelector::*ComplexMatcherMemFn)(MachineOperand &) const;
// CHECK-NEXT: typedef void(MyTargetInstructionSelector::*CustomRendererFn)(MachineInstrBuilder &, const MachineInstr &, int) const;
// CHECK-NEXT: const ISelInfoTy<PredicateBitset, ComplexMatcherMemFn, CustomRendererFn> ISelInfo;
// CHECK-NEXT: static MyTargetInstructionSelector::ComplexMatcherMemFn ComplexPredicateFns[];
// CHECK-NEXT: static MyTargetInstructionSelector::CustomRendererFn CustomRenderers[];
// CHECK-NEXT: bool testImmPredicate_I64(unsigned PredicateID, int64_t Imm) const override;
// CHECK-NEXT: bool testImmPredicate_APInt(unsigned PredicateID, const APInt &Imm) const override;
// CHECK-NEXT: bool testImmPredicate_APFloat(unsigned PredicateID, const APFloat &Imm) const override;
// CHECK-NEXT: const int64_t *getMatchTable() const override;
// CHECK-NEXT: bool testMIPredicate_MI(unsigned PredicateID, const MachineInstr &MI, const std::array<const MachineOperand *, 3> &Operands) const override;
// CHECK-NEXT: #endif // ifdef GET_GLOBALISEL_TEMPORARIES_DECL
// CHECK-LABEL: #ifdef GET_GLOBALISEL_TEMPORARIES_INIT
// CHECK-NEXT: , State(2),
// CHECK-NEXT: ISelInfo(TypeObjects, NumTypeObjects, FeatureBitsets, ComplexPredicateFns, CustomRenderers)
// CHECK-NEXT: #endif // ifdef GET_GLOBALISEL_TEMPORARIES_INIT
// CHECK-LABEL: enum SubtargetFeatureBits : uint8_t {
// CHECK-NEXT: Feature_HasABit = 0,
// CHECK-NEXT: Feature_HasBBit = 1,
// CHECK-NEXT: Feature_HasCBit = 2,
// CHECK-NEXT: };
// CHECK-LABEL: PredicateBitset MyTargetInstructionSelector::
// CHECK-NEXT: computeAvailableModuleFeatures(const MyTargetSubtarget *Subtarget) const {
// CHECK-NEXT: PredicateBitset Features;
// CHECK-NEXT: if (Subtarget->hasA())
// CHECK-NEXT: Features.set(Feature_HasABit);
// CHECK-NEXT: if (Subtarget->hasB())
// CHECK-NEXT: Features.set(Feature_HasBBit);
// CHECK-NEXT: return Features;
// CHECK-NEXT: }
// CHECK-LABEL: PredicateBitset MyTargetInstructionSelector::
// CHECK-NEXT: computeAvailableFunctionFeatures(const MyTargetSubtarget *Subtarget, const MachineFunction *MF) const {
// CHECK-NEXT: PredicateBitset Features;
// CHECK-NEXT: if (Subtarget->hasC())
// CHECK-NEXT: Features.set(Feature_HasCBit);
// CHECK-NEXT: return Features;
// CHECK-NEXT: }
// CHECK-LABEL: // LLT Objects.
// CHECK-NEXT: enum {
// CHECK-NEXT: GILLT_p0s32
// CHECK-NEXT: GILLT_s32,
// CHECK-NEXT: }
// CHECK-NEXT: const static size_t NumTypeObjects = 2;
// CHECK-NEXT: const static LLT TypeObjects[] = {
// CHECK-NEXT: LLT::pointer(0, 32),
// CHECK-NEXT: LLT::scalar(32),
// CHECK-NEXT: };
// CHECK-LABEL: // Feature bitsets.
// CHECK-NEXT: enum {
// CHECK-NEXT: GIFBS_Invalid,
// CHECK-NEXT: GIFBS_HasA,
// CHECK-NEXT: GIFBS_HasA_HasB_HasC,
// CHECK-NEXT: }
// CHECK-NEXT: const static PredicateBitset FeatureBitsets[] {
// CHECK-NEXT: {}, // GIFBS_Invalid
// CHECK-NEXT: {Feature_HasABit, },
// CHECK-NEXT: {Feature_HasABit, Feature_HasBBit, Feature_HasCBit, },
// CHECK-NEXT: };
// CHECK-LABEL: // ComplexPattern predicates.
// CHECK-NEXT: enum {
// CHECK-NEXT: GICP_Invalid,
// CHECK-NEXT: GICP_gi_complex,
// CHECK-NEXT: GICP_gi_complex_rr,
// CHECK-NEXT: };
// CHECK-LABEL: // PatFrag predicates.
// CHECK-NEXT: enum {
// CHECK-NEXT: GIPFP_I64_Predicate_cimm8 = GIPFP_I64_Invalid + 1,
// CHECK-NEXT: GIPFP_I64_Predicate_simm8,
// CHECK-NEXT: };
// CHECK-NEXT: bool MyTargetInstructionSelector::testImmPredicate_I64(unsigned PredicateID, int64_t Imm) const {
// CHECK-NEXT: switch (PredicateID) {
// CHECK-NEXT: case GIPFP_I64_Predicate_cimm8: {
// CHECK-NEXT: return isInt<8>(Imm);
// CHECK-NEXT: llvm_unreachable("ImmediateCode should have returned");
// CHECK-NEXT: return false;
// CHECK-NEXT: }
// CHECK-NEXT: case GIPFP_I64_Predicate_simm8: {
// CHECK-NEXT: return isInt<8>(Imm);
// CHECK-NEXT: llvm_unreachable("ImmediateCode should have returned");
// CHECK-NEXT: return false;
// CHECK-NEXT: }
// CHECK-NEXT: }
// CHECK-NEXT: llvm_unreachable("Unknown predicate");
// CHECK-NEXT: return false;
// CHECK-NEXT: }
// CHECK-LABEL: // PatFrag predicates.
// CHECK-NEXT: enum {
// CHECK-NEXT: GIPFP_APFloat_Predicate_fpimmz = GIPFP_APFloat_Invalid + 1,
// CHECK-NEXT: };
// CHECK-NEXT: bool MyTargetInstructionSelector::testImmPredicate_APFloat(unsigned PredicateID, const APFloat & Imm) const {
// CHECK-NEXT: switch (PredicateID) {
// CHECK-NEXT: case GIPFP_APFloat_Predicate_fpimmz: {
// CHECK-NEXT: return Imm->isExactlyValue(0.0);
// CHECK-NEXT: llvm_unreachable("ImmediateCode should have returned");
// CHECK-NEXT: return false;
// CHECK-NEXT: }
// CHECK-NEXT: }
// CHECK-NEXT: llvm_unreachable("Unknown predicate");
// CHECK-NEXT: return false;
// CHECK-NEXT: }
// CHECK-LABEL: // PatFrag predicates.
// CHECK-NEXT: enum {
// CHECK-NEXT: GIPFP_APInt_Predicate_simm9 = GIPFP_APInt_Invalid + 1,
// CHECK-NEXT: };
// CHECK-NEXT: bool MyTargetInstructionSelector::testImmPredicate_APInt(unsigned PredicateID, const APInt & Imm) const {
// CHECK-NEXT: switch (PredicateID) {
// CHECK-NEXT: case GIPFP_APInt_Predicate_simm9: {
// CHECK-NEXT: return isInt<9>(Imm->getSExtValue());
// CHECK-NEXT: llvm_unreachable("ImmediateCode should have returned");
// CHECK-NEXT: return false;
// CHECK-NEXT: }
// CHECK-NEXT: }
// CHECK-NEXT: llvm_unreachable("Unknown predicate");
// CHECK-NEXT: return false;
// CHECK-NEXT: }
// CHECK-LABEL: MyTargetInstructionSelector::ComplexMatcherMemFn
// CHECK-NEXT: MyTargetInstructionSelector::ComplexPredicateFns[] = {
// CHECK-NEXT: nullptr, // GICP_Invalid
// CHECK-NEXT: &MyTargetInstructionSelector::selectComplexPattern, // gi_complex
// CHECK-NEXT: &MyTargetInstructionSelector::selectComplexPatternRR, // gi_complex_rr
// CHECK-NEXT: }
// CHECK-LABEL: // Custom renderers.
// CHECK-NEXT: enum {
// CHECK-NEXT: GICR_Invalid,
// CHECK-NEXT: GICR_renderImm,
// CHECK-NEXT: };
// CHECK-NEXT: MyTargetInstructionSelector::CustomRendererFn
// CHECK-NEXT: MyTargetInstructionSelector::CustomRenderers[] = {
// CHECK-NEXT: nullptr, // GICR_Invalid
// CHECK-NEXT: &MyTargetInstructionSelector::renderImm,
// CHECK-NEXT: };
2017-11-16 01:46:35 +01:00
// CHECK: bool MyTargetInstructionSelector::selectImpl(MachineInstr &I, CodeGenCoverage &CoverageInfo) const {
// CHECK-NEXT: MachineFunction &MF = *I.getParent()->getParent();
// CHECK-NEXT: MachineRegisterInfo &MRI = MF.getRegInfo();
// CHECK-NEXT: const PredicateBitset AvailableFeatures = getAvailableFeatures();
// CHECK-NEXT: NewMIVector OutMIs;
// CHECK-NEXT: State.MIs.clear();
// CHECK-NEXT: State.MIs.push_back(&I);
// CHECK: if (executeMatchTable(*this, OutMIs, State, ISelInfo, getMatchTable(), TII, MRI, TRI, RBI, AvailableFeatures, CoverageInfo)) {
// CHECK-NEXT: return true;
// CHECK-NEXT: }
// CHECK: const int64_t *
// CHECK-LABEL: MyTargetInstructionSelector::getMatchTable() const {
// CHECK-NEXT: MatchTable0[] = {
//===- Test a pattern with multiple ComplexPatterns in multiple instrs ----===//
//
// R19O-NEXT: GIM_SwitchOpcode, /*MI*/0, /*[*/{{[0-9]+}}, {{[0-9]+}}, /*)*//*default:*//*Label [[DEFAULT_NUM:[0-9]+]]*/ [[DEFAULT:[0-9]+]],
// R19O-NEXT: /*TargetOpcode::G_ADD*//*Label [[CASE_ADD_NUM:[0-9]+]]*/ [[CASE_ADD:[0-9]+]],
// R19O: /*TargetOpcode::G_SELECT*//*Label [[CASE_SELECT_NUM:[0-9]+]]*/ [[CASE_SELECT:[0-9]+]],
// R19O: // Label [[CASE_ADD_NUM]]: @[[CASE_ADD]]
// R19O: // Label [[CASE_SELECT_NUM]]: @[[CASE_SELECT]]
// R19O-NEXT: GIM_Try, /*On fail goto*//*Label [[GROUP_NUM:[0-9]+]]*/ [[GROUP:[0-9]+]],
// R19O-NEXT: GIM_CheckType, /*MI*/0, /*Op*/0, /*Type*/GILLT_s32,
// R19O-NEXT: GIM_CheckType, /*MI*/0, /*Op*/1, /*Type*/GILLT_s32,
// R19O-NEXT: GIM_CheckType, /*MI*/0, /*Op*/2, /*Type*/GILLT_s32,
// R19O-NEXT: GIM_CheckType, /*MI*/0, /*Op*/3, /*Type*/GILLT_s32,
//
// R19C-NEXT: GIM_Try, /*On fail goto*//*Label [[LABEL_NUM:[0-9]+]]*/ [[LABEL:[0-9]+]],
//
// R19O-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/0, /*RC*/MyTarget::GPR32RegClassID,
// R19O-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/1, /*RC*/MyTarget::GPR32RegClassID,
// R19N-NEXT: GIM_CheckNumOperands, /*MI*/0, /*Expected*/4,
// R19N-NEXT: GIM_CheckOpcode, /*MI*/0, TargetOpcode::G_SELECT,
// R19N-NEXT: // MIs[0] dst
// R19N-NEXT: GIM_CheckType, /*MI*/0, /*Op*/0, /*Type*/GILLT_s32,
// R19N-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/0, /*RC*/MyTarget::GPR32RegClassID,
// R19N-NEXT: // MIs[0] src1
// R19N-NEXT: GIM_CheckType, /*MI*/0, /*Op*/1, /*Type*/GILLT_s32,
// R19N-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/1, /*RC*/MyTarget::GPR32RegClassID,
// R19N-NEXT: // MIs[0] complex_rr:src2a:src2b
// R19N-NEXT: GIM_CheckType, /*MI*/0, /*Op*/2, /*Type*/GILLT_s32,
//
// R19N-NEXT: GIM_CheckComplexPattern, /*MI*/0, /*Op*/2, /*Renderer*/0, GICP_gi_complex_rr,
// R19N-NEXT: // MIs[0] Operand 3
// R19N-NEXT: GIM_CheckType, /*MI*/0, /*Op*/3, /*Type*/GILLT_s32,
// R19C-NEXT: GIM_RecordInsn, /*DefineMI*/1, /*MI*/0, /*OpIdx*/3, // MIs[1]
// R19N-NEXT: GIM_CheckNumOperands, /*MI*/1, /*Expected*/4,
// R19C-NEXT: GIM_CheckOpcode, /*MI*/1, TargetOpcode::G_SELECT,
// R19N-NEXT: // MIs[1] Operand 0
// R19N-NEXT: GIM_CheckType, /*MI*/1, /*Op*/0, /*Type*/GILLT_s32,
// R19N-NEXT: // MIs[1] src3
// R19C-NEXT: GIM_CheckType, /*MI*/1, /*Op*/1, /*Type*/GILLT_s32,
// R19O-NEXT: GIM_CheckType, /*MI*/1, /*Op*/2, /*Type*/GILLT_s32,
// R19O-NEXT: GIM_CheckType, /*MI*/1, /*Op*/3, /*Type*/GILLT_s32,
// R19N-NEXT: GIM_CheckRegBankForClass, /*MI*/1, /*Op*/1, /*RC*/MyTarget::GPR32RegClassID,
// R19N-NEXT: // MIs[1] src4
// R19N-NEXT: GIM_CheckType, /*MI*/1, /*Op*/2, /*Type*/GILLT_s32,
// R19N-NEXT: GIM_CheckComplexPattern, /*MI*/1, /*Op*/2, /*Renderer*/1, GICP_gi_complex,
// R19N-NEXT: // MIs[1] complex:src5a:src5b
// R19N-NEXT: GIM_CheckType, /*MI*/1, /*Op*/3, /*Type*/GILLT_s32,
// R19N-NEXT: GIM_CheckComplexPattern, /*MI*/1, /*Op*/3, /*Renderer*/2, GICP_gi_complex,
// R19O-NEXT: GIM_CheckRegBankForClass, /*MI*/1, /*Op*/1, /*RC*/MyTarget::GPR32RegClassID,
// R19C-NEXT: GIM_CheckIsSafeToFold, /*InsnID*/1,
// R19O-NEXT: GIM_CheckComplexPattern, /*MI*/0, /*Op*/2, /*Renderer*/0, GICP_gi_complex_rr,
// R19O-NEXT: GIM_CheckComplexPattern, /*MI*/1, /*Op*/2, /*Renderer*/1, GICP_gi_complex,
// R19O-NEXT: GIM_CheckComplexPattern, /*MI*/1, /*Op*/3, /*Renderer*/2, GICP_gi_complex,
// R19C-NEXT: // (select:{ *:[i32] } GPR32:{ *:[i32] }:$src1, (complex_rr:{ *:[i32] } GPR32:{ *:[i32] }:$src2a, GPR32:{ *:[i32] }:$src2b), (select:{ *:[i32] } GPR32:{ *:[i32] }:$src3, complex:{ *:[i32] }:$src4, (complex:{ *:[i32] } i32imm:{ *:[i32] }:$src5a, i32imm:{ *:[i32] }:$src5b))) => (INSN3:{ *:[i32] } GPR32:{ *:[i32] }:$src1, GPR32:{ *:[i32] }:$src2b, GPR32:{ *:[i32] }:$src2a, (INSN4:{ *:[i32] } GPR32:{ *:[i32] }:$src3, complex:{ *:[i32] }:$src4, i32imm:{ *:[i32] }:$src5a, i32imm:{ *:[i32] }:$src5b))
// R19C-NEXT: GIR_MakeTempReg, /*TempRegID*/0, /*TypeID*/GILLT_s32,
// R19C-NEXT: GIR_BuildMI, /*InsnID*/1, /*Opcode*/MyTarget::INSN4,
// R19C-NEXT: GIR_AddTempRegister, /*InsnID*/1, /*TempRegID*/0, /*TempRegFlags*/RegState::Define,
// R19C-NEXT: GIR_Copy, /*NewInsnID*/1, /*OldInsnID*/1, /*OpIdx*/1, // src3
// R19C-NEXT: GIR_ComplexRenderer, /*InsnID*/1, /*RendererID*/1,
// R19C-NEXT: GIR_ComplexSubOperandRenderer, /*InsnID*/1, /*RendererID*/2, /*SubOperand*/0, // src5a
// R19C-NEXT: GIR_ComplexSubOperandRenderer, /*InsnID*/1, /*RendererID*/2, /*SubOperand*/1, // src5b
// R19C-NEXT: GIR_ConstrainSelectedInstOperands, /*InsnID*/1,
// R19C-NEXT: GIR_BuildMI, /*InsnID*/0, /*Opcode*/MyTarget::INSN3,
// R19C-NEXT: GIR_Copy, /*NewInsnID*/0, /*OldInsnID*/0, /*OpIdx*/0, // dst
// R19C-NEXT: GIR_Copy, /*NewInsnID*/0, /*OldInsnID*/0, /*OpIdx*/1, // src1
// R19C-NEXT: GIR_ComplexSubOperandRenderer, /*InsnID*/0, /*RendererID*/0, /*SubOperand*/1, // src2b
// R19C-NEXT: GIR_ComplexSubOperandRenderer, /*InsnID*/0, /*RendererID*/0, /*SubOperand*/0, // src2a
// R19C-NEXT: GIR_AddTempRegister, /*InsnID*/0, /*TempRegID*/0, /*TempRegFlags*/0,
// R19C-NEXT: GIR_EraseFromParent, /*InsnID*/0,
// R19C-NEXT: GIR_ConstrainSelectedInstOperands, /*InsnID*/0,
// R19C-NEXT: // GIR_Coverage, 19,
// R19C-NEXT: GIR_Done,
// R19C-NEXT: // Label [[LABEL_NUM]]: @[[LABEL]]
//
// R19O: // Label [[GROUP_NUM]]: @[[GROUP]]
// R19O-NEXT: GIM_Reject,
// R19O: // Label [[DEFAULT_NUM]]: @[[DEFAULT]]
// R19O-NEXT: GIM_Reject,
// R19O-NEXT: };
def INSN3 : I<(outs GPR32:$dst),
(ins GPR32Op:$src1, GPR32:$src2a, GPR32:$src2b, GPR32:$scr), []>;
def INSN4 : I<(outs GPR32:$scr),
(ins GPR32:$src3, complex:$src4, i32imm:$src5a, i32imm:$src5b), []>;
def : Pat<(select GPR32:$src1, (complex_rr GPR32:$src2a, GPR32:$src2b),
(select GPR32:$src3,
complex:$src4,
(complex i32imm:$src5a, i32imm:$src5b))),
(INSN3 GPR32:$src1, GPR32:$src2b, GPR32:$src2a,
(INSN4 GPR32:$src3, complex:$src4, i32imm:$src5a,
i32imm:$src5b))>;
// R21O-NEXT: GIM_SwitchOpcode, /*MI*/0, /*[*/{{[0-9]+}}, {{[0-9]+}}, /*)*//*default:*//*Label [[DEFAULT_NUM:[0-9]+]]*/ [[DEFAULT:[0-9]+]],
// R21O-NEXT: /*TargetOpcode::G_ADD*//*Label [[CASE_ADD_NUM:[0-9]+]]*/ [[CASE_ADD:[0-9]+]],
// R21O: /*TargetOpcode::G_SELECT*//*Label [[CASE_SELECT_NUM:[0-9]+]]*/ [[CASE_SELECT:[0-9]+]],
// R21O: // Label [[CASE_ADD_NUM]]: @[[CASE_ADD]]
// R21O: // Label [[CASE_SELECT_NUM]]: @[[CASE_SELECT]]
// R21O-NEXT: GIM_Try, /*On fail goto*//*Label [[GROUP_NUM:[0-9]+]]*/ [[GROUP:[0-9]+]],
// R21O-NEXT: GIM_CheckType, /*MI*/0, /*Op*/0, /*Type*/GILLT_s32,
// R21O-NEXT: GIM_CheckType, /*MI*/0, /*Op*/1, /*Type*/GILLT_s32,
// R21O-NEXT: GIM_CheckType, /*MI*/0, /*Op*/2, /*Type*/GILLT_s32,
// R21O-NEXT: GIM_CheckType, /*MI*/0, /*Op*/3, /*Type*/GILLT_s32,
//
// R21C-NEXT: GIM_Try, /*On fail goto*//*Label [[PREV_NUM:[0-9]+]]*/ [[PREV:[0-9]+]], // Rule ID 19 //
// R21C-NOT: GIR_Done,
// R21C: // GIR_Coverage, 19,
// R21C-NEXT: GIR_Done,
// R21C-NEXT: // Label [[PREV_NUM]]: @[[PREV]]
// R21C-NEXT: GIM_Try, /*On fail goto*//*Label [[LABEL_NUM:[0-9]+]]*/ [[LABEL:[0-9]+]], // Rule ID 21 //
//
// R21O-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/0, /*RC*/MyTarget::GPR32RegClassID,
// R21O-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/1, /*RC*/MyTarget::GPR32RegClassID,
// R21N-NEXT: GIM_CheckNumOperands, /*MI*/0, /*Expected*/4,
// R21N-NEXT: GIM_CheckOpcode, /*MI*/0, TargetOpcode::G_SELECT,
// R21N-NEXT: // MIs[0] dst
// R21N-NEXT: GIM_CheckType, /*MI*/0, /*Op*/0, /*Type*/GILLT_s32,
// R21N-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/0, /*RC*/MyTarget::GPR32RegClassID,
// R21N-NEXT: // MIs[0] src1
// R21N-NEXT: GIM_CheckType, /*MI*/0, /*Op*/1, /*Type*/GILLT_s32,
// R21N-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/1, /*RC*/MyTarget::GPR32RegClassID,
// R21N-NEXT: // MIs[0] src2
// R21N-NEXT: GIM_CheckType, /*MI*/0, /*Op*/2, /*Type*/GILLT_s32,
//
// R21O-NEXT: GIM_CheckCxxInsnPredicate, /*MI*/0, /*FnId*/GIPFP_MI_Predicate_frag,
// R21C-NEXT: GIM_CheckComplexPattern, /*MI*/0, /*Op*/2, /*Renderer*/0, GICP_gi_complex,
// R21N-NEXT: // MIs[0] src3
// R21N-NEXT: GIM_CheckType, /*MI*/0, /*Op*/3, /*Type*/GILLT_s32,
// R21C-NEXT: GIM_CheckComplexPattern, /*MI*/0, /*Op*/3, /*Renderer*/1, GICP_gi_complex,
// R21N-NEXT: GIM_CheckCxxInsnPredicate, /*MI*/0, /*FnId*/GIPFP_MI_Predicate_frag,
// R21C-NEXT: // (select:{ *:[i32] } GPR32:{ *:[i32] }:$src1, complex:{ *:[i32] }:$src2, complex:{ *:[i32] }:$src3)<<P:Predicate_frag>> => (INSN2:{ *:[i32] } GPR32:{ *:[i32] }:$src1, complex:{ *:[i32] }:$src3, complex:{ *:[i32] }:$src2)
// R21C-NEXT: GIR_BuildMI, /*InsnID*/0, /*Opcode*/MyTarget::INSN2,
// R21C-NEXT: GIR_Copy, /*NewInsnID*/0, /*OldInsnID*/0, /*OpIdx*/0, // dst
// R21C-NEXT: GIR_Copy, /*NewInsnID*/0, /*OldInsnID*/0, /*OpIdx*/1, // src1
// R21C-NEXT: GIR_ComplexRenderer, /*InsnID*/0, /*RendererID*/1,
// R21C-NEXT: GIR_ComplexRenderer, /*InsnID*/0, /*RendererID*/0,
// R21C-NEXT: GIR_MergeMemOperands, /*InsnID*/0, /*MergeInsnID's*/0, GIU_MergeMemOperands_EndOfList,
// R21C-NEXT: GIR_EraseFromParent, /*InsnID*/0,
// R21C-NEXT: GIR_ConstrainSelectedInstOperands, /*InsnID*/0,
// R21C-NEXT: // GIR_Coverage, 21,
// R21C-NEXT: GIR_Done,
// R21C-NEXT: // Label [[LABEL_NUM]]: @[[LABEL]]
//
// R21O-NEXT: GIM_Reject,
// R21O-NEXT: // Label [[GROUP_NUM]]: @[[GROUP]]
// R21O-NEXT: GIM_Reject,
// R21O: // Label [[DEFAULT_NUM]]: @[[DEFAULT]]
// R21O-NEXT: GIM_Reject,
// R21O-NEXT: };
//===- Test a pattern with ComplexPattern operands. -----------------------===//
//
// R20O-NEXT: GIM_SwitchOpcode, /*MI*/0, /*[*/{{[0-9]+}}, {{[0-9]+}}, /*)*//*default:*//*Label [[DEFAULT_NUM:[0-9]+]]*/ [[DEFAULT:[0-9]+]],
// R20O-NEXT: /*TargetOpcode::G_ADD*//*Label [[CASE_ADD_NUM:[0-9]+]]*/ [[CASE_ADD:[0-9]+]],
// R20O: /*TargetOpcode::G_SUB*//*Label [[CASE_SUB_NUM:[0-9]+]]*/ [[CASE_SUB:[0-9]+]],
// R20O: // Label [[CASE_ADD_NUM]]: @[[CASE_ADD]]
// R20O: // Label [[CASE_SUB_NUM]]: @[[CASE_SUB]]
// R20O-NEXT: GIM_Try, /*On fail goto*//*Label [[GROUP_NUM:[0-9]+]]*/ [[GROUP:[0-9]+]],
// R20O-NEXT: GIM_CheckType, /*MI*/0, /*Op*/0, /*Type*/GILLT_s32,
// R20O-NEXT: GIM_CheckType, /*MI*/0, /*Op*/1, /*Type*/GILLT_s32,
// R20O-NEXT: GIM_CheckType, /*MI*/0, /*Op*/2, /*Type*/GILLT_s32,
// R20O-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/0, /*RC*/MyTarget::GPR32RegClassID,
//
// R20N: GIM_Try, /*On fail goto*//*Label [[PREV_NUM:[0-9]+]]*/ [[PREV:[0-9]+]], // Rule ID 21 //
// R20N: // Label [[PREV_NUM]]: @[[PREV]]
//
// R20C-NEXT: GIM_Try, /*On fail goto*//*Label [[LABEL_NUM:[0-9]+]]*/ [[LABEL:[0-9]+]], // Rule ID 20 //
//
// R20N-NEXT: GIM_CheckNumOperands, /*MI*/0, /*Expected*/3,
// R20N-NEXT: GIM_CheckOpcode, /*MI*/0, TargetOpcode::G_SUB,
// R20N-NEXT: // MIs[0] dst
// R20N-NEXT: GIM_CheckType, /*MI*/0, /*Op*/0, /*Type*/GILLT_s32,
// R20N-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/0, /*RC*/MyTarget::GPR32RegClassID,
// R20N-NEXT: // MIs[0] src1
// R20N-NEXT: GIM_CheckType, /*MI*/0, /*Op*/1, /*Type*/GILLT_s32,
//
// R20N-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/1, /*RC*/MyTarget::GPR32RegClassID,
// R20N-NEXT: // MIs[0] src2
// R20N-NEXT: GIM_CheckType, /*MI*/0, /*Op*/2, /*Type*/GILLT_s32,
// R20O-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/1, /*RC*/MyTarget::GPR32RegClassID,
// R20C-NEXT: GIM_CheckComplexPattern, /*MI*/0, /*Op*/2, /*Renderer*/0, GICP_gi_complex,
// R20C-NEXT: // (sub:{ *:[i32] } GPR32:{ *:[i32] }:$src1, complex:{ *:[i32] }:$src2) => (INSN1:{ *:[i32] } GPR32:{ *:[i32] }:$src1, complex:{ *:[i32] }:$src2)
// R20C-NEXT: GIR_BuildMI, /*InsnID*/0, /*Opcode*/MyTarget::INSN1,
// R20C-NEXT: GIR_Copy, /*NewInsnID*/0, /*OldInsnID*/0, /*OpIdx*/0, // dst
// R20C-NEXT: GIR_Copy, /*NewInsnID*/0, /*OldInsnID*/0, /*OpIdx*/1, // src1
// R20C-NEXT: GIR_ComplexRenderer, /*InsnID*/0, /*RendererID*/0,
// R20C-NEXT: GIR_EraseFromParent, /*InsnID*/0,
// R20C-NEXT: GIR_ConstrainSelectedInstOperands, /*InsnID*/0,
// R20C-NEXT: // GIR_Coverage, 20,
// R20C-NEXT: GIR_Done,
// R20C-NEXT: // Label [[LABEL_NUM]]: @[[LABEL]]
//
// R20O: // Label [[GROUP_NUM]]: @[[GROUP]]
// R20O-NEXT: GIM_Reject,
// R20O: // Label [[DEFAULT_NUM]]: @[[DEFAULT]]
// R20O-NEXT: GIM_Reject,
// R20O-NEXT: };
def INSN1 : I<(outs GPR32:$dst), (ins GPR32:$src1, complex:$src2), []>;
def : Pat<(sub GPR32:$src1, complex:$src2), (INSN1 GPR32:$src1, complex:$src2)>;
//===- Test a pattern with multiple ComplexPattern operands. --------------===//
//
def : GINodeEquiv<G_SELECT, select>;
let mayLoad = 1 in {
def INSN2 : I<(outs GPR32:$dst), (ins GPR32Op:$src1, complex:$src2, complex:$src3), []>;
}
def frag : PatFrag<(ops node:$a, node:$b, node:$c),
(select node:$a, node:$b, node:$c),
[{ return true; // C++ code }]> {
let GISelPredicateCode = [{ return true; // C++ code }];
}
def : Pat<(frag GPR32:$src1, complex:$src2, complex:$src3),
(INSN2 GPR32:$src1, complex:$src3, complex:$src2)>;
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 16:43:30 +01:00
//===- Test a more complex multi-instruction match. -----------------------===//
//
// R00O-NEXT: GIM_SwitchOpcode, /*MI*/0, /*[*/{{[0-9]+}}, {{[0-9]+}}, /*)*//*default:*//*Label [[DEFAULT_NUM:[0-9]+]]*/ [[DEFAULT:[0-9]+]],
// R00O-NEXT: /*TargetOpcode::G_ADD*//*Label [[CASE_ADD_NUM:[0-9]+]]*/ [[CASE_ADD:[0-9]+]],
// R00O: /*TargetOpcode::G_SUB*//*Label [[CASE_SUB_NUM:[0-9]+]]*/ [[CASE_SUB:[0-9]+]],
// R00O: // Label [[CASE_ADD_NUM]]: @[[CASE_ADD]]
// R00O: // Label [[CASE_SUB_NUM]]: @[[CASE_SUB]]
// R00O-NEXT: GIM_Try, /*On fail goto*//*Label [[GROUP_NUM:[0-9]+]]*/ [[GROUP:[0-9]+]],
// R00O-NEXT: GIM_CheckType, /*MI*/0, /*Op*/0, /*Type*/GILLT_s32,
// R00O-NEXT: GIM_CheckType, /*MI*/0, /*Op*/1, /*Type*/GILLT_s32,
// R00O-NEXT: GIM_CheckType, /*MI*/0, /*Op*/2, /*Type*/GILLT_s32,
// R00O-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/0, /*RC*/MyTarget::GPR32RegClassID,
//
// R00C: GIM_Try, /*On fail goto*//*Label [[PREV_NUM:[0-9]+]]*/ [[PREV:[0-9]+]], // Rule ID 20 //
// R00C: // Label [[PREV_NUM]]: @[[PREV]]
//
// R00C-NEXT: GIM_Try, /*On fail goto*//*Label [[LABEL_NUM:[0-9]+]]*/ [[LABEL:[0-9]+]], // Rule ID 0 //
// R00C-NEXT: GIM_CheckFeatures, GIFBS_HasA,
// R00N-NEXT: GIM_CheckNumOperands, /*MI*/0, /*Expected*/3,
// R00N-NEXT: GIM_CheckOpcode, /*MI*/0, TargetOpcode::G_SUB,
// R00N-NEXT: // MIs[0] dst
// R00N-NEXT: GIM_CheckType, /*MI*/0, /*Op*/0, /*Type*/GILLT_s32,
// R00N-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/0, /*RC*/MyTarget::GPR32RegClassID,
// R00N-NEXT: // MIs[0] Operand 1
// R00N-NEXT: GIM_CheckType, /*MI*/0, /*Op*/1, /*Type*/GILLT_s32,
// R00C-NEXT: GIM_RecordInsn, /*DefineMI*/1, /*MI*/0, /*OpIdx*/1, // MIs[1]
// R00N-NEXT: GIM_CheckNumOperands, /*MI*/1, /*Expected*/3,
// R00C-NEXT: GIM_CheckOpcode, /*MI*/1, TargetOpcode::G_SUB,
// R00N-NEXT: // MIs[1] Operand 0
// R00N-NEXT: GIM_CheckType, /*MI*/1, /*Op*/0, /*Type*/GILLT_s32,
// R00N-NEXT: // MIs[1] src1
// R00C-NEXT: GIM_CheckType, /*MI*/1, /*Op*/1, /*Type*/GILLT_s32,
// R00O-NEXT: GIM_CheckType, /*MI*/1, /*Op*/2, /*Type*/GILLT_s32,
// R00N-NEXT: GIM_CheckRegBankForClass, /*MI*/1, /*Op*/1, /*RC*/MyTarget::GPR32RegClassID,
// R00N-NEXT: // MIs[1] src2
// R00N-NEXT: GIM_CheckType, /*MI*/1, /*Op*/2, /*Type*/GILLT_s32,
// R00N-NEXT: GIM_CheckRegBankForClass, /*MI*/1, /*Op*/2, /*RC*/MyTarget::GPR32RegClassID,
// R00N-NEXT: // MIs[0] Operand 2
// R00N-NEXT: GIM_CheckType, /*MI*/0, /*Op*/2, /*Type*/GILLT_s32,
// R00O-NEXT: GIM_CheckRegBankForClass, /*MI*/1, /*Op*/1, /*RC*/MyTarget::GPR32RegClassID,
// R00O-NEXT: GIM_CheckRegBankForClass, /*MI*/1, /*Op*/2, /*RC*/MyTarget::GPR32RegClassID,
// R00C-NEXT: GIM_RecordInsn, /*DefineMI*/2, /*MI*/0, /*OpIdx*/2, // MIs[2]
// R00N-NEXT: GIM_CheckNumOperands, /*MI*/2, /*Expected*/3,
// R00C-NEXT: GIM_CheckOpcode, /*MI*/2, TargetOpcode::G_SUB,
// R00N-NEXT: // MIs[2] Operand 0
// R00N-NEXT: GIM_CheckType, /*MI*/2, /*Op*/0, /*Type*/GILLT_s32,
// R00N-NEXT: // MIs[2] src3
// R00C-NEXT: GIM_CheckType, /*MI*/2, /*Op*/1, /*Type*/GILLT_s32,
// R00O-NEXT: GIM_CheckType, /*MI*/2, /*Op*/2, /*Type*/GILLT_s32,
// R00N-NEXT: GIM_CheckRegBankForClass, /*MI*/2, /*Op*/1, /*RC*/MyTarget::GPR32RegClassID,
// R00N-NEXT: // MIs[2] src4
// R00N-NEXT: GIM_CheckType, /*MI*/2, /*Op*/2, /*Type*/GILLT_s32,
// R00N-NEXT: GIM_CheckRegBankForClass, /*MI*/2, /*Op*/2, /*RC*/MyTarget::GPR32RegClassID,
// R00O-NEXT: GIM_CheckRegBankForClass, /*MI*/2, /*Op*/1, /*RC*/MyTarget::GPR32RegClassID,
// R00O-NEXT: GIM_CheckRegBankForClass, /*MI*/2, /*Op*/2, /*RC*/MyTarget::GPR32RegClassID,
// R00C-NEXT: GIM_CheckIsSafeToFold, /*InsnID*/1,
// R00C-NEXT: GIM_CheckIsSafeToFold, /*InsnID*/2,
// R00C-NEXT: // (sub:{ *:[i32] } (sub:{ *:[i32] } GPR32:{ *:[i32] }:$src1, GPR32:{ *:[i32] }:$src2), (sub:{ *:[i32] } GPR32:{ *:[i32] }:$src3, GPR32:{ *:[i32] }:$src4)) => (INSNBOB:{ *:[i32] } GPR32:{ *:[i32] }:$src1, GPR32:{ *:[i32] }:$src2, GPR32:{ *:[i32] }:$src3, GPR32:{ *:[i32] }:$src4)
// R00C-NEXT: GIR_BuildMI, /*InsnID*/0, /*Opcode*/MyTarget::INSNBOB,
// R00C-NEXT: GIR_Copy, /*NewInsnID*/0, /*OldInsnID*/0, /*OpIdx*/0, // dst
// R00C-NEXT: GIR_Copy, /*NewInsnID*/0, /*OldInsnID*/1, /*OpIdx*/1, // src1
// R00C-NEXT: GIR_Copy, /*NewInsnID*/0, /*OldInsnID*/1, /*OpIdx*/2, // src2
// R00C-NEXT: GIR_Copy, /*NewInsnID*/0, /*OldInsnID*/2, /*OpIdx*/1, // src3
// R00C-NEXT: GIR_Copy, /*NewInsnID*/0, /*OldInsnID*/2, /*OpIdx*/2, // src4
// R00C-NEXT: GIR_EraseFromParent, /*InsnID*/0,
// R00C-NEXT: GIR_ConstrainSelectedInstOperands, /*InsnID*/0,
// R00C-NEXT: // GIR_Coverage, 0,
// R00C-NEXT: GIR_Done,
// R00C-NEXT: // Label [[LABEL_NUM]]: @[[LABEL]]
//
// R00O-NEXT: GIM_Reject,
// R00O-NEXT: // Label [[GROUP_NUM]]: @[[GROUP]]
// R00O-NEXT: GIM_Reject,
// R00O: // Label [[DEFAULT_NUM]]: @[[DEFAULT]]
// R00O-NEXT: GIM_Reject,
// R00O-NEXT: };
def INSNBOB : I<(outs GPR32:$dst), (ins GPR32:$src1, GPR32:$src2, GPR32:$src3, GPR32:$src4),
[(set GPR32:$dst,
(sub (sub GPR32:$src1, GPR32:$src2), (sub GPR32:$src3, GPR32:$src4)))]>,
Requires<[HasA]>;
//===- Test a simple pattern with an intrinsic. ---------------------------===//
//
// R01O-NEXT: GIM_SwitchOpcode, /*MI*/0, /*[*/{{[0-9]+}}, {{[0-9]+}}, /*)*//*default:*//*Label [[DEFAULT_NUM:[0-9]+]]*/ [[DEFAULT:[0-9]+]],
// R01O-NEXT: /*TargetOpcode::G_ADD*//*Label [[CASE_ADD_NUM:[0-9]+]]*/ [[CASE_ADD:[0-9]+]],
// R01O: /*TargetOpcode::G_INTRINSIC*//*Label [[CASE_INTRINSIC_NUM:[0-9]+]]*/ [[CASE_INTRINSIC:[0-9]+]],
// R01O: // Label [[CASE_ADD_NUM]]: @[[CASE_ADD]]
// R01O: // Label [[CASE_INTRINSIC_NUM]]: @[[CASE_INTRINSIC]]
//
// R01N: GIM_Try, /*On fail goto*//*Label [[PREV_NUM:[0-9]+]]*/ [[PREV:[0-9]+]], // Rule ID 0 //
// R01N: // Label [[PREV_NUM]]: @[[PREV]]
//
// R01C-NEXT: GIM_Try, /*On fail goto*//*Label [[LABEL_NUM:[0-9]+]]*/ [[LABEL:[0-9]+]], // Rule ID 1 //
// R01C-NEXT: GIM_CheckNumOperands, /*MI*/0, /*Expected*/3,
//
// R01O-NEXT: GIM_CheckIntrinsicID, /*MI*/0, /*Op*/1, Intrinsic::mytarget_nop,
// R01O-NEXT: GIM_CheckType, /*MI*/0, /*Op*/0, /*Type*/GILLT_s32,
// R01O-NEXT: GIM_CheckType, /*MI*/0, /*Op*/2, /*Type*/GILLT_s32,
// R01O-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/0, /*RC*/MyTarget::GPR32RegClassID,
//
// R01N-NEXT: GIM_CheckOpcode, /*MI*/0, TargetOpcode::G_INTRINSIC,
// R01N-NEXT: // MIs[0] dst
// R01N-NEXT: GIM_CheckType, /*MI*/0, /*Op*/0, /*Type*/GILLT_s32,
// R01N-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/0, /*RC*/MyTarget::GPR32RegClassID,
// R01N-NEXT: // MIs[0] Operand 1
// R01N-NEXT: GIM_CheckIntrinsicID, /*MI*/0, /*Op*/1, Intrinsic::mytarget_nop,
// R01N-NEXT: // MIs[0] src1
// R01N-NEXT: GIM_CheckType, /*MI*/0, /*Op*/2, /*Type*/GILLT_s32,
//
// R01C-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/2, /*RC*/MyTarget::GPR32RegClassID,
// R01C-NEXT: // (intrinsic_wo_chain:{ *:[i32] } [[ID:[0-9]+]]:{ *:[iPTR] }, GPR32:{ *:[i32] }:$src1) => (MOV:{ *:[i32] } GPR32:{ *:[i32] }:$src1)
// R01C-NEXT: GIR_BuildMI, /*InsnID*/0, /*Opcode*/MyTarget::MOV,
// R01C-NEXT: GIR_Copy, /*NewInsnID*/0, /*OldInsnID*/0, /*OpIdx*/0, // dst
// R01C-NEXT: GIR_Copy, /*NewInsnID*/0, /*OldInsnID*/0, /*OpIdx*/2, // src1
// R01C-NEXT: GIR_EraseFromParent, /*InsnID*/0,
// R01C-NEXT: GIR_ConstrainSelectedInstOperands, /*InsnID*/0,
// R01C-NEXT: // GIR_Coverage, 1,
// R01C-NEXT: GIR_Done,
// R01C-NEXT: // Label [[LABEL_NUM]]: @[[LABEL]]
//
// R01O-NEXT: GIM_Reject,
// R01O: // Label [[DEFAULT_NUM]]: @[[DEFAULT]]
// R01O-NEXT: GIM_Reject,
def MOV : I<(outs GPR32:$dst), (ins GPR32:$src1),
[(set GPR32:$dst, (int_mytarget_nop GPR32:$src1))]>;
//===- Test a simple pattern with a default operand. ----------------------===//
//
// R02O-NEXT: GIM_SwitchOpcode, /*MI*/0, /*[*/{{[0-9]+}}, {{[0-9]+}}, /*)*//*default:*//*Label [[DEFAULT_NUM:[0-9]+]]*/ [[DEFAULT:[0-9]+]],
// R02O-NEXT: /*TargetOpcode::G_ADD*//*Label [[CASE_ADD_NUM:[0-9]+]]*/ [[CASE_ADD:[0-9]+]],
// R02O: /*TargetOpcode::G_XOR*//*Label [[CASE_XOR_NUM:[0-9]+]]*/ [[CASE_XOR:[0-9]+]],
// R02O: // Label [[CASE_ADD_NUM]]: @[[CASE_ADD]]
// R02O: // Label [[CASE_XOR_NUM]]: @[[CASE_XOR]]
// R02O-NEXT: GIM_Try, /*On fail goto*//*Label [[GROUP_NUM:[0-9]+]]*/ [[GROUP:[0-9]+]],
// R02O-NEXT: GIM_CheckType, /*MI*/0, /*Op*/0, /*Type*/GILLT_s32,
// R02O-NEXT: GIM_CheckType, /*MI*/0, /*Op*/1, /*Type*/GILLT_s32,
// R02O-NEXT: GIM_CheckType, /*MI*/0, /*Op*/2, /*Type*/GILLT_s32,
// R02O-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/0, /*RC*/MyTarget::GPR32RegClassID,
// R02O-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/1, /*RC*/MyTarget::GPR32RegClassID,
//
// R02N: GIM_Try, /*On fail goto*//*Label [[PREV_NUM:[0-9]+]]*/ [[PREV:[0-9]+]], // Rule ID 1 //
// R02N: // Label [[PREV_NUM]]: @[[PREV]]
//
// R02C-NEXT: GIM_Try, /*On fail goto*//*Label [[LABEL_NUM:[0-9]+]]*/ [[LABEL:[0-9]+]], // Rule ID 2 //
//
// R02N-NEXT: GIM_CheckNumOperands, /*MI*/0, /*Expected*/3,
// R02N-NEXT: GIM_CheckOpcode, /*MI*/0, TargetOpcode::G_XOR,
// R02N-NEXT: // MIs[0] dst
// R02N-NEXT: GIM_CheckType, /*MI*/0, /*Op*/0, /*Type*/GILLT_s32,
// R02N-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/0, /*RC*/MyTarget::GPR32RegClassID,
// R02N-NEXT: // MIs[0] src1
// R02N-NEXT: GIM_CheckType, /*MI*/0, /*Op*/1, /*Type*/GILLT_s32,
// R02N-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/1, /*RC*/MyTarget::GPR32RegClassID,
// R02N-NEXT: // MIs[0] Operand 2
// R02N-NEXT: GIM_CheckType, /*MI*/0, /*Op*/2, /*Type*/GILLT_s32,
//
// R02C-NEXT: GIM_CheckConstantInt, /*MI*/0, /*Op*/2, -2
// R02C-NEXT: // (xor:{ *:[i32] } GPR32:{ *:[i32] }:$src1, -2:{ *:[i32] }) => (XORI:{ *:[i32] } GPR32:{ *:[i32] }:$src1)
// R02C-NEXT: GIR_BuildMI, /*InsnID*/0, /*Opcode*/MyTarget::XORI,
// R02C-NEXT: GIR_Copy, /*NewInsnID*/0, /*OldInsnID*/0, /*OpIdx*/0, // dst
// R02C-NEXT: GIR_AddImm, /*InsnID*/0, /*Imm*/-1,
// R02C-NEXT: GIR_Copy, /*NewInsnID*/0, /*OldInsnID*/0, /*OpIdx*/1, // src1
// R02C-NEXT: GIR_EraseFromParent, /*InsnID*/0,
// R02C-NEXT: GIR_ConstrainSelectedInstOperands, /*InsnID*/0,
// R02C-NEXT: // GIR_Coverage, 2,
// R02C-NEXT: GIR_Done,
// R02C-NEXT: // Label [[LABEL_NUM]]: @[[LABEL]]
//
// R02O: // Label [[DEFAULT_NUM]]: @[[DEFAULT]]
// R02O-NEXT: GIM_Reject,
// The -2 is just to distinguish it from the 'not' case below.
def XORI : I<(outs GPR32:$dst), (ins m1:$src2, GPR32:$src1),
[(set GPR32:$dst, (xor GPR32:$src1, -2))]>;
//===- Test a simple pattern with a default register operand. -------------===//
//
// NOOPT-NEXT: GIM_Try, /*On fail goto*//*Label [[LABEL_NUM:[0-9]+]]*/ [[LABEL:[0-9]+]],
// NOOPT-NEXT: GIM_CheckNumOperands, /*MI*/0, /*Expected*/3,
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-18 20:47:41 +01:00
// NOOPT-NEXT: GIM_CheckOpcode, /*MI*/0, TargetOpcode::G_XOR,
// NOOPT-NEXT: // MIs[0] dst
// NOOPT-NEXT: GIM_CheckType, /*MI*/0, /*Op*/0, /*Type*/GILLT_s32,
// NOOPT-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/0, /*RC*/MyTarget::GPR32RegClassID,
// NOOPT-NEXT: // MIs[0] src1
// NOOPT-NEXT: GIM_CheckType, /*MI*/0, /*Op*/1, /*Type*/GILLT_s32,
// NOOPT-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/1, /*RC*/MyTarget::GPR32RegClassID,
// NOOPT-NEXT: // MIs[0] Operand 2
// NOOPT-NEXT: GIM_CheckType, /*MI*/0, /*Op*/2, /*Type*/GILLT_s32,
// NOOPT-NEXT: GIM_CheckConstantInt, /*MI*/0, /*Op*/2, -3
// NOOPT-NEXT: // (xor:{ *:[i32] } GPR32:{ *:[i32] }:$src1, -3:{ *:[i32] }) => (XOR:{ *:[i32] } GPR32:{ *:[i32] }:$src1)
// NOOPT-NEXT: GIR_BuildMI, /*InsnID*/0, /*Opcode*/MyTarget::XOR,
// NOOPT-NEXT: GIR_Copy, /*NewInsnID*/0, /*OldInsnID*/0, /*OpIdx*/0, // dst
// NOOPT-NEXT: GIR_AddRegister, /*InsnID*/0, MyTarget::R0,
// NOOPT-NEXT: GIR_Copy, /*NewInsnID*/0, /*OldInsnID*/0, /*OpIdx*/1, // src1
// NOOPT-NEXT: GIR_EraseFromParent, /*InsnID*/0,
// NOOPT-NEXT: GIR_ConstrainSelectedInstOperands, /*InsnID*/0,
// NOOPT-NEXT: // GIR_Coverage, 3,
// NOOPT-NEXT: GIR_Done,
// NOOPT-NEXT: // Label [[LABEL_NUM]]: @[[LABEL]]
// The -3 is just to distinguish it from the 'not' case below and the other default op case above.
def XOR : I<(outs GPR32:$dst), (ins Z:$src2, GPR32:$src1),
[(set GPR32:$dst, (xor GPR32:$src1, -3))]>;
//===- Test a simple pattern with a multiple default operands. ------------===//
//
// NOOPT-NEXT: GIM_Try, /*On fail goto*//*Label [[LABEL_NUM:[0-9]+]]*/ [[LABEL:[0-9]+]],
// NOOPT-NEXT: GIM_CheckNumOperands, /*MI*/0, /*Expected*/3,
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-18 20:47:41 +01:00
// NOOPT-NEXT: GIM_CheckOpcode, /*MI*/0, TargetOpcode::G_XOR,
// NOOPT-NEXT: // MIs[0] dst
// NOOPT-NEXT: GIM_CheckType, /*MI*/0, /*Op*/0, /*Type*/GILLT_s32,
// NOOPT-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/0, /*RC*/MyTarget::GPR32RegClassID,
// NOOPT-NEXT: // MIs[0] src1
// NOOPT-NEXT: GIM_CheckType, /*MI*/0, /*Op*/1, /*Type*/GILLT_s32,
// NOOPT-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/1, /*RC*/MyTarget::GPR32RegClassID,
// NOOPT-NEXT: // MIs[0] Operand 2
// NOOPT-NEXT: GIM_CheckType, /*MI*/0, /*Op*/2, /*Type*/GILLT_s32,
// NOOPT-NEXT: GIM_CheckConstantInt, /*MI*/0, /*Op*/2, -4
// NOOPT-NEXT: // (xor:{ *:[i32] } GPR32:{ *:[i32] }:$src1, -4:{ *:[i32] }) => (XORlike:{ *:[i32] } GPR32:{ *:[i32] }:$src1)
// NOOPT-NEXT: GIR_BuildMI, /*InsnID*/0, /*Opcode*/MyTarget::XORlike,
// NOOPT-NEXT: GIR_Copy, /*NewInsnID*/0, /*OldInsnID*/0, /*OpIdx*/0, // dst
// NOOPT-NEXT: GIR_AddImm, /*InsnID*/0, /*Imm*/-1,
// NOOPT-NEXT: GIR_AddRegister, /*InsnID*/0, MyTarget::R0,
// NOOPT-NEXT: GIR_Copy, /*NewInsnID*/0, /*OldInsnID*/0, /*OpIdx*/1, // src1
// NOOPT-NEXT: GIR_EraseFromParent, /*InsnID*/0,
// NOOPT-NEXT: GIR_ConstrainSelectedInstOperands, /*InsnID*/0,
// NOOPT-NEXT: // GIR_Coverage, 4,
// NOOPT-NEXT: GIR_Done,
// NOOPT-NEXT: // Label [[LABEL_NUM]]: @[[LABEL]]
// The -4 is just to distinguish it from the other 'not' cases.
def XORlike : I<(outs GPR32:$dst), (ins m1Z:$src2, GPR32:$src1),
[(set GPR32:$dst, (xor GPR32:$src1, -4))]>;
//===- Test a simple pattern with multiple operands with defaults. --------===//
//
// NOOPT-NEXT: GIM_Try, /*On fail goto*//*Label [[LABEL_NUM:[0-9]+]]*/ [[LABEL:[0-9]+]],
// NOOPT-NEXT: GIM_CheckNumOperands, /*MI*/0, /*Expected*/3,
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-18 20:47:41 +01:00
// NOOPT-NEXT: GIM_CheckOpcode, /*MI*/0, TargetOpcode::G_XOR,
// NOOPT-NEXT: // MIs[0] dst
// NOOPT-NEXT: GIM_CheckType, /*MI*/0, /*Op*/0, /*Type*/GILLT_s32,
// NOOPT-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/0, /*RC*/MyTarget::GPR32RegClassID,
// NOOPT-NEXT: // MIs[0] src1
// NOOPT-NEXT: GIM_CheckType, /*MI*/0, /*Op*/1, /*Type*/GILLT_s32,
// NOOPT-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/1, /*RC*/MyTarget::GPR32RegClassID,
// NOOPT-NEXT: // MIs[0] Operand 2
// NOOPT-NEXT: GIM_CheckType, /*MI*/0, /*Op*/2, /*Type*/GILLT_s32,
// NOOPT-NEXT: GIM_CheckConstantInt, /*MI*/0, /*Op*/2, -5,
// NOOPT-NEXT: // (xor:{ *:[i32] } GPR32:{ *:[i32] }:$src1, -5:{ *:[i32] }) => (XORManyDefaults:{ *:[i32] } GPR32:{ *:[i32] }:$src1)
// NOOPT-NEXT: GIR_BuildMI, /*InsnID*/0, /*Opcode*/MyTarget::XORManyDefaults,
// NOOPT-NEXT: GIR_Copy, /*NewInsnID*/0, /*OldInsnID*/0, /*OpIdx*/0, // dst
// NOOPT-NEXT: GIR_AddImm, /*InsnID*/0, /*Imm*/-1,
// NOOPT-NEXT: GIR_AddRegister, /*InsnID*/0, MyTarget::R0,
// NOOPT-NEXT: GIR_AddRegister, /*InsnID*/0, MyTarget::R0,
// NOOPT-NEXT: GIR_Copy, /*NewInsnID*/0, /*OldInsnID*/0, /*OpIdx*/1, // src1
// NOOPT-NEXT: GIR_EraseFromParent, /*InsnID*/0,
// NOOPT-NEXT: GIR_ConstrainSelectedInstOperands, /*InsnID*/0,
// NOOPT-NEXT: // GIR_Coverage, 5,
// NOOPT-NEXT: GIR_Done,
// NOOPT-NEXT: // Label [[LABEL_NUM]]: @[[LABEL]]
// The -5 is just to distinguish it from the other cases.
def XORManyDefaults : I<(outs GPR32:$dst), (ins m1Z:$src3, Z:$src2, GPR32:$src1),
[(set GPR32:$dst, (xor GPR32:$src1, -5))]>;
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 16:43:30 +01:00
//===- Test a simple pattern with constant immediate operands. ------------===//
//
// This must precede the 3-register variants because constant immediates have
// priority over register banks.
//
// NOOPT-NEXT: GIM_Try, /*On fail goto*//*Label [[LABEL_NUM:[0-9]+]]*/ [[LABEL:[0-9]+]],
// NOOPT-NEXT: GIM_CheckNumOperands, /*MI*/0, /*Expected*/3,
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-18 20:47:41 +01:00
// NOOPT-NEXT: GIM_CheckOpcode, /*MI*/0, TargetOpcode::G_XOR,
// NOOPT-NEXT: // MIs[0] dst
// NOOPT-NEXT: GIM_CheckType, /*MI*/0, /*Op*/0, /*Type*/GILLT_s32,
// NOOPT-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/0, /*RC*/MyTarget::GPR32RegClassID,
// NOOPT-NEXT: // MIs[0] Wm
// NOOPT-NEXT: GIM_CheckType, /*MI*/0, /*Op*/1, /*Type*/GILLT_s32,
// NOOPT-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/1, /*RC*/MyTarget::GPR32RegClassID,
// NOOPT-NEXT: // MIs[0] Operand 2
// NOOPT-NEXT: GIM_CheckType, /*MI*/0, /*Op*/2, /*Type*/GILLT_s32,
// NOOPT-NEXT: GIM_CheckConstantInt, /*MI*/0, /*Op*/2, -1,
// NOOPT-NEXT: // (xor:{ *:[i32] } GPR32:{ *:[i32] }:$Wm, -1:{ *:[i32] }) => (ORN:{ *:[i32] } R0:{ *:[i32] }, GPR32:{ *:[i32] }:$Wm)
// NOOPT-NEXT: GIR_BuildMI, /*InsnID*/0, /*Opcode*/MyTarget::ORN,
// NOOPT-NEXT: GIR_Copy, /*NewInsnID*/0, /*OldInsnID*/0, /*OpIdx*/0, // dst
// NOOPT-NEXT: GIR_AddRegister, /*InsnID*/0, MyTarget::R0,
// NOOPT-NEXT: GIR_Copy, /*NewInsnID*/0, /*OldInsnID*/0, /*OpIdx*/1, // Wm
// NOOPT-NEXT: GIR_EraseFromParent, /*InsnID*/0,
// NOOPT-NEXT: GIR_ConstrainSelectedInstOperands, /*InsnID*/0,
// NOOPT-NEXT: // GIR_Coverage, 22,
// NOOPT-NEXT: GIR_Done,
// NOOPT-NEXT: // Label [[LABEL_NUM]]: @[[LABEL]]
[globalisel] Decouple src pattern operands from dst pattern operands. Summary: This isn't testable for AArch64 by itself so this patch also adds support for constant immediates in the pattern and physical register uses in the result. The new IntOperandMatcher matches the constant in patterns such as '(set $rd:GPR32, (G_XOR $rs:GPR32, -1))'. It's always safe to fold immediates into an instruction so this is the first rule that will match across multiple BB's. The Renderer hierarchy is responsible for adding operands to the result instruction. Renderers can copy operands (CopyRenderer) or add physical registers (in particular %wzr and %xzr) to the result instruction in any order (OperandMatchers now import the operand names from SelectionDAG to allow renderers to access any operand). This allows us to emit the result instruction for: %1 = G_XOR %0, -1 --> %1 = ORNWrr %wzr, %0 %1 = G_XOR -1, %0 --> %1 = ORNWrr %wzr, %0 although the latter is untested since the matcher/importer has not been taught about commutativity yet. Added BuildMIAction which can build new instructions and mutate them where possible. W.r.t the mutation aspect, MatchActions are now told the name of an instruction they can recycle and BuildMIAction will emit mutation code when the renderers are appropriate. They are appropriate when all operands are rendered using CopyRenderer and the indices are the same as the matcher. This currently assumes that all operands have at least one matcher. Finally, this change also fixes a crash in AArch64InstructionSelector::select() caused by an immediate operand passing isImm() rather than isCImm(). This was uncovered by the other changes and was detected by existing tests. Depends on D29711 Reviewers: t.p.northover, ab, qcolombet, rovka, aditya_nandakumar, javed.absar Reviewed By: rovka Subscribers: aemerson, dberris, kristof.beyls, llvm-commits Differential Revision: https://reviews.llvm.org/D29712 llvm-svn: 296131
2017-02-24 16:43:30 +01:00
def ORN : I<(outs GPR32:$dst), (ins GPR32:$src1, GPR32:$src2), []>;
def : Pat<(not GPR32:$Wm), (ORN R0, GPR32:$Wm)>;
//===- Test a nested instruction match. -----------------------------------===//
//
// NOOPT-NEXT: GIM_Try, /*On fail goto*//*Label [[LABEL_NUM:[0-9]+]]*/ [[LABEL:[0-9]+]],
// NOOPT-NEXT: GIM_CheckFeatures, GIFBS_HasA,
// NOOPT-NEXT: GIM_CheckNumOperands, /*MI*/0, /*Expected*/3,
// NOOPT-NEXT: GIM_CheckOpcode, /*MI*/0, TargetOpcode::G_MUL,
// NOOPT-NEXT: // MIs[0] dst
// NOOPT-NEXT: GIM_CheckType, /*MI*/0, /*Op*/0, /*Type*/GILLT_s32,
// NOOPT-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/0, /*RC*/MyTarget::GPR32RegClassID,
// NOOPT-NEXT: // MIs[0] Operand 1
// NOOPT-NEXT: GIM_CheckType, /*MI*/0, /*Op*/1, /*Type*/GILLT_s32,
// NOOPT-NEXT: GIM_RecordInsn, /*DefineMI*/1, /*MI*/0, /*OpIdx*/1, // MIs[1]
// NOOPT-NEXT: GIM_CheckNumOperands, /*MI*/1, /*Expected*/3,
// NOOPT-NEXT: GIM_CheckOpcode, /*MI*/1, TargetOpcode::G_ADD,
// NOOPT-NEXT: // MIs[1] Operand 0
// NOOPT-NEXT: GIM_CheckType, /*MI*/1, /*Op*/0, /*Type*/GILLT_s32,
// NOOPT-NEXT: // MIs[1] src1
// NOOPT-NEXT: GIM_CheckType, /*MI*/1, /*Op*/1, /*Type*/GILLT_s32,
// NOOPT-NEXT: GIM_CheckRegBankForClass, /*MI*/1, /*Op*/1, /*RC*/MyTarget::GPR32RegClassID,
// NOOPT-NEXT: // MIs[1] src2
// NOOPT-NEXT: GIM_CheckType, /*MI*/1, /*Op*/2, /*Type*/GILLT_s32,
// NOOPT-NEXT: GIM_CheckRegBankForClass, /*MI*/1, /*Op*/2, /*RC*/MyTarget::GPR32RegClassID,
// NOOPT-NEXT: // MIs[0] src3
// NOOPT-NEXT: GIM_CheckType, /*MI*/0, /*Op*/2, /*Type*/GILLT_s32,
// NOOPT-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/2, /*RC*/MyTarget::GPR32RegClassID,
// NOOPT-NEXT: GIM_CheckIsSafeToFold, /*InsnID*/1,
// NOOPT-NEXT: // (mul:{ *:[i32] } (add:{ *:[i32] } GPR32:{ *:[i32] }:$src1, GPR32:{ *:[i32] }:$src2), GPR32:{ *:[i32] }:$src3) => (MULADD:{ *:[i32] } GPR32:{ *:[i32] }:$src1, GPR32:{ *:[i32] }:$src2, GPR32:{ *:[i32] }:$src3)
// NOOPT-NEXT: GIR_BuildMI, /*InsnID*/0, /*Opcode*/MyTarget::MULADD,
// NOOPT-NEXT: GIR_Copy, /*NewInsnID*/0, /*OldInsnID*/0, /*OpIdx*/0, // dst
// NOOPT-NEXT: GIR_Copy, /*NewInsnID*/0, /*OldInsnID*/1, /*OpIdx*/1, // src1
// NOOPT-NEXT: GIR_Copy, /*NewInsnID*/0, /*OldInsnID*/1, /*OpIdx*/2, // src2
// NOOPT-NEXT: GIR_Copy, /*NewInsnID*/0, /*OldInsnID*/0, /*OpIdx*/2, // src3
// NOOPT-NEXT: GIR_EraseFromParent, /*InsnID*/0,
// NOOPT-NEXT: GIR_ConstrainSelectedInstOperands, /*InsnID*/0,
// NOOPT-NEXT: // GIR_Coverage, 6,
// NOOPT-NEXT: GIR_Done,
// NOOPT-NEXT: // Label [[LABEL_NUM]]: @[[LABEL]]
// We also get a second rule by commutativity.
//
// NOOPT-NEXT: GIM_Try, /*On fail goto*//*Label [[LABEL_NUM:[0-9]+]]*/ [[LABEL:[0-9]+]],
// NOOPT-NEXT: GIM_CheckFeatures, GIFBS_HasA,
// NOOPT-NEXT: GIM_CheckNumOperands, /*MI*/0, /*Expected*/3,
// NOOPT-NEXT: GIM_CheckOpcode, /*MI*/0, TargetOpcode::G_MUL,
// NOOPT-NEXT: // MIs[0] dst
// NOOPT-NEXT: GIM_CheckType, /*MI*/0, /*Op*/0, /*Type*/GILLT_s32,
// NOOPT-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/0, /*RC*/MyTarget::GPR32RegClassID,
// NOOPT-NEXT: // MIs[0] src3
// NOOPT-NEXT: GIM_CheckType, /*MI*/0, /*Op*/1, /*Type*/GILLT_s32,
// NOOPT-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/1, /*RC*/MyTarget::GPR32RegClassID,
// NOOPT-NEXT: // MIs[0] Operand 2
// NOOPT-NEXT: GIM_CheckType, /*MI*/0, /*Op*/2, /*Type*/GILLT_s32,
// NOOPT-NEXT: GIM_RecordInsn, /*DefineMI*/1, /*MI*/0, /*OpIdx*/2,
// NOOPT-NEXT: GIM_CheckNumOperands, /*MI*/1, /*Expected*/3,
// NOOPT-NEXT: GIM_CheckOpcode, /*MI*/1, TargetOpcode::G_ADD,
// NOOPT-NEXT: // MIs[1] Operand 0
// NOOPT-NEXT: GIM_CheckType, /*MI*/1, /*Op*/0, /*Type*/GILLT_s32,
// NOOPT-NEXT: // MIs[1] src1
// NOOPT-NEXT: GIM_CheckType, /*MI*/1, /*Op*/1, /*Type*/GILLT_s32,
// NOOPT-NEXT: GIM_CheckRegBankForClass, /*MI*/1, /*Op*/1, /*RC*/MyTarget::GPR32RegClassID,
// NOOPT-NEXT: // MIs[1] src2
// NOOPT-NEXT: GIM_CheckType, /*MI*/1, /*Op*/2, /*Type*/GILLT_s32,
// NOOPT-NEXT: GIM_CheckRegBankForClass, /*MI*/1, /*Op*/2, /*RC*/MyTarget::GPR32RegClassID,
// NOOPT-NEXT: GIM_CheckIsSafeToFold, /*InsnID*/1,
// NOOPT-NEXT: // (mul:{ *:[i32] } GPR32:{ *:[i32] }:$src3, (add:{ *:[i32] } GPR32:{ *:[i32] }:$src1, GPR32:{ *:[i32] }:$src2)) => (MULADD:{ *:[i32] } GPR32:{ *:[i32] }:$src1, GPR32:{ *:[i32] }:$src2, GPR32:{ *:[i32] }:$src3)
// NOOPT-NEXT: GIR_BuildMI, /*InsnID*/0, /*Opcode*/MyTarget::MULADD,
// NOOPT-NEXT: GIR_Copy, /*NewInsnID*/0, /*OldInsnID*/0, /*OpIdx*/0, // dst
// NOOPT-NEXT: GIR_Copy, /*NewInsnID*/0, /*OldInsnID*/1, /*OpIdx*/1, // src1
// NOOPT-NEXT: GIR_Copy, /*NewInsnID*/0, /*OldInsnID*/1, /*OpIdx*/2, // src2
// NOOPT-NEXT: GIR_Copy, /*NewInsnID*/0, /*OldInsnID*/0, /*OpIdx*/1, // src3
// NOOPT-NEXT: GIR_EraseFromParent, /*InsnID*/0,
// NOOPT-NEXT: GIR_ConstrainSelectedInstOperands, /*InsnID*/0,
// NOOPT-NEXT: // GIR_Coverage, 26,
// NOOPT-NEXT: GIR_Done,
// NOOPT-NEXT: // Label [[LABEL_NUM]]: @[[LABEL]]
def MULADD : I<(outs GPR32:$dst), (ins GPR32:$src1, GPR32:$src2, GPR32:$src3),
[(set GPR32:$dst,
(mul (add GPR32:$src1, GPR32:$src2), GPR32:$src3))]>,
Requires<[HasA]>;
//===- Test a simple pattern with just a specific leaf immediate. ---------===//
//
// NOOPT-NEXT: GIM_Try, /*On fail goto*//*Label [[LABEL_NUM:[0-9]+]]*/ [[LABEL:[0-9]+]],
// NOOPT-NEXT: GIM_CheckNumOperands, /*MI*/0, /*Expected*/2,
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-18 20:47:41 +01:00
// NOOPT-NEXT: GIM_CheckOpcode, /*MI*/0, TargetOpcode::G_CONSTANT,
// NOOPT-NEXT: // MIs[0] dst
// NOOPT-NEXT: GIM_CheckType, /*MI*/0, /*Op*/0, /*Type*/GILLT_s32,
// NOOPT-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/0, /*RC*/MyTarget::GPR32RegClassID,
// NOOPT-NEXT: // MIs[0] Operand 1
// NOOPT-NEXT: GIM_CheckLiteralInt, /*MI*/0, /*Op*/1, 1,
// NOOPT-NEXT: // 1:{ *:[i32] } => (MOV1:{ *:[i32] })
// NOOPT-NEXT: GIR_BuildMI, /*InsnID*/0, /*Opcode*/MyTarget::MOV1,
// NOOPT-NEXT: GIR_Copy, /*NewInsnID*/0, /*OldInsnID*/0, /*OpIdx*/0, // dst
// NOOPT-NEXT: GIR_EraseFromParent, /*InsnID*/0,
// NOOPT-NEXT: GIR_ConstrainSelectedInstOperands, /*InsnID*/0,
// NOOPT-NEXT: // GIR_Coverage, 7,
// NOOPT-NEXT: GIR_Done,
// NOOPT-NEXT: // Label [[LABEL_NUM]]: @[[LABEL]]
def MOV1 : I<(outs GPR32:$dst), (ins), [(set GPR32:$dst, 1)]>;
//===- Test a simple pattern with a leaf immediate and a predicate. -------===//
//
// NOOPT-NEXT: GIM_Try, /*On fail goto*//*Label [[LABEL_NUM:[0-9]+]]*/ [[LABEL:[0-9]+]],
// NOOPT-NEXT: GIM_CheckNumOperands, /*MI*/0, /*Expected*/2,
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-18 20:47:41 +01:00
// NOOPT-NEXT: GIM_CheckOpcode, /*MI*/0, TargetOpcode::G_CONSTANT,
// NOOPT-NEXT: GIM_CheckI64ImmPredicate, /*MI*/0, /*Predicate*/GIPFP_I64_Predicate_simm8,
// NOOPT-NEXT: // MIs[0] dst
// NOOPT-NEXT: GIM_CheckType, /*MI*/0, /*Op*/0, /*Type*/GILLT_s32,
// NOOPT-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/0, /*RC*/MyTarget::GPR32RegClassID,
// NOOPT-NEXT: // MIs[0] Operand 1
// NOOPT-NEXT: // No operand predicates
// NOOPT-NEXT: // (imm:{ *:[i32] })<<P:Predicate_simm8>>:$imm => (MOVimm8:{ *:[i32] } (imm:{ *:[i32] }):$imm)
// NOOPT-NEXT: GIR_BuildMI, /*InsnID*/0, /*Opcode*/MyTarget::MOVimm8,
// NOOPT-NEXT: GIR_Copy, /*NewInsnID*/0, /*OldInsnID*/0, /*OpIdx*/0, // dst
// NOOPT-NEXT: GIR_CopyConstantAsSImm, /*NewInsnID*/0, /*OldInsnID*/0, // imm
// NOOPT-NEXT: GIR_EraseFromParent, /*InsnID*/0,
// NOOPT-NEXT: GIR_ConstrainSelectedInstOperands, /*InsnID*/0,
// NOOPT-NEXT: // GIR_Coverage, 8,
// NOOPT-NEXT: GIR_Done,
// NOOPT-NEXT: // Label [[LABEL_NUM]]: @[[LABEL]]
def simm8 : ImmLeaf<i32, [{ return isInt<8>(Imm); }]>;
def MOVimm8 : I<(outs GPR32:$dst), (ins i32imm:$imm), [(set GPR32:$dst, simm8:$imm)]>;
//===- Same again but use an IntImmLeaf. ----------------------------------===//
//
// NOOPT-NEXT: GIM_Try, /*On fail goto*//*Label [[LABEL_NUM:[0-9]+]]*/ [[LABEL:[0-9]+]],
// NOOPT-NEXT: GIM_CheckNumOperands, /*MI*/0, /*Expected*/2,
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-18 20:47:41 +01:00
// NOOPT-NEXT: GIM_CheckOpcode, /*MI*/0, TargetOpcode::G_CONSTANT,
// NOOPT-NEXT: GIM_CheckAPIntImmPredicate, /*MI*/0, /*Predicate*/GIPFP_APInt_Predicate_simm9,
// NOOPT-NEXT: // MIs[0] dst
// NOOPT-NEXT: GIM_CheckType, /*MI*/0, /*Op*/0, /*Type*/GILLT_s32,
// NOOPT-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/0, /*RC*/MyTarget::GPR32RegClassID,
// NOOPT-NEXT: // MIs[0] Operand 1
// NOOPT-NEXT: // No operand predicates
// NOOPT-NEXT: // (imm:{ *:[i32] })<<P:Predicate_simm9>>:$imm => (MOVimm9:{ *:[i32] } (imm:{ *:[i32] }):$imm)
// NOOPT-NEXT: GIR_BuildMI, /*InsnID*/0, /*Opcode*/MyTarget::MOVimm9,
// NOOPT-NEXT: GIR_Copy, /*NewInsnID*/0, /*OldInsnID*/0, /*OpIdx*/0, // dst
// NOOPT-NEXT: GIR_CopyConstantAsSImm, /*NewInsnID*/0, /*OldInsnID*/0, // imm
// NOOPT-NEXT: GIR_EraseFromParent, /*InsnID*/0,
// NOOPT-NEXT: GIR_ConstrainSelectedInstOperands, /*InsnID*/0,
// NOOPT-NEXT: // GIR_Coverage, 9,
// NOOPT-NEXT: GIR_Done,
// NOOPT-NEXT: // Label [[LABEL_NUM]]: @[[LABEL]]
def simm9 : IntImmLeaf<i32, [{ return isInt<9>(Imm->getSExtValue()); }]>;
def MOVimm9 : I<(outs GPR32:$dst), (ins i32imm:$imm), [(set GPR32:$dst, simm9:$imm)]>;
//===- Test a pattern with a custom renderer. -----------------------------===//
//
// NOOPT-NEXT: GIM_Try, /*On fail goto*//*Label [[LABEL_NUM:[0-9]+]]*/ [[LABEL:[0-9]+]],
// NOOPT-NEXT: GIM_CheckNumOperands, /*MI*/0, /*Expected*/2,
// NOOPT-NEXT: GIM_CheckOpcode, /*MI*/0, TargetOpcode::G_CONSTANT,
// NOOPT-NEXT: GIM_CheckI64ImmPredicate, /*MI*/0, /*Predicate*/GIPFP_I64_Predicate_cimm8,
// NOOPT-NEXT: // MIs[0] dst
// NOOPT-NEXT: GIM_CheckType, /*MI*/0, /*Op*/0, /*Type*/GILLT_s32,
// NOOPT-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/0, /*RC*/MyTarget::GPR32RegClassID,
// NOOPT-NEXT: // MIs[0] Operand 1
// NOOPT-NEXT: // No operand predicates
// NOOPT-NEXT: // (imm:{ *:[i32] })<<P:Predicate_cimm8>><<X:cimm8_xform>>:$imm => (MOVcimm8:{ *:[i32] } (cimm8_xform:{ *:[i32] } (imm:{ *:[i32] }):$imm))
// NOOPT-NEXT: GIR_BuildMI, /*InsnID*/0, /*Opcode*/MyTarget::MOVcimm8,
// NOOPT-NEXT: GIR_Copy, /*NewInsnID*/0, /*OldInsnID*/0, /*OpIdx*/0, // dst
// NOOPT-NEXT: GIR_CustomRenderer, /*InsnID*/0, /*OldInsnID*/0, /*Renderer*/GICR_renderImm, // imm
// NOOPT-NEXT: GIR_EraseFromParent, /*InsnID*/0,
// NOOPT-NEXT: GIR_ConstrainSelectedInstOperands, /*InsnID*/0,
// NOOPT-NEXT: // GIR_Coverage, 10,
// NOOPT-NEXT: GIR_Done,
// NOOPT-NEXT: // Label [[LABEL_NUM]]: @[[LABEL]]
def MOVcimm8 : I<(outs GPR32:$dst), (ins i32imm:$imm), [(set GPR32:$dst, cimm8:$imm)]>;
//===- Test a simple pattern with a FP immediate and a predicate. ---------===//
//
// NOOPT-NEXT: GIM_Try, /*On fail goto*//*Label [[LABEL_NUM:[0-9]+]]*/ [[LABEL:[0-9]+]],
// NOOPT-NEXT: GIM_CheckNumOperands, /*MI*/0, /*Expected*/2,
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-18 20:47:41 +01:00
// NOOPT-NEXT: GIM_CheckOpcode, /*MI*/0, TargetOpcode::G_FCONSTANT,
// NOOPT-NEXT: GIM_CheckAPFloatImmPredicate, /*MI*/0, /*Predicate*/GIPFP_APFloat_Predicate_fpimmz,
// NOOPT-NEXT: // MIs[0] dst
// NOOPT-NEXT: GIM_CheckType, /*MI*/0, /*Op*/0, /*Type*/GILLT_s32,
// NOOPT-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/0, /*RC*/MyTarget::FPR32RegClassID,
// NOOPT-NEXT: // MIs[0] Operand 1
// NOOPT-NEXT: // No operand predicates
// NOOPT-NEXT: // (fpimm:{ *:[f32] })<<P:Predicate_fpimmz>>:$imm => (MOVfpimmz:{ *:[f32] } (fpimm:{ *:[f32] }):$imm)
// NOOPT-NEXT: GIR_BuildMI, /*InsnID*/0, /*Opcode*/MyTarget::MOVfpimmz,
// NOOPT-NEXT: GIR_Copy, /*NewInsnID*/0, /*OldInsnID*/0, /*OpIdx*/0, // dst
// NOOPT-NEXT: GIR_CopyFConstantAsFPImm, /*NewInsnID*/0, /*OldInsnID*/0, // imm
// NOOPT-NEXT: GIR_EraseFromParent, /*InsnID*/0,
// NOOPT-NEXT: GIR_ConstrainSelectedInstOperands, /*InsnID*/0,
// NOOPT-NEXT: // GIR_Coverage, 17,
// NOOPT-NEXT: GIR_Done,
// NOOPT-NEXT: // Label [[LABEL_NUM]]: @[[LABEL]]
//===- Test a simple pattern with inferred pointer operands. ---------------===//
//
// NOOPT-NEXT: GIM_Try, /*On fail goto*//*Label [[LABEL_NUM:[0-9]+]]*/ [[LABEL:[0-9]+]],
// NOOPT-NEXT: GIM_CheckNumOperands, /*MI*/0, /*Expected*/2,
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-18 20:47:41 +01:00
// NOOPT-NEXT: GIM_CheckOpcode, /*MI*/0, TargetOpcode::G_LOAD,
// NOOPT-NEXT: GIM_CheckMemorySizeEqualToLLT, /*MI*/0, /*MMO*/0, /*OpIdx*/0,
// NOOPT-NEXT: GIM_CheckAtomicOrdering, /*MI*/0, /*Order*/(int64_t)AtomicOrdering::NotAtomic,
// NOOPT-NEXT: // MIs[0] dst
// NOOPT-NEXT: GIM_CheckType, /*MI*/0, /*Op*/0, /*Type*/GILLT_s32,
// NOOPT-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/0, /*RC*/MyTarget::GPR32RegClassID,
// NOOPT-NEXT: // MIs[0] src1
// NOOPT-NEXT: GIM_CheckPointerToAny, /*MI*/0, /*Op*/1, /*SizeInBits*/32,
// NOOPT-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/1, /*RC*/MyTarget::GPR32RegClassID,
// NOOPT-NEXT: // (ld:{ *:[i32] } GPR32:{ *:[i32] }:$src1)<<P:Predicate_unindexedload>><<P:Predicate_load>> => (LOAD:{ *:[i32] } GPR32:{ *:[i32] }:$src1)
// NOOPT-NEXT: GIR_MutateOpcode, /*InsnID*/0, /*RecycleInsnID*/0, /*Opcode*/MyTarget::LOAD,
// NOOPT-NEXT: GIR_ConstrainSelectedInstOperands, /*InsnID*/0,
// NOOPT-NEXT: // GIR_Coverage, 11,
// NOOPT-NEXT: GIR_Done,
// NOOPT-NEXT: // Label [[LABEL_NUM]]: @[[LABEL]]
def LOAD : I<(outs GPR32:$dst), (ins GPR32:$src1),
[(set GPR32:$dst, (load GPR32:$src1))]>;
//===- Test a simple pattern with explicit pointer operands. ---------------===//
// NOOPT-NEXT: GIM_Try, /*On fail goto*//*Label [[LABEL_NUM:[0-9]+]]*/ [[LABEL:[0-9]+]],
// NOOPT-NEXT: GIM_CheckNumOperands, /*MI*/0, /*Expected*/2,
// NOOPT-NEXT: GIM_CheckOpcode, /*MI*/0, TargetOpcode::G_LOAD,
// NOOPT-NEXT: GIM_CheckMemorySizeEqualToLLT, /*MI*/0, /*MMO*/0, /*OpIdx*/0,
// NOOPT-NEXT: GIM_CheckAtomicOrdering, /*MI*/0, /*Order*/(int64_t)AtomicOrdering::NotAtomic,
// NOOPT-NEXT: // MIs[0] dst
// NOOPT-NEXT: GIM_CheckType, /*MI*/0, /*Op*/0, /*Type*/GILLT_p0s32,
// NOOPT-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/0, /*RC*/MyTarget::GPR32RegClassID,
// NOOPT-NEXT: // MIs[0] src
// NOOPT-NEXT: GIM_CheckPointerToAny, /*MI*/0, /*Op*/1, /*SizeInBits*/32,
// NOOPT-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/1, /*RC*/MyTarget::GPR32RegClassID,
// NOOPT-NEXT: // (ld:{ *:[i32] } GPR32:{ *:[i32] }:$src)<<P:Predicate_unindexedload>><<P:Predicate_load>> => (LOAD:{ *:[i32] } GPR32:{ *:[i32] }:$src)
// NOOPT-NEXT: GIR_MutateOpcode, /*InsnID*/0, /*RecycleInsnID*/0, /*Opcode*/MyTarget::LOAD,
// NOOPT-NEXT: GIR_ConstrainSelectedInstOperands, /*InsnID*/0,
// NOOPT-NEXT: // GIR_Coverage, 23,
// NOOPT-NEXT: GIR_Done,
// NOOPT-NEXT: // Label [[LABEL_NUM]]: @[[LABEL]]
def : Pat<(load GPR32:$src),
(p0 (LOAD GPR32:$src))>;
[globalisel] Update GlobalISel emitter to match new representation of extending loads Summary: Previously, a extending load was represented at (G_*EXT (G_LOAD x)). This had a few drawbacks: * G_LOAD had to be legal for all sizes you could extend from, even if registers didn't naturally hold those sizes. * All sizes you could extend from had to be allocatable just in case the extend went missing (e.g. by optimization). * At minimum, G_*EXT and G_TRUNC had to be legal for these sizes. As we improve optimization of extends and truncates, this legality requirement would spread without considerable care w.r.t when certain combines were permitted. * The SelectionDAG importer required some ugly and fragile pattern rewriting to translate patterns into this style. This patch changes the representation to: * (G_[SZ]EXTLOAD x) * (G_LOAD x) any-extends when MMO.getSize() * 8 < ResultTy.getSizeInBits() which resolves these issues by allowing targets to work entirely in their native register sizes, and by having a more direct translation from SelectionDAG patterns. Each extending load can be lowered by the legalizer into separate extends and loads, however a target that supports s1 will need the any-extending load to extend to at least s8 since LLVM does not represent memory accesses smaller than 8 bit. The legalizer can widenScalar G_LOAD into an any-extending load but sign/zero-extending loads need help from something else like a combiner pass. A follow-up patch that adds combiner helpers for for this will follow. The new representation requires that the MMO correctly reflect the memory access so this has been corrected in a couple tests. I've also moved the extending loads to their own tests since they are (mostly) separate opcodes now. Additionally, the re-write appears to have invalidated two tests from select-with-no-legality-check.mir since the matcher table no longer contains loads that result in s1's and they aren't legal in AArch64 anymore. Depends on D45540 Reviewers: ab, aditya_nandakumar, bogner, rtereshin, volkan, rovka, javed.absar Reviewed By: rtereshin Subscribers: javed.absar, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D45541 llvm-svn: 331601
2018-05-05 22:53:24 +02:00
//===- Test a simple pattern with a sextload -------------------------------===//
//
// NOOPT-NEXT: GIM_Try, /*On fail goto*//*Label [[LABEL_NUM:[0-9]+]]*/ [[LABEL:[0-9]+]],
// NOOPT-NEXT: GIM_CheckNumOperands, /*MI*/0, /*Expected*/2,
[globalisel] Update GlobalISel emitter to match new representation of extending loads Summary: Previously, a extending load was represented at (G_*EXT (G_LOAD x)). This had a few drawbacks: * G_LOAD had to be legal for all sizes you could extend from, even if registers didn't naturally hold those sizes. * All sizes you could extend from had to be allocatable just in case the extend went missing (e.g. by optimization). * At minimum, G_*EXT and G_TRUNC had to be legal for these sizes. As we improve optimization of extends and truncates, this legality requirement would spread without considerable care w.r.t when certain combines were permitted. * The SelectionDAG importer required some ugly and fragile pattern rewriting to translate patterns into this style. This patch changes the representation to: * (G_[SZ]EXTLOAD x) * (G_LOAD x) any-extends when MMO.getSize() * 8 < ResultTy.getSizeInBits() which resolves these issues by allowing targets to work entirely in their native register sizes, and by having a more direct translation from SelectionDAG patterns. Each extending load can be lowered by the legalizer into separate extends and loads, however a target that supports s1 will need the any-extending load to extend to at least s8 since LLVM does not represent memory accesses smaller than 8 bit. The legalizer can widenScalar G_LOAD into an any-extending load but sign/zero-extending loads need help from something else like a combiner pass. A follow-up patch that adds combiner helpers for for this will follow. The new representation requires that the MMO correctly reflect the memory access so this has been corrected in a couple tests. I've also moved the extending loads to their own tests since they are (mostly) separate opcodes now. Additionally, the re-write appears to have invalidated two tests from select-with-no-legality-check.mir since the matcher table no longer contains loads that result in s1's and they aren't legal in AArch64 anymore. Depends on D45540 Reviewers: ab, aditya_nandakumar, bogner, rtereshin, volkan, rovka, javed.absar Reviewed By: rtereshin Subscribers: javed.absar, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D45541 llvm-svn: 331601
2018-05-05 22:53:24 +02:00
// NOOPT-NEXT: GIM_CheckOpcode, /*MI*/0, TargetOpcode::G_SEXTLOAD,
// NOOPT-NEXT: GIM_CheckMemorySizeEqualTo, /*MI*/0, /*MMO*/0, /*Size*/2,
// NOOPT-NEXT: GIM_CheckAtomicOrdering, /*MI*/0, /*Order*/(int64_t)AtomicOrdering::NotAtomic,
// NOOPT-NEXT: // MIs[0] dst
// NOOPT-NEXT: GIM_CheckType, /*MI*/0, /*Op*/0, /*Type*/GILLT_s32,
// NOOPT-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/0, /*RC*/MyTarget::GPR32RegClassID,
// NOOPT-NEXT: // MIs[0] src1
// NOOPT-NEXT: GIM_CheckPointerToAny, /*MI*/0, /*Op*/1, /*SizeInBits*/32,
// NOOPT-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/1, /*RC*/MyTarget::GPR32RegClassID,
// NOOPT-NEXT: // (ld:{ *:[i32] } GPR32:{ *:[i32] }:$src1)<<P:Predicate_unindexedload>><<P:Predicate_sextload>><<P:Predicate_sextloadi16>> => (SEXTLOAD:{ *:[i32] } GPR32:{ *:[i32] }:$src1)
// NOOPT-NEXT: GIR_MutateOpcode, /*InsnID*/0, /*RecycleInsnID*/0, /*Opcode*/MyTarget::SEXTLOAD,
// NOOPT-NEXT: GIR_ConstrainSelectedInstOperands, /*InsnID*/0,
// NOOPT-NEXT: // GIR_Coverage, 12,
// NOOPT-NEXT: GIR_Done,
// NOOPT-NEXT: // Label [[LABEL_NUM]]: @[[LABEL]]
[globalisel] Update GlobalISel emitter to match new representation of extending loads Summary: Previously, a extending load was represented at (G_*EXT (G_LOAD x)). This had a few drawbacks: * G_LOAD had to be legal for all sizes you could extend from, even if registers didn't naturally hold those sizes. * All sizes you could extend from had to be allocatable just in case the extend went missing (e.g. by optimization). * At minimum, G_*EXT and G_TRUNC had to be legal for these sizes. As we improve optimization of extends and truncates, this legality requirement would spread without considerable care w.r.t when certain combines were permitted. * The SelectionDAG importer required some ugly and fragile pattern rewriting to translate patterns into this style. This patch changes the representation to: * (G_[SZ]EXTLOAD x) * (G_LOAD x) any-extends when MMO.getSize() * 8 < ResultTy.getSizeInBits() which resolves these issues by allowing targets to work entirely in their native register sizes, and by having a more direct translation from SelectionDAG patterns. Each extending load can be lowered by the legalizer into separate extends and loads, however a target that supports s1 will need the any-extending load to extend to at least s8 since LLVM does not represent memory accesses smaller than 8 bit. The legalizer can widenScalar G_LOAD into an any-extending load but sign/zero-extending loads need help from something else like a combiner pass. A follow-up patch that adds combiner helpers for for this will follow. The new representation requires that the MMO correctly reflect the memory access so this has been corrected in a couple tests. I've also moved the extending loads to their own tests since they are (mostly) separate opcodes now. Additionally, the re-write appears to have invalidated two tests from select-with-no-legality-check.mir since the matcher table no longer contains loads that result in s1's and they aren't legal in AArch64 anymore. Depends on D45540 Reviewers: ab, aditya_nandakumar, bogner, rtereshin, volkan, rovka, javed.absar Reviewed By: rtereshin Subscribers: javed.absar, llvm-commits, kristof.beyls Differential Revision: https://reviews.llvm.org/D45541 llvm-svn: 331601
2018-05-05 22:53:24 +02:00
def SEXTLOAD : I<(outs GPR32:$dst), (ins GPR32:$src1),
[(set GPR32:$dst, (sextloadi16 GPR32:$src1))]>;
//===- Test a simple pattern with regclass operands. ----------------------===//
//
// NOOPT-NEXT: GIM_Try, /*On fail goto*//*Label [[LABEL_NUM:[0-9]+]]*/ [[LABEL:[0-9]+]],
// NOOPT-NEXT: GIM_CheckNumOperands, /*MI*/0, /*Expected*/3,
// NOOPT-NEXT: GIM_CheckOpcode, /*MI*/0, TargetOpcode::G_ADD,
// NOOPT-NEXT: // MIs[0] dst
// NOOPT-NEXT: GIM_CheckType, /*MI*/0, /*Op*/0, /*Type*/GILLT_s32,
// NOOPT-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/0, /*RC*/MyTarget::GPR32RegClassID,
// NOOPT-NEXT: // MIs[0] src1
// NOOPT-NEXT: GIM_CheckType, /*MI*/0, /*Op*/1, /*Type*/GILLT_s32,
// NOOPT-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/1, /*RC*/MyTarget::GPR32RegClassID
// NOOPT-NEXT: // MIs[0] src2
// NOOPT-NEXT: GIM_CheckType, /*MI*/0, /*Op*/2, /*Type*/GILLT_s32,
// NOOPT-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/2, /*RC*/MyTarget::GPR32RegClassID,
// NOOPT-NEXT: // (add:{ *:[i32] } GPR32:{ *:[i32] }:$src1, GPR32:{ *:[i32] }:$src2) => (ADD:{ *:[i32] } GPR32:{ *:[i32] }:$src1, GPR32:{ *:[i32] }:$src2)
// NOOPT-NEXT: GIR_MutateOpcode, /*InsnID*/0, /*RecycleInsnID*/0, /*Opcode*/MyTarget::ADD,
// NOOPT-NEXT: GIR_ConstrainSelectedInstOperands, /*InsnID*/0,
// NOOPT-NEXT: // GIR_Coverage, 13,
// NOOPT-NEXT: GIR_Done,
// NOOPT-NEXT: // Label [[LABEL_NUM]]: @[[LABEL]]
def ADD : I<(outs GPR32:$dst), (ins GPR32:$src1, GPR32:$src2),
[(set GPR32:$dst, (add GPR32:$src1, GPR32:$src2))]>;
//===- Test a pattern with a tied operand in the matcher ------------------===//
//
// NOOPT-NEXT: GIM_Try, /*On fail goto*//*Label [[LABEL_NUM:[0-9]+]]*/ [[LABEL:[0-9]+]],
// NOOPT-NEXT: GIM_CheckNumOperands, /*MI*/0, /*Expected*/3,
// NOOPT-NEXT: GIM_CheckOpcode, /*MI*/0, TargetOpcode::G_ADD,
// NOOPT-NEXT: // MIs[0] dst
// NOOPT-NEXT: GIM_CheckType, /*MI*/0, /*Op*/0, /*Type*/GILLT_s32,
// NOOPT-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/0, /*RC*/MyTarget::GPR32RegClassID,
// NOOPT-NEXT: // MIs[0] src{{$}}
// NOOPT-NEXT: GIM_CheckType, /*MI*/0, /*Op*/1, /*Type*/GILLT_s32,
// NOOPT-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/1, /*RC*/MyTarget::GPR32RegClassID,
// NOOPT-NEXT: // MIs[0] src{{$}}
// NOOPT-NEXT: GIM_CheckIsSameOperand, /*MI*/0, /*OpIdx*/2, /*OtherMI*/0, /*OtherOpIdx*/1,
// NOOPT-NEXT: // (add:{ *:[i32] } GPR32:{ *:[i32] }:$src, GPR32:{ *:[i32] }:$src) => (DOUBLE:{ *:[i32] } GPR32:{ *:[i32] }:$src)
// NOOPT-NEXT: GIR_BuildMI, /*InsnID*/0, /*Opcode*/MyTarget::DOUBLE,
// NOOPT-NEXT: GIR_Copy, /*NewInsnID*/0, /*OldInsnID*/0, /*OpIdx*/0, // dst
// NOOPT-NEXT: GIR_Copy, /*NewInsnID*/0, /*OldInsnID*/0, /*OpIdx*/1, // src
// NOOPT-NEXT: GIR_EraseFromParent, /*InsnID*/0,
// NOOPT-NEXT: GIR_ConstrainSelectedInstOperands, /*InsnID*/0,
// NOOPT-NEXT: // GIR_Coverage, 14,
// NOOPT-NEXT: GIR_Done,
// NOOPT-NEXT: // Label [[LABEL_NUM]]: @[[LABEL]]
def DOUBLE : I<(outs GPR32:$dst), (ins GPR32:$src), [(set GPR32:$dst, (add GPR32:$src, GPR32:$src))]>;
//===- Test a simple pattern with ValueType operands. ----------------------===//
//
// NOOPT-NEXT: GIM_Try, /*On fail goto*//*Label [[LABEL_NUM:[0-9]+]]*/ [[LABEL:[0-9]+]],
// NOOPT-NEXT: GIM_CheckNumOperands, /*MI*/0, /*Expected*/3,
// NOOPT-NEXT: GIM_CheckOpcode, /*MI*/0, TargetOpcode::G_ADD,
// NOOPT-NEXT: // MIs[0] dst
// NOOPT-NEXT: GIM_CheckType, /*MI*/0, /*Op*/0, /*Type*/GILLT_s32,
// NOOPT-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/0, /*RC*/MyTarget::GPR32RegClassID,
// NOOPT-NEXT: // MIs[0] src1
// NOOPT-NEXT: GIM_CheckType, /*MI*/0, /*Op*/1, /*Type*/GILLT_s32,
// NOOPT-NEXT: // MIs[0] src2
// NOOPT-NEXT: GIM_CheckType, /*MI*/0, /*Op*/2, /*Type*/GILLT_s32,
// NOOPT-NEXT: // (add:{ *:[i32] } i32:{ *:[i32] }:$src1, i32:{ *:[i32] }:$src2) => (ADD:{ *:[i32] } i32:{ *:[i32] }:$src1, i32:{ *:[i32] }:$src2)
// NOOPT-NEXT: GIR_MutateOpcode, /*InsnID*/0, /*RecycleInsnID*/0, /*Opcode*/MyTarget::ADD,
// NOOPT-NEXT: GIR_ConstrainSelectedInstOperands, /*InsnID*/0,
// NOOPT-NEXT: // GIR_Coverage, 24,
// NOOPT-NEXT: GIR_Done,
// NOOPT-NEXT: // Label [[LABEL_NUM]]: @[[LABEL]]
def : Pat<(add i32:$src1, i32:$src2),
(ADD i32:$src1, i32:$src2)>;
//===- Test another simple pattern with regclass operands. ----------------===//
//
// NOOPT-NEXT: GIM_Try, /*On fail goto*//*Label [[LABEL_NUM:[0-9]+]]*/ [[LABEL:[0-9]+]],
// NOOPT-NEXT: GIM_CheckFeatures, GIFBS_HasA_HasB_HasC,
// NOOPT-NEXT: GIM_CheckNumOperands, /*MI*/0, /*Expected*/3,
// NOOPT-NEXT: GIM_CheckOpcode, /*MI*/0, TargetOpcode::G_MUL,
// NOOPT-NEXT: // MIs[0] dst
// NOOPT-NEXT: GIM_CheckType, /*MI*/0, /*Op*/0, /*Type*/GILLT_s32,
// NOOPT-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/0, /*RC*/MyTarget::GPR32RegClassID,
// NOOPT-NEXT: // MIs[0] src1
// NOOPT-NEXT: GIM_CheckType, /*MI*/0, /*Op*/1, /*Type*/GILLT_s32,
// NOOPT-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/1, /*RC*/MyTarget::GPR32RegClassID,
// NOOPT-NEXT: // MIs[0] src2
// NOOPT-NEXT: GIM_CheckType, /*MI*/0, /*Op*/2, /*Type*/GILLT_s32,
// NOOPT-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/2, /*RC*/MyTarget::GPR32RegClassID,
// NOOPT-NEXT: // (mul:{ *:[i32] } GPR32:{ *:[i32] }:$src1, GPR32:{ *:[i32] }:$src2) => (MUL:{ *:[i32] } GPR32:{ *:[i32] }:$src2, GPR32:{ *:[i32] }:$src1)
// NOOPT-NEXT: GIR_BuildMI, /*InsnID*/0, /*Opcode*/MyTarget::MUL,
// NOOPT-NEXT: GIR_Copy, /*NewInsnID*/0, /*OldInsnID*/0, /*OpIdx*/0, // dst
// NOOPT-NEXT: GIR_Copy, /*NewInsnID*/0, /*OldInsnID*/0, /*OpIdx*/2, // src2
// NOOPT-NEXT: GIR_Copy, /*NewInsnID*/0, /*OldInsnID*/0, /*OpIdx*/1, // src1
// NOOPT-NEXT: GIR_EraseFromParent, /*InsnID*/0,
// NOOPT-NEXT: GIR_ConstrainSelectedInstOperands, /*InsnID*/0,
// NOOPT-NEXT: // GIR_Coverage, 15,
// NOOPT-NEXT: GIR_Done,
// NOOPT-NEXT: // Label [[LABEL_NUM]]: @[[LABEL]]
def MUL : I<(outs GPR32:$dst), (ins GPR32:$src2, GPR32:$src1),
[(set GPR32:$dst, (mul GPR32:$src1, GPR32:$src2))]>,
Requires<[HasA, HasB, HasC]>;
//===- Test a COPY_TO_REGCLASS --------------------------------------------===//
//
//
// NOOPT-NEXT: GIM_Try, /*On fail goto*//*Label [[LABEL_NUM:[0-9]+]]*/ [[LABEL:[0-9]+]],
// NOOPT-NEXT: GIM_CheckNumOperands, /*MI*/0, /*Expected*/2,
// NOOPT-NEXT: GIM_CheckOpcode, /*MI*/0, TargetOpcode::G_BITCAST,
// NOOPT-NEXT: // MIs[0] dst
// NOOPT-NEXT: GIM_CheckType, /*MI*/0, /*Op*/0, /*Type*/GILLT_s32,
// NOOPT-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/0, /*RC*/MyTarget::GPR32RegClassID,
// NOOPT-NEXT: // MIs[0] src1
// NOOPT-NEXT: GIM_CheckType, /*MI*/0, /*Op*/1, /*Type*/GILLT_s32,
// NOOPT-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/1, /*RC*/MyTarget::FPR32RegClassID,
// NOOPT-NEXT: // (bitconvert:{ *:[i32] } FPR32:{ *:[f32] }:$src1) => (COPY_TO_REGCLASS:{ *:[i32] } FPR32:{ *:[f32] }:$src1, GPR32:{ *:[i32] })
// NOOPT-NEXT: GIR_MutateOpcode, /*InsnID*/0, /*RecycleInsnID*/0, /*Opcode*/TargetOpcode::COPY,
// NOOPT-NEXT: GIR_ConstrainOperandRC, /*InsnID*/0, /*Op*/0, MyTarget::GPR32RegClassID,
// NOOPT-NEXT: // GIR_Coverage, 25,
// NOOPT-NEXT: GIR_Done,
// NOOPT-NEXT: // Label [[LABEL_NUM]]: @[[LABEL]]
def : Pat<(i32 (bitconvert FPR32:$src1)),
(COPY_TO_REGCLASS FPR32:$src1, GPR32)>;
//===- Test a simple pattern with just a leaf immediate. ------------------===//
//
// NOOPT-NEXT: GIM_Try, /*On fail goto*//*Label [[LABEL_NUM:[0-9]+]]*/ [[LABEL:[0-9]+]],
// NOOPT-NEXT: GIM_CheckNumOperands, /*MI*/0, /*Expected*/2,
// NOOPT-NEXT: GIM_CheckOpcode, /*MI*/0, TargetOpcode::G_CONSTANT,
// NOOPT-NEXT: // MIs[0] dst
// NOOPT-NEXT: GIM_CheckType, /*MI*/0, /*Op*/0, /*Type*/GILLT_s32,
// NOOPT-NEXT: GIM_CheckRegBankForClass, /*MI*/0, /*Op*/0, /*RC*/MyTarget::GPR32RegClassID,
// NOOPT-NEXT: // MIs[0] Operand 1
// NOOPT-NEXT: // No operand predicates
// NOOPT-NEXT: // (imm:{ *:[i32] }):$imm => (MOVimm:{ *:[i32] } (imm:{ *:[i32] }):$imm)
// NOOPT-NEXT: GIR_BuildMI, /*InsnID*/0, /*Opcode*/MyTarget::MOVimm,
// NOOPT-NEXT: GIR_Copy, /*NewInsnID*/0, /*OldInsnID*/0, /*OpIdx*/0, // dst
// NOOPT-NEXT: GIR_CopyConstantAsSImm, /*NewInsnID*/0, /*OldInsnID*/0, // imm
// NOOPT-NEXT: GIR_EraseFromParent, /*InsnID*/0,
// NOOPT-NEXT: GIR_ConstrainSelectedInstOperands, /*InsnID*/0,
// NOOPT-NEXT: // GIR_Coverage, 16,
// NOOPT-NEXT: GIR_Done,
// NOOPT-NEXT: // Label [[LABEL_NUM]]: @[[LABEL]]
def MOVimm : I<(outs GPR32:$dst), (ins i32imm:$imm), [(set GPR32:$dst, imm:$imm)]>;
def fpimmz : FPImmLeaf<f32, [{ return Imm->isExactlyValue(0.0); }]>;
def MOVfpimmz : I<(outs FPR32:$dst), (ins f32imm:$imm), [(set FPR32:$dst, fpimmz:$imm)]>;
//===- Test a pattern with an MBB operand. --------------------------------===//
//
// NOOPT-NEXT: GIM_Try, /*On fail goto*//*Label [[LABEL_NUM:[0-9]+]]*/ [[LABEL:[0-9]+]],
// NOOPT-NEXT: GIM_CheckNumOperands, /*MI*/0, /*Expected*/1,
[TableGen][GlobalISel] Optimize MatchTable for faster instruction selection *** Context *** Prior to this patchw, the table generated for matching instruction was straight forward but highly inefficient. Basically, each pattern generates its own set of self contained checks and actions. E.g., TableGen generated: // First pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDrr // Second pattern CheckNumOperand 3 CheckOpcode G_ADD ... Build ADDri // Third pattern CheckNumOperand 3 CheckOpcode G_SUB ... Build SUBrr *** Problem *** Because of that generation, a *lot* of check were redundant between each pattern and were checked every single time until we reach the pattern that matches. E.g., Taking the previous table, let say we are matching a G_SUB, that means we were going to check all the rules for G_ADD before looking at the G_SUB rule. In particular we are going to do: check 3 operands; PASS check G_ADD; FAIL ; Next rule check 3 operands; PASS (but we already knew that!) check G_ADD; FAIL (well it is still not true) ; Next rule check 3 operands; PASS (really!!) check G_SUB; PASS (at last :P) *** Proposed Solution *** This patch introduces a concept of group of rules (GroupMatcher) that share some predicates and only get checked once for the whole group. This patch only creates groups with one nesting level. Conceptually there is nothing preventing us for having deeper nest level. However, the current implementation is not smart enough to share the recording (aka capturing) of values. That limits its ability to do more sharing. For the given example the current patch will generate: // First group CheckOpcode G_ADD // First pattern CheckNumOperand 3 ... Build ADDrr // Second pattern CheckNumOperand 3 ... Build ADDri // Second group CheckOpcode G_SUB // Third pattern CheckNumOperand 3 ... Build SUBrr But if we allowed several nesting level, it could create a sub group for the checknumoperand 3. (We would need to call optimizeRules on the rules within a group.) *** Result *** With only one level of nesting, the instruction selection pass is up to 4x faster. For instance, one instruction now takes 500 checks, instead of 24k! With more nesting we could get in the tens I believe. Differential Revision: https://reviews.llvm.org/D39034 rdar://problem/34670699 llvm-svn: 321017
2017-12-18 20:47:41 +01:00
// NOOPT-NEXT: GIM_CheckOpcode, /*MI*/0, TargetOpcode::G_BR,
// NOOPT-NEXT: // MIs[0] target
// NOOPT-NEXT: GIM_CheckIsMBB, /*MI*/0, /*Op*/0,
// NOOPT-NEXT: // (br (bb:{ *:[Other] }):$target) => (BR (bb:{ *:[Other] }):$target)
// NOOPT-NEXT: GIR_MutateOpcode, /*InsnID*/0, /*RecycleInsnID*/0, /*Opcode*/MyTarget::BR,
// NOOPT-NEXT: GIR_ConstrainSelectedInstOperands, /*InsnID*/0,
// NOOPT-NEXT: // GIR_Coverage, 18,
// NOOPT-NEXT: GIR_Done,
// NOOPT-NEXT: // Label [[LABEL_NUM]]: @[[LABEL]]
def BR : I<(outs), (ins unknown:$target),
[(br bb:$target)]>;
// NOOPT-NEXT: GIM_Reject,
// NOOPT-NEXT: };
// NOOPT-NEXT: return MatchTable0;