LLVM 20.0.0git
PPCMacroFusion.cpp
Go to the documentation of this file.
1//===- PPCMacroFusion.cpp - PowerPC Macro Fusion --------------------------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9/// \file This file contains the PowerPC implementation of the DAG scheduling
10/// mutation to pair instructions back to back.
11//
12//===----------------------------------------------------------------------===//
13
14#include "PPC.h"
15#include "PPCSubtarget.h"
16#include "llvm/ADT/DenseSet.h"
19#include <optional>
20
21using namespace llvm;
22namespace {
23
24class FusionFeature {
25public:
26 typedef SmallDenseSet<unsigned> FusionOpSet;
27
28 enum FusionKind {
29 #define FUSION_KIND(KIND) FK_##KIND
30 #define FUSION_FEATURE(KIND, HAS_FEATURE, DEP_OP_IDX, OPSET1, OPSET2) \
31 FUSION_KIND(KIND),
32 #include "PPCMacroFusion.def"
33 FUSION_KIND(END)
34 };
35private:
36 // Each fusion feature is assigned with one fusion kind. All the
37 // instructions with the same fusion kind have the same fusion characteristic.
38 FusionKind Kd;
39 // True if this feature is enabled.
40 bool Supported;
41 // li rx, si
42 // load rt, ra, rx
43 // The dependent operand index in the second op(load). And the negative means
44 // it could be any one.
45 int DepOpIdx;
46 // The first fusion op set.
47 FusionOpSet OpSet1;
48 // The second fusion op set.
49 FusionOpSet OpSet2;
50public:
51 FusionFeature(FusionKind Kind, bool HasFeature, int Index,
52 const FusionOpSet &First, const FusionOpSet &Second) :
53 Kd(Kind), Supported(HasFeature), DepOpIdx(Index), OpSet1(First),
54 OpSet2(Second) {}
55
56 bool hasOp1(unsigned Opc) const { return OpSet1.contains(Opc); }
57 bool hasOp2(unsigned Opc) const { return OpSet2.contains(Opc); }
58 bool isSupported() const { return Supported; }
59 std::optional<unsigned> depOpIdx() const {
60 if (DepOpIdx < 0)
61 return std::nullopt;
62 return DepOpIdx;
63 }
64
65 FusionKind getKind() const { return Kd; }
66};
67
68static bool matchingRegOps(const MachineInstr &FirstMI,
69 int FirstMIOpIndex,
70 const MachineInstr &SecondMI,
71 int SecondMIOpIndex) {
72 const MachineOperand &Op1 = FirstMI.getOperand(FirstMIOpIndex);
73 const MachineOperand &Op2 = SecondMI.getOperand(SecondMIOpIndex);
74 if (!Op1.isReg() || !Op2.isReg())
75 return false;
76
77 return Op1.getReg() == Op2.getReg();
78}
79
80static bool matchingImmOps(const MachineInstr &MI,
81 int MIOpIndex,
82 int64_t Expect,
83 unsigned ExtendFrom = 64) {
84 const MachineOperand &Op = MI.getOperand(MIOpIndex);
85 if (!Op.isImm())
86 return false;
87 int64_t Imm = Op.getImm();
88 if (ExtendFrom < 64)
89 Imm = SignExtend64(Imm, ExtendFrom);
90 return Imm == Expect;
91}
92
93// Return true if the FirstMI meets the constraints of SecondMI according to
94// fusion specification.
95static bool checkOpConstraints(FusionFeature::FusionKind Kd,
96 const MachineInstr &FirstMI,
97 const MachineInstr &SecondMI) {
98 switch (Kd) {
99 // The hardware didn't require any specific check for the fused instructions'
100 // operands. Therefore, return true to indicate that, it is fusable.
101 default: return true;
102 // [addi rt,ra,si - lxvd2x xt,ra,rb] etc.
103 case FusionFeature::FK_AddiLoad: {
104 // lxvd2x(ra) cannot be zero
105 const MachineOperand &RA = SecondMI.getOperand(1);
106 if (!RA.isReg())
107 return true;
108
109 return RA.getReg().isVirtual() ||
110 (RA.getReg() != PPC::ZERO && RA.getReg() != PPC::ZERO8);
111 }
112 // [addis rt,ra,si - ld rt,ds(ra)] etc.
113 case FusionFeature::FK_AddisLoad: {
114 const MachineOperand &RT = SecondMI.getOperand(0);
115 if (!RT.isReg())
116 return true;
117
118 // Only check it for non-virtual register.
119 if (!RT.getReg().isVirtual())
120 // addis(rt) = ld(ra) = ld(rt)
121 // ld(rt) cannot be zero
122 if (!matchingRegOps(SecondMI, 0, SecondMI, 2) ||
123 (RT.getReg() == PPC::ZERO || RT.getReg() == PPC::ZERO8))
124 return false;
125
126 // addis(si) first 12 bits must be all 1s or all 0s
127 const MachineOperand &SI = FirstMI.getOperand(2);
128 if (!SI.isImm())
129 return true;
130 int64_t Imm = SI.getImm();
131 if (((Imm & 0xFFF0) != 0) && ((Imm & 0xFFF0) != 0xFFF0))
132 return false;
133
134 // If si = 1111111111110000 and the msb of the d/ds field of the load equals
135 // 1, then fusion does not occur.
136 if ((Imm & 0xFFF0) == 0xFFF0) {
137 const MachineOperand &D = SecondMI.getOperand(1);
138 if (!D.isImm())
139 return true;
140
141 // 14 bit for DS field, while 16 bit for D field.
142 int MSB = 15;
143 if (SecondMI.getOpcode() == PPC::LD)
144 MSB = 13;
145
146 return (D.getImm() & (1ULL << MSB)) == 0;
147 }
148 return true;
149 }
150
151 case FusionFeature::FK_SldiAdd:
152 return (matchingImmOps(FirstMI, 2, 3) && matchingImmOps(FirstMI, 3, 60)) ||
153 (matchingImmOps(FirstMI, 2, 6) && matchingImmOps(FirstMI, 3, 57));
154
155 // rldicl rx, ra, 1, 0 - xor
156 case FusionFeature::FK_RotateLeftXor:
157 return matchingImmOps(FirstMI, 2, 1) && matchingImmOps(FirstMI, 3, 0);
158
159 // rldicr rx, ra, 1, 63 - xor
160 case FusionFeature::FK_RotateRightXor:
161 return matchingImmOps(FirstMI, 2, 1) && matchingImmOps(FirstMI, 3, 63);
162
163 // We actually use CMPW* and CMPD*, 'l' doesn't exist as an operand in instr.
164
165 // { lbz,lbzx,lhz,lhzx,lwz,lwzx } - cmpi 0,1,rx,{ 0,1,-1 }
166 // { lbz,lbzx,lhz,lhzx,lwz,lwzx } - cmpli 0,L,rx,{ 0,1 }
167 case FusionFeature::FK_LoadCmp1:
168 // { ld,ldx } - cmpi 0,1,rx,{ 0,1,-1 }
169 // { ld,ldx } - cmpli 0,1,rx,{ 0,1 }
170 case FusionFeature::FK_LoadCmp2: {
171 const MachineOperand &BT = SecondMI.getOperand(0);
172 if (!BT.isReg() || (!BT.getReg().isVirtual() && BT.getReg() != PPC::CR0))
173 return false;
174 if (SecondMI.getOpcode() == PPC::CMPDI &&
175 matchingImmOps(SecondMI, 2, -1, 16))
176 return true;
177 return matchingImmOps(SecondMI, 2, 0) || matchingImmOps(SecondMI, 2, 1);
178 }
179
180 // { lha,lhax,lwa,lwax } - cmpi 0,L,rx,{ 0,1,-1 }
181 case FusionFeature::FK_LoadCmp3: {
182 const MachineOperand &BT = SecondMI.getOperand(0);
183 if (!BT.isReg() || (!BT.getReg().isVirtual() && BT.getReg() != PPC::CR0))
184 return false;
185 return matchingImmOps(SecondMI, 2, 0) || matchingImmOps(SecondMI, 2, 1) ||
186 matchingImmOps(SecondMI, 2, -1, 16);
187 }
188
189 // mtctr - { bcctr,bcctrl }
190 case FusionFeature::FK_ZeroMoveCTR:
191 // ( mtctr rx ) is alias of ( mtspr 9, rx )
192 return (FirstMI.getOpcode() != PPC::MTSPR &&
193 FirstMI.getOpcode() != PPC::MTSPR8) ||
194 matchingImmOps(FirstMI, 0, 9);
195
196 // mtlr - { bclr,bclrl }
197 case FusionFeature::FK_ZeroMoveLR:
198 // ( mtlr rx ) is alias of ( mtspr 8, rx )
199 return (FirstMI.getOpcode() != PPC::MTSPR &&
200 FirstMI.getOpcode() != PPC::MTSPR8) ||
201 matchingImmOps(FirstMI, 0, 8);
202
203 // addis rx,ra,si - addi rt,rx,SI, SI >= 0
204 case FusionFeature::FK_AddisAddi: {
205 const MachineOperand &RA = FirstMI.getOperand(1);
206 const MachineOperand &SI = SecondMI.getOperand(2);
207 if (!SI.isImm() || !RA.isReg())
208 return false;
209 if (RA.getReg() == PPC::ZERO || RA.getReg() == PPC::ZERO8)
210 return false;
211 return SignExtend64(SI.getImm(), 16) >= 0;
212 }
213
214 // addi rx,ra,si - addis rt,rx,SI, ra > 0, SI >= 2
215 case FusionFeature::FK_AddiAddis: {
216 const MachineOperand &RA = FirstMI.getOperand(1);
217 const MachineOperand &SI = FirstMI.getOperand(2);
218 if (!SI.isImm() || !RA.isReg())
219 return false;
220 if (RA.getReg() == PPC::ZERO || RA.getReg() == PPC::ZERO8)
221 return false;
222 int64_t ExtendedSI = SignExtend64(SI.getImm(), 16);
223 return ExtendedSI >= 2;
224 }
225 }
226
227 llvm_unreachable("All the cases should have been handled");
228 return true;
229}
230
231/// Check if the instr pair, FirstMI and SecondMI, should be fused together.
232/// Given SecondMI, when FirstMI is unspecified, then check if SecondMI may be
233/// part of a fused pair at all.
235 const TargetSubtargetInfo &TSI,
236 const MachineInstr *FirstMI,
237 const MachineInstr &SecondMI) {
238 // We use the PPC namespace to avoid the need to prefix opcodes with PPC:: in
239 // the def file.
240 using namespace PPC;
241
242 const PPCSubtarget &ST = static_cast<const PPCSubtarget&>(TSI);
243 static const FusionFeature FusionFeatures[] = {
244 #define FUSION_FEATURE(KIND, HAS_FEATURE, DEP_OP_IDX, OPSET1, OPSET2) { \
245 FusionFeature::FUSION_KIND(KIND), ST.HAS_FEATURE(), DEP_OP_IDX, { OPSET1 },\
246 { OPSET2 } },
247 #include "PPCMacroFusion.def"
248 };
249 #undef FUSION_KIND
250
251 for (auto &Feature : FusionFeatures) {
252 // Skip if the feature is not supported.
253 if (!Feature.isSupported())
254 continue;
255
256 // Only when the SecondMI is fusable, we are starting to look for the
257 // fusable FirstMI.
258 if (Feature.hasOp2(SecondMI.getOpcode())) {
259 // If FirstMI == nullptr, that means, we're only checking whether SecondMI
260 // can be fused at all.
261 if (!FirstMI)
262 return true;
263
264 // Checking if the FirstMI is fusable with the SecondMI.
265 if (!Feature.hasOp1(FirstMI->getOpcode()))
266 continue;
267
268 auto DepOpIdx = Feature.depOpIdx();
269 if (DepOpIdx) {
270 // Checking if the result of the FirstMI is the desired operand of the
271 // SecondMI if the DepOpIdx is set. Otherwise, ignore it.
272 if (!matchingRegOps(*FirstMI, 0, SecondMI, *DepOpIdx))
273 return false;
274 }
275
276 // Checking more on the instruction operands.
277 if (checkOpConstraints(Feature.getKind(), *FirstMI, SecondMI))
278 return true;
279 }
280 }
281
282 return false;
283}
284
285} // end anonymous namespace
286
287namespace llvm {
288
289std::unique_ptr<ScheduleDAGMutation> createPowerPCMacroFusionDAGMutation() {
291}
292
293} // end namespace llvm
static bool shouldScheduleAdjacent(const TargetInstrInfo &TII, const TargetSubtargetInfo &TSI, const MachineInstr *FirstMI, const MachineInstr &SecondMI)
Check if the instr pair, FirstMI and SecondMI, should be fused together.
static GCRegistry::Add< StatepointGC > D("statepoint-example", "an example strategy for statepoint")
This file defines the DenseSet and SmallDenseSet classes.
const HexagonInstrInfo * TII
IRTranslator LLVM IR MI
#define FUSION_KIND(KIND)
SI optimize exec mask operations pre RA
This class represents an Operation in the Expression.
Representation of each machine instruction.
Definition: MachineInstr.h:69
unsigned getOpcode() const
Returns the opcode of this MachineInstr.
Definition: MachineInstr.h:569
const MachineOperand & getOperand(unsigned i) const
Definition: MachineInstr.h:579
MachineOperand class - Representation of each machine instruction operand.
bool isReg() const
isReg - Tests if this is a MO_Register operand.
Register getReg() const
getReg - Returns the register number.
constexpr bool isVirtual() const
Return true if the specified register number is in the virtual register namespace.
Definition: Register.h:91
Implements a dense probed hash-table based set with some number of buckets stored inline.
Definition: DenseSet.h:290
TargetInstrInfo - Interface to description of machine instruction set.
TargetSubtargetInfo - Generic base class for all target subtargets.
#define llvm_unreachable(msg)
Marks that the current location is not supposed to be reachable.
This is an optimization pass for GlobalISel generic memory operations.
Definition: AddressRanges.h:18
std::unique_ptr< ScheduleDAGMutation > createMacroFusionDAGMutation(ArrayRef< MacroFusionPredTy > Predicates, bool BranchOnly=false)
Create a DAG scheduling mutation to pair instructions back to back for instructions that benefit acco...
@ First
Helpers to iterate all locations in the MemoryEffectsBase class.
std::unique_ptr< ScheduleDAGMutation > createPowerPCMacroFusionDAGMutation()
Note that you have to add: DAG.addMutation(createPowerPCMacroFusionDAGMutation()); to PPCPassConfig::...
constexpr int64_t SignExtend64(uint64_t x)
Sign-extend the number in the bottom B bits of X to a 64-bit integer.
Definition: MathExtras.h:573
static bool shouldScheduleAdjacent(const TargetInstrInfo &TII, const TargetSubtargetInfo &TSI, const MachineInstr *FirstMI, const MachineInstr &SecondMI)
Check if the instr pair, FirstMI and SecondMI, should be fused together.