1 //===---------- AArch64CollectLOH.cpp - AArch64 collect LOH pass --*- C++ -*-=//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains a pass that collect the Linker Optimization Hint (LOH).
11 // This pass should be run at the very end of the compilation flow, just before
13 // To be useful for the linker, the LOH must be printed into the assembly file.
15 // A LOH describes a sequence of instructions that may be optimized by the
17 // This same sequence cannot be optimized by the compiler because some of
18 // the information will be known at link time.
19 // For instance, consider the following sequence:
20 // L1: adrp xA, sym@PAGE
21 // L2: add xB, xA, sym@PAGEOFF
22 // L3: ldr xC, [xB, #imm]
23 // This sequence can be turned into:
24 // A literal load if sym@PAGE + sym@PAGEOFF + #imm - address(L3) is < 1MB:
25 // L3: ldr xC, sym+#imm
26 // It may also be turned into either the following more efficient
28 // - If sym@PAGEOFF + #imm fits the encoding space of L3.
29 // L1: adrp xA, sym@PAGE
30 // L3: ldr xC, [xB, sym@PAGEOFF + #imm]
31 // - If sym@PAGE + sym@PAGEOFF - address(L1) < 1MB:
33 // L3: ldr xC, [xB, #imm]
35 // To be valid a LOH must meet all the requirements needed by all the related
36 // possible linker transformations.
37 // For instance, using the running example, the constraints to emit
38 // ".loh AdrpAddLdr" are:
39 // - L1, L2, and L3 instructions are of the expected type, i.e.,
40 // respectively ADRP, ADD (immediate), and LD.
41 // - The result of L1 is used only by L2.
42 // - The register argument (xA) used in the ADD instruction is defined
44 // - The result of L2 is used only by L3.
45 // - The base address (xB) in L3 is defined only L2.
46 // - The ADRP in L1 and the ADD in L2 must reference the same symbol using
47 // @PAGE/@PAGEOFF with no additional constants
49 // Currently supported LOHs are:
50 // * So called non-ADRP-related:
51 // - .loh AdrpAddLdr L1, L2, L3:
52 // L1: adrp xA, sym@PAGE
53 // L2: add xB, xA, sym@PAGEOFF
54 // L3: ldr xC, [xB, #imm]
55 // - .loh AdrpLdrGotLdr L1, L2, L3:
56 // L1: adrp xA, sym@GOTPAGE
57 // L2: ldr xB, [xA, sym@GOTPAGEOFF]
58 // L3: ldr xC, [xB, #imm]
59 // - .loh AdrpLdr L1, L3:
60 // L1: adrp xA, sym@PAGE
61 // L3: ldr xC, [xA, sym@PAGEOFF]
62 // - .loh AdrpAddStr L1, L2, L3:
63 // L1: adrp xA, sym@PAGE
64 // L2: add xB, xA, sym@PAGEOFF
65 // L3: str xC, [xB, #imm]
66 // - .loh AdrpLdrGotStr L1, L2, L3:
67 // L1: adrp xA, sym@GOTPAGE
68 // L2: ldr xB, [xA, sym@GOTPAGEOFF]
69 // L3: str xC, [xB, #imm]
70 // - .loh AdrpAdd L1, L2:
71 // L1: adrp xA, sym@PAGE
72 // L2: add xB, xA, sym@PAGEOFF
73 // For all these LOHs, L1, L2, L3 form a simple chain:
74 // L1 result is used only by L2 and L2 result by L3.
75 // L3 LOH-related argument is defined only by L2 and L2 LOH-related argument
77 // All these LOHs aim at using more efficient load/store patterns by folding
78 // some instructions used to compute the address directly into the load/store.
80 // * So called ADRP-related:
81 // - .loh AdrpAdrp L2, L1:
82 // L2: ADRP xA, sym1@PAGE
83 // L1: ADRP xA, sym2@PAGE
84 // L2 dominates L1 and xA is not redifined between L2 and L1
85 // This LOH aims at getting rid of redundant ADRP instructions.
87 // The overall design for emitting the LOHs is:
88 // 1. AArch64CollectLOH (this pass) records the LOHs in the AArch64FunctionInfo.
89 // 2. AArch64AsmPrinter reads the LOHs from AArch64FunctionInfo and it:
90 // 1. Associates them a label.
91 // 2. Emits them in a MCStreamer (EmitLOHDirective).
92 // - The MCMachOStreamer records them into the MCAssembler.
93 // - The MCAsmStreamer prints them.
94 // - Other MCStreamers ignore them.
95 // 3. Closes the MCStreamer:
96 // - The MachObjectWriter gets them from the MCAssembler and writes
97 // them in the object file.
98 // - Other ObjectWriters ignore them.
99 //===----------------------------------------------------------------------===//
102 #include "AArch64InstrInfo.h"
103 #include "AArch64MachineFunctionInfo.h"
104 #include "llvm/ADT/BitVector.h"
105 #include "llvm/ADT/DenseMap.h"
106 #include "llvm/ADT/MapVector.h"
107 #include "llvm/ADT/SmallVector.h"
108 #include "llvm/ADT/Statistic.h"
109 #include "llvm/CodeGen/MachineBasicBlock.h"
110 #include "llvm/CodeGen/MachineFunctionPass.h"
111 #include "llvm/CodeGen/MachineInstr.h"
112 #include "llvm/CodeGen/TargetRegisterInfo.h"
113 #include "llvm/Support/Debug.h"
114 #include "llvm/Support/ErrorHandling.h"
115 #include "llvm/Support/raw_ostream.h"
116 #include "llvm/Target/TargetMachine.h"
117 using namespace llvm;
119 #define DEBUG_TYPE "aarch64-collect-loh"
121 STATISTIC(NumADRPSimpleCandidate,
122 "Number of simplifiable ADRP dominate by another");
123 STATISTIC(NumADDToSTR, "Number of simplifiable STR reachable by ADD");
124 STATISTIC(NumLDRToSTR, "Number of simplifiable STR reachable by LDR");
125 STATISTIC(NumADDToLDR, "Number of simplifiable LDR reachable by ADD");
126 STATISTIC(NumLDRToLDR, "Number of simplifiable LDR reachable by LDR");
127 STATISTIC(NumADRPToLDR, "Number of simplifiable LDR reachable by ADRP");
128 STATISTIC(NumADRSimpleCandidate, "Number of simplifiable ADRP + ADD");
130 #define AARCH64_COLLECT_LOH_NAME "AArch64 Collect Linker Optimization Hint (LOH)"
134 struct AArch64CollectLOH : public MachineFunctionPass {
136 AArch64CollectLOH() : MachineFunctionPass(ID) {}
138 bool runOnMachineFunction(MachineFunction &MF) override;
140 MachineFunctionProperties getRequiredProperties() const override {
141 return MachineFunctionProperties().set(
142 MachineFunctionProperties::Property::NoVRegs);
145 StringRef getPassName() const override { return AARCH64_COLLECT_LOH_NAME; }
147 void getAnalysisUsage(AnalysisUsage &AU) const override {
148 MachineFunctionPass::getAnalysisUsage(AU);
149 AU.setPreservesAll();
153 char AArch64CollectLOH::ID = 0;
155 } // end anonymous namespace.
157 INITIALIZE_PASS(AArch64CollectLOH, "aarch64-collect-loh",
158 AARCH64_COLLECT_LOH_NAME, false, false)
160 static bool canAddBePartOfLOH(const MachineInstr &MI) {
161 // Check immediate to see if the immediate is an address.
162 switch (MI.getOperand(2).getType()) {
165 case MachineOperand::MO_GlobalAddress:
166 case MachineOperand::MO_JumpTableIndex:
167 case MachineOperand::MO_ConstantPoolIndex:
168 case MachineOperand::MO_BlockAddress:
173 /// Answer the following question: Can Def be one of the definition
174 /// involved in a part of a LOH?
175 static bool canDefBePartOfLOH(const MachineInstr &MI) {
176 // Accept ADRP, ADDLow and LOADGot.
177 switch (MI.getOpcode()) {
182 case AArch64::ADDXri:
183 return canAddBePartOfLOH(MI);
184 case AArch64::LDRXui:
185 // Check immediate to see if the immediate is an address.
186 switch (MI.getOperand(2).getType()) {
189 case MachineOperand::MO_GlobalAddress:
190 return MI.getOperand(2).getTargetFlags() & AArch64II::MO_GOT;
195 /// Check whether the given instruction can the end of a LOH chain involving a
197 static bool isCandidateStore(const MachineInstr &MI, const MachineOperand &MO) {
198 switch (MI.getOpcode()) {
201 case AArch64::STRBBui:
202 case AArch64::STRHHui:
203 case AArch64::STRBui:
204 case AArch64::STRHui:
205 case AArch64::STRWui:
206 case AArch64::STRXui:
207 case AArch64::STRSui:
208 case AArch64::STRDui:
209 case AArch64::STRQui:
210 // We can only optimize the index operand.
211 // In case we have str xA, [xA, #imm], this is two different uses
212 // of xA and we cannot fold, otherwise the xA stored may be wrong,
213 // even if #imm == 0.
214 return MI.getOperandNo(&MO) == 1 &&
215 MI.getOperand(0).getReg() != MI.getOperand(1).getReg();
219 /// Check whether the given instruction can be the end of a LOH chain
220 /// involving a load.
221 static bool isCandidateLoad(const MachineInstr &MI) {
222 switch (MI.getOpcode()) {
225 case AArch64::LDRSBWui:
226 case AArch64::LDRSBXui:
227 case AArch64::LDRSHWui:
228 case AArch64::LDRSHXui:
229 case AArch64::LDRSWui:
230 case AArch64::LDRBui:
231 case AArch64::LDRHui:
232 case AArch64::LDRWui:
233 case AArch64::LDRXui:
234 case AArch64::LDRSui:
235 case AArch64::LDRDui:
236 case AArch64::LDRQui:
237 return !(MI.getOperand(2).getTargetFlags() & AArch64II::MO_GOT);
241 /// Check whether the given instruction can load a litteral.
242 static bool supportLoadFromLiteral(const MachineInstr &MI) {
243 switch (MI.getOpcode()) {
246 case AArch64::LDRSWui:
247 case AArch64::LDRWui:
248 case AArch64::LDRXui:
249 case AArch64::LDRSui:
250 case AArch64::LDRDui:
251 case AArch64::LDRQui:
256 /// Number of GPR registers traked by mapRegToGPRIndex()
257 static const unsigned N_GPR_REGS = 31;
258 /// Map register number to index from 0-30.
259 static int mapRegToGPRIndex(MCPhysReg Reg) {
260 static_assert(AArch64::X28 - AArch64::X0 + 3 == N_GPR_REGS, "Number of GPRs");
261 static_assert(AArch64::W30 - AArch64::W0 + 1 == N_GPR_REGS, "Number of GPRs");
262 if (AArch64::X0 <= Reg && Reg <= AArch64::X28)
263 return Reg - AArch64::X0;
264 if (AArch64::W0 <= Reg && Reg <= AArch64::W30)
265 return Reg - AArch64::W0;
266 // TableGen gives "FP" and "LR" an index not adjacent to X28 so we have to
267 // handle them as special cases.
268 if (Reg == AArch64::FP)
270 if (Reg == AArch64::LR)
275 /// State tracked per register.
276 /// The main algorithm walks backwards over a basic block maintaining this
277 /// datastructure for each tracked general purpose register.
279 MCLOHType Type : 8; ///< "Best" type of LOH possible.
280 bool IsCandidate : 1; ///< Possible LOH candidate.
281 bool OneUser : 1; ///< Found exactly one user (yet).
282 bool MultiUsers : 1; ///< Found multiple users.
283 const MachineInstr *MI0; ///< First instruction involved in the LOH.
284 const MachineInstr *MI1; ///< Second instruction involved in the LOH
286 const MachineInstr *LastADRP; ///< Last ADRP in same register.
289 /// Update state \p Info given \p MI uses the tracked register.
290 static void handleUse(const MachineInstr &MI, const MachineOperand &MO,
292 // We have multiple uses if we already found one before.
293 if (Info.MultiUsers || Info.OneUser) {
294 Info.IsCandidate = false;
295 Info.MultiUsers = true;
300 // Start new LOHInfo if applicable.
301 if (isCandidateLoad(MI)) {
302 Info.Type = MCLOH_AdrpLdr;
303 Info.IsCandidate = true;
305 // Note that even this is AdrpLdr now, we can switch to a Ldr variant
307 } else if (isCandidateStore(MI, MO)) {
308 Info.Type = MCLOH_AdrpAddStr;
309 Info.IsCandidate = true;
312 } else if (MI.getOpcode() == AArch64::ADDXri) {
313 Info.Type = MCLOH_AdrpAdd;
314 Info.IsCandidate = true;
316 } else if (MI.getOpcode() == AArch64::LDRXui &&
317 MI.getOperand(2).getTargetFlags() & AArch64II::MO_GOT) {
318 Info.Type = MCLOH_AdrpLdrGot;
319 Info.IsCandidate = true;
324 /// Update state \p Info given the tracked register is clobbered.
325 static void handleClobber(LOHInfo &Info) {
326 Info.IsCandidate = false;
327 Info.OneUser = false;
328 Info.MultiUsers = false;
329 Info.LastADRP = nullptr;
332 /// Update state \p Info given that \p MI is possibly the middle instruction
333 /// of an LOH involving 3 instructions.
334 static bool handleMiddleInst(const MachineInstr &MI, LOHInfo &DefInfo,
336 if (!DefInfo.IsCandidate || (&DefInfo != &OpInfo && OpInfo.OneUser))
338 // Copy LOHInfo for dest register to LOHInfo for source register.
339 if (&DefInfo != &OpInfo) {
341 // Invalidate \p DefInfo because we track it in \p OpInfo now.
342 handleClobber(DefInfo);
344 DefInfo.LastADRP = nullptr;
346 // Advance state machine.
347 assert(OpInfo.IsCandidate && "Expect valid state");
348 if (MI.getOpcode() == AArch64::ADDXri && canAddBePartOfLOH(MI)) {
349 if (OpInfo.Type == MCLOH_AdrpLdr) {
350 OpInfo.Type = MCLOH_AdrpAddLdr;
351 OpInfo.IsCandidate = true;
354 } else if (OpInfo.Type == MCLOH_AdrpAddStr && OpInfo.MI1 == nullptr) {
355 OpInfo.Type = MCLOH_AdrpAddStr;
356 OpInfo.IsCandidate = true;
361 assert(MI.getOpcode() == AArch64::LDRXui && "Expect LDRXui");
362 assert((MI.getOperand(2).getTargetFlags() & AArch64II::MO_GOT) &&
363 "Expected GOT relocation");
364 if (OpInfo.Type == MCLOH_AdrpAddStr && OpInfo.MI1 == nullptr) {
365 OpInfo.Type = MCLOH_AdrpLdrGotStr;
366 OpInfo.IsCandidate = true;
369 } else if (OpInfo.Type == MCLOH_AdrpLdr) {
370 OpInfo.Type = MCLOH_AdrpLdrGotLdr;
371 OpInfo.IsCandidate = true;
379 /// Update state when seeing and ADRP instruction.
380 static void handleADRP(const MachineInstr &MI, AArch64FunctionInfo &AFI,
382 if (Info.LastADRP != nullptr) {
383 LLVM_DEBUG(dbgs() << "Adding MCLOH_AdrpAdrp:\n"
384 << '\t' << MI << '\t' << *Info.LastADRP);
385 AFI.addLOHDirective(MCLOH_AdrpAdrp, {&MI, Info.LastADRP});
386 ++NumADRPSimpleCandidate;
389 // Produce LOH directive if possible.
390 if (Info.IsCandidate) {
393 LLVM_DEBUG(dbgs() << "Adding MCLOH_AdrpAdd:\n"
394 << '\t' << MI << '\t' << *Info.MI0);
395 AFI.addLOHDirective(MCLOH_AdrpAdd, {&MI, Info.MI0});
396 ++NumADRSimpleCandidate;
399 if (supportLoadFromLiteral(*Info.MI0)) {
400 LLVM_DEBUG(dbgs() << "Adding MCLOH_AdrpLdr:\n"
401 << '\t' << MI << '\t' << *Info.MI0);
402 AFI.addLOHDirective(MCLOH_AdrpLdr, {&MI, Info.MI0});
406 case MCLOH_AdrpAddLdr:
407 LLVM_DEBUG(dbgs() << "Adding MCLOH_AdrpAddLdr:\n"
408 << '\t' << MI << '\t' << *Info.MI1 << '\t'
410 AFI.addLOHDirective(MCLOH_AdrpAddLdr, {&MI, Info.MI1, Info.MI0});
413 case MCLOH_AdrpAddStr:
414 if (Info.MI1 != nullptr) {
415 LLVM_DEBUG(dbgs() << "Adding MCLOH_AdrpAddStr:\n"
416 << '\t' << MI << '\t' << *Info.MI1 << '\t'
418 AFI.addLOHDirective(MCLOH_AdrpAddStr, {&MI, Info.MI1, Info.MI0});
422 case MCLOH_AdrpLdrGotLdr:
423 LLVM_DEBUG(dbgs() << "Adding MCLOH_AdrpLdrGotLdr:\n"
424 << '\t' << MI << '\t' << *Info.MI1 << '\t'
426 AFI.addLOHDirective(MCLOH_AdrpLdrGotLdr, {&MI, Info.MI1, Info.MI0});
429 case MCLOH_AdrpLdrGotStr:
430 LLVM_DEBUG(dbgs() << "Adding MCLOH_AdrpLdrGotStr:\n"
431 << '\t' << MI << '\t' << *Info.MI1 << '\t'
433 AFI.addLOHDirective(MCLOH_AdrpLdrGotStr, {&MI, Info.MI1, Info.MI0});
436 case MCLOH_AdrpLdrGot:
437 LLVM_DEBUG(dbgs() << "Adding MCLOH_AdrpLdrGot:\n"
438 << '\t' << MI << '\t' << *Info.MI0);
439 AFI.addLOHDirective(MCLOH_AdrpLdrGot, {&MI, Info.MI0});
442 llvm_unreachable("MCLOH_AdrpAdrp not used in state machine");
450 static void handleRegMaskClobber(const uint32_t *RegMask, MCPhysReg Reg,
452 if (!MachineOperand::clobbersPhysReg(RegMask, Reg))
454 int Idx = mapRegToGPRIndex(Reg);
456 handleClobber(LOHInfos[Idx]);
459 static void handleNormalInst(const MachineInstr &MI, LOHInfo *LOHInfos) {
460 // Handle defs and regmasks.
461 for (const MachineOperand &MO : MI.operands()) {
462 if (MO.isRegMask()) {
463 const uint32_t *RegMask = MO.getRegMask();
464 for (MCPhysReg Reg : AArch64::GPR32RegClass)
465 handleRegMaskClobber(RegMask, Reg, LOHInfos);
466 for (MCPhysReg Reg : AArch64::GPR64RegClass)
467 handleRegMaskClobber(RegMask, Reg, LOHInfos);
470 if (!MO.isReg() || !MO.isDef())
472 int Idx = mapRegToGPRIndex(MO.getReg());
475 handleClobber(LOHInfos[Idx]);
478 for (const MachineOperand &MO : MI.uses()) {
479 if (!MO.isReg() || !MO.readsReg())
481 int Idx = mapRegToGPRIndex(MO.getReg());
484 handleUse(MI, MO, LOHInfos[Idx]);
488 bool AArch64CollectLOH::runOnMachineFunction(MachineFunction &MF) {
489 if (skipFunction(MF.getFunction()))
492 LLVM_DEBUG(dbgs() << "********** AArch64 Collect LOH **********\n"
493 << "Looking in function " << MF.getName() << '\n');
495 LOHInfo LOHInfos[N_GPR_REGS];
496 AArch64FunctionInfo &AFI = *MF.getInfo<AArch64FunctionInfo>();
497 for (const MachineBasicBlock &MBB : MF) {
498 // Reset register tracking state.
499 memset(LOHInfos, 0, sizeof(LOHInfos));
500 // Live-out registers are used.
501 for (const MachineBasicBlock *Succ : MBB.successors()) {
502 for (const auto &LI : Succ->liveins()) {
503 int RegIdx = mapRegToGPRIndex(LI.PhysReg);
505 LOHInfos[RegIdx].OneUser = true;
509 // Walk the basic block backwards and update the per register state machine
511 for (const MachineInstr &MI : make_range(MBB.rbegin(), MBB.rend())) {
512 unsigned Opcode = MI.getOpcode();
514 case AArch64::ADDXri:
515 case AArch64::LDRXui:
516 if (canDefBePartOfLOH(MI)) {
517 const MachineOperand &Def = MI.getOperand(0);
518 const MachineOperand &Op = MI.getOperand(1);
519 assert(Def.isReg() && Def.isDef() && "Expected reg def");
520 assert(Op.isReg() && Op.isUse() && "Expected reg use");
521 int DefIdx = mapRegToGPRIndex(Def.getReg());
522 int OpIdx = mapRegToGPRIndex(Op.getReg());
523 if (DefIdx >= 0 && OpIdx >= 0 &&
524 handleMiddleInst(MI, LOHInfos[DefIdx], LOHInfos[OpIdx]))
529 const MachineOperand &Op0 = MI.getOperand(0);
530 int Idx = mapRegToGPRIndex(Op0.getReg());
532 handleADRP(MI, AFI, LOHInfos[Idx]);
537 handleNormalInst(MI, LOHInfos);
541 // Return "no change": The pass only collects information.
545 FunctionPass *llvm::createAArch64CollectLOHPass() {
546 return new AArch64CollectLOH();