1 | //===----- RISCVMergeBaseOffset.cpp - Optimise address calculations ------===// |
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 | // Merge the offset of address calculation into the offset field |
10 | // of instructions in a global address lowering sequence. |
11 | // |
12 | //===----------------------------------------------------------------------===// |
13 | |
14 | #include "RISCV.h" |
15 | #include "RISCVTargetMachine.h" |
16 | #include "llvm/CodeGen/MachineFunctionPass.h" |
17 | #include "llvm/CodeGen/Passes.h" |
18 | #include "llvm/MC/TargetRegistry.h" |
19 | #include "llvm/Support/Debug.h" |
20 | #include "llvm/Target/TargetOptions.h" |
21 | #include <optional> |
22 | using namespace llvm; |
23 | |
24 | #define DEBUG_TYPE "riscv-merge-base-offset" |
25 | #define RISCV_MERGE_BASE_OFFSET_NAME "RISC-V Merge Base Offset" |
26 | namespace { |
27 | |
28 | class RISCVMergeBaseOffsetOpt : public MachineFunctionPass { |
29 | const RISCVSubtarget *ST = nullptr; |
30 | MachineRegisterInfo *MRI; |
31 | |
32 | public: |
33 | static char ID; |
34 | bool runOnMachineFunction(MachineFunction &Fn) override; |
35 | bool detectFoldable(MachineInstr &Hi, MachineInstr *&Lo); |
36 | |
37 | bool detectAndFoldOffset(MachineInstr &Hi, MachineInstr &Lo); |
38 | void foldOffset(MachineInstr &Hi, MachineInstr &Lo, MachineInstr &Tail, |
39 | int64_t Offset); |
40 | bool foldLargeOffset(MachineInstr &Hi, MachineInstr &Lo, |
41 | MachineInstr &TailAdd, Register GSReg); |
42 | bool foldShiftedOffset(MachineInstr &Hi, MachineInstr &Lo, |
43 | MachineInstr &TailShXAdd, Register GSReg); |
44 | |
45 | bool foldIntoMemoryOps(MachineInstr &Hi, MachineInstr &Lo); |
46 | |
47 | RISCVMergeBaseOffsetOpt() : MachineFunctionPass(ID) {} |
48 | |
49 | MachineFunctionProperties getRequiredProperties() const override { |
50 | return MachineFunctionProperties().set( |
51 | MachineFunctionProperties::Property::IsSSA); |
52 | } |
53 | |
54 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
55 | AU.setPreservesCFG(); |
56 | MachineFunctionPass::getAnalysisUsage(AU); |
57 | } |
58 | |
59 | StringRef getPassName() const override { |
60 | return RISCV_MERGE_BASE_OFFSET_NAME; |
61 | } |
62 | }; |
63 | } // end anonymous namespace |
64 | |
65 | char RISCVMergeBaseOffsetOpt::ID = 0; |
66 | INITIALIZE_PASS(RISCVMergeBaseOffsetOpt, DEBUG_TYPE, |
67 | RISCV_MERGE_BASE_OFFSET_NAME, false, false) |
68 | |
69 | // Detect either of the patterns: |
70 | // |
71 | // 1. (medlow pattern): |
72 | // lui vreg1, %hi(s) |
73 | // addi vreg2, vreg1, %lo(s) |
74 | // |
75 | // 2. (medany pattern): |
76 | // .Lpcrel_hi1: |
77 | // auipc vreg1, %pcrel_hi(s) |
78 | // addi vreg2, vreg1, %pcrel_lo(.Lpcrel_hi1) |
79 | // |
80 | // The pattern is only accepted if: |
81 | // 1) The first instruction has only one use, which is the ADDI. |
82 | // 2) The address operands have the appropriate type, reflecting the |
83 | // lowering of a global address or constant pool using medlow or medany. |
84 | // 3) The offset value in the Global Address or Constant Pool is 0. |
85 | bool RISCVMergeBaseOffsetOpt::detectFoldable(MachineInstr &Hi, |
86 | MachineInstr *&Lo) { |
87 | if (Hi.getOpcode() != RISCV::LUI && Hi.getOpcode() != RISCV::AUIPC && |
88 | Hi.getOpcode() != RISCV::PseudoMovAddr) |
89 | return false; |
90 | |
91 | const MachineOperand &HiOp1 = Hi.getOperand(i: 1); |
92 | unsigned ExpectedFlags = |
93 | Hi.getOpcode() == RISCV::AUIPC ? RISCVII::MO_PCREL_HI : RISCVII::MO_HI; |
94 | if (HiOp1.getTargetFlags() != ExpectedFlags) |
95 | return false; |
96 | |
97 | if (!(HiOp1.isGlobal() || HiOp1.isCPI() || HiOp1.isBlockAddress()) || |
98 | HiOp1.getOffset() != 0) |
99 | return false; |
100 | |
101 | if (Hi.getOpcode() == RISCV::PseudoMovAddr) { |
102 | // Most of the code should handle it correctly without modification by |
103 | // setting Lo and Hi both point to PseudoMovAddr |
104 | Lo = &Hi; |
105 | } else { |
106 | Register HiDestReg = Hi.getOperand(i: 0).getReg(); |
107 | if (!MRI->hasOneUse(RegNo: HiDestReg)) |
108 | return false; |
109 | |
110 | Lo = &*MRI->use_instr_begin(RegNo: HiDestReg); |
111 | if (Lo->getOpcode() != RISCV::ADDI) |
112 | return false; |
113 | } |
114 | |
115 | const MachineOperand &LoOp2 = Lo->getOperand(i: 2); |
116 | if (Hi.getOpcode() == RISCV::LUI || Hi.getOpcode() == RISCV::PseudoMovAddr) { |
117 | if (LoOp2.getTargetFlags() != RISCVII::MO_LO || |
118 | !(LoOp2.isGlobal() || LoOp2.isCPI() || LoOp2.isBlockAddress()) || |
119 | LoOp2.getOffset() != 0) |
120 | return false; |
121 | } else { |
122 | assert(Hi.getOpcode() == RISCV::AUIPC); |
123 | if (LoOp2.getTargetFlags() != RISCVII::MO_PCREL_LO || |
124 | LoOp2.getType() != MachineOperand::MO_MCSymbol) |
125 | return false; |
126 | } |
127 | |
128 | if (HiOp1.isGlobal()) { |
129 | LLVM_DEBUG(dbgs() << " Found lowered global address: " |
130 | << *HiOp1.getGlobal() << "\n" ); |
131 | } else if (HiOp1.isBlockAddress()) { |
132 | LLVM_DEBUG(dbgs() << " Found lowered basic address: " |
133 | << *HiOp1.getBlockAddress() << "\n" ); |
134 | } else if (HiOp1.isCPI()) { |
135 | LLVM_DEBUG(dbgs() << " Found lowered constant pool: " << HiOp1.getIndex() |
136 | << "\n" ); |
137 | } |
138 | |
139 | return true; |
140 | } |
141 | |
142 | // Update the offset in Hi and Lo instructions. |
143 | // Delete the tail instruction and update all the uses to use the |
144 | // output from Lo. |
145 | void RISCVMergeBaseOffsetOpt::foldOffset(MachineInstr &Hi, MachineInstr &Lo, |
146 | MachineInstr &Tail, int64_t Offset) { |
147 | assert(isInt<32>(Offset) && "Unexpected offset" ); |
148 | // Put the offset back in Hi and the Lo |
149 | Hi.getOperand(i: 1).setOffset(Offset); |
150 | if (Hi.getOpcode() != RISCV::AUIPC) |
151 | Lo.getOperand(i: 2).setOffset(Offset); |
152 | // Delete the tail instruction. |
153 | MRI->constrainRegClass(Reg: Lo.getOperand(i: 0).getReg(), |
154 | RC: MRI->getRegClass(Reg: Tail.getOperand(i: 0).getReg())); |
155 | MRI->replaceRegWith(FromReg: Tail.getOperand(i: 0).getReg(), ToReg: Lo.getOperand(i: 0).getReg()); |
156 | Tail.eraseFromParent(); |
157 | LLVM_DEBUG(dbgs() << " Merged offset " << Offset << " into base.\n" |
158 | << " " << Hi << " " << Lo;); |
159 | } |
160 | |
161 | // Detect patterns for large offsets that are passed into an ADD instruction. |
162 | // If the pattern is found, updates the offset in Hi and Lo instructions |
163 | // and deletes TailAdd and the instructions that produced the offset. |
164 | // |
165 | // Base address lowering is of the form: |
166 | // Hi: lui vreg1, %hi(s) |
167 | // Lo: addi vreg2, vreg1, %lo(s) |
168 | // / \ |
169 | // / \ |
170 | // / \ |
171 | // / The large offset can be of two forms: \ |
172 | // 1) Offset that has non zero bits in lower 2) Offset that has non zero |
173 | // 12 bits and upper 20 bits bits in upper 20 bits only |
174 | // OffseLUI: lui vreg3, 4 |
175 | // OffsetTail: addi voff, vreg3, 188 OffsetTail: lui voff, 128 |
176 | // \ / |
177 | // \ / |
178 | // \ / |
179 | // \ / |
180 | // TailAdd: add vreg4, vreg2, voff |
181 | bool RISCVMergeBaseOffsetOpt::foldLargeOffset(MachineInstr &Hi, |
182 | MachineInstr &Lo, |
183 | MachineInstr &TailAdd, |
184 | Register GAReg) { |
185 | assert((TailAdd.getOpcode() == RISCV::ADD) && "Expected ADD instruction!" ); |
186 | Register Rs = TailAdd.getOperand(i: 1).getReg(); |
187 | Register Rt = TailAdd.getOperand(i: 2).getReg(); |
188 | Register Reg = Rs == GAReg ? Rt : Rs; |
189 | |
190 | // Can't fold if the register has more than one use. |
191 | if (!Reg.isVirtual() || !MRI->hasOneUse(RegNo: Reg)) |
192 | return false; |
193 | // This can point to an ADDI(W) or a LUI: |
194 | MachineInstr &OffsetTail = *MRI->getVRegDef(Reg); |
195 | if (OffsetTail.getOpcode() == RISCV::ADDI || |
196 | OffsetTail.getOpcode() == RISCV::ADDIW) { |
197 | // The offset value has non zero bits in both %hi and %lo parts. |
198 | // Detect an ADDI that feeds from a LUI instruction. |
199 | MachineOperand &AddiImmOp = OffsetTail.getOperand(i: 2); |
200 | if (AddiImmOp.getTargetFlags() != RISCVII::MO_None) |
201 | return false; |
202 | Register AddiReg = OffsetTail.getOperand(i: 1).getReg(); |
203 | int64_t OffLo = AddiImmOp.getImm(); |
204 | |
205 | // Handle rs1 of ADDI is X0. |
206 | if (AddiReg == RISCV::X0) { |
207 | LLVM_DEBUG(dbgs() << " Offset Instrs: " << OffsetTail); |
208 | foldOffset(Hi, Lo, Tail&: TailAdd, Offset: OffLo); |
209 | OffsetTail.eraseFromParent(); |
210 | return true; |
211 | } |
212 | |
213 | MachineInstr &OffsetLui = *MRI->getVRegDef(Reg: AddiReg); |
214 | MachineOperand &LuiImmOp = OffsetLui.getOperand(i: 1); |
215 | if (OffsetLui.getOpcode() != RISCV::LUI || |
216 | LuiImmOp.getTargetFlags() != RISCVII::MO_None || |
217 | !MRI->hasOneUse(RegNo: OffsetLui.getOperand(i: 0).getReg())) |
218 | return false; |
219 | int64_t Offset = SignExtend64<32>(x: LuiImmOp.getImm() << 12); |
220 | Offset += OffLo; |
221 | // RV32 ignores the upper 32 bits. ADDIW sign extends the result. |
222 | if (!ST->is64Bit() || OffsetTail.getOpcode() == RISCV::ADDIW) |
223 | Offset = SignExtend64<32>(x: Offset); |
224 | // We can only fold simm32 offsets. |
225 | if (!isInt<32>(x: Offset)) |
226 | return false; |
227 | LLVM_DEBUG(dbgs() << " Offset Instrs: " << OffsetTail |
228 | << " " << OffsetLui); |
229 | foldOffset(Hi, Lo, Tail&: TailAdd, Offset); |
230 | OffsetTail.eraseFromParent(); |
231 | OffsetLui.eraseFromParent(); |
232 | return true; |
233 | } else if (OffsetTail.getOpcode() == RISCV::LUI) { |
234 | // The offset value has all zero bits in the lower 12 bits. Only LUI |
235 | // exists. |
236 | LLVM_DEBUG(dbgs() << " Offset Instr: " << OffsetTail); |
237 | int64_t Offset = SignExtend64<32>(x: OffsetTail.getOperand(i: 1).getImm() << 12); |
238 | foldOffset(Hi, Lo, Tail&: TailAdd, Offset); |
239 | OffsetTail.eraseFromParent(); |
240 | return true; |
241 | } |
242 | return false; |
243 | } |
244 | |
245 | // Detect patterns for offsets that are passed into a SHXADD instruction. |
246 | // The offset has 1, 2, or 3 trailing zeros and fits in simm13, simm14, simm15. |
247 | // The constant is created with addi voff, x0, C, and shXadd is used to |
248 | // fill insert the trailing zeros and do the addition. |
249 | // If the pattern is found, updates the offset in Hi and Lo instructions |
250 | // and deletes TailShXAdd and the instructions that produced the offset. |
251 | // |
252 | // Hi: lui vreg1, %hi(s) |
253 | // Lo: addi vreg2, vreg1, %lo(s) |
254 | // OffsetTail: addi voff, x0, C |
255 | // TailAdd: shXadd vreg4, voff, vreg2 |
256 | bool RISCVMergeBaseOffsetOpt::foldShiftedOffset(MachineInstr &Hi, |
257 | MachineInstr &Lo, |
258 | MachineInstr &TailShXAdd, |
259 | Register GAReg) { |
260 | assert((TailShXAdd.getOpcode() == RISCV::SH1ADD || |
261 | TailShXAdd.getOpcode() == RISCV::SH2ADD || |
262 | TailShXAdd.getOpcode() == RISCV::SH3ADD) && |
263 | "Expected SHXADD instruction!" ); |
264 | |
265 | if (GAReg != TailShXAdd.getOperand(i: 2).getReg()) |
266 | return false; |
267 | |
268 | // The first source is the shifted operand. |
269 | Register Rs1 = TailShXAdd.getOperand(i: 1).getReg(); |
270 | |
271 | // Can't fold if the register has more than one use. |
272 | if (!Rs1.isVirtual() || !MRI->hasOneUse(RegNo: Rs1)) |
273 | return false; |
274 | // This can point to an ADDI X0, C. |
275 | MachineInstr &OffsetTail = *MRI->getVRegDef(Reg: Rs1); |
276 | if (OffsetTail.getOpcode() != RISCV::ADDI) |
277 | return false; |
278 | if (!OffsetTail.getOperand(i: 1).isReg() || |
279 | OffsetTail.getOperand(i: 1).getReg() != RISCV::X0 || |
280 | !OffsetTail.getOperand(i: 2).isImm()) |
281 | return false; |
282 | |
283 | int64_t Offset = OffsetTail.getOperand(i: 2).getImm(); |
284 | assert(isInt<12>(Offset) && "Unexpected offset" ); |
285 | |
286 | unsigned ShAmt; |
287 | switch (TailShXAdd.getOpcode()) { |
288 | default: llvm_unreachable("Unexpected opcode" ); |
289 | case RISCV::SH1ADD: ShAmt = 1; break; |
290 | case RISCV::SH2ADD: ShAmt = 2; break; |
291 | case RISCV::SH3ADD: ShAmt = 3; break; |
292 | } |
293 | |
294 | Offset = (uint64_t)Offset << ShAmt; |
295 | |
296 | LLVM_DEBUG(dbgs() << " Offset Instr: " << OffsetTail); |
297 | foldOffset(Hi, Lo, Tail&: TailShXAdd, Offset); |
298 | OffsetTail.eraseFromParent(); |
299 | return true; |
300 | } |
301 | |
302 | bool RISCVMergeBaseOffsetOpt::detectAndFoldOffset(MachineInstr &Hi, |
303 | MachineInstr &Lo) { |
304 | Register DestReg = Lo.getOperand(i: 0).getReg(); |
305 | |
306 | // Look for arithmetic instructions we can get an offset from. |
307 | // We might be able to remove the arithmetic instructions by folding the |
308 | // offset into the LUI+ADDI. |
309 | if (!MRI->hasOneUse(RegNo: DestReg)) |
310 | return false; |
311 | |
312 | // Lo has only one use. |
313 | MachineInstr &Tail = *MRI->use_instr_begin(RegNo: DestReg); |
314 | switch (Tail.getOpcode()) { |
315 | default: |
316 | LLVM_DEBUG(dbgs() << "Don't know how to get offset from this instr:" |
317 | << Tail); |
318 | break; |
319 | case RISCV::ADDI: { |
320 | // Offset is simply an immediate operand. |
321 | int64_t Offset = Tail.getOperand(i: 2).getImm(); |
322 | |
323 | // We might have two ADDIs in a row. |
324 | Register TailDestReg = Tail.getOperand(i: 0).getReg(); |
325 | if (MRI->hasOneUse(RegNo: TailDestReg)) { |
326 | MachineInstr &TailTail = *MRI->use_instr_begin(RegNo: TailDestReg); |
327 | if (TailTail.getOpcode() == RISCV::ADDI) { |
328 | Offset += TailTail.getOperand(i: 2).getImm(); |
329 | LLVM_DEBUG(dbgs() << " Offset Instrs: " << Tail << TailTail); |
330 | foldOffset(Hi, Lo, Tail&: TailTail, Offset); |
331 | Tail.eraseFromParent(); |
332 | return true; |
333 | } |
334 | } |
335 | |
336 | LLVM_DEBUG(dbgs() << " Offset Instr: " << Tail); |
337 | foldOffset(Hi, Lo, Tail, Offset); |
338 | return true; |
339 | } |
340 | case RISCV::ADD: |
341 | // The offset is too large to fit in the immediate field of ADDI. |
342 | // This can be in two forms: |
343 | // 1) LUI hi_Offset followed by: |
344 | // ADDI lo_offset |
345 | // This happens in case the offset has non zero bits in |
346 | // both hi 20 and lo 12 bits. |
347 | // 2) LUI (offset20) |
348 | // This happens in case the lower 12 bits of the offset are zeros. |
349 | return foldLargeOffset(Hi, Lo, TailAdd&: Tail, GAReg: DestReg); |
350 | case RISCV::SH1ADD: |
351 | case RISCV::SH2ADD: |
352 | case RISCV::SH3ADD: |
353 | // The offset is too large to fit in the immediate field of ADDI. |
354 | // It may be encoded as (SH2ADD (ADDI X0, C), DestReg) or |
355 | // (SH3ADD (ADDI X0, C), DestReg). |
356 | return foldShiftedOffset(Hi, Lo, TailShXAdd&: Tail, GAReg: DestReg); |
357 | } |
358 | |
359 | return false; |
360 | } |
361 | |
362 | bool RISCVMergeBaseOffsetOpt::foldIntoMemoryOps(MachineInstr &Hi, |
363 | MachineInstr &Lo) { |
364 | Register DestReg = Lo.getOperand(i: 0).getReg(); |
365 | |
366 | // If all the uses are memory ops with the same offset, we can transform: |
367 | // |
368 | // 1. (medlow pattern): |
369 | // Hi: lui vreg1, %hi(foo) ---> lui vreg1, %hi(foo+8) |
370 | // Lo: addi vreg2, vreg1, %lo(foo) ---> lw vreg3, lo(foo+8)(vreg1) |
371 | // Tail: lw vreg3, 8(vreg2) |
372 | // |
373 | // 2. (medany pattern): |
374 | // Hi: 1:auipc vreg1, %pcrel_hi(s) ---> auipc vreg1, %pcrel_hi(foo+8) |
375 | // Lo: addi vreg2, vreg1, %pcrel_lo(1b) ---> lw vreg3, %pcrel_lo(1b)(vreg1) |
376 | // Tail: lw vreg3, 8(vreg2) |
377 | |
378 | std::optional<int64_t> CommonOffset; |
379 | DenseMap<const MachineInstr *, SmallVector<unsigned>> |
380 | InlineAsmMemoryOpIndexesMap; |
381 | for (const MachineInstr &UseMI : MRI->use_instructions(Reg: DestReg)) { |
382 | switch (UseMI.getOpcode()) { |
383 | default: |
384 | LLVM_DEBUG(dbgs() << "Not a load or store instruction: " << UseMI); |
385 | return false; |
386 | case RISCV::LB: |
387 | case RISCV::LH: |
388 | case RISCV::LW: |
389 | case RISCV::LBU: |
390 | case RISCV::LHU: |
391 | case RISCV::LWU: |
392 | case RISCV::LD: |
393 | case RISCV::FLH: |
394 | case RISCV::FLW: |
395 | case RISCV::FLD: |
396 | case RISCV::SB: |
397 | case RISCV::SH: |
398 | case RISCV::SW: |
399 | case RISCV::SD: |
400 | case RISCV::FSH: |
401 | case RISCV::FSW: |
402 | case RISCV::FSD: { |
403 | if (UseMI.getOperand(i: 1).isFI()) |
404 | return false; |
405 | // Register defined by Lo should not be the value register. |
406 | if (DestReg == UseMI.getOperand(i: 0).getReg()) |
407 | return false; |
408 | assert(DestReg == UseMI.getOperand(1).getReg() && |
409 | "Expected base address use" ); |
410 | // All load/store instructions must use the same offset. |
411 | int64_t Offset = UseMI.getOperand(i: 2).getImm(); |
412 | if (CommonOffset && Offset != CommonOffset) |
413 | return false; |
414 | CommonOffset = Offset; |
415 | break; |
416 | } |
417 | case RISCV::INLINEASM: |
418 | case RISCV::INLINEASM_BR: { |
419 | SmallVector<unsigned> InlineAsmMemoryOpIndexes; |
420 | unsigned NumOps = 0; |
421 | for (unsigned I = InlineAsm::MIOp_FirstOperand; |
422 | I < UseMI.getNumOperands(); I += 1 + NumOps) { |
423 | const MachineOperand &FlagsMO = UseMI.getOperand(i: I); |
424 | // Should be an imm. |
425 | if (!FlagsMO.isImm()) |
426 | continue; |
427 | |
428 | const InlineAsm::Flag Flags(FlagsMO.getImm()); |
429 | NumOps = Flags.getNumOperandRegisters(); |
430 | |
431 | // Memory constraints have two operands. |
432 | if (NumOps != 2 || !Flags.isMemKind()) { |
433 | // If the register is used by something other than a memory contraint, |
434 | // we should not fold. |
435 | for (unsigned J = 0; J < NumOps; ++J) { |
436 | const MachineOperand &MO = UseMI.getOperand(i: I + 1 + J); |
437 | if (MO.isReg() && MO.getReg() == DestReg) |
438 | return false; |
439 | } |
440 | continue; |
441 | } |
442 | |
443 | // We can't do this for constraint A because AMO instructions don't have |
444 | // an immediate offset field. |
445 | if (Flags.getMemoryConstraintID() == InlineAsm::ConstraintCode::A) |
446 | return false; |
447 | |
448 | const MachineOperand &AddrMO = UseMI.getOperand(i: I + 1); |
449 | if (!AddrMO.isReg() || AddrMO.getReg() != DestReg) |
450 | continue; |
451 | |
452 | const MachineOperand &OffsetMO = UseMI.getOperand(i: I + 2); |
453 | if (!OffsetMO.isImm()) |
454 | continue; |
455 | |
456 | // All inline asm memory operands must use the same offset. |
457 | int64_t Offset = OffsetMO.getImm(); |
458 | if (CommonOffset && Offset != CommonOffset) |
459 | return false; |
460 | CommonOffset = Offset; |
461 | InlineAsmMemoryOpIndexes.push_back(Elt: I + 1); |
462 | } |
463 | InlineAsmMemoryOpIndexesMap.insert( |
464 | KV: std::make_pair(x: &UseMI, y&: InlineAsmMemoryOpIndexes)); |
465 | break; |
466 | } |
467 | } |
468 | } |
469 | |
470 | // We found a common offset. |
471 | // Update the offsets in global address lowering. |
472 | // We may have already folded some arithmetic so we need to add to any |
473 | // existing offset. |
474 | int64_t NewOffset = Hi.getOperand(i: 1).getOffset() + *CommonOffset; |
475 | // RV32 ignores the upper 32 bits. |
476 | if (!ST->is64Bit()) |
477 | NewOffset = SignExtend64<32>(x: NewOffset); |
478 | // We can only fold simm32 offsets. |
479 | if (!isInt<32>(x: NewOffset)) |
480 | return false; |
481 | |
482 | Hi.getOperand(i: 1).setOffset(NewOffset); |
483 | MachineOperand &ImmOp = Lo.getOperand(i: 2); |
484 | // Expand PseudoMovAddr into LUI |
485 | if (Hi.getOpcode() == RISCV::PseudoMovAddr) { |
486 | auto *TII = ST->getInstrInfo(); |
487 | Hi.setDesc(TII->get(Opcode: RISCV::LUI)); |
488 | Hi.removeOperand(OpNo: 2); |
489 | } |
490 | |
491 | if (Hi.getOpcode() != RISCV::AUIPC) |
492 | ImmOp.setOffset(NewOffset); |
493 | |
494 | // Update the immediate in the load/store instructions to add the offset. |
495 | for (MachineInstr &UseMI : |
496 | llvm::make_early_inc_range(Range: MRI->use_instructions(Reg: DestReg))) { |
497 | if (UseMI.getOpcode() == RISCV::INLINEASM || |
498 | UseMI.getOpcode() == RISCV::INLINEASM_BR) { |
499 | auto &InlineAsmMemoryOpIndexes = InlineAsmMemoryOpIndexesMap[&UseMI]; |
500 | for (unsigned I : InlineAsmMemoryOpIndexes) { |
501 | MachineOperand &MO = UseMI.getOperand(i: I + 1); |
502 | switch (ImmOp.getType()) { |
503 | case MachineOperand::MO_GlobalAddress: |
504 | MO.ChangeToGA(GV: ImmOp.getGlobal(), Offset: ImmOp.getOffset(), |
505 | TargetFlags: ImmOp.getTargetFlags()); |
506 | break; |
507 | case MachineOperand::MO_MCSymbol: |
508 | MO.ChangeToMCSymbol(Sym: ImmOp.getMCSymbol(), TargetFlags: ImmOp.getTargetFlags()); |
509 | MO.setOffset(ImmOp.getOffset()); |
510 | break; |
511 | case MachineOperand::MO_BlockAddress: |
512 | MO.ChangeToBA(BA: ImmOp.getBlockAddress(), Offset: ImmOp.getOffset(), |
513 | TargetFlags: ImmOp.getTargetFlags()); |
514 | break; |
515 | default: |
516 | report_fatal_error(reason: "unsupported machine operand type" ); |
517 | break; |
518 | } |
519 | } |
520 | } else { |
521 | UseMI.removeOperand(OpNo: 2); |
522 | UseMI.addOperand(Op: ImmOp); |
523 | } |
524 | } |
525 | |
526 | // Prevent Lo (originally PseudoMovAddr, which is also pointed by Hi) from |
527 | // being erased |
528 | if (&Lo == &Hi) |
529 | return true; |
530 | |
531 | MRI->replaceRegWith(FromReg: Lo.getOperand(i: 0).getReg(), ToReg: Hi.getOperand(i: 0).getReg()); |
532 | Lo.eraseFromParent(); |
533 | return true; |
534 | } |
535 | |
536 | bool RISCVMergeBaseOffsetOpt::runOnMachineFunction(MachineFunction &Fn) { |
537 | if (skipFunction(F: Fn.getFunction())) |
538 | return false; |
539 | |
540 | ST = &Fn.getSubtarget<RISCVSubtarget>(); |
541 | |
542 | bool MadeChange = false; |
543 | MRI = &Fn.getRegInfo(); |
544 | for (MachineBasicBlock &MBB : Fn) { |
545 | LLVM_DEBUG(dbgs() << "MBB: " << MBB.getName() << "\n" ); |
546 | for (MachineInstr &Hi : MBB) { |
547 | MachineInstr *Lo = nullptr; |
548 | if (!detectFoldable(Hi, Lo)) |
549 | continue; |
550 | MadeChange |= detectAndFoldOffset(Hi, Lo&: *Lo); |
551 | MadeChange |= foldIntoMemoryOps(Hi, Lo&: *Lo); |
552 | } |
553 | } |
554 | |
555 | return MadeChange; |
556 | } |
557 | |
558 | /// Returns an instance of the Merge Base Offset Optimization pass. |
559 | FunctionPass *llvm::createRISCVMergeBaseOffsetOptPass() { |
560 | return new RISCVMergeBaseOffsetOpt(); |
561 | } |
562 | |