1 | //===-- RISCVMakeCompressible.cpp - Make more instructions compressible ---===// |
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 | // This pass searches for instructions that are prevented from being compressed |
10 | // by one of the following: |
11 | // |
12 | // 1. The use of a single uncompressed register. |
13 | // 2. A base register + offset where the offset is too large to be compressed |
14 | // and the base register may or may not be compressed. |
15 | // |
16 | // |
17 | // For case 1, if a compressed register is available, then the uncompressed |
18 | // register is copied to the compressed register and its uses are replaced. |
19 | // |
20 | // For example, storing zero uses the uncompressible zero register: |
21 | // sw zero, 0(a0) # if zero |
22 | // sw zero, 8(a0) # if zero |
23 | // sw zero, 4(a0) # if zero |
24 | // sw zero, 24(a0) # if zero |
25 | // |
26 | // If a compressed register (e.g. a1) is available, the above can be transformed |
27 | // to the following to improve code size: |
28 | // li a1, 0 |
29 | // c.sw a1, 0(a0) |
30 | // c.sw a1, 8(a0) |
31 | // c.sw a1, 4(a0) |
32 | // c.sw a1, 24(a0) |
33 | // |
34 | // |
35 | // For case 2, if a compressed register is available, then the original base |
36 | // is copied and adjusted such that: |
37 | // |
38 | // new_base_register = base_register + adjustment |
39 | // base_register + large_offset = new_base_register + small_offset |
40 | // |
41 | // For example, the following offsets are too large for c.sw: |
42 | // lui a2, 983065 |
43 | // sw a1, -236(a2) |
44 | // sw a1, -240(a2) |
45 | // sw a1, -244(a2) |
46 | // sw a1, -248(a2) |
47 | // sw a1, -252(a2) |
48 | // sw a0, -256(a2) |
49 | // |
50 | // If a compressed register is available (e.g. a3), a new base could be created |
51 | // such that the addresses can accessed with a compressible offset, thus |
52 | // improving code size: |
53 | // lui a2, 983065 |
54 | // addi a3, a2, -256 |
55 | // c.sw a1, 20(a3) |
56 | // c.sw a1, 16(a3) |
57 | // c.sw a1, 12(a3) |
58 | // c.sw a1, 8(a3) |
59 | // c.sw a1, 4(a3) |
60 | // c.sw a0, 0(a3) |
61 | // |
62 | // |
63 | // This optimization is only applied if there are enough uses of the copied |
64 | // register for code size to be reduced. |
65 | // |
66 | //===----------------------------------------------------------------------===// |
67 | |
68 | #include "RISCV.h" |
69 | #include "RISCVSubtarget.h" |
70 | #include "llvm/CodeGen/Passes.h" |
71 | #include "llvm/CodeGen/RegisterScavenging.h" |
72 | #include "llvm/MC/TargetRegistry.h" |
73 | #include "llvm/Support/Debug.h" |
74 | |
75 | using namespace llvm; |
76 | |
77 | #define DEBUG_TYPE "riscv-make-compressible" |
78 | #define RISCV_COMPRESS_INSTRS_NAME "RISC-V Make Compressible" |
79 | |
80 | namespace { |
81 | |
82 | struct RISCVMakeCompressibleOpt : public MachineFunctionPass { |
83 | static char ID; |
84 | |
85 | bool runOnMachineFunction(MachineFunction &Fn) override; |
86 | |
87 | RISCVMakeCompressibleOpt() : MachineFunctionPass(ID) {} |
88 | |
89 | StringRef getPassName() const override { return RISCV_COMPRESS_INSTRS_NAME; } |
90 | }; |
91 | } // namespace |
92 | |
93 | char RISCVMakeCompressibleOpt::ID = 0; |
94 | INITIALIZE_PASS(RISCVMakeCompressibleOpt, "riscv-make-compressible" , |
95 | RISCV_COMPRESS_INSTRS_NAME, false, false) |
96 | |
97 | // Return log2(widthInBytes) of load/store done by Opcode. |
98 | static unsigned log2LdstWidth(unsigned Opcode) { |
99 | switch (Opcode) { |
100 | default: |
101 | llvm_unreachable("Unexpected opcode" ); |
102 | case RISCV::LBU: |
103 | case RISCV::SB: |
104 | return 0; |
105 | case RISCV::LH: |
106 | case RISCV::LHU: |
107 | case RISCV::SH: |
108 | return 1; |
109 | case RISCV::LW: |
110 | case RISCV::SW: |
111 | case RISCV::FLW: |
112 | case RISCV::FSW: |
113 | return 2; |
114 | case RISCV::LD: |
115 | case RISCV::SD: |
116 | case RISCV::FLD: |
117 | case RISCV::FSD: |
118 | return 3; |
119 | } |
120 | } |
121 | |
122 | // Return bit field size of immediate operand of Opcode. |
123 | static unsigned offsetMask(unsigned Opcode) { |
124 | switch (Opcode) { |
125 | default: |
126 | llvm_unreachable("Unexpected opcode" ); |
127 | case RISCV::LBU: |
128 | case RISCV::SB: |
129 | return maskTrailingOnes<unsigned>(N: 2U); |
130 | case RISCV::LH: |
131 | case RISCV::LHU: |
132 | case RISCV::SH: |
133 | return maskTrailingOnes<unsigned>(N: 1U); |
134 | case RISCV::LW: |
135 | case RISCV::SW: |
136 | case RISCV::FLW: |
137 | case RISCV::FSW: |
138 | case RISCV::LD: |
139 | case RISCV::SD: |
140 | case RISCV::FLD: |
141 | case RISCV::FSD: |
142 | return maskTrailingOnes<unsigned>(N: 5U); |
143 | } |
144 | } |
145 | |
146 | // Return a mask for the offset bits of a non-stack-pointer based compressed |
147 | // load/store. |
148 | static uint8_t compressedLDSTOffsetMask(unsigned Opcode) { |
149 | return offsetMask(Opcode) << log2LdstWidth(Opcode); |
150 | } |
151 | |
152 | // Return true if Offset fits within a compressed stack-pointer based |
153 | // load/store. |
154 | static bool compressibleSPOffset(int64_t Offset, unsigned Opcode) { |
155 | // Compressed sp-based loads and stores only work for 32/64 bits. |
156 | switch (log2LdstWidth(Opcode)) { |
157 | case 2: |
158 | return isShiftedUInt<6, 2>(x: Offset); |
159 | case 3: |
160 | return isShiftedUInt<6, 3>(x: Offset); |
161 | } |
162 | return false; |
163 | } |
164 | |
165 | // Given an offset for a load/store, return the adjustment required to the base |
166 | // register such that the address can be accessed with a compressible offset. |
167 | // This will return 0 if the offset is already compressible. |
168 | static int64_t getBaseAdjustForCompression(int64_t Offset, unsigned Opcode) { |
169 | // Return the excess bits that do not fit in a compressible offset. |
170 | return Offset & ~compressedLDSTOffsetMask(Opcode); |
171 | } |
172 | |
173 | // Return true if Reg is in a compressed register class. |
174 | static bool isCompressedReg(Register Reg) { |
175 | return RISCV::GPRCRegClass.contains(Reg) || |
176 | RISCV::FPR32CRegClass.contains(Reg) || |
177 | RISCV::FPR64CRegClass.contains(Reg); |
178 | } |
179 | |
180 | // Return true if MI is a load for which there exists a compressed version. |
181 | static bool isCompressibleLoad(const MachineInstr &MI) { |
182 | const RISCVSubtarget &STI = MI.getMF()->getSubtarget<RISCVSubtarget>(); |
183 | |
184 | switch (MI.getOpcode()) { |
185 | default: |
186 | return false; |
187 | case RISCV::LBU: |
188 | case RISCV::LH: |
189 | case RISCV::LHU: |
190 | return STI.hasStdExtZcb(); |
191 | case RISCV::LW: |
192 | case RISCV::LD: |
193 | return STI.hasStdExtCOrZca(); |
194 | case RISCV::FLW: |
195 | return !STI.is64Bit() && STI.hasStdExtCOrZcfOrZce(); |
196 | case RISCV::FLD: |
197 | return STI.hasStdExtCOrZcd(); |
198 | } |
199 | } |
200 | |
201 | // Return true if MI is a store for which there exists a compressed version. |
202 | static bool isCompressibleStore(const MachineInstr &MI) { |
203 | const RISCVSubtarget &STI = MI.getMF()->getSubtarget<RISCVSubtarget>(); |
204 | |
205 | switch (MI.getOpcode()) { |
206 | default: |
207 | return false; |
208 | case RISCV::SB: |
209 | case RISCV::SH: |
210 | return STI.hasStdExtZcb(); |
211 | case RISCV::SW: |
212 | case RISCV::SD: |
213 | return STI.hasStdExtCOrZca(); |
214 | case RISCV::FSW: |
215 | return !STI.is64Bit() && STI.hasStdExtCOrZcfOrZce(); |
216 | case RISCV::FSD: |
217 | return STI.hasStdExtCOrZcd(); |
218 | } |
219 | } |
220 | |
221 | // Find a single register and/or large offset which, if compressible, would |
222 | // allow the given instruction to be compressed. |
223 | // |
224 | // Possible return values: |
225 | // |
226 | // {Reg, 0} - Uncompressed Reg needs replacing with a compressed |
227 | // register. |
228 | // {Reg, N} - Reg needs replacing with a compressed register and |
229 | // N needs adding to the new register. (Reg may be |
230 | // compressed or uncompressed). |
231 | // {RISCV::NoRegister, 0} - No suitable optimization found for this |
232 | // instruction. |
233 | static RegImmPair getRegImmPairPreventingCompression(const MachineInstr &MI) { |
234 | const unsigned Opcode = MI.getOpcode(); |
235 | |
236 | if (isCompressibleLoad(MI) || isCompressibleStore(MI)) { |
237 | const MachineOperand &MOImm = MI.getOperand(i: 2); |
238 | if (!MOImm.isImm()) |
239 | return RegImmPair(RISCV::NoRegister, 0); |
240 | |
241 | int64_t Offset = MOImm.getImm(); |
242 | int64_t NewBaseAdjust = getBaseAdjustForCompression(Offset, Opcode); |
243 | Register Base = MI.getOperand(i: 1).getReg(); |
244 | |
245 | // Memory accesses via the stack pointer do not have a requirement for |
246 | // either of the registers to be compressible and can take a larger offset. |
247 | if (RISCV::SPRegClass.contains(Reg: Base)) { |
248 | if (!compressibleSPOffset(Offset, Opcode) && NewBaseAdjust) |
249 | return RegImmPair(Base, NewBaseAdjust); |
250 | } else { |
251 | Register SrcDest = MI.getOperand(i: 0).getReg(); |
252 | bool SrcDestCompressed = isCompressedReg(Reg: SrcDest); |
253 | bool BaseCompressed = isCompressedReg(Reg: Base); |
254 | |
255 | // If only Base and/or offset prevent compression, then return Base and |
256 | // any adjustment required to make the offset compressible. |
257 | if ((!BaseCompressed || NewBaseAdjust) && SrcDestCompressed) |
258 | return RegImmPair(Base, NewBaseAdjust); |
259 | |
260 | // For loads, we can only change the base register since dest is defined |
261 | // rather than used. |
262 | // |
263 | // For stores, we can change SrcDest (and Base if SrcDest == Base) but |
264 | // cannot resolve an uncompressible offset in this case. |
265 | if (isCompressibleStore(MI)) { |
266 | if (!SrcDestCompressed && (BaseCompressed || SrcDest == Base) && |
267 | !NewBaseAdjust) |
268 | return RegImmPair(SrcDest, NewBaseAdjust); |
269 | } |
270 | } |
271 | } |
272 | return RegImmPair(RISCV::NoRegister, 0); |
273 | } |
274 | |
275 | // Check all uses after FirstMI of the given register, keeping a vector of |
276 | // instructions that would be compressible if the given register (and offset if |
277 | // applicable) were compressible. |
278 | // |
279 | // If there are enough uses for this optimization to improve code size and a |
280 | // compressed register is available, return that compressed register. |
281 | static Register analyzeCompressibleUses(MachineInstr &FirstMI, |
282 | RegImmPair RegImm, |
283 | SmallVectorImpl<MachineInstr *> &MIs) { |
284 | MachineBasicBlock &MBB = *FirstMI.getParent(); |
285 | const TargetRegisterInfo *TRI = |
286 | MBB.getParent()->getSubtarget().getRegisterInfo(); |
287 | |
288 | for (MachineBasicBlock::instr_iterator I = FirstMI.getIterator(), |
289 | E = MBB.instr_end(); |
290 | I != E; ++I) { |
291 | MachineInstr &MI = *I; |
292 | |
293 | // Determine if this is an instruction which would benefit from using the |
294 | // new register. |
295 | RegImmPair CandidateRegImm = getRegImmPairPreventingCompression(MI); |
296 | if (CandidateRegImm.Reg == RegImm.Reg && CandidateRegImm.Imm == RegImm.Imm) |
297 | MIs.push_back(Elt: &MI); |
298 | |
299 | // If RegImm.Reg is modified by this instruction, then we cannot optimize |
300 | // past this instruction. If the register is already compressed, then it may |
301 | // possible to optimize a large offset in the current instruction - this |
302 | // will have been detected by the preceeding call to |
303 | // getRegImmPairPreventingCompression. |
304 | if (MI.modifiesRegister(Reg: RegImm.Reg, TRI)) |
305 | break; |
306 | } |
307 | |
308 | // Adjusting the base costs one new uncompressed addi and therefore three uses |
309 | // are required for a code size reduction. If no base adjustment is required, |
310 | // then copying the register costs one new c.mv (or c.li Rd, 0 for "copying" |
311 | // the zero register) and therefore two uses are required for a code size |
312 | // reduction. |
313 | if (MIs.size() < 2 || (RegImm.Imm != 0 && MIs.size() < 3)) |
314 | return RISCV::NoRegister; |
315 | |
316 | // Find a compressible register which will be available from the first |
317 | // instruction we care about to the last. |
318 | const TargetRegisterClass *RCToScavenge; |
319 | |
320 | // Work out the compressed register class from which to scavenge. |
321 | if (RISCV::GPRRegClass.contains(Reg: RegImm.Reg)) |
322 | RCToScavenge = &RISCV::GPRCRegClass; |
323 | else if (RISCV::FPR32RegClass.contains(Reg: RegImm.Reg)) |
324 | RCToScavenge = &RISCV::FPR32CRegClass; |
325 | else if (RISCV::FPR64RegClass.contains(Reg: RegImm.Reg)) |
326 | RCToScavenge = &RISCV::FPR64CRegClass; |
327 | else |
328 | return RISCV::NoRegister; |
329 | |
330 | RegScavenger RS; |
331 | RS.enterBasicBlockEnd(MBB); |
332 | RS.backward(I: std::next(x: MIs.back()->getIterator())); |
333 | return RS.scavengeRegisterBackwards(RC: *RCToScavenge, To: FirstMI.getIterator(), |
334 | /*RestoreAfter=*/false, /*SPAdj=*/0, |
335 | /*AllowSpill=*/false); |
336 | } |
337 | |
338 | // Update uses of the old register in the given instruction to the new register. |
339 | static void updateOperands(MachineInstr &MI, RegImmPair OldRegImm, |
340 | Register NewReg) { |
341 | unsigned Opcode = MI.getOpcode(); |
342 | |
343 | // If this pass is extended to support more instructions, the check for |
344 | // definedness may need to be strengthened. |
345 | assert((isCompressibleLoad(MI) || isCompressibleStore(MI)) && |
346 | "Unsupported instruction for this optimization." ); |
347 | |
348 | int SkipN = 0; |
349 | |
350 | // Skip the first (value) operand to a store instruction (except if the store |
351 | // offset is zero) in order to avoid an incorrect transformation. |
352 | // e.g. sd a0, 808(a0) to addi a2, a0, 768; sd a2, 40(a2) |
353 | if (isCompressibleStore(MI) && OldRegImm.Imm != 0) |
354 | SkipN = 1; |
355 | |
356 | // Update registers |
357 | for (MachineOperand &MO : drop_begin(RangeOrContainer: MI.operands(), N: SkipN)) |
358 | if (MO.isReg() && MO.getReg() == OldRegImm.Reg) { |
359 | // Do not update operands that define the old register. |
360 | // |
361 | // The new register was scavenged for the range of instructions that are |
362 | // being updated, therefore it should not be defined within this range |
363 | // except possibly in the final instruction. |
364 | if (MO.isDef()) { |
365 | assert(isCompressibleLoad(MI)); |
366 | continue; |
367 | } |
368 | // Update reg |
369 | MO.setReg(NewReg); |
370 | } |
371 | |
372 | // Update offset |
373 | MachineOperand &MOImm = MI.getOperand(i: 2); |
374 | int64_t NewOffset = MOImm.getImm() & compressedLDSTOffsetMask(Opcode); |
375 | MOImm.setImm(NewOffset); |
376 | } |
377 | |
378 | bool RISCVMakeCompressibleOpt::runOnMachineFunction(MachineFunction &Fn) { |
379 | // This is a size optimization. |
380 | if (skipFunction(F: Fn.getFunction()) || !Fn.getFunction().hasMinSize()) |
381 | return false; |
382 | |
383 | const RISCVSubtarget &STI = Fn.getSubtarget<RISCVSubtarget>(); |
384 | const RISCVInstrInfo &TII = *STI.getInstrInfo(); |
385 | |
386 | // This optimization only makes sense if compressed instructions are emitted. |
387 | if (!STI.hasStdExtCOrZca()) |
388 | return false; |
389 | |
390 | for (MachineBasicBlock &MBB : Fn) { |
391 | LLVM_DEBUG(dbgs() << "MBB: " << MBB.getName() << "\n" ); |
392 | for (MachineInstr &MI : MBB) { |
393 | // Determine if this instruction would otherwise be compressed if not for |
394 | // an uncompressible register or offset. |
395 | RegImmPair RegImm = getRegImmPairPreventingCompression(MI); |
396 | if (!RegImm.Reg && RegImm.Imm == 0) |
397 | continue; |
398 | |
399 | // Determine if there is a set of instructions for which replacing this |
400 | // register with a compressed register (and compressible offset if |
401 | // applicable) is possible and will allow compression. |
402 | SmallVector<MachineInstr *, 8> MIs; |
403 | Register NewReg = analyzeCompressibleUses(FirstMI&: MI, RegImm, MIs); |
404 | if (!NewReg) |
405 | continue; |
406 | |
407 | // Create the appropriate copy and/or offset. |
408 | if (RISCV::GPRRegClass.contains(Reg: RegImm.Reg)) { |
409 | assert(isInt<12>(RegImm.Imm)); |
410 | BuildMI(BB&: MBB, I&: MI, MIMD: MI.getDebugLoc(), MCID: TII.get(Opcode: RISCV::ADDI), DestReg: NewReg) |
411 | .addReg(RegNo: RegImm.Reg) |
412 | .addImm(Val: RegImm.Imm); |
413 | } else { |
414 | // If we are looking at replacing an FPR register we don't expect to |
415 | // have any offset. The only compressible FP instructions with an offset |
416 | // are loads and stores, for which the offset applies to the GPR operand |
417 | // not the FPR operand. |
418 | assert(RegImm.Imm == 0); |
419 | unsigned Opcode = RISCV::FPR32RegClass.contains(Reg: RegImm.Reg) |
420 | ? RISCV::FSGNJ_S |
421 | : RISCV::FSGNJ_D; |
422 | BuildMI(BB&: MBB, I&: MI, MIMD: MI.getDebugLoc(), MCID: TII.get(Opcode), DestReg: NewReg) |
423 | .addReg(RegNo: RegImm.Reg) |
424 | .addReg(RegNo: RegImm.Reg); |
425 | } |
426 | |
427 | // Update the set of instructions to use the compressed register and |
428 | // compressible offset instead. These instructions should now be |
429 | // compressible. |
430 | // TODO: Update all uses if RegImm.Imm == 0? Not just those that are |
431 | // expected to become compressible. |
432 | for (MachineInstr *UpdateMI : MIs) |
433 | updateOperands(MI&: *UpdateMI, OldRegImm: RegImm, NewReg); |
434 | } |
435 | } |
436 | return true; |
437 | } |
438 | |
439 | /// Returns an instance of the Make Compressible Optimization pass. |
440 | FunctionPass *llvm::createRISCVMakeCompressibleOptPass() { |
441 | return new RISCVMakeCompressibleOpt(); |
442 | } |
443 | |