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     initializeRISCVMakeCompressibleOptPass(*PassRegistry::getPassRegistry());
89   }
90 
91   StringRef getPassName() const override { return RISCV_COMPRESS_INSTRS_NAME; }
92 };
93 } // namespace
94 
95 char RISCVMakeCompressibleOpt::ID = 0;
96 INITIALIZE_PASS(RISCVMakeCompressibleOpt, "riscv-make-compressible",
97                 RISCV_COMPRESS_INSTRS_NAME, false, false)
98 
99 // Return log2(widthInBytes) of load/store done by Opcode.
100 static unsigned log2LdstWidth(unsigned Opcode) {
101   switch (Opcode) {
102   default:
103     llvm_unreachable("Unexpected opcode");
104   case RISCV::LW:
105   case RISCV::SW:
106   case RISCV::FLW:
107   case RISCV::FSW:
108     return 2;
109   case RISCV::LD:
110   case RISCV::SD:
111   case RISCV::FLD:
112   case RISCV::FSD:
113     return 3;
114   }
115 }
116 
117 // Return a mask for the offset bits of a non-stack-pointer based compressed
118 // load/store.
119 static uint8_t compressedLDSTOffsetMask(unsigned Opcode) {
120   return 0x1f << log2LdstWidth(Opcode);
121 }
122 
123 // Return true if Offset fits within a compressed stack-pointer based
124 // load/store.
125 static bool compressibleSPOffset(int64_t Offset, unsigned Opcode) {
126   return log2LdstWidth(Opcode) == 2 ? isShiftedUInt<6, 2>(Offset)
127                                     : isShiftedUInt<6, 3>(Offset);
128 }
129 
130 // Given an offset for a load/store, return the adjustment required to the base
131 // register such that the address can be accessed with a compressible offset.
132 // This will return 0 if the offset is already compressible.
133 static int64_t getBaseAdjustForCompression(int64_t Offset, unsigned Opcode) {
134   // Return the excess bits that do not fit in a compressible offset.
135   return Offset & ~compressedLDSTOffsetMask(Opcode);
136 }
137 
138 // Return true if Reg is in a compressed register class.
139 static bool isCompressedReg(Register Reg) {
140   return RISCV::GPRCRegClass.contains(Reg) ||
141          RISCV::FPR32CRegClass.contains(Reg) ||
142          RISCV::FPR64CRegClass.contains(Reg);
143 }
144 
145 // Return true if MI is a load for which there exists a compressed version.
146 static bool isCompressibleLoad(const MachineInstr &MI) {
147   const RISCVSubtarget &STI = MI.getMF()->getSubtarget<RISCVSubtarget>();
148   const unsigned Opcode = MI.getOpcode();
149 
150   return Opcode == RISCV::LW || (!STI.is64Bit() && Opcode == RISCV::FLW) ||
151          Opcode == RISCV::LD || Opcode == RISCV::FLD;
152 }
153 
154 // Return true if MI is a store for which there exists a compressed version.
155 static bool isCompressibleStore(const MachineInstr &MI) {
156   const RISCVSubtarget &STI = MI.getMF()->getSubtarget<RISCVSubtarget>();
157   const unsigned Opcode = MI.getOpcode();
158 
159   return Opcode == RISCV::SW || (!STI.is64Bit() && Opcode == RISCV::FSW) ||
160          Opcode == RISCV::SD || Opcode == RISCV::FSD;
161 }
162 
163 // Find a single register and/or large offset which, if compressible, would
164 // allow the given instruction to be compressed.
165 //
166 // Possible return values:
167 //
168 //   {Reg, 0}               - Uncompressed Reg needs replacing with a compressed
169 //                            register.
170 //   {Reg, N}               - Reg needs replacing with a compressed register and
171 //                            N needs adding to the new register. (Reg may be
172 //                            compressed or uncompressed).
173 //   {RISCV::NoRegister, 0} - No suitable optimization found for this
174 //   instruction.
175 static RegImmPair getRegImmPairPreventingCompression(const MachineInstr &MI) {
176   const unsigned Opcode = MI.getOpcode();
177 
178   if (isCompressibleLoad(MI) || isCompressibleStore(MI)) {
179     const MachineOperand &MOImm = MI.getOperand(2);
180     if (!MOImm.isImm())
181       return RegImmPair(RISCV::NoRegister, 0);
182 
183     int64_t Offset = MOImm.getImm();
184     int64_t NewBaseAdjust = getBaseAdjustForCompression(Offset, Opcode);
185     Register Base = MI.getOperand(1).getReg();
186 
187     // Memory accesses via the stack pointer do not have a requirement for
188     // either of the registers to be compressible and can take a larger offset.
189     if (RISCV::SPRegClass.contains(Base)) {
190       if (!compressibleSPOffset(Offset, Opcode) && NewBaseAdjust)
191         return RegImmPair(Base, NewBaseAdjust);
192     } else {
193       Register SrcDest = MI.getOperand(0).getReg();
194       bool SrcDestCompressed = isCompressedReg(SrcDest);
195       bool BaseCompressed = isCompressedReg(Base);
196 
197       // If only Base and/or offset prevent compression, then return Base and
198       // any adjustment required to make the offset compressible.
199       if ((!BaseCompressed || NewBaseAdjust) && SrcDestCompressed)
200         return RegImmPair(Base, NewBaseAdjust);
201 
202       // For loads, we can only change the base register since dest is defined
203       // rather than used.
204       //
205       // For stores, we can change SrcDest (and Base if SrcDest == Base) but
206       // cannot resolve an uncompressible offset in this case.
207       if (isCompressibleStore(MI)) {
208         if (!SrcDestCompressed && (BaseCompressed || SrcDest == Base) &&
209             !NewBaseAdjust)
210           return RegImmPair(SrcDest, NewBaseAdjust);
211       }
212     }
213   }
214   return RegImmPair(RISCV::NoRegister, 0);
215 }
216 
217 // Check all uses after FirstMI of the given register, keeping a vector of
218 // instructions that would be compressible if the given register (and offset if
219 // applicable) were compressible.
220 //
221 // If there are enough uses for this optimization to improve code size and a
222 // compressed register is available, return that compressed register.
223 static Register analyzeCompressibleUses(MachineInstr &FirstMI,
224                                         RegImmPair RegImm,
225                                         SmallVectorImpl<MachineInstr *> &MIs) {
226   MachineBasicBlock &MBB = *FirstMI.getParent();
227   const TargetRegisterInfo *TRI =
228       MBB.getParent()->getSubtarget().getRegisterInfo();
229 
230   for (MachineBasicBlock::instr_iterator I = FirstMI.getIterator(),
231                                          E = MBB.instr_end();
232        I != E; ++I) {
233     MachineInstr &MI = *I;
234 
235     // Determine if this is an instruction which would benefit from using the
236     // new register.
237     RegImmPair CandidateRegImm = getRegImmPairPreventingCompression(MI);
238     if (CandidateRegImm.Reg == RegImm.Reg && CandidateRegImm.Imm == RegImm.Imm)
239       MIs.push_back(&MI);
240 
241     // If RegImm.Reg is modified by this instruction, then we cannot optimize
242     // past this instruction. If the register is already compressed, then it may
243     // possible to optimize a large offset in the current instruction - this
244     // will have been detected by the preceeding call to
245     // getRegImmPairPreventingCompression.
246     if (MI.modifiesRegister(RegImm.Reg, TRI))
247       break;
248   }
249 
250   // Adjusting the base costs one new uncompressed addi and therefore three uses
251   // are required for a code size reduction. If no base adjustment is required,
252   // then copying the register costs one new c.mv (or c.li Rd, 0 for "copying"
253   // the zero register) and therefore two uses are required for a code size
254   // reduction.
255   if (MIs.size() < 2 || (RegImm.Imm != 0 && MIs.size() < 3))
256     return RISCV::NoRegister;
257 
258   // Find a compressible register which will be available from the first
259   // instruction we care about to the last.
260   const TargetRegisterClass *RCToScavenge;
261 
262   // Work out the compressed register class from which to scavenge.
263   if (RISCV::GPRRegClass.contains(RegImm.Reg))
264     RCToScavenge = &RISCV::GPRCRegClass;
265   else if (RISCV::FPR32RegClass.contains(RegImm.Reg))
266     RCToScavenge = &RISCV::FPR32CRegClass;
267   else if (RISCV::FPR64RegClass.contains(RegImm.Reg))
268     RCToScavenge = &RISCV::FPR64CRegClass;
269   else
270     return RISCV::NoRegister;
271 
272   RegScavenger RS;
273   RS.enterBasicBlockEnd(MBB);
274   RS.backward(MIs.back()->getIterator());
275   return RS.scavengeRegisterBackwards(*RCToScavenge, FirstMI.getIterator(),
276                                       /*RestoreAfter=*/false, /*SPAdj=*/0,
277                                       /*AllowSpill=*/false);
278 }
279 
280 // Update uses of the old register in the given instruction to the new register.
281 static void updateOperands(MachineInstr &MI, RegImmPair OldRegImm,
282                            Register NewReg) {
283   unsigned Opcode = MI.getOpcode();
284 
285   // If this pass is extended to support more instructions, the check for
286   // definedness may need to be strengthened.
287   assert((isCompressibleLoad(MI) || isCompressibleStore(MI)) &&
288          "Unsupported instruction for this optimization.");
289 
290   int SkipN = 0;
291 
292   // Skip the first (value) operand to a store instruction (except if the store
293   // offset is zero) in order to avoid an incorrect transformation.
294   // e.g. sd a0, 808(a0) to addi a2, a0, 768; sd a2, 40(a2)
295   if (isCompressibleStore(MI) && OldRegImm.Imm != 0)
296     SkipN = 1;
297 
298   // Update registers
299   for (MachineOperand &MO : drop_begin(MI.operands(), SkipN))
300     if (MO.isReg() && MO.getReg() == OldRegImm.Reg) {
301       // Do not update operands that define the old register.
302       //
303       // The new register was scavenged for the range of instructions that are
304       // being updated, therefore it should not be defined within this range
305       // except possibly in the final instruction.
306       if (MO.isDef()) {
307         assert(isCompressibleLoad(MI));
308         continue;
309       }
310       // Update reg
311       MO.setReg(NewReg);
312     }
313 
314   // Update offset
315   MachineOperand &MOImm = MI.getOperand(2);
316   int64_t NewOffset = MOImm.getImm() & compressedLDSTOffsetMask(Opcode);
317   MOImm.setImm(NewOffset);
318 }
319 
320 bool RISCVMakeCompressibleOpt::runOnMachineFunction(MachineFunction &Fn) {
321   // This is a size optimization.
322   if (skipFunction(Fn.getFunction()) || !Fn.getFunction().hasMinSize())
323     return false;
324 
325   const RISCVSubtarget &STI = Fn.getSubtarget<RISCVSubtarget>();
326   const RISCVInstrInfo &TII = *STI.getInstrInfo();
327 
328   // This optimization only makes sense if compressed instructions are emitted.
329   // FIXME: Support Zca, Zcf, Zcd granularity.
330   if (!STI.hasStdExtC())
331     return false;
332 
333   for (MachineBasicBlock &MBB : Fn) {
334     LLVM_DEBUG(dbgs() << "MBB: " << MBB.getName() << "\n");
335     for (MachineInstr &MI : MBB) {
336       // Determine if this instruction would otherwise be compressed if not for
337       // an uncompressible register or offset.
338       RegImmPair RegImm = getRegImmPairPreventingCompression(MI);
339       if (!RegImm.Reg && RegImm.Imm == 0)
340         continue;
341 
342       // Determine if there is a set of instructions for which replacing this
343       // register with a compressed register (and compressible offset if
344       // applicable) is possible and will allow compression.
345       SmallVector<MachineInstr *, 8> MIs;
346       Register NewReg = analyzeCompressibleUses(MI, RegImm, MIs);
347       if (!NewReg)
348         continue;
349 
350       // Create the appropriate copy and/or offset.
351       if (RISCV::GPRRegClass.contains(RegImm.Reg)) {
352         assert(isInt<12>(RegImm.Imm));
353         BuildMI(MBB, MI, MI.getDebugLoc(), TII.get(RISCV::ADDI), NewReg)
354             .addReg(RegImm.Reg)
355             .addImm(RegImm.Imm);
356       } else {
357         // If we are looking at replacing an FPR register we don't expect to
358         // have any offset. The only compressible FP instructions with an offset
359         // are loads and stores, for which the offset applies to the GPR operand
360         // not the FPR operand.
361         assert(RegImm.Imm == 0);
362         unsigned Opcode = RISCV::FPR32RegClass.contains(RegImm.Reg)
363                               ? RISCV::FSGNJ_S
364                               : RISCV::FSGNJ_D;
365         BuildMI(MBB, MI, MI.getDebugLoc(), TII.get(Opcode), NewReg)
366             .addReg(RegImm.Reg)
367             .addReg(RegImm.Reg);
368       }
369 
370       // Update the set of instructions to use the compressed register and
371       // compressible offset instead. These instructions should now be
372       // compressible.
373       // TODO: Update all uses if RegImm.Imm == 0? Not just those that are
374       // expected to become compressible.
375       for (MachineInstr *UpdateMI : MIs)
376         updateOperands(*UpdateMI, RegImm, NewReg);
377     }
378   }
379   return true;
380 }
381 
382 /// Returns an instance of the Make Compressible Optimization pass.
383 FunctionPass *llvm::createRISCVMakeCompressibleOptPass() {
384   return new RISCVMakeCompressibleOpt();
385 }
386