xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp (revision ba3c1f5972d7b90feb6e6da47905ff2757e0fe57)
1 //===- SIInstrInfo.cpp - SI Instruction Information  ----------------------===//
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
10 /// SI Implementation of TargetInstrInfo.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "SIInstrInfo.h"
15 #include "AMDGPU.h"
16 #include "AMDGPUInstrInfo.h"
17 #include "GCNHazardRecognizer.h"
18 #include "GCNSubtarget.h"
19 #include "SIMachineFunctionInfo.h"
20 #include "llvm/Analysis/ValueTracking.h"
21 #include "llvm/CodeGen/LiveIntervals.h"
22 #include "llvm/CodeGen/LiveVariables.h"
23 #include "llvm/CodeGen/MachineDominators.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineScheduler.h"
26 #include "llvm/CodeGen/RegisterScavenging.h"
27 #include "llvm/CodeGen/ScheduleDAG.h"
28 #include "llvm/IR/DiagnosticInfo.h"
29 #include "llvm/IR/IntrinsicsAMDGPU.h"
30 #include "llvm/MC/MCContext.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Target/TargetMachine.h"
33 
34 using namespace llvm;
35 
36 #define DEBUG_TYPE "si-instr-info"
37 
38 #define GET_INSTRINFO_CTOR_DTOR
39 #include "AMDGPUGenInstrInfo.inc"
40 
41 namespace llvm {
42 namespace AMDGPU {
43 #define GET_D16ImageDimIntrinsics_IMPL
44 #define GET_ImageDimIntrinsicTable_IMPL
45 #define GET_RsrcIntrinsics_IMPL
46 #include "AMDGPUGenSearchableTables.inc"
47 }
48 }
49 
50 
51 // Must be at least 4 to be able to branch over minimum unconditional branch
52 // code. This is only for making it possible to write reasonably small tests for
53 // long branches.
54 static cl::opt<unsigned>
55 BranchOffsetBits("amdgpu-s-branch-bits", cl::ReallyHidden, cl::init(16),
56                  cl::desc("Restrict range of branch instructions (DEBUG)"));
57 
58 static cl::opt<bool> Fix16BitCopies(
59   "amdgpu-fix-16-bit-physreg-copies",
60   cl::desc("Fix copies between 32 and 16 bit registers by extending to 32 bit"),
61   cl::init(true),
62   cl::ReallyHidden);
63 
64 SIInstrInfo::SIInstrInfo(const GCNSubtarget &ST)
65   : AMDGPUGenInstrInfo(AMDGPU::ADJCALLSTACKUP, AMDGPU::ADJCALLSTACKDOWN),
66     RI(ST), ST(ST) {
67   SchedModel.init(&ST);
68 }
69 
70 //===----------------------------------------------------------------------===//
71 // TargetInstrInfo callbacks
72 //===----------------------------------------------------------------------===//
73 
74 static unsigned getNumOperandsNoGlue(SDNode *Node) {
75   unsigned N = Node->getNumOperands();
76   while (N && Node->getOperand(N - 1).getValueType() == MVT::Glue)
77     --N;
78   return N;
79 }
80 
81 /// Returns true if both nodes have the same value for the given
82 ///        operand \p Op, or if both nodes do not have this operand.
83 static bool nodesHaveSameOperandValue(SDNode *N0, SDNode* N1, unsigned OpName) {
84   unsigned Opc0 = N0->getMachineOpcode();
85   unsigned Opc1 = N1->getMachineOpcode();
86 
87   int Op0Idx = AMDGPU::getNamedOperandIdx(Opc0, OpName);
88   int Op1Idx = AMDGPU::getNamedOperandIdx(Opc1, OpName);
89 
90   if (Op0Idx == -1 && Op1Idx == -1)
91     return true;
92 
93 
94   if ((Op0Idx == -1 && Op1Idx != -1) ||
95       (Op1Idx == -1 && Op0Idx != -1))
96     return false;
97 
98   // getNamedOperandIdx returns the index for the MachineInstr's operands,
99   // which includes the result as the first operand. We are indexing into the
100   // MachineSDNode's operands, so we need to skip the result operand to get
101   // the real index.
102   --Op0Idx;
103   --Op1Idx;
104 
105   return N0->getOperand(Op0Idx) == N1->getOperand(Op1Idx);
106 }
107 
108 bool SIInstrInfo::isReallyTriviallyReMaterializable(
109     const MachineInstr &MI) const {
110   if (isVOP1(MI) || isVOP2(MI) || isVOP3(MI) || isSDWA(MI) || isSALU(MI)) {
111     // Normally VALU use of exec would block the rematerialization, but that
112     // is OK in this case to have an implicit exec read as all VALU do.
113     // We really want all of the generic logic for this except for this.
114 
115     // Another potential implicit use is mode register. The core logic of
116     // the RA will not attempt rematerialization if mode is set anywhere
117     // in the function, otherwise it is safe since mode is not changed.
118 
119     // There is difference to generic method which does not allow
120     // rematerialization if there are virtual register uses. We allow this,
121     // therefore this method includes SOP instructions as well.
122     return !MI.hasImplicitDef() &&
123            MI.getNumImplicitOperands() == MI.getDesc().implicit_uses().size() &&
124            !MI.mayRaiseFPException();
125   }
126 
127   return false;
128 }
129 
130 // Returns true if the scalar result of a VALU instruction depends on exec.
131 static bool resultDependsOnExec(const MachineInstr &MI) {
132   // Ignore comparisons which are only used masked with exec.
133   // This allows some hoisting/sinking of VALU comparisons.
134   if (MI.isCompare()) {
135     const MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
136     Register DstReg = MI.getOperand(0).getReg();
137     if (!DstReg.isVirtual())
138       return true;
139     for (MachineInstr &Use : MRI.use_nodbg_instructions(DstReg)) {
140       switch (Use.getOpcode()) {
141       case AMDGPU::S_AND_SAVEEXEC_B32:
142       case AMDGPU::S_AND_SAVEEXEC_B64:
143         break;
144       case AMDGPU::S_AND_B32:
145       case AMDGPU::S_AND_B64:
146         if (!Use.readsRegister(AMDGPU::EXEC))
147           return true;
148         break;
149       default:
150         return true;
151       }
152     }
153     return false;
154   }
155 
156   switch (MI.getOpcode()) {
157   default:
158     break;
159   case AMDGPU::V_READFIRSTLANE_B32:
160     return true;
161   }
162 
163   return false;
164 }
165 
166 bool SIInstrInfo::isIgnorableUse(const MachineOperand &MO) const {
167   // Any implicit use of exec by VALU is not a real register read.
168   return MO.getReg() == AMDGPU::EXEC && MO.isImplicit() &&
169          isVALU(*MO.getParent()) && !resultDependsOnExec(*MO.getParent());
170 }
171 
172 bool SIInstrInfo::areLoadsFromSameBasePtr(SDNode *Load0, SDNode *Load1,
173                                           int64_t &Offset0,
174                                           int64_t &Offset1) const {
175   if (!Load0->isMachineOpcode() || !Load1->isMachineOpcode())
176     return false;
177 
178   unsigned Opc0 = Load0->getMachineOpcode();
179   unsigned Opc1 = Load1->getMachineOpcode();
180 
181   // Make sure both are actually loads.
182   if (!get(Opc0).mayLoad() || !get(Opc1).mayLoad())
183     return false;
184 
185   if (isDS(Opc0) && isDS(Opc1)) {
186 
187     // FIXME: Handle this case:
188     if (getNumOperandsNoGlue(Load0) != getNumOperandsNoGlue(Load1))
189       return false;
190 
191     // Check base reg.
192     if (Load0->getOperand(0) != Load1->getOperand(0))
193       return false;
194 
195     // Skip read2 / write2 variants for simplicity.
196     // TODO: We should report true if the used offsets are adjacent (excluded
197     // st64 versions).
198     int Offset0Idx = AMDGPU::getNamedOperandIdx(Opc0, AMDGPU::OpName::offset);
199     int Offset1Idx = AMDGPU::getNamedOperandIdx(Opc1, AMDGPU::OpName::offset);
200     if (Offset0Idx == -1 || Offset1Idx == -1)
201       return false;
202 
203     // XXX - be careful of dataless loads
204     // getNamedOperandIdx returns the index for MachineInstrs.  Since they
205     // include the output in the operand list, but SDNodes don't, we need to
206     // subtract the index by one.
207     Offset0Idx -= get(Opc0).NumDefs;
208     Offset1Idx -= get(Opc1).NumDefs;
209     Offset0 = cast<ConstantSDNode>(Load0->getOperand(Offset0Idx))->getZExtValue();
210     Offset1 = cast<ConstantSDNode>(Load1->getOperand(Offset1Idx))->getZExtValue();
211     return true;
212   }
213 
214   if (isSMRD(Opc0) && isSMRD(Opc1)) {
215     // Skip time and cache invalidation instructions.
216     if (!AMDGPU::hasNamedOperand(Opc0, AMDGPU::OpName::sbase) ||
217         !AMDGPU::hasNamedOperand(Opc1, AMDGPU::OpName::sbase))
218       return false;
219 
220     unsigned NumOps = getNumOperandsNoGlue(Load0);
221     if (NumOps != getNumOperandsNoGlue(Load1))
222       return false;
223 
224     // Check base reg.
225     if (Load0->getOperand(0) != Load1->getOperand(0))
226       return false;
227 
228     // Match register offsets, if both register and immediate offsets present.
229     assert(NumOps == 4 || NumOps == 5);
230     if (NumOps == 5 && Load0->getOperand(1) != Load1->getOperand(1))
231       return false;
232 
233     const ConstantSDNode *Load0Offset =
234         dyn_cast<ConstantSDNode>(Load0->getOperand(NumOps - 3));
235     const ConstantSDNode *Load1Offset =
236         dyn_cast<ConstantSDNode>(Load1->getOperand(NumOps - 3));
237 
238     if (!Load0Offset || !Load1Offset)
239       return false;
240 
241     Offset0 = Load0Offset->getZExtValue();
242     Offset1 = Load1Offset->getZExtValue();
243     return true;
244   }
245 
246   // MUBUF and MTBUF can access the same addresses.
247   if ((isMUBUF(Opc0) || isMTBUF(Opc0)) && (isMUBUF(Opc1) || isMTBUF(Opc1))) {
248 
249     // MUBUF and MTBUF have vaddr at different indices.
250     if (!nodesHaveSameOperandValue(Load0, Load1, AMDGPU::OpName::soffset) ||
251         !nodesHaveSameOperandValue(Load0, Load1, AMDGPU::OpName::vaddr) ||
252         !nodesHaveSameOperandValue(Load0, Load1, AMDGPU::OpName::srsrc))
253       return false;
254 
255     int OffIdx0 = AMDGPU::getNamedOperandIdx(Opc0, AMDGPU::OpName::offset);
256     int OffIdx1 = AMDGPU::getNamedOperandIdx(Opc1, AMDGPU::OpName::offset);
257 
258     if (OffIdx0 == -1 || OffIdx1 == -1)
259       return false;
260 
261     // getNamedOperandIdx returns the index for MachineInstrs.  Since they
262     // include the output in the operand list, but SDNodes don't, we need to
263     // subtract the index by one.
264     OffIdx0 -= get(Opc0).NumDefs;
265     OffIdx1 -= get(Opc1).NumDefs;
266 
267     SDValue Off0 = Load0->getOperand(OffIdx0);
268     SDValue Off1 = Load1->getOperand(OffIdx1);
269 
270     // The offset might be a FrameIndexSDNode.
271     if (!isa<ConstantSDNode>(Off0) || !isa<ConstantSDNode>(Off1))
272       return false;
273 
274     Offset0 = cast<ConstantSDNode>(Off0)->getZExtValue();
275     Offset1 = cast<ConstantSDNode>(Off1)->getZExtValue();
276     return true;
277   }
278 
279   return false;
280 }
281 
282 static bool isStride64(unsigned Opc) {
283   switch (Opc) {
284   case AMDGPU::DS_READ2ST64_B32:
285   case AMDGPU::DS_READ2ST64_B64:
286   case AMDGPU::DS_WRITE2ST64_B32:
287   case AMDGPU::DS_WRITE2ST64_B64:
288     return true;
289   default:
290     return false;
291   }
292 }
293 
294 bool SIInstrInfo::getMemOperandsWithOffsetWidth(
295     const MachineInstr &LdSt, SmallVectorImpl<const MachineOperand *> &BaseOps,
296     int64_t &Offset, bool &OffsetIsScalable, unsigned &Width,
297     const TargetRegisterInfo *TRI) const {
298   if (!LdSt.mayLoadOrStore())
299     return false;
300 
301   unsigned Opc = LdSt.getOpcode();
302   OffsetIsScalable = false;
303   const MachineOperand *BaseOp, *OffsetOp;
304   int DataOpIdx;
305 
306   if (isDS(LdSt)) {
307     BaseOp = getNamedOperand(LdSt, AMDGPU::OpName::addr);
308     OffsetOp = getNamedOperand(LdSt, AMDGPU::OpName::offset);
309     if (OffsetOp) {
310       // Normal, single offset LDS instruction.
311       if (!BaseOp) {
312         // DS_CONSUME/DS_APPEND use M0 for the base address.
313         // TODO: find the implicit use operand for M0 and use that as BaseOp?
314         return false;
315       }
316       BaseOps.push_back(BaseOp);
317       Offset = OffsetOp->getImm();
318       // Get appropriate operand, and compute width accordingly.
319       DataOpIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::vdst);
320       if (DataOpIdx == -1)
321         DataOpIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::data0);
322       Width = getOpSize(LdSt, DataOpIdx);
323     } else {
324       // The 2 offset instructions use offset0 and offset1 instead. We can treat
325       // these as a load with a single offset if the 2 offsets are consecutive.
326       // We will use this for some partially aligned loads.
327       const MachineOperand *Offset0Op =
328           getNamedOperand(LdSt, AMDGPU::OpName::offset0);
329       const MachineOperand *Offset1Op =
330           getNamedOperand(LdSt, AMDGPU::OpName::offset1);
331 
332       unsigned Offset0 = Offset0Op->getImm();
333       unsigned Offset1 = Offset1Op->getImm();
334       if (Offset0 + 1 != Offset1)
335         return false;
336 
337       // Each of these offsets is in element sized units, so we need to convert
338       // to bytes of the individual reads.
339 
340       unsigned EltSize;
341       if (LdSt.mayLoad())
342         EltSize = TRI->getRegSizeInBits(*getOpRegClass(LdSt, 0)) / 16;
343       else {
344         assert(LdSt.mayStore());
345         int Data0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::data0);
346         EltSize = TRI->getRegSizeInBits(*getOpRegClass(LdSt, Data0Idx)) / 8;
347       }
348 
349       if (isStride64(Opc))
350         EltSize *= 64;
351 
352       BaseOps.push_back(BaseOp);
353       Offset = EltSize * Offset0;
354       // Get appropriate operand(s), and compute width accordingly.
355       DataOpIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::vdst);
356       if (DataOpIdx == -1) {
357         DataOpIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::data0);
358         Width = getOpSize(LdSt, DataOpIdx);
359         DataOpIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::data1);
360         Width += getOpSize(LdSt, DataOpIdx);
361       } else {
362         Width = getOpSize(LdSt, DataOpIdx);
363       }
364     }
365     return true;
366   }
367 
368   if (isMUBUF(LdSt) || isMTBUF(LdSt)) {
369     const MachineOperand *RSrc = getNamedOperand(LdSt, AMDGPU::OpName::srsrc);
370     if (!RSrc) // e.g. BUFFER_WBINVL1_VOL
371       return false;
372     BaseOps.push_back(RSrc);
373     BaseOp = getNamedOperand(LdSt, AMDGPU::OpName::vaddr);
374     if (BaseOp && !BaseOp->isFI())
375       BaseOps.push_back(BaseOp);
376     const MachineOperand *OffsetImm =
377         getNamedOperand(LdSt, AMDGPU::OpName::offset);
378     Offset = OffsetImm->getImm();
379     const MachineOperand *SOffset =
380         getNamedOperand(LdSt, AMDGPU::OpName::soffset);
381     if (SOffset) {
382       if (SOffset->isReg())
383         BaseOps.push_back(SOffset);
384       else
385         Offset += SOffset->getImm();
386     }
387     // Get appropriate operand, and compute width accordingly.
388     DataOpIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::vdst);
389     if (DataOpIdx == -1)
390       DataOpIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::vdata);
391     if (DataOpIdx == -1) // LDS DMA
392       return false;
393     Width = getOpSize(LdSt, DataOpIdx);
394     return true;
395   }
396 
397   if (isMIMG(LdSt)) {
398     int SRsrcIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::srsrc);
399     BaseOps.push_back(&LdSt.getOperand(SRsrcIdx));
400     int VAddr0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::vaddr0);
401     if (VAddr0Idx >= 0) {
402       // GFX10 possible NSA encoding.
403       for (int I = VAddr0Idx; I < SRsrcIdx; ++I)
404         BaseOps.push_back(&LdSt.getOperand(I));
405     } else {
406       BaseOps.push_back(getNamedOperand(LdSt, AMDGPU::OpName::vaddr));
407     }
408     Offset = 0;
409     // Get appropriate operand, and compute width accordingly.
410     DataOpIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::vdata);
411     Width = getOpSize(LdSt, DataOpIdx);
412     return true;
413   }
414 
415   if (isSMRD(LdSt)) {
416     BaseOp = getNamedOperand(LdSt, AMDGPU::OpName::sbase);
417     if (!BaseOp) // e.g. S_MEMTIME
418       return false;
419     BaseOps.push_back(BaseOp);
420     OffsetOp = getNamedOperand(LdSt, AMDGPU::OpName::offset);
421     Offset = OffsetOp ? OffsetOp->getImm() : 0;
422     // Get appropriate operand, and compute width accordingly.
423     DataOpIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::sdst);
424     Width = getOpSize(LdSt, DataOpIdx);
425     return true;
426   }
427 
428   if (isFLAT(LdSt)) {
429     // Instructions have either vaddr or saddr or both or none.
430     BaseOp = getNamedOperand(LdSt, AMDGPU::OpName::vaddr);
431     if (BaseOp)
432       BaseOps.push_back(BaseOp);
433     BaseOp = getNamedOperand(LdSt, AMDGPU::OpName::saddr);
434     if (BaseOp)
435       BaseOps.push_back(BaseOp);
436     Offset = getNamedOperand(LdSt, AMDGPU::OpName::offset)->getImm();
437     // Get appropriate operand, and compute width accordingly.
438     DataOpIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::vdst);
439     if (DataOpIdx == -1)
440       DataOpIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::vdata);
441     if (DataOpIdx == -1) // LDS DMA
442       return false;
443     Width = getOpSize(LdSt, DataOpIdx);
444     return true;
445   }
446 
447   return false;
448 }
449 
450 static bool memOpsHaveSameBasePtr(const MachineInstr &MI1,
451                                   ArrayRef<const MachineOperand *> BaseOps1,
452                                   const MachineInstr &MI2,
453                                   ArrayRef<const MachineOperand *> BaseOps2) {
454   // Only examine the first "base" operand of each instruction, on the
455   // assumption that it represents the real base address of the memory access.
456   // Other operands are typically offsets or indices from this base address.
457   if (BaseOps1.front()->isIdenticalTo(*BaseOps2.front()))
458     return true;
459 
460   if (!MI1.hasOneMemOperand() || !MI2.hasOneMemOperand())
461     return false;
462 
463   auto MO1 = *MI1.memoperands_begin();
464   auto MO2 = *MI2.memoperands_begin();
465   if (MO1->getAddrSpace() != MO2->getAddrSpace())
466     return false;
467 
468   auto Base1 = MO1->getValue();
469   auto Base2 = MO2->getValue();
470   if (!Base1 || !Base2)
471     return false;
472   Base1 = getUnderlyingObject(Base1);
473   Base2 = getUnderlyingObject(Base2);
474 
475   if (isa<UndefValue>(Base1) || isa<UndefValue>(Base2))
476     return false;
477 
478   return Base1 == Base2;
479 }
480 
481 bool SIInstrInfo::shouldClusterMemOps(ArrayRef<const MachineOperand *> BaseOps1,
482                                       ArrayRef<const MachineOperand *> BaseOps2,
483                                       unsigned NumLoads,
484                                       unsigned NumBytes) const {
485   // If the mem ops (to be clustered) do not have the same base ptr, then they
486   // should not be clustered
487   if (!BaseOps1.empty() && !BaseOps2.empty()) {
488     const MachineInstr &FirstLdSt = *BaseOps1.front()->getParent();
489     const MachineInstr &SecondLdSt = *BaseOps2.front()->getParent();
490     if (!memOpsHaveSameBasePtr(FirstLdSt, BaseOps1, SecondLdSt, BaseOps2))
491       return false;
492   } else if (!BaseOps1.empty() || !BaseOps2.empty()) {
493     // If only one base op is empty, they do not have the same base ptr
494     return false;
495   }
496 
497   // In order to avoid register pressure, on an average, the number of DWORDS
498   // loaded together by all clustered mem ops should not exceed 8. This is an
499   // empirical value based on certain observations and performance related
500   // experiments.
501   // The good thing about this heuristic is - it avoids clustering of too many
502   // sub-word loads, and also avoids clustering of wide loads. Below is the
503   // brief summary of how the heuristic behaves for various `LoadSize`.
504   // (1) 1 <= LoadSize <= 4: cluster at max 8 mem ops
505   // (2) 5 <= LoadSize <= 8: cluster at max 4 mem ops
506   // (3) 9 <= LoadSize <= 12: cluster at max 2 mem ops
507   // (4) 13 <= LoadSize <= 16: cluster at max 2 mem ops
508   // (5) LoadSize >= 17: do not cluster
509   const unsigned LoadSize = NumBytes / NumLoads;
510   const unsigned NumDWORDs = ((LoadSize + 3) / 4) * NumLoads;
511   return NumDWORDs <= 8;
512 }
513 
514 // FIXME: This behaves strangely. If, for example, you have 32 load + stores,
515 // the first 16 loads will be interleaved with the stores, and the next 16 will
516 // be clustered as expected. It should really split into 2 16 store batches.
517 //
518 // Loads are clustered until this returns false, rather than trying to schedule
519 // groups of stores. This also means we have to deal with saying different
520 // address space loads should be clustered, and ones which might cause bank
521 // conflicts.
522 //
523 // This might be deprecated so it might not be worth that much effort to fix.
524 bool SIInstrInfo::shouldScheduleLoadsNear(SDNode *Load0, SDNode *Load1,
525                                           int64_t Offset0, int64_t Offset1,
526                                           unsigned NumLoads) const {
527   assert(Offset1 > Offset0 &&
528          "Second offset should be larger than first offset!");
529   // If we have less than 16 loads in a row, and the offsets are within 64
530   // bytes, then schedule together.
531 
532   // A cacheline is 64 bytes (for global memory).
533   return (NumLoads <= 16 && (Offset1 - Offset0) < 64);
534 }
535 
536 static void reportIllegalCopy(const SIInstrInfo *TII, MachineBasicBlock &MBB,
537                               MachineBasicBlock::iterator MI,
538                               const DebugLoc &DL, MCRegister DestReg,
539                               MCRegister SrcReg, bool KillSrc,
540                               const char *Msg = "illegal SGPR to VGPR copy") {
541   MachineFunction *MF = MBB.getParent();
542   DiagnosticInfoUnsupported IllegalCopy(MF->getFunction(), Msg, DL, DS_Error);
543   LLVMContext &C = MF->getFunction().getContext();
544   C.diagnose(IllegalCopy);
545 
546   BuildMI(MBB, MI, DL, TII->get(AMDGPU::SI_ILLEGAL_COPY), DestReg)
547     .addReg(SrcReg, getKillRegState(KillSrc));
548 }
549 
550 /// Handle copying from SGPR to AGPR, or from AGPR to AGPR on GFX908. It is not
551 /// possible to have a direct copy in these cases on GFX908, so an intermediate
552 /// VGPR copy is required.
553 static void indirectCopyToAGPR(const SIInstrInfo &TII,
554                                MachineBasicBlock &MBB,
555                                MachineBasicBlock::iterator MI,
556                                const DebugLoc &DL, MCRegister DestReg,
557                                MCRegister SrcReg, bool KillSrc,
558                                RegScavenger &RS, bool RegsOverlap,
559                                Register ImpDefSuperReg = Register(),
560                                Register ImpUseSuperReg = Register()) {
561   assert((TII.getSubtarget().hasMAIInsts() &&
562           !TII.getSubtarget().hasGFX90AInsts()) &&
563          "Expected GFX908 subtarget.");
564 
565   assert((AMDGPU::SReg_32RegClass.contains(SrcReg) ||
566           AMDGPU::AGPR_32RegClass.contains(SrcReg)) &&
567          "Source register of the copy should be either an SGPR or an AGPR.");
568 
569   assert(AMDGPU::AGPR_32RegClass.contains(DestReg) &&
570          "Destination register of the copy should be an AGPR.");
571 
572   const SIRegisterInfo &RI = TII.getRegisterInfo();
573 
574   // First try to find defining accvgpr_write to avoid temporary registers.
575   // In the case of copies of overlapping AGPRs, we conservatively do not
576   // reuse previous accvgpr_writes. Otherwise, we may incorrectly pick up
577   // an accvgpr_write used for this same copy due to implicit-defs
578   if (!RegsOverlap) {
579     for (auto Def = MI, E = MBB.begin(); Def != E; ) {
580       --Def;
581       if (!Def->definesRegister(SrcReg, &RI))
582         continue;
583       if (Def->getOpcode() != AMDGPU::V_ACCVGPR_WRITE_B32_e64)
584         break;
585 
586       MachineOperand &DefOp = Def->getOperand(1);
587       assert(DefOp.isReg() || DefOp.isImm());
588 
589       if (DefOp.isReg()) {
590         bool SafeToPropagate = true;
591         // Check that register source operand is not clobbered before MI.
592         // Immediate operands are always safe to propagate.
593         for (auto I = Def; I != MI && SafeToPropagate; ++I)
594           if (I->modifiesRegister(DefOp.getReg(), &RI))
595             SafeToPropagate = false;
596 
597         if (!SafeToPropagate)
598           break;
599 
600         DefOp.setIsKill(false);
601       }
602 
603       MachineInstrBuilder Builder =
604         BuildMI(MBB, MI, DL, TII.get(AMDGPU::V_ACCVGPR_WRITE_B32_e64), DestReg)
605         .add(DefOp);
606       if (ImpDefSuperReg)
607         Builder.addReg(ImpDefSuperReg, RegState::Define | RegState::Implicit);
608 
609       if (ImpUseSuperReg) {
610         Builder.addReg(ImpUseSuperReg,
611                       getKillRegState(KillSrc) | RegState::Implicit);
612       }
613 
614       return;
615     }
616   }
617 
618   RS.enterBasicBlock(MBB);
619   RS.forward(MI);
620 
621   // Ideally we want to have three registers for a long reg_sequence copy
622   // to hide 2 waitstates between v_mov_b32 and accvgpr_write.
623   unsigned MaxVGPRs = RI.getRegPressureLimit(&AMDGPU::VGPR_32RegClass,
624                                              *MBB.getParent());
625 
626   // Registers in the sequence are allocated contiguously so we can just
627   // use register number to pick one of three round-robin temps.
628   unsigned RegNo = (DestReg - AMDGPU::AGPR0) % 3;
629   Register Tmp =
630       MBB.getParent()->getInfo<SIMachineFunctionInfo>()->getVGPRForAGPRCopy();
631   assert(MBB.getParent()->getRegInfo().isReserved(Tmp) &&
632          "VGPR used for an intermediate copy should have been reserved.");
633 
634   // Only loop through if there are any free registers left, otherwise
635   // scavenger may report a fatal error without emergency spill slot
636   // or spill with the slot.
637   while (RegNo-- && RS.FindUnusedReg(&AMDGPU::VGPR_32RegClass)) {
638     Register Tmp2 = RS.scavengeRegister(&AMDGPU::VGPR_32RegClass, 0);
639     if (!Tmp2 || RI.getHWRegIndex(Tmp2) >= MaxVGPRs)
640       break;
641     Tmp = Tmp2;
642     RS.setRegUsed(Tmp);
643   }
644 
645   // Insert copy to temporary VGPR.
646   unsigned TmpCopyOp = AMDGPU::V_MOV_B32_e32;
647   if (AMDGPU::AGPR_32RegClass.contains(SrcReg)) {
648     TmpCopyOp = AMDGPU::V_ACCVGPR_READ_B32_e64;
649   } else {
650     assert(AMDGPU::SReg_32RegClass.contains(SrcReg));
651   }
652 
653   MachineInstrBuilder UseBuilder = BuildMI(MBB, MI, DL, TII.get(TmpCopyOp), Tmp)
654     .addReg(SrcReg, getKillRegState(KillSrc));
655   if (ImpUseSuperReg) {
656     UseBuilder.addReg(ImpUseSuperReg,
657                       getKillRegState(KillSrc) | RegState::Implicit);
658   }
659 
660   MachineInstrBuilder DefBuilder
661     = BuildMI(MBB, MI, DL, TII.get(AMDGPU::V_ACCVGPR_WRITE_B32_e64), DestReg)
662     .addReg(Tmp, RegState::Kill);
663 
664   if (ImpDefSuperReg)
665     DefBuilder.addReg(ImpDefSuperReg, RegState::Define | RegState::Implicit);
666 }
667 
668 static void expandSGPRCopy(const SIInstrInfo &TII, MachineBasicBlock &MBB,
669                            MachineBasicBlock::iterator MI, const DebugLoc &DL,
670                            MCRegister DestReg, MCRegister SrcReg, bool KillSrc,
671                            const TargetRegisterClass *RC, bool Forward) {
672   const SIRegisterInfo &RI = TII.getRegisterInfo();
673   ArrayRef<int16_t> BaseIndices = RI.getRegSplitParts(RC, 4);
674   MachineBasicBlock::iterator I = MI;
675   MachineInstr *FirstMI = nullptr, *LastMI = nullptr;
676 
677   for (unsigned Idx = 0; Idx < BaseIndices.size(); ++Idx) {
678     int16_t SubIdx = BaseIndices[Idx];
679     Register Reg = RI.getSubReg(DestReg, SubIdx);
680     unsigned Opcode = AMDGPU::S_MOV_B32;
681 
682     // Is SGPR aligned? If so try to combine with next.
683     Register Src = RI.getSubReg(SrcReg, SubIdx);
684     bool AlignedDest = ((Reg - AMDGPU::SGPR0) % 2) == 0;
685     bool AlignedSrc = ((Src - AMDGPU::SGPR0) % 2) == 0;
686     if (AlignedDest && AlignedSrc && (Idx + 1 < BaseIndices.size())) {
687       // Can use SGPR64 copy
688       unsigned Channel = RI.getChannelFromSubReg(SubIdx);
689       SubIdx = RI.getSubRegFromChannel(Channel, 2);
690       Opcode = AMDGPU::S_MOV_B64;
691       Idx++;
692     }
693 
694     LastMI = BuildMI(MBB, I, DL, TII.get(Opcode), RI.getSubReg(DestReg, SubIdx))
695                  .addReg(RI.getSubReg(SrcReg, SubIdx))
696                  .addReg(SrcReg, RegState::Implicit);
697 
698     if (!FirstMI)
699       FirstMI = LastMI;
700 
701     if (!Forward)
702       I--;
703   }
704 
705   assert(FirstMI && LastMI);
706   if (!Forward)
707     std::swap(FirstMI, LastMI);
708 
709   FirstMI->addOperand(
710       MachineOperand::CreateReg(DestReg, true /*IsDef*/, true /*IsImp*/));
711 
712   if (KillSrc)
713     LastMI->addRegisterKilled(SrcReg, &RI);
714 }
715 
716 void SIInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
717                               MachineBasicBlock::iterator MI,
718                               const DebugLoc &DL, MCRegister DestReg,
719                               MCRegister SrcReg, bool KillSrc) const {
720   const TargetRegisterClass *RC = RI.getPhysRegBaseClass(DestReg);
721 
722   // FIXME: This is hack to resolve copies between 16 bit and 32 bit
723   // registers until all patterns are fixed.
724   if (Fix16BitCopies &&
725       ((RI.getRegSizeInBits(*RC) == 16) ^
726        (RI.getRegSizeInBits(*RI.getPhysRegBaseClass(SrcReg)) == 16))) {
727     MCRegister &RegToFix = (RI.getRegSizeInBits(*RC) == 16) ? DestReg : SrcReg;
728     MCRegister Super = RI.get32BitRegister(RegToFix);
729     assert(RI.getSubReg(Super, AMDGPU::lo16) == RegToFix);
730     RegToFix = Super;
731 
732     if (DestReg == SrcReg) {
733       // Insert empty bundle since ExpandPostRA expects an instruction here.
734       BuildMI(MBB, MI, DL, get(AMDGPU::BUNDLE));
735       return;
736     }
737 
738     RC = RI.getPhysRegBaseClass(DestReg);
739   }
740 
741   if (RC == &AMDGPU::VGPR_32RegClass) {
742     assert(AMDGPU::VGPR_32RegClass.contains(SrcReg) ||
743            AMDGPU::SReg_32RegClass.contains(SrcReg) ||
744            AMDGPU::AGPR_32RegClass.contains(SrcReg));
745     unsigned Opc = AMDGPU::AGPR_32RegClass.contains(SrcReg) ?
746                      AMDGPU::V_ACCVGPR_READ_B32_e64 : AMDGPU::V_MOV_B32_e32;
747     BuildMI(MBB, MI, DL, get(Opc), DestReg)
748       .addReg(SrcReg, getKillRegState(KillSrc));
749     return;
750   }
751 
752   if (RC == &AMDGPU::SReg_32_XM0RegClass ||
753       RC == &AMDGPU::SReg_32RegClass) {
754     if (SrcReg == AMDGPU::SCC) {
755       BuildMI(MBB, MI, DL, get(AMDGPU::S_CSELECT_B32), DestReg)
756           .addImm(1)
757           .addImm(0);
758       return;
759     }
760 
761     if (DestReg == AMDGPU::VCC_LO) {
762       if (AMDGPU::SReg_32RegClass.contains(SrcReg)) {
763         BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B32), AMDGPU::VCC_LO)
764           .addReg(SrcReg, getKillRegState(KillSrc));
765       } else {
766         // FIXME: Hack until VReg_1 removed.
767         assert(AMDGPU::VGPR_32RegClass.contains(SrcReg));
768         BuildMI(MBB, MI, DL, get(AMDGPU::V_CMP_NE_U32_e32))
769           .addImm(0)
770           .addReg(SrcReg, getKillRegState(KillSrc));
771       }
772 
773       return;
774     }
775 
776     if (!AMDGPU::SReg_32RegClass.contains(SrcReg)) {
777       reportIllegalCopy(this, MBB, MI, DL, DestReg, SrcReg, KillSrc);
778       return;
779     }
780 
781     BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B32), DestReg)
782             .addReg(SrcReg, getKillRegState(KillSrc));
783     return;
784   }
785 
786   if (RC == &AMDGPU::SReg_64RegClass) {
787     if (SrcReg == AMDGPU::SCC) {
788       BuildMI(MBB, MI, DL, get(AMDGPU::S_CSELECT_B64), DestReg)
789           .addImm(1)
790           .addImm(0);
791       return;
792     }
793 
794     if (DestReg == AMDGPU::VCC) {
795       if (AMDGPU::SReg_64RegClass.contains(SrcReg)) {
796         BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B64), AMDGPU::VCC)
797           .addReg(SrcReg, getKillRegState(KillSrc));
798       } else {
799         // FIXME: Hack until VReg_1 removed.
800         assert(AMDGPU::VGPR_32RegClass.contains(SrcReg));
801         BuildMI(MBB, MI, DL, get(AMDGPU::V_CMP_NE_U32_e32))
802           .addImm(0)
803           .addReg(SrcReg, getKillRegState(KillSrc));
804       }
805 
806       return;
807     }
808 
809     if (!AMDGPU::SReg_64RegClass.contains(SrcReg)) {
810       reportIllegalCopy(this, MBB, MI, DL, DestReg, SrcReg, KillSrc);
811       return;
812     }
813 
814     BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B64), DestReg)
815             .addReg(SrcReg, getKillRegState(KillSrc));
816     return;
817   }
818 
819   if (DestReg == AMDGPU::SCC) {
820     // Copying 64-bit or 32-bit sources to SCC barely makes sense,
821     // but SelectionDAG emits such copies for i1 sources.
822     if (AMDGPU::SReg_64RegClass.contains(SrcReg)) {
823       // This copy can only be produced by patterns
824       // with explicit SCC, which are known to be enabled
825       // only for subtargets with S_CMP_LG_U64 present.
826       assert(ST.hasScalarCompareEq64());
827       BuildMI(MBB, MI, DL, get(AMDGPU::S_CMP_LG_U64))
828           .addReg(SrcReg, getKillRegState(KillSrc))
829           .addImm(0);
830     } else {
831       assert(AMDGPU::SReg_32RegClass.contains(SrcReg));
832       BuildMI(MBB, MI, DL, get(AMDGPU::S_CMP_LG_U32))
833           .addReg(SrcReg, getKillRegState(KillSrc))
834           .addImm(0);
835     }
836 
837     return;
838   }
839 
840   if (RC == &AMDGPU::AGPR_32RegClass) {
841     if (AMDGPU::VGPR_32RegClass.contains(SrcReg) ||
842         (ST.hasGFX90AInsts() && AMDGPU::SReg_32RegClass.contains(SrcReg))) {
843       BuildMI(MBB, MI, DL, get(AMDGPU::V_ACCVGPR_WRITE_B32_e64), DestReg)
844         .addReg(SrcReg, getKillRegState(KillSrc));
845       return;
846     }
847 
848     if (AMDGPU::AGPR_32RegClass.contains(SrcReg) && ST.hasGFX90AInsts()) {
849       BuildMI(MBB, MI, DL, get(AMDGPU::V_ACCVGPR_MOV_B32), DestReg)
850         .addReg(SrcReg, getKillRegState(KillSrc));
851       return;
852     }
853 
854     // FIXME: Pass should maintain scavenger to avoid scan through the block on
855     // every AGPR spill.
856     RegScavenger RS;
857     const bool Overlap = RI.regsOverlap(SrcReg, DestReg);
858     indirectCopyToAGPR(*this, MBB, MI, DL, DestReg, SrcReg, KillSrc, RS, Overlap);
859     return;
860   }
861 
862   const unsigned Size = RI.getRegSizeInBits(*RC);
863   if (Size == 16) {
864     assert(AMDGPU::VGPR_LO16RegClass.contains(SrcReg) ||
865            AMDGPU::VGPR_HI16RegClass.contains(SrcReg) ||
866            AMDGPU::SReg_LO16RegClass.contains(SrcReg) ||
867            AMDGPU::AGPR_LO16RegClass.contains(SrcReg));
868 
869     bool IsSGPRDst = AMDGPU::SReg_LO16RegClass.contains(DestReg);
870     bool IsSGPRSrc = AMDGPU::SReg_LO16RegClass.contains(SrcReg);
871     bool IsAGPRDst = AMDGPU::AGPR_LO16RegClass.contains(DestReg);
872     bool IsAGPRSrc = AMDGPU::AGPR_LO16RegClass.contains(SrcReg);
873     bool DstLow = AMDGPU::VGPR_LO16RegClass.contains(DestReg) ||
874                   AMDGPU::SReg_LO16RegClass.contains(DestReg) ||
875                   AMDGPU::AGPR_LO16RegClass.contains(DestReg);
876     bool SrcLow = AMDGPU::VGPR_LO16RegClass.contains(SrcReg) ||
877                   AMDGPU::SReg_LO16RegClass.contains(SrcReg) ||
878                   AMDGPU::AGPR_LO16RegClass.contains(SrcReg);
879     MCRegister NewDestReg = RI.get32BitRegister(DestReg);
880     MCRegister NewSrcReg = RI.get32BitRegister(SrcReg);
881 
882     if (IsSGPRDst) {
883       if (!IsSGPRSrc) {
884         reportIllegalCopy(this, MBB, MI, DL, DestReg, SrcReg, KillSrc);
885         return;
886       }
887 
888       BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B32), NewDestReg)
889         .addReg(NewSrcReg, getKillRegState(KillSrc));
890       return;
891     }
892 
893     if (IsAGPRDst || IsAGPRSrc) {
894       if (!DstLow || !SrcLow) {
895         reportIllegalCopy(this, MBB, MI, DL, DestReg, SrcReg, KillSrc,
896                           "Cannot use hi16 subreg with an AGPR!");
897       }
898 
899       copyPhysReg(MBB, MI, DL, NewDestReg, NewSrcReg, KillSrc);
900       return;
901     }
902 
903     if (IsSGPRSrc && !ST.hasSDWAScalar()) {
904       if (!DstLow || !SrcLow) {
905         reportIllegalCopy(this, MBB, MI, DL, DestReg, SrcReg, KillSrc,
906                           "Cannot use hi16 subreg on VI!");
907       }
908 
909       BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), NewDestReg)
910         .addReg(NewSrcReg, getKillRegState(KillSrc));
911       return;
912     }
913 
914     auto MIB = BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_sdwa), NewDestReg)
915       .addImm(0) // src0_modifiers
916       .addReg(NewSrcReg)
917       .addImm(0) // clamp
918       .addImm(DstLow ? AMDGPU::SDWA::SdwaSel::WORD_0
919                      : AMDGPU::SDWA::SdwaSel::WORD_1)
920       .addImm(AMDGPU::SDWA::DstUnused::UNUSED_PRESERVE)
921       .addImm(SrcLow ? AMDGPU::SDWA::SdwaSel::WORD_0
922                      : AMDGPU::SDWA::SdwaSel::WORD_1)
923       .addReg(NewDestReg, RegState::Implicit | RegState::Undef);
924     // First implicit operand is $exec.
925     MIB->tieOperands(0, MIB->getNumOperands() - 1);
926     return;
927   }
928 
929   const TargetRegisterClass *SrcRC = RI.getPhysRegBaseClass(SrcReg);
930   if (RC == RI.getVGPR64Class() && (SrcRC == RC || RI.isSGPRClass(SrcRC))) {
931     if (ST.hasMovB64()) {
932       BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B64_e32), DestReg)
933         .addReg(SrcReg, getKillRegState(KillSrc));
934       return;
935     }
936     if (ST.hasPackedFP32Ops()) {
937       BuildMI(MBB, MI, DL, get(AMDGPU::V_PK_MOV_B32), DestReg)
938         .addImm(SISrcMods::OP_SEL_1)
939         .addReg(SrcReg)
940         .addImm(SISrcMods::OP_SEL_0 | SISrcMods::OP_SEL_1)
941         .addReg(SrcReg)
942         .addImm(0) // op_sel_lo
943         .addImm(0) // op_sel_hi
944         .addImm(0) // neg_lo
945         .addImm(0) // neg_hi
946         .addImm(0) // clamp
947         .addReg(SrcReg, getKillRegState(KillSrc) | RegState::Implicit);
948       return;
949     }
950   }
951 
952   const bool Forward = RI.getHWRegIndex(DestReg) <= RI.getHWRegIndex(SrcReg);
953   if (RI.isSGPRClass(RC)) {
954     if (!RI.isSGPRClass(SrcRC)) {
955       reportIllegalCopy(this, MBB, MI, DL, DestReg, SrcReg, KillSrc);
956       return;
957     }
958     const bool CanKillSuperReg = KillSrc && !RI.regsOverlap(SrcReg, DestReg);
959     expandSGPRCopy(*this, MBB, MI, DL, DestReg, SrcReg, CanKillSuperReg, RC,
960                    Forward);
961     return;
962   }
963 
964   unsigned EltSize = 4;
965   unsigned Opcode = AMDGPU::V_MOV_B32_e32;
966   if (RI.isAGPRClass(RC)) {
967     if (ST.hasGFX90AInsts() && RI.isAGPRClass(SrcRC))
968       Opcode = AMDGPU::V_ACCVGPR_MOV_B32;
969     else if (RI.hasVGPRs(SrcRC) ||
970              (ST.hasGFX90AInsts() && RI.isSGPRClass(SrcRC)))
971       Opcode = AMDGPU::V_ACCVGPR_WRITE_B32_e64;
972     else
973       Opcode = AMDGPU::INSTRUCTION_LIST_END;
974   } else if (RI.hasVGPRs(RC) && RI.isAGPRClass(SrcRC)) {
975     Opcode = AMDGPU::V_ACCVGPR_READ_B32_e64;
976   } else if ((Size % 64 == 0) && RI.hasVGPRs(RC) &&
977              (RI.isProperlyAlignedRC(*RC) &&
978               (SrcRC == RC || RI.isSGPRClass(SrcRC)))) {
979     // TODO: In 96-bit case, could do a 64-bit mov and then a 32-bit mov.
980     if (ST.hasMovB64()) {
981       Opcode = AMDGPU::V_MOV_B64_e32;
982       EltSize = 8;
983     } else if (ST.hasPackedFP32Ops()) {
984       Opcode = AMDGPU::V_PK_MOV_B32;
985       EltSize = 8;
986     }
987   }
988 
989   // For the cases where we need an intermediate instruction/temporary register
990   // (destination is an AGPR), we need a scavenger.
991   //
992   // FIXME: The pass should maintain this for us so we don't have to re-scan the
993   // whole block for every handled copy.
994   std::unique_ptr<RegScavenger> RS;
995   if (Opcode == AMDGPU::INSTRUCTION_LIST_END)
996     RS.reset(new RegScavenger());
997 
998   ArrayRef<int16_t> SubIndices = RI.getRegSplitParts(RC, EltSize);
999 
1000   // If there is an overlap, we can't kill the super-register on the last
1001   // instruction, since it will also kill the components made live by this def.
1002   const bool Overlap = RI.regsOverlap(SrcReg, DestReg);
1003   const bool CanKillSuperReg = KillSrc && !Overlap;
1004 
1005   for (unsigned Idx = 0; Idx < SubIndices.size(); ++Idx) {
1006     unsigned SubIdx;
1007     if (Forward)
1008       SubIdx = SubIndices[Idx];
1009     else
1010       SubIdx = SubIndices[SubIndices.size() - Idx - 1];
1011 
1012     bool IsFirstSubreg = Idx == 0;
1013     bool UseKill = CanKillSuperReg && Idx == SubIndices.size() - 1;
1014 
1015     if (Opcode == AMDGPU::INSTRUCTION_LIST_END) {
1016       Register ImpDefSuper = IsFirstSubreg ? Register(DestReg) : Register();
1017       Register ImpUseSuper = SrcReg;
1018       indirectCopyToAGPR(*this, MBB, MI, DL, RI.getSubReg(DestReg, SubIdx),
1019                          RI.getSubReg(SrcReg, SubIdx), UseKill, *RS, Overlap,
1020                          ImpDefSuper, ImpUseSuper);
1021     } else if (Opcode == AMDGPU::V_PK_MOV_B32) {
1022       Register DstSubReg = RI.getSubReg(DestReg, SubIdx);
1023       Register SrcSubReg = RI.getSubReg(SrcReg, SubIdx);
1024       MachineInstrBuilder MIB =
1025         BuildMI(MBB, MI, DL, get(AMDGPU::V_PK_MOV_B32), DstSubReg)
1026         .addImm(SISrcMods::OP_SEL_1)
1027         .addReg(SrcSubReg)
1028         .addImm(SISrcMods::OP_SEL_0 | SISrcMods::OP_SEL_1)
1029         .addReg(SrcSubReg)
1030         .addImm(0) // op_sel_lo
1031         .addImm(0) // op_sel_hi
1032         .addImm(0) // neg_lo
1033         .addImm(0) // neg_hi
1034         .addImm(0) // clamp
1035         .addReg(SrcReg, getKillRegState(UseKill) | RegState::Implicit);
1036       if (IsFirstSubreg)
1037         MIB.addReg(DestReg, RegState::Define | RegState::Implicit);
1038     } else {
1039       MachineInstrBuilder Builder =
1040         BuildMI(MBB, MI, DL, get(Opcode), RI.getSubReg(DestReg, SubIdx))
1041         .addReg(RI.getSubReg(SrcReg, SubIdx));
1042       if (IsFirstSubreg)
1043         Builder.addReg(DestReg, RegState::Define | RegState::Implicit);
1044 
1045       Builder.addReg(SrcReg, getKillRegState(UseKill) | RegState::Implicit);
1046     }
1047   }
1048 }
1049 
1050 int SIInstrInfo::commuteOpcode(unsigned Opcode) const {
1051   int NewOpc;
1052 
1053   // Try to map original to commuted opcode
1054   NewOpc = AMDGPU::getCommuteRev(Opcode);
1055   if (NewOpc != -1)
1056     // Check if the commuted (REV) opcode exists on the target.
1057     return pseudoToMCOpcode(NewOpc) != -1 ? NewOpc : -1;
1058 
1059   // Try to map commuted to original opcode
1060   NewOpc = AMDGPU::getCommuteOrig(Opcode);
1061   if (NewOpc != -1)
1062     // Check if the original (non-REV) opcode exists on the target.
1063     return pseudoToMCOpcode(NewOpc) != -1 ? NewOpc : -1;
1064 
1065   return Opcode;
1066 }
1067 
1068 void SIInstrInfo::materializeImmediate(MachineBasicBlock &MBB,
1069                                        MachineBasicBlock::iterator MI,
1070                                        const DebugLoc &DL, Register DestReg,
1071                                        int64_t Value) const {
1072   MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
1073   const TargetRegisterClass *RegClass = MRI.getRegClass(DestReg);
1074   if (RegClass == &AMDGPU::SReg_32RegClass ||
1075       RegClass == &AMDGPU::SGPR_32RegClass ||
1076       RegClass == &AMDGPU::SReg_32_XM0RegClass ||
1077       RegClass == &AMDGPU::SReg_32_XM0_XEXECRegClass) {
1078     BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B32), DestReg)
1079       .addImm(Value);
1080     return;
1081   }
1082 
1083   if (RegClass == &AMDGPU::SReg_64RegClass ||
1084       RegClass == &AMDGPU::SGPR_64RegClass ||
1085       RegClass == &AMDGPU::SReg_64_XEXECRegClass) {
1086     BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B64), DestReg)
1087       .addImm(Value);
1088     return;
1089   }
1090 
1091   if (RegClass == &AMDGPU::VGPR_32RegClass) {
1092     BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DestReg)
1093       .addImm(Value);
1094     return;
1095   }
1096   if (RegClass->hasSuperClassEq(&AMDGPU::VReg_64RegClass)) {
1097     BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B64_PSEUDO), DestReg)
1098       .addImm(Value);
1099     return;
1100   }
1101 
1102   unsigned EltSize = 4;
1103   unsigned Opcode = AMDGPU::V_MOV_B32_e32;
1104   if (RI.isSGPRClass(RegClass)) {
1105     if (RI.getRegSizeInBits(*RegClass) > 32) {
1106       Opcode =  AMDGPU::S_MOV_B64;
1107       EltSize = 8;
1108     } else {
1109       Opcode = AMDGPU::S_MOV_B32;
1110       EltSize = 4;
1111     }
1112   }
1113 
1114   ArrayRef<int16_t> SubIndices = RI.getRegSplitParts(RegClass, EltSize);
1115   for (unsigned Idx = 0; Idx < SubIndices.size(); ++Idx) {
1116     int64_t IdxValue = Idx == 0 ? Value : 0;
1117 
1118     MachineInstrBuilder Builder = BuildMI(MBB, MI, DL,
1119       get(Opcode), RI.getSubReg(DestReg, SubIndices[Idx]));
1120     Builder.addImm(IdxValue);
1121   }
1122 }
1123 
1124 const TargetRegisterClass *
1125 SIInstrInfo::getPreferredSelectRegClass(unsigned Size) const {
1126   return &AMDGPU::VGPR_32RegClass;
1127 }
1128 
1129 void SIInstrInfo::insertVectorSelect(MachineBasicBlock &MBB,
1130                                      MachineBasicBlock::iterator I,
1131                                      const DebugLoc &DL, Register DstReg,
1132                                      ArrayRef<MachineOperand> Cond,
1133                                      Register TrueReg,
1134                                      Register FalseReg) const {
1135   MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
1136   const TargetRegisterClass *BoolXExecRC =
1137     RI.getRegClass(AMDGPU::SReg_1_XEXECRegClassID);
1138   assert(MRI.getRegClass(DstReg) == &AMDGPU::VGPR_32RegClass &&
1139          "Not a VGPR32 reg");
1140 
1141   if (Cond.size() == 1) {
1142     Register SReg = MRI.createVirtualRegister(BoolXExecRC);
1143     BuildMI(MBB, I, DL, get(AMDGPU::COPY), SReg)
1144       .add(Cond[0]);
1145     BuildMI(MBB, I, DL, get(AMDGPU::V_CNDMASK_B32_e64), DstReg)
1146       .addImm(0)
1147       .addReg(FalseReg)
1148       .addImm(0)
1149       .addReg(TrueReg)
1150       .addReg(SReg);
1151   } else if (Cond.size() == 2) {
1152     assert(Cond[0].isImm() && "Cond[0] is not an immediate");
1153     switch (Cond[0].getImm()) {
1154     case SIInstrInfo::SCC_TRUE: {
1155       Register SReg = MRI.createVirtualRegister(BoolXExecRC);
1156       BuildMI(MBB, I, DL, get(ST.isWave32() ? AMDGPU::S_CSELECT_B32
1157                                             : AMDGPU::S_CSELECT_B64), SReg)
1158         .addImm(1)
1159         .addImm(0);
1160       BuildMI(MBB, I, DL, get(AMDGPU::V_CNDMASK_B32_e64), DstReg)
1161         .addImm(0)
1162         .addReg(FalseReg)
1163         .addImm(0)
1164         .addReg(TrueReg)
1165         .addReg(SReg);
1166       break;
1167     }
1168     case SIInstrInfo::SCC_FALSE: {
1169       Register SReg = MRI.createVirtualRegister(BoolXExecRC);
1170       BuildMI(MBB, I, DL, get(ST.isWave32() ? AMDGPU::S_CSELECT_B32
1171                                             : AMDGPU::S_CSELECT_B64), SReg)
1172         .addImm(0)
1173         .addImm(1);
1174       BuildMI(MBB, I, DL, get(AMDGPU::V_CNDMASK_B32_e64), DstReg)
1175         .addImm(0)
1176         .addReg(FalseReg)
1177         .addImm(0)
1178         .addReg(TrueReg)
1179         .addReg(SReg);
1180       break;
1181     }
1182     case SIInstrInfo::VCCNZ: {
1183       MachineOperand RegOp = Cond[1];
1184       RegOp.setImplicit(false);
1185       Register SReg = MRI.createVirtualRegister(BoolXExecRC);
1186       BuildMI(MBB, I, DL, get(AMDGPU::COPY), SReg)
1187         .add(RegOp);
1188       BuildMI(MBB, I, DL, get(AMDGPU::V_CNDMASK_B32_e64), DstReg)
1189           .addImm(0)
1190           .addReg(FalseReg)
1191           .addImm(0)
1192           .addReg(TrueReg)
1193           .addReg(SReg);
1194       break;
1195     }
1196     case SIInstrInfo::VCCZ: {
1197       MachineOperand RegOp = Cond[1];
1198       RegOp.setImplicit(false);
1199       Register SReg = MRI.createVirtualRegister(BoolXExecRC);
1200       BuildMI(MBB, I, DL, get(AMDGPU::COPY), SReg)
1201         .add(RegOp);
1202       BuildMI(MBB, I, DL, get(AMDGPU::V_CNDMASK_B32_e64), DstReg)
1203           .addImm(0)
1204           .addReg(TrueReg)
1205           .addImm(0)
1206           .addReg(FalseReg)
1207           .addReg(SReg);
1208       break;
1209     }
1210     case SIInstrInfo::EXECNZ: {
1211       Register SReg = MRI.createVirtualRegister(BoolXExecRC);
1212       Register SReg2 = MRI.createVirtualRegister(RI.getBoolRC());
1213       BuildMI(MBB, I, DL, get(ST.isWave32() ? AMDGPU::S_OR_SAVEEXEC_B32
1214                                             : AMDGPU::S_OR_SAVEEXEC_B64), SReg2)
1215         .addImm(0);
1216       BuildMI(MBB, I, DL, get(ST.isWave32() ? AMDGPU::S_CSELECT_B32
1217                                             : AMDGPU::S_CSELECT_B64), SReg)
1218         .addImm(1)
1219         .addImm(0);
1220       BuildMI(MBB, I, DL, get(AMDGPU::V_CNDMASK_B32_e64), DstReg)
1221         .addImm(0)
1222         .addReg(FalseReg)
1223         .addImm(0)
1224         .addReg(TrueReg)
1225         .addReg(SReg);
1226       break;
1227     }
1228     case SIInstrInfo::EXECZ: {
1229       Register SReg = MRI.createVirtualRegister(BoolXExecRC);
1230       Register SReg2 = MRI.createVirtualRegister(RI.getBoolRC());
1231       BuildMI(MBB, I, DL, get(ST.isWave32() ? AMDGPU::S_OR_SAVEEXEC_B32
1232                                             : AMDGPU::S_OR_SAVEEXEC_B64), SReg2)
1233         .addImm(0);
1234       BuildMI(MBB, I, DL, get(ST.isWave32() ? AMDGPU::S_CSELECT_B32
1235                                             : AMDGPU::S_CSELECT_B64), SReg)
1236         .addImm(0)
1237         .addImm(1);
1238       BuildMI(MBB, I, DL, get(AMDGPU::V_CNDMASK_B32_e64), DstReg)
1239         .addImm(0)
1240         .addReg(FalseReg)
1241         .addImm(0)
1242         .addReg(TrueReg)
1243         .addReg(SReg);
1244       llvm_unreachable("Unhandled branch predicate EXECZ");
1245       break;
1246     }
1247     default:
1248       llvm_unreachable("invalid branch predicate");
1249     }
1250   } else {
1251     llvm_unreachable("Can only handle Cond size 1 or 2");
1252   }
1253 }
1254 
1255 Register SIInstrInfo::insertEQ(MachineBasicBlock *MBB,
1256                                MachineBasicBlock::iterator I,
1257                                const DebugLoc &DL,
1258                                Register SrcReg, int Value) const {
1259   MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
1260   Register Reg = MRI.createVirtualRegister(RI.getBoolRC());
1261   BuildMI(*MBB, I, DL, get(AMDGPU::V_CMP_EQ_I32_e64), Reg)
1262     .addImm(Value)
1263     .addReg(SrcReg);
1264 
1265   return Reg;
1266 }
1267 
1268 Register SIInstrInfo::insertNE(MachineBasicBlock *MBB,
1269                                MachineBasicBlock::iterator I,
1270                                const DebugLoc &DL,
1271                                Register SrcReg, int Value) const {
1272   MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
1273   Register Reg = MRI.createVirtualRegister(RI.getBoolRC());
1274   BuildMI(*MBB, I, DL, get(AMDGPU::V_CMP_NE_I32_e64), Reg)
1275     .addImm(Value)
1276     .addReg(SrcReg);
1277 
1278   return Reg;
1279 }
1280 
1281 unsigned SIInstrInfo::getMovOpcode(const TargetRegisterClass *DstRC) const {
1282 
1283   if (RI.isAGPRClass(DstRC))
1284     return AMDGPU::COPY;
1285   if (RI.getRegSizeInBits(*DstRC) == 32) {
1286     return RI.isSGPRClass(DstRC) ? AMDGPU::S_MOV_B32 : AMDGPU::V_MOV_B32_e32;
1287   } else if (RI.getRegSizeInBits(*DstRC) == 64 && RI.isSGPRClass(DstRC)) {
1288     return AMDGPU::S_MOV_B64;
1289   } else if (RI.getRegSizeInBits(*DstRC) == 64 && !RI.isSGPRClass(DstRC)) {
1290     return  AMDGPU::V_MOV_B64_PSEUDO;
1291   }
1292   return AMDGPU::COPY;
1293 }
1294 
1295 const MCInstrDesc &
1296 SIInstrInfo::getIndirectGPRIDXPseudo(unsigned VecSize,
1297                                      bool IsIndirectSrc) const {
1298   if (IsIndirectSrc) {
1299     if (VecSize <= 32) // 4 bytes
1300       return get(AMDGPU::V_INDIRECT_REG_READ_GPR_IDX_B32_V1);
1301     if (VecSize <= 64) // 8 bytes
1302       return get(AMDGPU::V_INDIRECT_REG_READ_GPR_IDX_B32_V2);
1303     if (VecSize <= 96) // 12 bytes
1304       return get(AMDGPU::V_INDIRECT_REG_READ_GPR_IDX_B32_V3);
1305     if (VecSize <= 128) // 16 bytes
1306       return get(AMDGPU::V_INDIRECT_REG_READ_GPR_IDX_B32_V4);
1307     if (VecSize <= 160) // 20 bytes
1308       return get(AMDGPU::V_INDIRECT_REG_READ_GPR_IDX_B32_V5);
1309     if (VecSize <= 256) // 32 bytes
1310       return get(AMDGPU::V_INDIRECT_REG_READ_GPR_IDX_B32_V8);
1311     if (VecSize <= 288) // 36 bytes
1312       return get(AMDGPU::V_INDIRECT_REG_READ_GPR_IDX_B32_V9);
1313     if (VecSize <= 320) // 40 bytes
1314       return get(AMDGPU::V_INDIRECT_REG_READ_GPR_IDX_B32_V10);
1315     if (VecSize <= 352) // 44 bytes
1316       return get(AMDGPU::V_INDIRECT_REG_READ_GPR_IDX_B32_V11);
1317     if (VecSize <= 384) // 48 bytes
1318       return get(AMDGPU::V_INDIRECT_REG_READ_GPR_IDX_B32_V12);
1319     if (VecSize <= 512) // 64 bytes
1320       return get(AMDGPU::V_INDIRECT_REG_READ_GPR_IDX_B32_V16);
1321     if (VecSize <= 1024) // 128 bytes
1322       return get(AMDGPU::V_INDIRECT_REG_READ_GPR_IDX_B32_V32);
1323 
1324     llvm_unreachable("unsupported size for IndirectRegReadGPRIDX pseudos");
1325   }
1326 
1327   if (VecSize <= 32) // 4 bytes
1328     return get(AMDGPU::V_INDIRECT_REG_WRITE_GPR_IDX_B32_V1);
1329   if (VecSize <= 64) // 8 bytes
1330     return get(AMDGPU::V_INDIRECT_REG_WRITE_GPR_IDX_B32_V2);
1331   if (VecSize <= 96) // 12 bytes
1332     return get(AMDGPU::V_INDIRECT_REG_WRITE_GPR_IDX_B32_V3);
1333   if (VecSize <= 128) // 16 bytes
1334     return get(AMDGPU::V_INDIRECT_REG_WRITE_GPR_IDX_B32_V4);
1335   if (VecSize <= 160) // 20 bytes
1336     return get(AMDGPU::V_INDIRECT_REG_WRITE_GPR_IDX_B32_V5);
1337   if (VecSize <= 256) // 32 bytes
1338     return get(AMDGPU::V_INDIRECT_REG_WRITE_GPR_IDX_B32_V8);
1339   if (VecSize <= 288) // 36 bytes
1340     return get(AMDGPU::V_INDIRECT_REG_WRITE_GPR_IDX_B32_V9);
1341   if (VecSize <= 320) // 40 bytes
1342     return get(AMDGPU::V_INDIRECT_REG_WRITE_GPR_IDX_B32_V10);
1343   if (VecSize <= 352) // 44 bytes
1344     return get(AMDGPU::V_INDIRECT_REG_WRITE_GPR_IDX_B32_V11);
1345   if (VecSize <= 384) // 48 bytes
1346     return get(AMDGPU::V_INDIRECT_REG_WRITE_GPR_IDX_B32_V12);
1347   if (VecSize <= 512) // 64 bytes
1348     return get(AMDGPU::V_INDIRECT_REG_WRITE_GPR_IDX_B32_V16);
1349   if (VecSize <= 1024) // 128 bytes
1350     return get(AMDGPU::V_INDIRECT_REG_WRITE_GPR_IDX_B32_V32);
1351 
1352   llvm_unreachable("unsupported size for IndirectRegWriteGPRIDX pseudos");
1353 }
1354 
1355 static unsigned getIndirectVGPRWriteMovRelPseudoOpc(unsigned VecSize) {
1356   if (VecSize <= 32) // 4 bytes
1357     return AMDGPU::V_INDIRECT_REG_WRITE_MOVREL_B32_V1;
1358   if (VecSize <= 64) // 8 bytes
1359     return AMDGPU::V_INDIRECT_REG_WRITE_MOVREL_B32_V2;
1360   if (VecSize <= 96) // 12 bytes
1361     return AMDGPU::V_INDIRECT_REG_WRITE_MOVREL_B32_V3;
1362   if (VecSize <= 128) // 16 bytes
1363     return AMDGPU::V_INDIRECT_REG_WRITE_MOVREL_B32_V4;
1364   if (VecSize <= 160) // 20 bytes
1365     return AMDGPU::V_INDIRECT_REG_WRITE_MOVREL_B32_V5;
1366   if (VecSize <= 256) // 32 bytes
1367     return AMDGPU::V_INDIRECT_REG_WRITE_MOVREL_B32_V8;
1368   if (VecSize <= 288) // 36 bytes
1369     return AMDGPU::V_INDIRECT_REG_WRITE_MOVREL_B32_V9;
1370   if (VecSize <= 320) // 40 bytes
1371     return AMDGPU::V_INDIRECT_REG_WRITE_MOVREL_B32_V10;
1372   if (VecSize <= 352) // 44 bytes
1373     return AMDGPU::V_INDIRECT_REG_WRITE_MOVREL_B32_V11;
1374   if (VecSize <= 384) // 48 bytes
1375     return AMDGPU::V_INDIRECT_REG_WRITE_MOVREL_B32_V12;
1376   if (VecSize <= 512) // 64 bytes
1377     return AMDGPU::V_INDIRECT_REG_WRITE_MOVREL_B32_V16;
1378   if (VecSize <= 1024) // 128 bytes
1379     return AMDGPU::V_INDIRECT_REG_WRITE_MOVREL_B32_V32;
1380 
1381   llvm_unreachable("unsupported size for IndirectRegWrite pseudos");
1382 }
1383 
1384 static unsigned getIndirectSGPRWriteMovRelPseudo32(unsigned VecSize) {
1385   if (VecSize <= 32) // 4 bytes
1386     return AMDGPU::S_INDIRECT_REG_WRITE_MOVREL_B32_V1;
1387   if (VecSize <= 64) // 8 bytes
1388     return AMDGPU::S_INDIRECT_REG_WRITE_MOVREL_B32_V2;
1389   if (VecSize <= 96) // 12 bytes
1390     return AMDGPU::S_INDIRECT_REG_WRITE_MOVREL_B32_V3;
1391   if (VecSize <= 128) // 16 bytes
1392     return AMDGPU::S_INDIRECT_REG_WRITE_MOVREL_B32_V4;
1393   if (VecSize <= 160) // 20 bytes
1394     return AMDGPU::S_INDIRECT_REG_WRITE_MOVREL_B32_V5;
1395   if (VecSize <= 256) // 32 bytes
1396     return AMDGPU::S_INDIRECT_REG_WRITE_MOVREL_B32_V8;
1397   if (VecSize <= 512) // 64 bytes
1398     return AMDGPU::S_INDIRECT_REG_WRITE_MOVREL_B32_V16;
1399   if (VecSize <= 1024) // 128 bytes
1400     return AMDGPU::S_INDIRECT_REG_WRITE_MOVREL_B32_V32;
1401 
1402   llvm_unreachable("unsupported size for IndirectRegWrite pseudos");
1403 }
1404 
1405 static unsigned getIndirectSGPRWriteMovRelPseudo64(unsigned VecSize) {
1406   if (VecSize <= 64) // 8 bytes
1407     return AMDGPU::S_INDIRECT_REG_WRITE_MOVREL_B64_V1;
1408   if (VecSize <= 128) // 16 bytes
1409     return AMDGPU::S_INDIRECT_REG_WRITE_MOVREL_B64_V2;
1410   if (VecSize <= 256) // 32 bytes
1411     return AMDGPU::S_INDIRECT_REG_WRITE_MOVREL_B64_V4;
1412   if (VecSize <= 512) // 64 bytes
1413     return AMDGPU::S_INDIRECT_REG_WRITE_MOVREL_B64_V8;
1414   if (VecSize <= 1024) // 128 bytes
1415     return AMDGPU::S_INDIRECT_REG_WRITE_MOVREL_B64_V16;
1416 
1417   llvm_unreachable("unsupported size for IndirectRegWrite pseudos");
1418 }
1419 
1420 const MCInstrDesc &
1421 SIInstrInfo::getIndirectRegWriteMovRelPseudo(unsigned VecSize, unsigned EltSize,
1422                                              bool IsSGPR) const {
1423   if (IsSGPR) {
1424     switch (EltSize) {
1425     case 32:
1426       return get(getIndirectSGPRWriteMovRelPseudo32(VecSize));
1427     case 64:
1428       return get(getIndirectSGPRWriteMovRelPseudo64(VecSize));
1429     default:
1430       llvm_unreachable("invalid reg indexing elt size");
1431     }
1432   }
1433 
1434   assert(EltSize == 32 && "invalid reg indexing elt size");
1435   return get(getIndirectVGPRWriteMovRelPseudoOpc(VecSize));
1436 }
1437 
1438 static unsigned getSGPRSpillSaveOpcode(unsigned Size) {
1439   switch (Size) {
1440   case 4:
1441     return AMDGPU::SI_SPILL_S32_SAVE;
1442   case 8:
1443     return AMDGPU::SI_SPILL_S64_SAVE;
1444   case 12:
1445     return AMDGPU::SI_SPILL_S96_SAVE;
1446   case 16:
1447     return AMDGPU::SI_SPILL_S128_SAVE;
1448   case 20:
1449     return AMDGPU::SI_SPILL_S160_SAVE;
1450   case 24:
1451     return AMDGPU::SI_SPILL_S192_SAVE;
1452   case 28:
1453     return AMDGPU::SI_SPILL_S224_SAVE;
1454   case 32:
1455     return AMDGPU::SI_SPILL_S256_SAVE;
1456   case 36:
1457     return AMDGPU::SI_SPILL_S288_SAVE;
1458   case 40:
1459     return AMDGPU::SI_SPILL_S320_SAVE;
1460   case 44:
1461     return AMDGPU::SI_SPILL_S352_SAVE;
1462   case 48:
1463     return AMDGPU::SI_SPILL_S384_SAVE;
1464   case 64:
1465     return AMDGPU::SI_SPILL_S512_SAVE;
1466   case 128:
1467     return AMDGPU::SI_SPILL_S1024_SAVE;
1468   default:
1469     llvm_unreachable("unknown register size");
1470   }
1471 }
1472 
1473 static unsigned getVGPRSpillSaveOpcode(unsigned Size) {
1474   switch (Size) {
1475   case 4:
1476     return AMDGPU::SI_SPILL_V32_SAVE;
1477   case 8:
1478     return AMDGPU::SI_SPILL_V64_SAVE;
1479   case 12:
1480     return AMDGPU::SI_SPILL_V96_SAVE;
1481   case 16:
1482     return AMDGPU::SI_SPILL_V128_SAVE;
1483   case 20:
1484     return AMDGPU::SI_SPILL_V160_SAVE;
1485   case 24:
1486     return AMDGPU::SI_SPILL_V192_SAVE;
1487   case 28:
1488     return AMDGPU::SI_SPILL_V224_SAVE;
1489   case 32:
1490     return AMDGPU::SI_SPILL_V256_SAVE;
1491   case 36:
1492     return AMDGPU::SI_SPILL_V288_SAVE;
1493   case 40:
1494     return AMDGPU::SI_SPILL_V320_SAVE;
1495   case 44:
1496     return AMDGPU::SI_SPILL_V352_SAVE;
1497   case 48:
1498     return AMDGPU::SI_SPILL_V384_SAVE;
1499   case 64:
1500     return AMDGPU::SI_SPILL_V512_SAVE;
1501   case 128:
1502     return AMDGPU::SI_SPILL_V1024_SAVE;
1503   default:
1504     llvm_unreachable("unknown register size");
1505   }
1506 }
1507 
1508 static unsigned getAGPRSpillSaveOpcode(unsigned Size) {
1509   switch (Size) {
1510   case 4:
1511     return AMDGPU::SI_SPILL_A32_SAVE;
1512   case 8:
1513     return AMDGPU::SI_SPILL_A64_SAVE;
1514   case 12:
1515     return AMDGPU::SI_SPILL_A96_SAVE;
1516   case 16:
1517     return AMDGPU::SI_SPILL_A128_SAVE;
1518   case 20:
1519     return AMDGPU::SI_SPILL_A160_SAVE;
1520   case 24:
1521     return AMDGPU::SI_SPILL_A192_SAVE;
1522   case 28:
1523     return AMDGPU::SI_SPILL_A224_SAVE;
1524   case 32:
1525     return AMDGPU::SI_SPILL_A256_SAVE;
1526   case 36:
1527     return AMDGPU::SI_SPILL_A288_SAVE;
1528   case 40:
1529     return AMDGPU::SI_SPILL_A320_SAVE;
1530   case 44:
1531     return AMDGPU::SI_SPILL_A352_SAVE;
1532   case 48:
1533     return AMDGPU::SI_SPILL_A384_SAVE;
1534   case 64:
1535     return AMDGPU::SI_SPILL_A512_SAVE;
1536   case 128:
1537     return AMDGPU::SI_SPILL_A1024_SAVE;
1538   default:
1539     llvm_unreachable("unknown register size");
1540   }
1541 }
1542 
1543 static unsigned getAVSpillSaveOpcode(unsigned Size) {
1544   switch (Size) {
1545   case 4:
1546     return AMDGPU::SI_SPILL_AV32_SAVE;
1547   case 8:
1548     return AMDGPU::SI_SPILL_AV64_SAVE;
1549   case 12:
1550     return AMDGPU::SI_SPILL_AV96_SAVE;
1551   case 16:
1552     return AMDGPU::SI_SPILL_AV128_SAVE;
1553   case 20:
1554     return AMDGPU::SI_SPILL_AV160_SAVE;
1555   case 24:
1556     return AMDGPU::SI_SPILL_AV192_SAVE;
1557   case 28:
1558     return AMDGPU::SI_SPILL_AV224_SAVE;
1559   case 32:
1560     return AMDGPU::SI_SPILL_AV256_SAVE;
1561   case 36:
1562     return AMDGPU::SI_SPILL_AV288_SAVE;
1563   case 40:
1564     return AMDGPU::SI_SPILL_AV320_SAVE;
1565   case 44:
1566     return AMDGPU::SI_SPILL_AV352_SAVE;
1567   case 48:
1568     return AMDGPU::SI_SPILL_AV384_SAVE;
1569   case 64:
1570     return AMDGPU::SI_SPILL_AV512_SAVE;
1571   case 128:
1572     return AMDGPU::SI_SPILL_AV1024_SAVE;
1573   default:
1574     llvm_unreachable("unknown register size");
1575   }
1576 }
1577 
1578 void SIInstrInfo::storeRegToStackSlot(
1579     MachineBasicBlock &MBB, MachineBasicBlock::iterator MI, Register SrcReg,
1580     bool isKill, int FrameIndex, const TargetRegisterClass *RC,
1581     const TargetRegisterInfo *TRI, Register VReg) const {
1582   MachineFunction *MF = MBB.getParent();
1583   SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();
1584   MachineFrameInfo &FrameInfo = MF->getFrameInfo();
1585   const DebugLoc &DL = MBB.findDebugLoc(MI);
1586 
1587   MachinePointerInfo PtrInfo
1588     = MachinePointerInfo::getFixedStack(*MF, FrameIndex);
1589   MachineMemOperand *MMO = MF->getMachineMemOperand(
1590       PtrInfo, MachineMemOperand::MOStore, FrameInfo.getObjectSize(FrameIndex),
1591       FrameInfo.getObjectAlign(FrameIndex));
1592   unsigned SpillSize = TRI->getSpillSize(*RC);
1593 
1594   MachineRegisterInfo &MRI = MF->getRegInfo();
1595   if (RI.isSGPRClass(RC)) {
1596     MFI->setHasSpilledSGPRs();
1597     assert(SrcReg != AMDGPU::M0 && "m0 should not be spilled");
1598     assert(SrcReg != AMDGPU::EXEC_LO && SrcReg != AMDGPU::EXEC_HI &&
1599            SrcReg != AMDGPU::EXEC && "exec should not be spilled");
1600 
1601     // We are only allowed to create one new instruction when spilling
1602     // registers, so we need to use pseudo instruction for spilling SGPRs.
1603     const MCInstrDesc &OpDesc = get(getSGPRSpillSaveOpcode(SpillSize));
1604 
1605     // The SGPR spill/restore instructions only work on number sgprs, so we need
1606     // to make sure we are using the correct register class.
1607     if (SrcReg.isVirtual() && SpillSize == 4) {
1608       MRI.constrainRegClass(SrcReg, &AMDGPU::SReg_32_XM0_XEXECRegClass);
1609     }
1610 
1611     BuildMI(MBB, MI, DL, OpDesc)
1612       .addReg(SrcReg, getKillRegState(isKill)) // data
1613       .addFrameIndex(FrameIndex)               // addr
1614       .addMemOperand(MMO)
1615       .addReg(MFI->getStackPtrOffsetReg(), RegState::Implicit);
1616 
1617     if (RI.spillSGPRToVGPR())
1618       FrameInfo.setStackID(FrameIndex, TargetStackID::SGPRSpill);
1619     return;
1620   }
1621 
1622   unsigned Opcode = RI.isVectorSuperClass(RC)
1623                         ? getAVSpillSaveOpcode(SpillSize)
1624                         : RI.isAGPRClass(RC)
1625                               ? getAGPRSpillSaveOpcode(SpillSize)
1626                               : getVGPRSpillSaveOpcode(SpillSize);
1627   MFI->setHasSpilledVGPRs();
1628 
1629   BuildMI(MBB, MI, DL, get(Opcode))
1630     .addReg(SrcReg, getKillRegState(isKill)) // data
1631     .addFrameIndex(FrameIndex)               // addr
1632     .addReg(MFI->getStackPtrOffsetReg())     // scratch_offset
1633     .addImm(0)                               // offset
1634     .addMemOperand(MMO);
1635 }
1636 
1637 static unsigned getSGPRSpillRestoreOpcode(unsigned Size) {
1638   switch (Size) {
1639   case 4:
1640     return AMDGPU::SI_SPILL_S32_RESTORE;
1641   case 8:
1642     return AMDGPU::SI_SPILL_S64_RESTORE;
1643   case 12:
1644     return AMDGPU::SI_SPILL_S96_RESTORE;
1645   case 16:
1646     return AMDGPU::SI_SPILL_S128_RESTORE;
1647   case 20:
1648     return AMDGPU::SI_SPILL_S160_RESTORE;
1649   case 24:
1650     return AMDGPU::SI_SPILL_S192_RESTORE;
1651   case 28:
1652     return AMDGPU::SI_SPILL_S224_RESTORE;
1653   case 32:
1654     return AMDGPU::SI_SPILL_S256_RESTORE;
1655   case 36:
1656     return AMDGPU::SI_SPILL_S288_RESTORE;
1657   case 40:
1658     return AMDGPU::SI_SPILL_S320_RESTORE;
1659   case 44:
1660     return AMDGPU::SI_SPILL_S352_RESTORE;
1661   case 48:
1662     return AMDGPU::SI_SPILL_S384_RESTORE;
1663   case 64:
1664     return AMDGPU::SI_SPILL_S512_RESTORE;
1665   case 128:
1666     return AMDGPU::SI_SPILL_S1024_RESTORE;
1667   default:
1668     llvm_unreachable("unknown register size");
1669   }
1670 }
1671 
1672 static unsigned getVGPRSpillRestoreOpcode(unsigned Size) {
1673   switch (Size) {
1674   case 4:
1675     return AMDGPU::SI_SPILL_V32_RESTORE;
1676   case 8:
1677     return AMDGPU::SI_SPILL_V64_RESTORE;
1678   case 12:
1679     return AMDGPU::SI_SPILL_V96_RESTORE;
1680   case 16:
1681     return AMDGPU::SI_SPILL_V128_RESTORE;
1682   case 20:
1683     return AMDGPU::SI_SPILL_V160_RESTORE;
1684   case 24:
1685     return AMDGPU::SI_SPILL_V192_RESTORE;
1686   case 28:
1687     return AMDGPU::SI_SPILL_V224_RESTORE;
1688   case 32:
1689     return AMDGPU::SI_SPILL_V256_RESTORE;
1690   case 36:
1691     return AMDGPU::SI_SPILL_V288_RESTORE;
1692   case 40:
1693     return AMDGPU::SI_SPILL_V320_RESTORE;
1694   case 44:
1695     return AMDGPU::SI_SPILL_V352_RESTORE;
1696   case 48:
1697     return AMDGPU::SI_SPILL_V384_RESTORE;
1698   case 64:
1699     return AMDGPU::SI_SPILL_V512_RESTORE;
1700   case 128:
1701     return AMDGPU::SI_SPILL_V1024_RESTORE;
1702   default:
1703     llvm_unreachable("unknown register size");
1704   }
1705 }
1706 
1707 static unsigned getAGPRSpillRestoreOpcode(unsigned Size) {
1708   switch (Size) {
1709   case 4:
1710     return AMDGPU::SI_SPILL_A32_RESTORE;
1711   case 8:
1712     return AMDGPU::SI_SPILL_A64_RESTORE;
1713   case 12:
1714     return AMDGPU::SI_SPILL_A96_RESTORE;
1715   case 16:
1716     return AMDGPU::SI_SPILL_A128_RESTORE;
1717   case 20:
1718     return AMDGPU::SI_SPILL_A160_RESTORE;
1719   case 24:
1720     return AMDGPU::SI_SPILL_A192_RESTORE;
1721   case 28:
1722     return AMDGPU::SI_SPILL_A224_RESTORE;
1723   case 32:
1724     return AMDGPU::SI_SPILL_A256_RESTORE;
1725   case 36:
1726     return AMDGPU::SI_SPILL_A288_RESTORE;
1727   case 40:
1728     return AMDGPU::SI_SPILL_A320_RESTORE;
1729   case 44:
1730     return AMDGPU::SI_SPILL_A352_RESTORE;
1731   case 48:
1732     return AMDGPU::SI_SPILL_A384_RESTORE;
1733   case 64:
1734     return AMDGPU::SI_SPILL_A512_RESTORE;
1735   case 128:
1736     return AMDGPU::SI_SPILL_A1024_RESTORE;
1737   default:
1738     llvm_unreachable("unknown register size");
1739   }
1740 }
1741 
1742 static unsigned getAVSpillRestoreOpcode(unsigned Size) {
1743   switch (Size) {
1744   case 4:
1745     return AMDGPU::SI_SPILL_AV32_RESTORE;
1746   case 8:
1747     return AMDGPU::SI_SPILL_AV64_RESTORE;
1748   case 12:
1749     return AMDGPU::SI_SPILL_AV96_RESTORE;
1750   case 16:
1751     return AMDGPU::SI_SPILL_AV128_RESTORE;
1752   case 20:
1753     return AMDGPU::SI_SPILL_AV160_RESTORE;
1754   case 24:
1755     return AMDGPU::SI_SPILL_AV192_RESTORE;
1756   case 28:
1757     return AMDGPU::SI_SPILL_AV224_RESTORE;
1758   case 32:
1759     return AMDGPU::SI_SPILL_AV256_RESTORE;
1760   case 36:
1761     return AMDGPU::SI_SPILL_AV288_RESTORE;
1762   case 40:
1763     return AMDGPU::SI_SPILL_AV320_RESTORE;
1764   case 44:
1765     return AMDGPU::SI_SPILL_AV352_RESTORE;
1766   case 48:
1767     return AMDGPU::SI_SPILL_AV384_RESTORE;
1768   case 64:
1769     return AMDGPU::SI_SPILL_AV512_RESTORE;
1770   case 128:
1771     return AMDGPU::SI_SPILL_AV1024_RESTORE;
1772   default:
1773     llvm_unreachable("unknown register size");
1774   }
1775 }
1776 
1777 void SIInstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
1778                                        MachineBasicBlock::iterator MI,
1779                                        Register DestReg, int FrameIndex,
1780                                        const TargetRegisterClass *RC,
1781                                        const TargetRegisterInfo *TRI,
1782                                        Register VReg) const {
1783   MachineFunction *MF = MBB.getParent();
1784   SIMachineFunctionInfo *MFI = MF->getInfo<SIMachineFunctionInfo>();
1785   MachineFrameInfo &FrameInfo = MF->getFrameInfo();
1786   const DebugLoc &DL = MBB.findDebugLoc(MI);
1787   unsigned SpillSize = TRI->getSpillSize(*RC);
1788 
1789   MachinePointerInfo PtrInfo
1790     = MachinePointerInfo::getFixedStack(*MF, FrameIndex);
1791 
1792   MachineMemOperand *MMO = MF->getMachineMemOperand(
1793       PtrInfo, MachineMemOperand::MOLoad, FrameInfo.getObjectSize(FrameIndex),
1794       FrameInfo.getObjectAlign(FrameIndex));
1795 
1796   if (RI.isSGPRClass(RC)) {
1797     MFI->setHasSpilledSGPRs();
1798     assert(DestReg != AMDGPU::M0 && "m0 should not be reloaded into");
1799     assert(DestReg != AMDGPU::EXEC_LO && DestReg != AMDGPU::EXEC_HI &&
1800            DestReg != AMDGPU::EXEC && "exec should not be spilled");
1801 
1802     // FIXME: Maybe this should not include a memoperand because it will be
1803     // lowered to non-memory instructions.
1804     const MCInstrDesc &OpDesc = get(getSGPRSpillRestoreOpcode(SpillSize));
1805     if (DestReg.isVirtual() && SpillSize == 4) {
1806       MachineRegisterInfo &MRI = MF->getRegInfo();
1807       MRI.constrainRegClass(DestReg, &AMDGPU::SReg_32_XM0_XEXECRegClass);
1808     }
1809 
1810     if (RI.spillSGPRToVGPR())
1811       FrameInfo.setStackID(FrameIndex, TargetStackID::SGPRSpill);
1812     BuildMI(MBB, MI, DL, OpDesc, DestReg)
1813       .addFrameIndex(FrameIndex) // addr
1814       .addMemOperand(MMO)
1815       .addReg(MFI->getStackPtrOffsetReg(), RegState::Implicit);
1816 
1817     return;
1818   }
1819 
1820   unsigned Opcode = RI.isVectorSuperClass(RC)
1821                         ? getAVSpillRestoreOpcode(SpillSize)
1822                         : RI.isAGPRClass(RC)
1823                               ? getAGPRSpillRestoreOpcode(SpillSize)
1824                               : getVGPRSpillRestoreOpcode(SpillSize);
1825   BuildMI(MBB, MI, DL, get(Opcode), DestReg)
1826       .addFrameIndex(FrameIndex)           // vaddr
1827       .addReg(MFI->getStackPtrOffsetReg()) // scratch_offset
1828       .addImm(0)                           // offset
1829       .addMemOperand(MMO);
1830 }
1831 
1832 void SIInstrInfo::insertNoop(MachineBasicBlock &MBB,
1833                              MachineBasicBlock::iterator MI) const {
1834   insertNoops(MBB, MI, 1);
1835 }
1836 
1837 void SIInstrInfo::insertNoops(MachineBasicBlock &MBB,
1838                               MachineBasicBlock::iterator MI,
1839                               unsigned Quantity) const {
1840   DebugLoc DL = MBB.findDebugLoc(MI);
1841   while (Quantity > 0) {
1842     unsigned Arg = std::min(Quantity, 8u);
1843     Quantity -= Arg;
1844     BuildMI(MBB, MI, DL, get(AMDGPU::S_NOP)).addImm(Arg - 1);
1845   }
1846 }
1847 
1848 void SIInstrInfo::insertReturn(MachineBasicBlock &MBB) const {
1849   auto MF = MBB.getParent();
1850   SIMachineFunctionInfo *Info = MF->getInfo<SIMachineFunctionInfo>();
1851 
1852   assert(Info->isEntryFunction());
1853 
1854   if (MBB.succ_empty()) {
1855     bool HasNoTerminator = MBB.getFirstTerminator() == MBB.end();
1856     if (HasNoTerminator) {
1857       if (Info->returnsVoid()) {
1858         BuildMI(MBB, MBB.end(), DebugLoc(), get(AMDGPU::S_ENDPGM)).addImm(0);
1859       } else {
1860         BuildMI(MBB, MBB.end(), DebugLoc(), get(AMDGPU::SI_RETURN_TO_EPILOG));
1861       }
1862     }
1863   }
1864 }
1865 
1866 unsigned SIInstrInfo::getNumWaitStates(const MachineInstr &MI) {
1867   switch (MI.getOpcode()) {
1868   default:
1869     if (MI.isMetaInstruction())
1870       return 0;
1871     return 1; // FIXME: Do wait states equal cycles?
1872 
1873   case AMDGPU::S_NOP:
1874     return MI.getOperand(0).getImm() + 1;
1875   // SI_RETURN_TO_EPILOG is a fallthrough to code outside of the function. The
1876   // hazard, even if one exist, won't really be visible. Should we handle it?
1877   }
1878 }
1879 
1880 bool SIInstrInfo::expandPostRAPseudo(MachineInstr &MI) const {
1881   const SIRegisterInfo *TRI = ST.getRegisterInfo();
1882   MachineBasicBlock &MBB = *MI.getParent();
1883   DebugLoc DL = MBB.findDebugLoc(MI);
1884   switch (MI.getOpcode()) {
1885   default: return TargetInstrInfo::expandPostRAPseudo(MI);
1886   case AMDGPU::S_MOV_B64_term:
1887     // This is only a terminator to get the correct spill code placement during
1888     // register allocation.
1889     MI.setDesc(get(AMDGPU::S_MOV_B64));
1890     break;
1891 
1892   case AMDGPU::S_MOV_B32_term:
1893     // This is only a terminator to get the correct spill code placement during
1894     // register allocation.
1895     MI.setDesc(get(AMDGPU::S_MOV_B32));
1896     break;
1897 
1898   case AMDGPU::S_XOR_B64_term:
1899     // This is only a terminator to get the correct spill code placement during
1900     // register allocation.
1901     MI.setDesc(get(AMDGPU::S_XOR_B64));
1902     break;
1903 
1904   case AMDGPU::S_XOR_B32_term:
1905     // This is only a terminator to get the correct spill code placement during
1906     // register allocation.
1907     MI.setDesc(get(AMDGPU::S_XOR_B32));
1908     break;
1909   case AMDGPU::S_OR_B64_term:
1910     // This is only a terminator to get the correct spill code placement during
1911     // register allocation.
1912     MI.setDesc(get(AMDGPU::S_OR_B64));
1913     break;
1914   case AMDGPU::S_OR_B32_term:
1915     // This is only a terminator to get the correct spill code placement during
1916     // register allocation.
1917     MI.setDesc(get(AMDGPU::S_OR_B32));
1918     break;
1919 
1920   case AMDGPU::S_ANDN2_B64_term:
1921     // This is only a terminator to get the correct spill code placement during
1922     // register allocation.
1923     MI.setDesc(get(AMDGPU::S_ANDN2_B64));
1924     break;
1925 
1926   case AMDGPU::S_ANDN2_B32_term:
1927     // This is only a terminator to get the correct spill code placement during
1928     // register allocation.
1929     MI.setDesc(get(AMDGPU::S_ANDN2_B32));
1930     break;
1931 
1932   case AMDGPU::S_AND_B64_term:
1933     // This is only a terminator to get the correct spill code placement during
1934     // register allocation.
1935     MI.setDesc(get(AMDGPU::S_AND_B64));
1936     break;
1937 
1938   case AMDGPU::S_AND_B32_term:
1939     // This is only a terminator to get the correct spill code placement during
1940     // register allocation.
1941     MI.setDesc(get(AMDGPU::S_AND_B32));
1942     break;
1943 
1944   case AMDGPU::V_MOV_B64_PSEUDO: {
1945     Register Dst = MI.getOperand(0).getReg();
1946     Register DstLo = RI.getSubReg(Dst, AMDGPU::sub0);
1947     Register DstHi = RI.getSubReg(Dst, AMDGPU::sub1);
1948 
1949     const MachineOperand &SrcOp = MI.getOperand(1);
1950     // FIXME: Will this work for 64-bit floating point immediates?
1951     assert(!SrcOp.isFPImm());
1952     if (ST.hasMovB64()) {
1953       MI.setDesc(get(AMDGPU::V_MOV_B64_e32));
1954       if (SrcOp.isReg() || isInlineConstant(MI, 1) ||
1955           isUInt<32>(SrcOp.getImm()))
1956         break;
1957     }
1958     if (SrcOp.isImm()) {
1959       APInt Imm(64, SrcOp.getImm());
1960       APInt Lo(32, Imm.getLoBits(32).getZExtValue());
1961       APInt Hi(32, Imm.getHiBits(32).getZExtValue());
1962       if (ST.hasPackedFP32Ops() && Lo == Hi && isInlineConstant(Lo)) {
1963         BuildMI(MBB, MI, DL, get(AMDGPU::V_PK_MOV_B32), Dst)
1964           .addImm(SISrcMods::OP_SEL_1)
1965           .addImm(Lo.getSExtValue())
1966           .addImm(SISrcMods::OP_SEL_1)
1967           .addImm(Lo.getSExtValue())
1968           .addImm(0)  // op_sel_lo
1969           .addImm(0)  // op_sel_hi
1970           .addImm(0)  // neg_lo
1971           .addImm(0)  // neg_hi
1972           .addImm(0); // clamp
1973       } else {
1974         BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DstLo)
1975           .addImm(Lo.getSExtValue())
1976           .addReg(Dst, RegState::Implicit | RegState::Define);
1977         BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DstHi)
1978           .addImm(Hi.getSExtValue())
1979           .addReg(Dst, RegState::Implicit | RegState::Define);
1980       }
1981     } else {
1982       assert(SrcOp.isReg());
1983       if (ST.hasPackedFP32Ops() &&
1984           !RI.isAGPR(MBB.getParent()->getRegInfo(), SrcOp.getReg())) {
1985         BuildMI(MBB, MI, DL, get(AMDGPU::V_PK_MOV_B32), Dst)
1986           .addImm(SISrcMods::OP_SEL_1) // src0_mod
1987           .addReg(SrcOp.getReg())
1988           .addImm(SISrcMods::OP_SEL_0 | SISrcMods::OP_SEL_1) // src1_mod
1989           .addReg(SrcOp.getReg())
1990           .addImm(0)  // op_sel_lo
1991           .addImm(0)  // op_sel_hi
1992           .addImm(0)  // neg_lo
1993           .addImm(0)  // neg_hi
1994           .addImm(0); // clamp
1995       } else {
1996         BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DstLo)
1997           .addReg(RI.getSubReg(SrcOp.getReg(), AMDGPU::sub0))
1998           .addReg(Dst, RegState::Implicit | RegState::Define);
1999         BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), DstHi)
2000           .addReg(RI.getSubReg(SrcOp.getReg(), AMDGPU::sub1))
2001           .addReg(Dst, RegState::Implicit | RegState::Define);
2002       }
2003     }
2004     MI.eraseFromParent();
2005     break;
2006   }
2007   case AMDGPU::V_MOV_B64_DPP_PSEUDO: {
2008     expandMovDPP64(MI);
2009     break;
2010   }
2011   case AMDGPU::S_MOV_B64_IMM_PSEUDO: {
2012     const MachineOperand &SrcOp = MI.getOperand(1);
2013     assert(!SrcOp.isFPImm());
2014     APInt Imm(64, SrcOp.getImm());
2015     if (Imm.isIntN(32) || isInlineConstant(Imm)) {
2016       MI.setDesc(get(AMDGPU::S_MOV_B64));
2017       break;
2018     }
2019 
2020     Register Dst = MI.getOperand(0).getReg();
2021     Register DstLo = RI.getSubReg(Dst, AMDGPU::sub0);
2022     Register DstHi = RI.getSubReg(Dst, AMDGPU::sub1);
2023 
2024     APInt Lo(32, Imm.getLoBits(32).getZExtValue());
2025     APInt Hi(32, Imm.getHiBits(32).getZExtValue());
2026     BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B32), DstLo)
2027       .addImm(Lo.getSExtValue())
2028       .addReg(Dst, RegState::Implicit | RegState::Define);
2029     BuildMI(MBB, MI, DL, get(AMDGPU::S_MOV_B32), DstHi)
2030       .addImm(Hi.getSExtValue())
2031       .addReg(Dst, RegState::Implicit | RegState::Define);
2032     MI.eraseFromParent();
2033     break;
2034   }
2035   case AMDGPU::V_SET_INACTIVE_B32: {
2036     unsigned NotOpc = ST.isWave32() ? AMDGPU::S_NOT_B32 : AMDGPU::S_NOT_B64;
2037     unsigned Exec = ST.isWave32() ? AMDGPU::EXEC_LO : AMDGPU::EXEC;
2038     // FIXME: We may possibly optimize the COPY once we find ways to make LLVM
2039     // optimizations (mainly Register Coalescer) aware of WWM register liveness.
2040     BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), MI.getOperand(0).getReg())
2041         .add(MI.getOperand(1));
2042     auto FirstNot = BuildMI(MBB, MI, DL, get(NotOpc), Exec).addReg(Exec);
2043     FirstNot->addRegisterDead(AMDGPU::SCC, TRI); // SCC is overwritten
2044     BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_e32), MI.getOperand(0).getReg())
2045       .add(MI.getOperand(2));
2046     BuildMI(MBB, MI, DL, get(NotOpc), Exec)
2047       .addReg(Exec);
2048     MI.eraseFromParent();
2049     break;
2050   }
2051   case AMDGPU::V_SET_INACTIVE_B64: {
2052     unsigned NotOpc = ST.isWave32() ? AMDGPU::S_NOT_B32 : AMDGPU::S_NOT_B64;
2053     unsigned Exec = ST.isWave32() ? AMDGPU::EXEC_LO : AMDGPU::EXEC;
2054     MachineInstr *Copy = BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B64_PSEUDO),
2055                                  MI.getOperand(0).getReg())
2056                              .add(MI.getOperand(1));
2057     expandPostRAPseudo(*Copy);
2058     auto FirstNot = BuildMI(MBB, MI, DL, get(NotOpc), Exec).addReg(Exec);
2059     FirstNot->addRegisterDead(AMDGPU::SCC, TRI); // SCC is overwritten
2060     Copy = BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B64_PSEUDO),
2061                    MI.getOperand(0).getReg())
2062                .add(MI.getOperand(2));
2063     expandPostRAPseudo(*Copy);
2064     BuildMI(MBB, MI, DL, get(NotOpc), Exec)
2065       .addReg(Exec);
2066     MI.eraseFromParent();
2067     break;
2068   }
2069   case AMDGPU::V_INDIRECT_REG_WRITE_MOVREL_B32_V1:
2070   case AMDGPU::V_INDIRECT_REG_WRITE_MOVREL_B32_V2:
2071   case AMDGPU::V_INDIRECT_REG_WRITE_MOVREL_B32_V3:
2072   case AMDGPU::V_INDIRECT_REG_WRITE_MOVREL_B32_V4:
2073   case AMDGPU::V_INDIRECT_REG_WRITE_MOVREL_B32_V5:
2074   case AMDGPU::V_INDIRECT_REG_WRITE_MOVREL_B32_V8:
2075   case AMDGPU::V_INDIRECT_REG_WRITE_MOVREL_B32_V9:
2076   case AMDGPU::V_INDIRECT_REG_WRITE_MOVREL_B32_V10:
2077   case AMDGPU::V_INDIRECT_REG_WRITE_MOVREL_B32_V11:
2078   case AMDGPU::V_INDIRECT_REG_WRITE_MOVREL_B32_V12:
2079   case AMDGPU::V_INDIRECT_REG_WRITE_MOVREL_B32_V16:
2080   case AMDGPU::V_INDIRECT_REG_WRITE_MOVREL_B32_V32:
2081   case AMDGPU::S_INDIRECT_REG_WRITE_MOVREL_B32_V1:
2082   case AMDGPU::S_INDIRECT_REG_WRITE_MOVREL_B32_V2:
2083   case AMDGPU::S_INDIRECT_REG_WRITE_MOVREL_B32_V3:
2084   case AMDGPU::S_INDIRECT_REG_WRITE_MOVREL_B32_V4:
2085   case AMDGPU::S_INDIRECT_REG_WRITE_MOVREL_B32_V5:
2086   case AMDGPU::S_INDIRECT_REG_WRITE_MOVREL_B32_V8:
2087   case AMDGPU::S_INDIRECT_REG_WRITE_MOVREL_B32_V16:
2088   case AMDGPU::S_INDIRECT_REG_WRITE_MOVREL_B32_V32:
2089   case AMDGPU::S_INDIRECT_REG_WRITE_MOVREL_B64_V1:
2090   case AMDGPU::S_INDIRECT_REG_WRITE_MOVREL_B64_V2:
2091   case AMDGPU::S_INDIRECT_REG_WRITE_MOVREL_B64_V4:
2092   case AMDGPU::S_INDIRECT_REG_WRITE_MOVREL_B64_V8:
2093   case AMDGPU::S_INDIRECT_REG_WRITE_MOVREL_B64_V16: {
2094     const TargetRegisterClass *EltRC = getOpRegClass(MI, 2);
2095 
2096     unsigned Opc;
2097     if (RI.hasVGPRs(EltRC)) {
2098       Opc = AMDGPU::V_MOVRELD_B32_e32;
2099     } else {
2100       Opc = RI.getRegSizeInBits(*EltRC) == 64 ? AMDGPU::S_MOVRELD_B64
2101                                               : AMDGPU::S_MOVRELD_B32;
2102     }
2103 
2104     const MCInstrDesc &OpDesc = get(Opc);
2105     Register VecReg = MI.getOperand(0).getReg();
2106     bool IsUndef = MI.getOperand(1).isUndef();
2107     unsigned SubReg = MI.getOperand(3).getImm();
2108     assert(VecReg == MI.getOperand(1).getReg());
2109 
2110     MachineInstrBuilder MIB =
2111       BuildMI(MBB, MI, DL, OpDesc)
2112         .addReg(RI.getSubReg(VecReg, SubReg), RegState::Undef)
2113         .add(MI.getOperand(2))
2114         .addReg(VecReg, RegState::ImplicitDefine)
2115         .addReg(VecReg, RegState::Implicit | (IsUndef ? RegState::Undef : 0));
2116 
2117     const int ImpDefIdx =
2118         OpDesc.getNumOperands() + OpDesc.implicit_uses().size();
2119     const int ImpUseIdx = ImpDefIdx + 1;
2120     MIB->tieOperands(ImpDefIdx, ImpUseIdx);
2121     MI.eraseFromParent();
2122     break;
2123   }
2124   case AMDGPU::V_INDIRECT_REG_WRITE_GPR_IDX_B32_V1:
2125   case AMDGPU::V_INDIRECT_REG_WRITE_GPR_IDX_B32_V2:
2126   case AMDGPU::V_INDIRECT_REG_WRITE_GPR_IDX_B32_V3:
2127   case AMDGPU::V_INDIRECT_REG_WRITE_GPR_IDX_B32_V4:
2128   case AMDGPU::V_INDIRECT_REG_WRITE_GPR_IDX_B32_V5:
2129   case AMDGPU::V_INDIRECT_REG_WRITE_GPR_IDX_B32_V8:
2130   case AMDGPU::V_INDIRECT_REG_WRITE_GPR_IDX_B32_V9:
2131   case AMDGPU::V_INDIRECT_REG_WRITE_GPR_IDX_B32_V10:
2132   case AMDGPU::V_INDIRECT_REG_WRITE_GPR_IDX_B32_V11:
2133   case AMDGPU::V_INDIRECT_REG_WRITE_GPR_IDX_B32_V12:
2134   case AMDGPU::V_INDIRECT_REG_WRITE_GPR_IDX_B32_V16:
2135   case AMDGPU::V_INDIRECT_REG_WRITE_GPR_IDX_B32_V32: {
2136     assert(ST.useVGPRIndexMode());
2137     Register VecReg = MI.getOperand(0).getReg();
2138     bool IsUndef = MI.getOperand(1).isUndef();
2139     Register Idx = MI.getOperand(3).getReg();
2140     Register SubReg = MI.getOperand(4).getImm();
2141 
2142     MachineInstr *SetOn = BuildMI(MBB, MI, DL, get(AMDGPU::S_SET_GPR_IDX_ON))
2143                               .addReg(Idx)
2144                               .addImm(AMDGPU::VGPRIndexMode::DST_ENABLE);
2145     SetOn->getOperand(3).setIsUndef();
2146 
2147     const MCInstrDesc &OpDesc = get(AMDGPU::V_MOV_B32_indirect_write);
2148     MachineInstrBuilder MIB =
2149         BuildMI(MBB, MI, DL, OpDesc)
2150             .addReg(RI.getSubReg(VecReg, SubReg), RegState::Undef)
2151             .add(MI.getOperand(2))
2152             .addReg(VecReg, RegState::ImplicitDefine)
2153             .addReg(VecReg,
2154                     RegState::Implicit | (IsUndef ? RegState::Undef : 0));
2155 
2156     const int ImpDefIdx =
2157         OpDesc.getNumOperands() + OpDesc.implicit_uses().size();
2158     const int ImpUseIdx = ImpDefIdx + 1;
2159     MIB->tieOperands(ImpDefIdx, ImpUseIdx);
2160 
2161     MachineInstr *SetOff = BuildMI(MBB, MI, DL, get(AMDGPU::S_SET_GPR_IDX_OFF));
2162 
2163     finalizeBundle(MBB, SetOn->getIterator(), std::next(SetOff->getIterator()));
2164 
2165     MI.eraseFromParent();
2166     break;
2167   }
2168   case AMDGPU::V_INDIRECT_REG_READ_GPR_IDX_B32_V1:
2169   case AMDGPU::V_INDIRECT_REG_READ_GPR_IDX_B32_V2:
2170   case AMDGPU::V_INDIRECT_REG_READ_GPR_IDX_B32_V3:
2171   case AMDGPU::V_INDIRECT_REG_READ_GPR_IDX_B32_V4:
2172   case AMDGPU::V_INDIRECT_REG_READ_GPR_IDX_B32_V5:
2173   case AMDGPU::V_INDIRECT_REG_READ_GPR_IDX_B32_V8:
2174   case AMDGPU::V_INDIRECT_REG_READ_GPR_IDX_B32_V9:
2175   case AMDGPU::V_INDIRECT_REG_READ_GPR_IDX_B32_V10:
2176   case AMDGPU::V_INDIRECT_REG_READ_GPR_IDX_B32_V11:
2177   case AMDGPU::V_INDIRECT_REG_READ_GPR_IDX_B32_V12:
2178   case AMDGPU::V_INDIRECT_REG_READ_GPR_IDX_B32_V16:
2179   case AMDGPU::V_INDIRECT_REG_READ_GPR_IDX_B32_V32: {
2180     assert(ST.useVGPRIndexMode());
2181     Register Dst = MI.getOperand(0).getReg();
2182     Register VecReg = MI.getOperand(1).getReg();
2183     bool IsUndef = MI.getOperand(1).isUndef();
2184     Register Idx = MI.getOperand(2).getReg();
2185     Register SubReg = MI.getOperand(3).getImm();
2186 
2187     MachineInstr *SetOn = BuildMI(MBB, MI, DL, get(AMDGPU::S_SET_GPR_IDX_ON))
2188                               .addReg(Idx)
2189                               .addImm(AMDGPU::VGPRIndexMode::SRC0_ENABLE);
2190     SetOn->getOperand(3).setIsUndef();
2191 
2192     BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_indirect_read))
2193         .addDef(Dst)
2194         .addReg(RI.getSubReg(VecReg, SubReg), RegState::Undef)
2195         .addReg(VecReg, RegState::Implicit | (IsUndef ? RegState::Undef : 0));
2196 
2197     MachineInstr *SetOff = BuildMI(MBB, MI, DL, get(AMDGPU::S_SET_GPR_IDX_OFF));
2198 
2199     finalizeBundle(MBB, SetOn->getIterator(), std::next(SetOff->getIterator()));
2200 
2201     MI.eraseFromParent();
2202     break;
2203   }
2204   case AMDGPU::SI_PC_ADD_REL_OFFSET: {
2205     MachineFunction &MF = *MBB.getParent();
2206     Register Reg = MI.getOperand(0).getReg();
2207     Register RegLo = RI.getSubReg(Reg, AMDGPU::sub0);
2208     Register RegHi = RI.getSubReg(Reg, AMDGPU::sub1);
2209 
2210     // Create a bundle so these instructions won't be re-ordered by the
2211     // post-RA scheduler.
2212     MIBundleBuilder Bundler(MBB, MI);
2213     Bundler.append(BuildMI(MF, DL, get(AMDGPU::S_GETPC_B64), Reg));
2214 
2215     // Add 32-bit offset from this instruction to the start of the
2216     // constant data.
2217     Bundler.append(BuildMI(MF, DL, get(AMDGPU::S_ADD_U32), RegLo)
2218                        .addReg(RegLo)
2219                        .add(MI.getOperand(1)));
2220 
2221     MachineInstrBuilder MIB = BuildMI(MF, DL, get(AMDGPU::S_ADDC_U32), RegHi)
2222                                   .addReg(RegHi);
2223     MIB.add(MI.getOperand(2));
2224 
2225     Bundler.append(MIB);
2226     finalizeBundle(MBB, Bundler.begin());
2227 
2228     MI.eraseFromParent();
2229     break;
2230   }
2231   case AMDGPU::ENTER_STRICT_WWM: {
2232     // This only gets its own opcode so that SIPreAllocateWWMRegs can tell when
2233     // Whole Wave Mode is entered.
2234     MI.setDesc(get(ST.isWave32() ? AMDGPU::S_OR_SAVEEXEC_B32
2235                                  : AMDGPU::S_OR_SAVEEXEC_B64));
2236     break;
2237   }
2238   case AMDGPU::ENTER_STRICT_WQM: {
2239     // This only gets its own opcode so that SIPreAllocateWWMRegs can tell when
2240     // STRICT_WQM is entered.
2241     const unsigned Exec = ST.isWave32() ? AMDGPU::EXEC_LO : AMDGPU::EXEC;
2242     const unsigned WQMOp = ST.isWave32() ? AMDGPU::S_WQM_B32 : AMDGPU::S_WQM_B64;
2243     const unsigned MovOp = ST.isWave32() ? AMDGPU::S_MOV_B32 : AMDGPU::S_MOV_B64;
2244     BuildMI(MBB, MI, DL, get(MovOp), MI.getOperand(0).getReg()).addReg(Exec);
2245     BuildMI(MBB, MI, DL, get(WQMOp), Exec).addReg(Exec);
2246 
2247     MI.eraseFromParent();
2248     break;
2249   }
2250   case AMDGPU::EXIT_STRICT_WWM:
2251   case AMDGPU::EXIT_STRICT_WQM: {
2252     // This only gets its own opcode so that SIPreAllocateWWMRegs can tell when
2253     // WWM/STICT_WQM is exited.
2254     MI.setDesc(get(ST.isWave32() ? AMDGPU::S_MOV_B32 : AMDGPU::S_MOV_B64));
2255     break;
2256   }
2257   case AMDGPU::ENTER_PSEUDO_WM:
2258   case AMDGPU::EXIT_PSEUDO_WM: {
2259     // These do nothing.
2260     MI.eraseFromParent();
2261     break;
2262   }
2263   case AMDGPU::SI_RETURN: {
2264     const MachineFunction *MF = MBB.getParent();
2265     const GCNSubtarget &ST = MF->getSubtarget<GCNSubtarget>();
2266     const SIRegisterInfo *TRI = ST.getRegisterInfo();
2267     // Hiding the return address use with SI_RETURN may lead to extra kills in
2268     // the function and missing live-ins. We are fine in practice because callee
2269     // saved register handling ensures the register value is restored before
2270     // RET, but we need the undef flag here to appease the MachineVerifier
2271     // liveness checks.
2272     MachineInstrBuilder MIB =
2273         BuildMI(MBB, MI, DL, get(AMDGPU::S_SETPC_B64_return))
2274             .addReg(TRI->getReturnAddressReg(*MF), RegState::Undef);
2275 
2276     MIB.copyImplicitOps(MI);
2277     MI.eraseFromParent();
2278     break;
2279   }
2280   }
2281   return true;
2282 }
2283 
2284 std::pair<MachineInstr*, MachineInstr*>
2285 SIInstrInfo::expandMovDPP64(MachineInstr &MI) const {
2286   assert (MI.getOpcode() == AMDGPU::V_MOV_B64_DPP_PSEUDO);
2287 
2288   if (ST.hasMovB64() &&
2289       AMDGPU::isLegal64BitDPPControl(
2290         getNamedOperand(MI, AMDGPU::OpName::dpp_ctrl)->getImm())) {
2291     MI.setDesc(get(AMDGPU::V_MOV_B64_dpp));
2292     return std::pair(&MI, nullptr);
2293   }
2294 
2295   MachineBasicBlock &MBB = *MI.getParent();
2296   DebugLoc DL = MBB.findDebugLoc(MI);
2297   MachineFunction *MF = MBB.getParent();
2298   MachineRegisterInfo &MRI = MF->getRegInfo();
2299   Register Dst = MI.getOperand(0).getReg();
2300   unsigned Part = 0;
2301   MachineInstr *Split[2];
2302 
2303   for (auto Sub : { AMDGPU::sub0, AMDGPU::sub1 }) {
2304     auto MovDPP = BuildMI(MBB, MI, DL, get(AMDGPU::V_MOV_B32_dpp));
2305     if (Dst.isPhysical()) {
2306       MovDPP.addDef(RI.getSubReg(Dst, Sub));
2307     } else {
2308       assert(MRI.isSSA());
2309       auto Tmp = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
2310       MovDPP.addDef(Tmp);
2311     }
2312 
2313     for (unsigned I = 1; I <= 2; ++I) { // old and src operands.
2314       const MachineOperand &SrcOp = MI.getOperand(I);
2315       assert(!SrcOp.isFPImm());
2316       if (SrcOp.isImm()) {
2317         APInt Imm(64, SrcOp.getImm());
2318         Imm.ashrInPlace(Part * 32);
2319         MovDPP.addImm(Imm.getLoBits(32).getZExtValue());
2320       } else {
2321         assert(SrcOp.isReg());
2322         Register Src = SrcOp.getReg();
2323         if (Src.isPhysical())
2324           MovDPP.addReg(RI.getSubReg(Src, Sub));
2325         else
2326           MovDPP.addReg(Src, SrcOp.isUndef() ? RegState::Undef : 0, Sub);
2327       }
2328     }
2329 
2330     for (const MachineOperand &MO : llvm::drop_begin(MI.explicit_operands(), 3))
2331       MovDPP.addImm(MO.getImm());
2332 
2333     Split[Part] = MovDPP;
2334     ++Part;
2335   }
2336 
2337   if (Dst.isVirtual())
2338     BuildMI(MBB, MI, DL, get(AMDGPU::REG_SEQUENCE), Dst)
2339       .addReg(Split[0]->getOperand(0).getReg())
2340       .addImm(AMDGPU::sub0)
2341       .addReg(Split[1]->getOperand(0).getReg())
2342       .addImm(AMDGPU::sub1);
2343 
2344   MI.eraseFromParent();
2345   return std::pair(Split[0], Split[1]);
2346 }
2347 
2348 bool SIInstrInfo::swapSourceModifiers(MachineInstr &MI,
2349                                       MachineOperand &Src0,
2350                                       unsigned Src0OpName,
2351                                       MachineOperand &Src1,
2352                                       unsigned Src1OpName) const {
2353   MachineOperand *Src0Mods = getNamedOperand(MI, Src0OpName);
2354   if (!Src0Mods)
2355     return false;
2356 
2357   MachineOperand *Src1Mods = getNamedOperand(MI, Src1OpName);
2358   assert(Src1Mods &&
2359          "All commutable instructions have both src0 and src1 modifiers");
2360 
2361   int Src0ModsVal = Src0Mods->getImm();
2362   int Src1ModsVal = Src1Mods->getImm();
2363 
2364   Src1Mods->setImm(Src0ModsVal);
2365   Src0Mods->setImm(Src1ModsVal);
2366   return true;
2367 }
2368 
2369 static MachineInstr *swapRegAndNonRegOperand(MachineInstr &MI,
2370                                              MachineOperand &RegOp,
2371                                              MachineOperand &NonRegOp) {
2372   Register Reg = RegOp.getReg();
2373   unsigned SubReg = RegOp.getSubReg();
2374   bool IsKill = RegOp.isKill();
2375   bool IsDead = RegOp.isDead();
2376   bool IsUndef = RegOp.isUndef();
2377   bool IsDebug = RegOp.isDebug();
2378 
2379   if (NonRegOp.isImm())
2380     RegOp.ChangeToImmediate(NonRegOp.getImm());
2381   else if (NonRegOp.isFI())
2382     RegOp.ChangeToFrameIndex(NonRegOp.getIndex());
2383   else if (NonRegOp.isGlobal()) {
2384     RegOp.ChangeToGA(NonRegOp.getGlobal(), NonRegOp.getOffset(),
2385                      NonRegOp.getTargetFlags());
2386   } else
2387     return nullptr;
2388 
2389   // Make sure we don't reinterpret a subreg index in the target flags.
2390   RegOp.setTargetFlags(NonRegOp.getTargetFlags());
2391 
2392   NonRegOp.ChangeToRegister(Reg, false, false, IsKill, IsDead, IsUndef, IsDebug);
2393   NonRegOp.setSubReg(SubReg);
2394 
2395   return &MI;
2396 }
2397 
2398 MachineInstr *SIInstrInfo::commuteInstructionImpl(MachineInstr &MI, bool NewMI,
2399                                                   unsigned Src0Idx,
2400                                                   unsigned Src1Idx) const {
2401   assert(!NewMI && "this should never be used");
2402 
2403   unsigned Opc = MI.getOpcode();
2404   int CommutedOpcode = commuteOpcode(Opc);
2405   if (CommutedOpcode == -1)
2406     return nullptr;
2407 
2408   assert(AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0) ==
2409            static_cast<int>(Src0Idx) &&
2410          AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src1) ==
2411            static_cast<int>(Src1Idx) &&
2412          "inconsistency with findCommutedOpIndices");
2413 
2414   MachineOperand &Src0 = MI.getOperand(Src0Idx);
2415   MachineOperand &Src1 = MI.getOperand(Src1Idx);
2416 
2417   MachineInstr *CommutedMI = nullptr;
2418   if (Src0.isReg() && Src1.isReg()) {
2419     if (isOperandLegal(MI, Src1Idx, &Src0)) {
2420       // Be sure to copy the source modifiers to the right place.
2421       CommutedMI
2422         = TargetInstrInfo::commuteInstructionImpl(MI, NewMI, Src0Idx, Src1Idx);
2423     }
2424 
2425   } else if (Src0.isReg() && !Src1.isReg()) {
2426     // src0 should always be able to support any operand type, so no need to
2427     // check operand legality.
2428     CommutedMI = swapRegAndNonRegOperand(MI, Src0, Src1);
2429   } else if (!Src0.isReg() && Src1.isReg()) {
2430     if (isOperandLegal(MI, Src1Idx, &Src0))
2431       CommutedMI = swapRegAndNonRegOperand(MI, Src1, Src0);
2432   } else {
2433     // FIXME: Found two non registers to commute. This does happen.
2434     return nullptr;
2435   }
2436 
2437   if (CommutedMI) {
2438     swapSourceModifiers(MI, Src0, AMDGPU::OpName::src0_modifiers,
2439                         Src1, AMDGPU::OpName::src1_modifiers);
2440 
2441     CommutedMI->setDesc(get(CommutedOpcode));
2442   }
2443 
2444   return CommutedMI;
2445 }
2446 
2447 // This needs to be implemented because the source modifiers may be inserted
2448 // between the true commutable operands, and the base
2449 // TargetInstrInfo::commuteInstruction uses it.
2450 bool SIInstrInfo::findCommutedOpIndices(const MachineInstr &MI,
2451                                         unsigned &SrcOpIdx0,
2452                                         unsigned &SrcOpIdx1) const {
2453   return findCommutedOpIndices(MI.getDesc(), SrcOpIdx0, SrcOpIdx1);
2454 }
2455 
2456 bool SIInstrInfo::findCommutedOpIndices(const MCInstrDesc &Desc,
2457                                         unsigned &SrcOpIdx0,
2458                                         unsigned &SrcOpIdx1) const {
2459   if (!Desc.isCommutable())
2460     return false;
2461 
2462   unsigned Opc = Desc.getOpcode();
2463   int Src0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0);
2464   if (Src0Idx == -1)
2465     return false;
2466 
2467   int Src1Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src1);
2468   if (Src1Idx == -1)
2469     return false;
2470 
2471   return fixCommutedOpIndices(SrcOpIdx0, SrcOpIdx1, Src0Idx, Src1Idx);
2472 }
2473 
2474 bool SIInstrInfo::isBranchOffsetInRange(unsigned BranchOp,
2475                                         int64_t BrOffset) const {
2476   // BranchRelaxation should never have to check s_setpc_b64 because its dest
2477   // block is unanalyzable.
2478   assert(BranchOp != AMDGPU::S_SETPC_B64);
2479 
2480   // Convert to dwords.
2481   BrOffset /= 4;
2482 
2483   // The branch instructions do PC += signext(SIMM16 * 4) + 4, so the offset is
2484   // from the next instruction.
2485   BrOffset -= 1;
2486 
2487   return isIntN(BranchOffsetBits, BrOffset);
2488 }
2489 
2490 MachineBasicBlock *SIInstrInfo::getBranchDestBlock(
2491   const MachineInstr &MI) const {
2492   if (MI.getOpcode() == AMDGPU::S_SETPC_B64) {
2493     // This would be a difficult analysis to perform, but can always be legal so
2494     // there's no need to analyze it.
2495     return nullptr;
2496   }
2497 
2498   return MI.getOperand(0).getMBB();
2499 }
2500 
2501 bool SIInstrInfo::hasDivergentBranch(const MachineBasicBlock *MBB) const {
2502   for (const MachineInstr &MI : MBB->terminators()) {
2503     if (MI.getOpcode() == AMDGPU::SI_NON_UNIFORM_BRCOND_PSEUDO ||
2504         MI.getOpcode() == AMDGPU::SI_IF || MI.getOpcode() == AMDGPU::SI_ELSE ||
2505         MI.getOpcode() == AMDGPU::SI_LOOP)
2506       return true;
2507   }
2508   return false;
2509 }
2510 
2511 void SIInstrInfo::insertIndirectBranch(MachineBasicBlock &MBB,
2512                                        MachineBasicBlock &DestBB,
2513                                        MachineBasicBlock &RestoreBB,
2514                                        const DebugLoc &DL, int64_t BrOffset,
2515                                        RegScavenger *RS) const {
2516   assert(RS && "RegScavenger required for long branching");
2517   assert(MBB.empty() &&
2518          "new block should be inserted for expanding unconditional branch");
2519   assert(MBB.pred_size() == 1);
2520   assert(RestoreBB.empty() &&
2521          "restore block should be inserted for restoring clobbered registers");
2522 
2523   MachineFunction *MF = MBB.getParent();
2524   MachineRegisterInfo &MRI = MF->getRegInfo();
2525 
2526   // FIXME: Virtual register workaround for RegScavenger not working with empty
2527   // blocks.
2528   Register PCReg = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
2529 
2530   auto I = MBB.end();
2531 
2532   // We need to compute the offset relative to the instruction immediately after
2533   // s_getpc_b64. Insert pc arithmetic code before last terminator.
2534   MachineInstr *GetPC = BuildMI(MBB, I, DL, get(AMDGPU::S_GETPC_B64), PCReg);
2535 
2536   auto &MCCtx = MF->getContext();
2537   MCSymbol *PostGetPCLabel =
2538       MCCtx.createTempSymbol("post_getpc", /*AlwaysAddSuffix=*/true);
2539   GetPC->setPostInstrSymbol(*MF, PostGetPCLabel);
2540 
2541   MCSymbol *OffsetLo =
2542       MCCtx.createTempSymbol("offset_lo", /*AlwaysAddSuffix=*/true);
2543   MCSymbol *OffsetHi =
2544       MCCtx.createTempSymbol("offset_hi", /*AlwaysAddSuffix=*/true);
2545   BuildMI(MBB, I, DL, get(AMDGPU::S_ADD_U32))
2546       .addReg(PCReg, RegState::Define, AMDGPU::sub0)
2547       .addReg(PCReg, 0, AMDGPU::sub0)
2548       .addSym(OffsetLo, MO_FAR_BRANCH_OFFSET);
2549   BuildMI(MBB, I, DL, get(AMDGPU::S_ADDC_U32))
2550       .addReg(PCReg, RegState::Define, AMDGPU::sub1)
2551       .addReg(PCReg, 0, AMDGPU::sub1)
2552       .addSym(OffsetHi, MO_FAR_BRANCH_OFFSET);
2553 
2554   // Insert the indirect branch after the other terminator.
2555   BuildMI(&MBB, DL, get(AMDGPU::S_SETPC_B64))
2556     .addReg(PCReg);
2557 
2558   // FIXME: If spilling is necessary, this will fail because this scavenger has
2559   // no emergency stack slots. It is non-trivial to spill in this situation,
2560   // because the restore code needs to be specially placed after the
2561   // jump. BranchRelaxation then needs to be made aware of the newly inserted
2562   // block.
2563   //
2564   // If a spill is needed for the pc register pair, we need to insert a spill
2565   // restore block right before the destination block, and insert a short branch
2566   // into the old destination block's fallthrough predecessor.
2567   // e.g.:
2568   //
2569   // s_cbranch_scc0 skip_long_branch:
2570   //
2571   // long_branch_bb:
2572   //   spill s[8:9]
2573   //   s_getpc_b64 s[8:9]
2574   //   s_add_u32 s8, s8, restore_bb
2575   //   s_addc_u32 s9, s9, 0
2576   //   s_setpc_b64 s[8:9]
2577   //
2578   // skip_long_branch:
2579   //   foo;
2580   //
2581   // .....
2582   //
2583   // dest_bb_fallthrough_predecessor:
2584   // bar;
2585   // s_branch dest_bb
2586   //
2587   // restore_bb:
2588   //  restore s[8:9]
2589   //  fallthrough dest_bb
2590   ///
2591   // dest_bb:
2592   //   buzz;
2593 
2594   RS->enterBasicBlockEnd(MBB);
2595   Register Scav = RS->scavengeRegisterBackwards(
2596       AMDGPU::SReg_64RegClass, MachineBasicBlock::iterator(GetPC),
2597       /* RestoreAfter */ false, 0, /* AllowSpill */ false);
2598   if (Scav) {
2599     RS->setRegUsed(Scav);
2600     MRI.replaceRegWith(PCReg, Scav);
2601     MRI.clearVirtRegs();
2602   } else {
2603     // As SGPR needs VGPR to be spilled, we reuse the slot of temporary VGPR for
2604     // SGPR spill.
2605     const GCNSubtarget &ST = MF->getSubtarget<GCNSubtarget>();
2606     const SIRegisterInfo *TRI = ST.getRegisterInfo();
2607     TRI->spillEmergencySGPR(GetPC, RestoreBB, AMDGPU::SGPR0_SGPR1, RS);
2608     MRI.replaceRegWith(PCReg, AMDGPU::SGPR0_SGPR1);
2609     MRI.clearVirtRegs();
2610   }
2611 
2612   MCSymbol *DestLabel = Scav ? DestBB.getSymbol() : RestoreBB.getSymbol();
2613   // Now, the distance could be defined.
2614   auto *Offset = MCBinaryExpr::createSub(
2615       MCSymbolRefExpr::create(DestLabel, MCCtx),
2616       MCSymbolRefExpr::create(PostGetPCLabel, MCCtx), MCCtx);
2617   // Add offset assignments.
2618   auto *Mask = MCConstantExpr::create(0xFFFFFFFFULL, MCCtx);
2619   OffsetLo->setVariableValue(MCBinaryExpr::createAnd(Offset, Mask, MCCtx));
2620   auto *ShAmt = MCConstantExpr::create(32, MCCtx);
2621   OffsetHi->setVariableValue(MCBinaryExpr::createAShr(Offset, ShAmt, MCCtx));
2622 }
2623 
2624 unsigned SIInstrInfo::getBranchOpcode(SIInstrInfo::BranchPredicate Cond) {
2625   switch (Cond) {
2626   case SIInstrInfo::SCC_TRUE:
2627     return AMDGPU::S_CBRANCH_SCC1;
2628   case SIInstrInfo::SCC_FALSE:
2629     return AMDGPU::S_CBRANCH_SCC0;
2630   case SIInstrInfo::VCCNZ:
2631     return AMDGPU::S_CBRANCH_VCCNZ;
2632   case SIInstrInfo::VCCZ:
2633     return AMDGPU::S_CBRANCH_VCCZ;
2634   case SIInstrInfo::EXECNZ:
2635     return AMDGPU::S_CBRANCH_EXECNZ;
2636   case SIInstrInfo::EXECZ:
2637     return AMDGPU::S_CBRANCH_EXECZ;
2638   default:
2639     llvm_unreachable("invalid branch predicate");
2640   }
2641 }
2642 
2643 SIInstrInfo::BranchPredicate SIInstrInfo::getBranchPredicate(unsigned Opcode) {
2644   switch (Opcode) {
2645   case AMDGPU::S_CBRANCH_SCC0:
2646     return SCC_FALSE;
2647   case AMDGPU::S_CBRANCH_SCC1:
2648     return SCC_TRUE;
2649   case AMDGPU::S_CBRANCH_VCCNZ:
2650     return VCCNZ;
2651   case AMDGPU::S_CBRANCH_VCCZ:
2652     return VCCZ;
2653   case AMDGPU::S_CBRANCH_EXECNZ:
2654     return EXECNZ;
2655   case AMDGPU::S_CBRANCH_EXECZ:
2656     return EXECZ;
2657   default:
2658     return INVALID_BR;
2659   }
2660 }
2661 
2662 bool SIInstrInfo::analyzeBranchImpl(MachineBasicBlock &MBB,
2663                                     MachineBasicBlock::iterator I,
2664                                     MachineBasicBlock *&TBB,
2665                                     MachineBasicBlock *&FBB,
2666                                     SmallVectorImpl<MachineOperand> &Cond,
2667                                     bool AllowModify) const {
2668   if (I->getOpcode() == AMDGPU::S_BRANCH) {
2669     // Unconditional Branch
2670     TBB = I->getOperand(0).getMBB();
2671     return false;
2672   }
2673 
2674   MachineBasicBlock *CondBB = nullptr;
2675 
2676   if (I->getOpcode() == AMDGPU::SI_NON_UNIFORM_BRCOND_PSEUDO) {
2677     CondBB = I->getOperand(1).getMBB();
2678     Cond.push_back(I->getOperand(0));
2679   } else {
2680     BranchPredicate Pred = getBranchPredicate(I->getOpcode());
2681     if (Pred == INVALID_BR)
2682       return true;
2683 
2684     CondBB = I->getOperand(0).getMBB();
2685     Cond.push_back(MachineOperand::CreateImm(Pred));
2686     Cond.push_back(I->getOperand(1)); // Save the branch register.
2687   }
2688   ++I;
2689 
2690   if (I == MBB.end()) {
2691     // Conditional branch followed by fall-through.
2692     TBB = CondBB;
2693     return false;
2694   }
2695 
2696   if (I->getOpcode() == AMDGPU::S_BRANCH) {
2697     TBB = CondBB;
2698     FBB = I->getOperand(0).getMBB();
2699     return false;
2700   }
2701 
2702   return true;
2703 }
2704 
2705 bool SIInstrInfo::analyzeBranch(MachineBasicBlock &MBB, MachineBasicBlock *&TBB,
2706                                 MachineBasicBlock *&FBB,
2707                                 SmallVectorImpl<MachineOperand> &Cond,
2708                                 bool AllowModify) const {
2709   MachineBasicBlock::iterator I = MBB.getFirstTerminator();
2710   auto E = MBB.end();
2711   if (I == E)
2712     return false;
2713 
2714   // Skip over the instructions that are artificially terminators for special
2715   // exec management.
2716   while (I != E && !I->isBranch() && !I->isReturn()) {
2717     switch (I->getOpcode()) {
2718     case AMDGPU::S_MOV_B64_term:
2719     case AMDGPU::S_XOR_B64_term:
2720     case AMDGPU::S_OR_B64_term:
2721     case AMDGPU::S_ANDN2_B64_term:
2722     case AMDGPU::S_AND_B64_term:
2723     case AMDGPU::S_MOV_B32_term:
2724     case AMDGPU::S_XOR_B32_term:
2725     case AMDGPU::S_OR_B32_term:
2726     case AMDGPU::S_ANDN2_B32_term:
2727     case AMDGPU::S_AND_B32_term:
2728       break;
2729     case AMDGPU::SI_IF:
2730     case AMDGPU::SI_ELSE:
2731     case AMDGPU::SI_KILL_I1_TERMINATOR:
2732     case AMDGPU::SI_KILL_F32_COND_IMM_TERMINATOR:
2733       // FIXME: It's messy that these need to be considered here at all.
2734       return true;
2735     default:
2736       llvm_unreachable("unexpected non-branch terminator inst");
2737     }
2738 
2739     ++I;
2740   }
2741 
2742   if (I == E)
2743     return false;
2744 
2745   return analyzeBranchImpl(MBB, I, TBB, FBB, Cond, AllowModify);
2746 }
2747 
2748 unsigned SIInstrInfo::removeBranch(MachineBasicBlock &MBB,
2749                                    int *BytesRemoved) const {
2750   unsigned Count = 0;
2751   unsigned RemovedSize = 0;
2752   for (MachineInstr &MI : llvm::make_early_inc_range(MBB.terminators())) {
2753     // Skip over artificial terminators when removing instructions.
2754     if (MI.isBranch() || MI.isReturn()) {
2755       RemovedSize += getInstSizeInBytes(MI);
2756       MI.eraseFromParent();
2757       ++Count;
2758     }
2759   }
2760 
2761   if (BytesRemoved)
2762     *BytesRemoved = RemovedSize;
2763 
2764   return Count;
2765 }
2766 
2767 // Copy the flags onto the implicit condition register operand.
2768 static void preserveCondRegFlags(MachineOperand &CondReg,
2769                                  const MachineOperand &OrigCond) {
2770   CondReg.setIsUndef(OrigCond.isUndef());
2771   CondReg.setIsKill(OrigCond.isKill());
2772 }
2773 
2774 unsigned SIInstrInfo::insertBranch(MachineBasicBlock &MBB,
2775                                    MachineBasicBlock *TBB,
2776                                    MachineBasicBlock *FBB,
2777                                    ArrayRef<MachineOperand> Cond,
2778                                    const DebugLoc &DL,
2779                                    int *BytesAdded) const {
2780   if (!FBB && Cond.empty()) {
2781     BuildMI(&MBB, DL, get(AMDGPU::S_BRANCH))
2782       .addMBB(TBB);
2783     if (BytesAdded)
2784       *BytesAdded = ST.hasOffset3fBug() ? 8 : 4;
2785     return 1;
2786   }
2787 
2788   if(Cond.size() == 1 && Cond[0].isReg()) {
2789      BuildMI(&MBB, DL, get(AMDGPU::SI_NON_UNIFORM_BRCOND_PSEUDO))
2790        .add(Cond[0])
2791        .addMBB(TBB);
2792      return 1;
2793   }
2794 
2795   assert(TBB && Cond[0].isImm());
2796 
2797   unsigned Opcode
2798     = getBranchOpcode(static_cast<BranchPredicate>(Cond[0].getImm()));
2799 
2800   if (!FBB) {
2801     Cond[1].isUndef();
2802     MachineInstr *CondBr =
2803       BuildMI(&MBB, DL, get(Opcode))
2804       .addMBB(TBB);
2805 
2806     // Copy the flags onto the implicit condition register operand.
2807     preserveCondRegFlags(CondBr->getOperand(1), Cond[1]);
2808     fixImplicitOperands(*CondBr);
2809 
2810     if (BytesAdded)
2811       *BytesAdded = ST.hasOffset3fBug() ? 8 : 4;
2812     return 1;
2813   }
2814 
2815   assert(TBB && FBB);
2816 
2817   MachineInstr *CondBr =
2818     BuildMI(&MBB, DL, get(Opcode))
2819     .addMBB(TBB);
2820   fixImplicitOperands(*CondBr);
2821   BuildMI(&MBB, DL, get(AMDGPU::S_BRANCH))
2822     .addMBB(FBB);
2823 
2824   MachineOperand &CondReg = CondBr->getOperand(1);
2825   CondReg.setIsUndef(Cond[1].isUndef());
2826   CondReg.setIsKill(Cond[1].isKill());
2827 
2828   if (BytesAdded)
2829     *BytesAdded = ST.hasOffset3fBug() ? 16 : 8;
2830 
2831   return 2;
2832 }
2833 
2834 bool SIInstrInfo::reverseBranchCondition(
2835   SmallVectorImpl<MachineOperand> &Cond) const {
2836   if (Cond.size() != 2) {
2837     return true;
2838   }
2839 
2840   if (Cond[0].isImm()) {
2841     Cond[0].setImm(-Cond[0].getImm());
2842     return false;
2843   }
2844 
2845   return true;
2846 }
2847 
2848 bool SIInstrInfo::canInsertSelect(const MachineBasicBlock &MBB,
2849                                   ArrayRef<MachineOperand> Cond,
2850                                   Register DstReg, Register TrueReg,
2851                                   Register FalseReg, int &CondCycles,
2852                                   int &TrueCycles, int &FalseCycles) const {
2853   switch (Cond[0].getImm()) {
2854   case VCCNZ:
2855   case VCCZ: {
2856     const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2857     const TargetRegisterClass *RC = MRI.getRegClass(TrueReg);
2858     if (MRI.getRegClass(FalseReg) != RC)
2859       return false;
2860 
2861     int NumInsts = AMDGPU::getRegBitWidth(RC->getID()) / 32;
2862     CondCycles = TrueCycles = FalseCycles = NumInsts; // ???
2863 
2864     // Limit to equal cost for branch vs. N v_cndmask_b32s.
2865     return RI.hasVGPRs(RC) && NumInsts <= 6;
2866   }
2867   case SCC_TRUE:
2868   case SCC_FALSE: {
2869     // FIXME: We could insert for VGPRs if we could replace the original compare
2870     // with a vector one.
2871     const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2872     const TargetRegisterClass *RC = MRI.getRegClass(TrueReg);
2873     if (MRI.getRegClass(FalseReg) != RC)
2874       return false;
2875 
2876     int NumInsts = AMDGPU::getRegBitWidth(RC->getID()) / 32;
2877 
2878     // Multiples of 8 can do s_cselect_b64
2879     if (NumInsts % 2 == 0)
2880       NumInsts /= 2;
2881 
2882     CondCycles = TrueCycles = FalseCycles = NumInsts; // ???
2883     return RI.isSGPRClass(RC);
2884   }
2885   default:
2886     return false;
2887   }
2888 }
2889 
2890 void SIInstrInfo::insertSelect(MachineBasicBlock &MBB,
2891                                MachineBasicBlock::iterator I, const DebugLoc &DL,
2892                                Register DstReg, ArrayRef<MachineOperand> Cond,
2893                                Register TrueReg, Register FalseReg) const {
2894   BranchPredicate Pred = static_cast<BranchPredicate>(Cond[0].getImm());
2895   if (Pred == VCCZ || Pred == SCC_FALSE) {
2896     Pred = static_cast<BranchPredicate>(-Pred);
2897     std::swap(TrueReg, FalseReg);
2898   }
2899 
2900   MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
2901   const TargetRegisterClass *DstRC = MRI.getRegClass(DstReg);
2902   unsigned DstSize = RI.getRegSizeInBits(*DstRC);
2903 
2904   if (DstSize == 32) {
2905     MachineInstr *Select;
2906     if (Pred == SCC_TRUE) {
2907       Select = BuildMI(MBB, I, DL, get(AMDGPU::S_CSELECT_B32), DstReg)
2908         .addReg(TrueReg)
2909         .addReg(FalseReg);
2910     } else {
2911       // Instruction's operands are backwards from what is expected.
2912       Select = BuildMI(MBB, I, DL, get(AMDGPU::V_CNDMASK_B32_e32), DstReg)
2913         .addReg(FalseReg)
2914         .addReg(TrueReg);
2915     }
2916 
2917     preserveCondRegFlags(Select->getOperand(3), Cond[1]);
2918     return;
2919   }
2920 
2921   if (DstSize == 64 && Pred == SCC_TRUE) {
2922     MachineInstr *Select =
2923       BuildMI(MBB, I, DL, get(AMDGPU::S_CSELECT_B64), DstReg)
2924       .addReg(TrueReg)
2925       .addReg(FalseReg);
2926 
2927     preserveCondRegFlags(Select->getOperand(3), Cond[1]);
2928     return;
2929   }
2930 
2931   static const int16_t Sub0_15[] = {
2932     AMDGPU::sub0, AMDGPU::sub1, AMDGPU::sub2, AMDGPU::sub3,
2933     AMDGPU::sub4, AMDGPU::sub5, AMDGPU::sub6, AMDGPU::sub7,
2934     AMDGPU::sub8, AMDGPU::sub9, AMDGPU::sub10, AMDGPU::sub11,
2935     AMDGPU::sub12, AMDGPU::sub13, AMDGPU::sub14, AMDGPU::sub15,
2936   };
2937 
2938   static const int16_t Sub0_15_64[] = {
2939     AMDGPU::sub0_sub1, AMDGPU::sub2_sub3,
2940     AMDGPU::sub4_sub5, AMDGPU::sub6_sub7,
2941     AMDGPU::sub8_sub9, AMDGPU::sub10_sub11,
2942     AMDGPU::sub12_sub13, AMDGPU::sub14_sub15,
2943   };
2944 
2945   unsigned SelOp = AMDGPU::V_CNDMASK_B32_e32;
2946   const TargetRegisterClass *EltRC = &AMDGPU::VGPR_32RegClass;
2947   const int16_t *SubIndices = Sub0_15;
2948   int NElts = DstSize / 32;
2949 
2950   // 64-bit select is only available for SALU.
2951   // TODO: Split 96-bit into 64-bit and 32-bit, not 3x 32-bit.
2952   if (Pred == SCC_TRUE) {
2953     if (NElts % 2) {
2954       SelOp = AMDGPU::S_CSELECT_B32;
2955       EltRC = &AMDGPU::SGPR_32RegClass;
2956     } else {
2957       SelOp = AMDGPU::S_CSELECT_B64;
2958       EltRC = &AMDGPU::SGPR_64RegClass;
2959       SubIndices = Sub0_15_64;
2960       NElts /= 2;
2961     }
2962   }
2963 
2964   MachineInstrBuilder MIB = BuildMI(
2965     MBB, I, DL, get(AMDGPU::REG_SEQUENCE), DstReg);
2966 
2967   I = MIB->getIterator();
2968 
2969   SmallVector<Register, 8> Regs;
2970   for (int Idx = 0; Idx != NElts; ++Idx) {
2971     Register DstElt = MRI.createVirtualRegister(EltRC);
2972     Regs.push_back(DstElt);
2973 
2974     unsigned SubIdx = SubIndices[Idx];
2975 
2976     MachineInstr *Select;
2977     if (SelOp == AMDGPU::V_CNDMASK_B32_e32) {
2978       Select =
2979         BuildMI(MBB, I, DL, get(SelOp), DstElt)
2980         .addReg(FalseReg, 0, SubIdx)
2981         .addReg(TrueReg, 0, SubIdx);
2982     } else {
2983       Select =
2984         BuildMI(MBB, I, DL, get(SelOp), DstElt)
2985         .addReg(TrueReg, 0, SubIdx)
2986         .addReg(FalseReg, 0, SubIdx);
2987     }
2988 
2989     preserveCondRegFlags(Select->getOperand(3), Cond[1]);
2990     fixImplicitOperands(*Select);
2991 
2992     MIB.addReg(DstElt)
2993        .addImm(SubIdx);
2994   }
2995 }
2996 
2997 bool SIInstrInfo::isFoldableCopy(const MachineInstr &MI) {
2998   switch (MI.getOpcode()) {
2999   case AMDGPU::V_MOV_B32_e32:
3000   case AMDGPU::V_MOV_B32_e64:
3001   case AMDGPU::V_MOV_B64_PSEUDO:
3002   case AMDGPU::V_MOV_B64_e32:
3003   case AMDGPU::V_MOV_B64_e64:
3004   case AMDGPU::S_MOV_B32:
3005   case AMDGPU::S_MOV_B64:
3006   case AMDGPU::COPY:
3007   case AMDGPU::V_ACCVGPR_WRITE_B32_e64:
3008   case AMDGPU::V_ACCVGPR_READ_B32_e64:
3009   case AMDGPU::V_ACCVGPR_MOV_B32:
3010     return true;
3011   default:
3012     return false;
3013   }
3014 }
3015 
3016 static constexpr unsigned ModifierOpNames[] = {
3017     AMDGPU::OpName::src0_modifiers, AMDGPU::OpName::src1_modifiers,
3018     AMDGPU::OpName::src2_modifiers, AMDGPU::OpName::clamp,
3019     AMDGPU::OpName::omod,           AMDGPU::OpName::op_sel};
3020 
3021 void SIInstrInfo::removeModOperands(MachineInstr &MI) const {
3022   unsigned Opc = MI.getOpcode();
3023   for (unsigned Name : reverse(ModifierOpNames)) {
3024     int Idx = AMDGPU::getNamedOperandIdx(Opc, Name);
3025     if (Idx >= 0)
3026       MI.removeOperand(Idx);
3027   }
3028 }
3029 
3030 bool SIInstrInfo::FoldImmediate(MachineInstr &UseMI, MachineInstr &DefMI,
3031                                 Register Reg, MachineRegisterInfo *MRI) const {
3032   if (!MRI->hasOneNonDBGUse(Reg))
3033     return false;
3034 
3035   switch (DefMI.getOpcode()) {
3036   default:
3037     return false;
3038   case AMDGPU::S_MOV_B64:
3039     // TODO: We could fold 64-bit immediates, but this get complicated
3040     // when there are sub-registers.
3041     return false;
3042 
3043   case AMDGPU::V_MOV_B32_e32:
3044   case AMDGPU::S_MOV_B32:
3045   case AMDGPU::V_ACCVGPR_WRITE_B32_e64:
3046     break;
3047   }
3048 
3049   const MachineOperand *ImmOp = getNamedOperand(DefMI, AMDGPU::OpName::src0);
3050   assert(ImmOp);
3051   // FIXME: We could handle FrameIndex values here.
3052   if (!ImmOp->isImm())
3053     return false;
3054 
3055   unsigned Opc = UseMI.getOpcode();
3056   if (Opc == AMDGPU::COPY) {
3057     Register DstReg = UseMI.getOperand(0).getReg();
3058     bool Is16Bit = getOpSize(UseMI, 0) == 2;
3059     bool isVGPRCopy = RI.isVGPR(*MRI, DstReg);
3060     unsigned NewOpc = isVGPRCopy ? AMDGPU::V_MOV_B32_e32 : AMDGPU::S_MOV_B32;
3061     APInt Imm(32, ImmOp->getImm());
3062 
3063     if (UseMI.getOperand(1).getSubReg() == AMDGPU::hi16)
3064       Imm = Imm.ashr(16);
3065 
3066     if (RI.isAGPR(*MRI, DstReg)) {
3067       if (!isInlineConstant(Imm))
3068         return false;
3069       NewOpc = AMDGPU::V_ACCVGPR_WRITE_B32_e64;
3070     }
3071 
3072     if (Is16Bit) {
3073       if (isVGPRCopy)
3074         return false; // Do not clobber vgpr_hi16
3075 
3076       if (DstReg.isVirtual() && UseMI.getOperand(0).getSubReg() != AMDGPU::lo16)
3077         return false;
3078 
3079       UseMI.getOperand(0).setSubReg(0);
3080       if (DstReg.isPhysical()) {
3081         DstReg = RI.get32BitRegister(DstReg);
3082         UseMI.getOperand(0).setReg(DstReg);
3083       }
3084       assert(UseMI.getOperand(1).getReg().isVirtual());
3085     }
3086 
3087     UseMI.setDesc(get(NewOpc));
3088     UseMI.getOperand(1).ChangeToImmediate(Imm.getSExtValue());
3089     UseMI.addImplicitDefUseOperands(*UseMI.getParent()->getParent());
3090     return true;
3091   }
3092 
3093   if (Opc == AMDGPU::V_MAD_F32_e64 || Opc == AMDGPU::V_MAC_F32_e64 ||
3094       Opc == AMDGPU::V_MAD_F16_e64 || Opc == AMDGPU::V_MAC_F16_e64 ||
3095       Opc == AMDGPU::V_FMA_F32_e64 || Opc == AMDGPU::V_FMAC_F32_e64 ||
3096       Opc == AMDGPU::V_FMA_F16_e64 || Opc == AMDGPU::V_FMAC_F16_e64 ||
3097       Opc == AMDGPU::V_FMAC_F16_t16_e64) {
3098     // Don't fold if we are using source or output modifiers. The new VOP2
3099     // instructions don't have them.
3100     if (hasAnyModifiersSet(UseMI))
3101       return false;
3102 
3103     // If this is a free constant, there's no reason to do this.
3104     // TODO: We could fold this here instead of letting SIFoldOperands do it
3105     // later.
3106     MachineOperand *Src0 = getNamedOperand(UseMI, AMDGPU::OpName::src0);
3107 
3108     // Any src operand can be used for the legality check.
3109     if (isInlineConstant(UseMI, *Src0, *ImmOp))
3110       return false;
3111 
3112     bool IsF32 = Opc == AMDGPU::V_MAD_F32_e64 || Opc == AMDGPU::V_MAC_F32_e64 ||
3113                  Opc == AMDGPU::V_FMA_F32_e64 || Opc == AMDGPU::V_FMAC_F32_e64;
3114     bool IsFMA =
3115         Opc == AMDGPU::V_FMA_F32_e64 || Opc == AMDGPU::V_FMAC_F32_e64 ||
3116         Opc == AMDGPU::V_FMA_F16_e64 || Opc == AMDGPU::V_FMAC_F16_e64 ||
3117         Opc == AMDGPU::V_FMAC_F16_t16_e64;
3118     MachineOperand *Src1 = getNamedOperand(UseMI, AMDGPU::OpName::src1);
3119     MachineOperand *Src2 = getNamedOperand(UseMI, AMDGPU::OpName::src2);
3120 
3121     // Multiplied part is the constant: Use v_madmk_{f16, f32}.
3122     // We should only expect these to be on src0 due to canonicalization.
3123     if (Src0->isReg() && Src0->getReg() == Reg) {
3124       if (!Src1->isReg() || RI.isSGPRClass(MRI->getRegClass(Src1->getReg())))
3125         return false;
3126 
3127       if (!Src2->isReg() || RI.isSGPRClass(MRI->getRegClass(Src2->getReg())))
3128         return false;
3129 
3130       unsigned NewOpc =
3131           IsFMA ? (IsF32                    ? AMDGPU::V_FMAMK_F32
3132                    : ST.hasTrue16BitInsts() ? AMDGPU::V_FMAMK_F16_t16
3133                                             : AMDGPU::V_FMAMK_F16)
3134                 : (IsF32 ? AMDGPU::V_MADMK_F32 : AMDGPU::V_MADMK_F16);
3135       if (pseudoToMCOpcode(NewOpc) == -1)
3136         return false;
3137 
3138       // We need to swap operands 0 and 1 since madmk constant is at operand 1.
3139 
3140       const int64_t Imm = ImmOp->getImm();
3141 
3142       // FIXME: This would be a lot easier if we could return a new instruction
3143       // instead of having to modify in place.
3144 
3145       Register Src1Reg = Src1->getReg();
3146       unsigned Src1SubReg = Src1->getSubReg();
3147       Src0->setReg(Src1Reg);
3148       Src0->setSubReg(Src1SubReg);
3149       Src0->setIsKill(Src1->isKill());
3150 
3151       if (Opc == AMDGPU::V_MAC_F32_e64 || Opc == AMDGPU::V_MAC_F16_e64 ||
3152           Opc == AMDGPU::V_FMAC_F32_e64 || Opc == AMDGPU::V_FMAC_F16_t16_e64 ||
3153           Opc == AMDGPU::V_FMAC_F16_e64)
3154         UseMI.untieRegOperand(
3155             AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src2));
3156 
3157       Src1->ChangeToImmediate(Imm);
3158 
3159       removeModOperands(UseMI);
3160       UseMI.setDesc(get(NewOpc));
3161 
3162       bool DeleteDef = MRI->use_nodbg_empty(Reg);
3163       if (DeleteDef)
3164         DefMI.eraseFromParent();
3165 
3166       return true;
3167     }
3168 
3169     // Added part is the constant: Use v_madak_{f16, f32}.
3170     if (Src2->isReg() && Src2->getReg() == Reg) {
3171       // Not allowed to use constant bus for another operand.
3172       // We can however allow an inline immediate as src0.
3173       bool Src0Inlined = false;
3174       if (Src0->isReg()) {
3175         // Try to inline constant if possible.
3176         // If the Def moves immediate and the use is single
3177         // We are saving VGPR here.
3178         MachineInstr *Def = MRI->getUniqueVRegDef(Src0->getReg());
3179         if (Def && Def->isMoveImmediate() &&
3180           isInlineConstant(Def->getOperand(1)) &&
3181           MRI->hasOneUse(Src0->getReg())) {
3182           Src0->ChangeToImmediate(Def->getOperand(1).getImm());
3183           Src0Inlined = true;
3184         } else if ((Src0->getReg().isPhysical() &&
3185                     (ST.getConstantBusLimit(Opc) <= 1 &&
3186                      RI.isSGPRClass(RI.getPhysRegBaseClass(Src0->getReg())))) ||
3187                    (Src0->getReg().isVirtual() &&
3188                     (ST.getConstantBusLimit(Opc) <= 1 &&
3189                      RI.isSGPRClass(MRI->getRegClass(Src0->getReg())))))
3190           return false;
3191           // VGPR is okay as Src0 - fallthrough
3192       }
3193 
3194       if (Src1->isReg() && !Src0Inlined ) {
3195         // We have one slot for inlinable constant so far - try to fill it
3196         MachineInstr *Def = MRI->getUniqueVRegDef(Src1->getReg());
3197         if (Def && Def->isMoveImmediate() &&
3198             isInlineConstant(Def->getOperand(1)) &&
3199             MRI->hasOneUse(Src1->getReg()) &&
3200             commuteInstruction(UseMI)) {
3201             Src0->ChangeToImmediate(Def->getOperand(1).getImm());
3202         } else if ((Src1->getReg().isPhysical() &&
3203                     RI.isSGPRClass(RI.getPhysRegBaseClass(Src1->getReg()))) ||
3204                    (Src1->getReg().isVirtual() &&
3205                     RI.isSGPRClass(MRI->getRegClass(Src1->getReg()))))
3206           return false;
3207           // VGPR is okay as Src1 - fallthrough
3208       }
3209 
3210       unsigned NewOpc =
3211           IsFMA ? (IsF32                    ? AMDGPU::V_FMAAK_F32
3212                    : ST.hasTrue16BitInsts() ? AMDGPU::V_FMAAK_F16_t16
3213                                             : AMDGPU::V_FMAAK_F16)
3214                 : (IsF32 ? AMDGPU::V_MADAK_F32 : AMDGPU::V_MADAK_F16);
3215       if (pseudoToMCOpcode(NewOpc) == -1)
3216         return false;
3217 
3218       const int64_t Imm = ImmOp->getImm();
3219 
3220       // FIXME: This would be a lot easier if we could return a new instruction
3221       // instead of having to modify in place.
3222 
3223       if (Opc == AMDGPU::V_MAC_F32_e64 || Opc == AMDGPU::V_MAC_F16_e64 ||
3224           Opc == AMDGPU::V_FMAC_F32_e64 || Opc == AMDGPU::V_FMAC_F16_t16_e64 ||
3225           Opc == AMDGPU::V_FMAC_F16_e64)
3226         UseMI.untieRegOperand(
3227             AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src2));
3228 
3229       // ChangingToImmediate adds Src2 back to the instruction.
3230       Src2->ChangeToImmediate(Imm);
3231 
3232       // These come before src2.
3233       removeModOperands(UseMI);
3234       UseMI.setDesc(get(NewOpc));
3235       // It might happen that UseMI was commuted
3236       // and we now have SGPR as SRC1. If so 2 inlined
3237       // constant and SGPR are illegal.
3238       legalizeOperands(UseMI);
3239 
3240       bool DeleteDef = MRI->use_nodbg_empty(Reg);
3241       if (DeleteDef)
3242         DefMI.eraseFromParent();
3243 
3244       return true;
3245     }
3246   }
3247 
3248   return false;
3249 }
3250 
3251 static bool
3252 memOpsHaveSameBaseOperands(ArrayRef<const MachineOperand *> BaseOps1,
3253                            ArrayRef<const MachineOperand *> BaseOps2) {
3254   if (BaseOps1.size() != BaseOps2.size())
3255     return false;
3256   for (size_t I = 0, E = BaseOps1.size(); I < E; ++I) {
3257     if (!BaseOps1[I]->isIdenticalTo(*BaseOps2[I]))
3258       return false;
3259   }
3260   return true;
3261 }
3262 
3263 static bool offsetsDoNotOverlap(int WidthA, int OffsetA,
3264                                 int WidthB, int OffsetB) {
3265   int LowOffset = OffsetA < OffsetB ? OffsetA : OffsetB;
3266   int HighOffset = OffsetA < OffsetB ? OffsetB : OffsetA;
3267   int LowWidth = (LowOffset == OffsetA) ? WidthA : WidthB;
3268   return LowOffset + LowWidth <= HighOffset;
3269 }
3270 
3271 bool SIInstrInfo::checkInstOffsetsDoNotOverlap(const MachineInstr &MIa,
3272                                                const MachineInstr &MIb) const {
3273   SmallVector<const MachineOperand *, 4> BaseOps0, BaseOps1;
3274   int64_t Offset0, Offset1;
3275   unsigned Dummy0, Dummy1;
3276   bool Offset0IsScalable, Offset1IsScalable;
3277   if (!getMemOperandsWithOffsetWidth(MIa, BaseOps0, Offset0, Offset0IsScalable,
3278                                      Dummy0, &RI) ||
3279       !getMemOperandsWithOffsetWidth(MIb, BaseOps1, Offset1, Offset1IsScalable,
3280                                      Dummy1, &RI))
3281     return false;
3282 
3283   if (!memOpsHaveSameBaseOperands(BaseOps0, BaseOps1))
3284     return false;
3285 
3286   if (!MIa.hasOneMemOperand() || !MIb.hasOneMemOperand()) {
3287     // FIXME: Handle ds_read2 / ds_write2.
3288     return false;
3289   }
3290   unsigned Width0 = MIa.memoperands().front()->getSize();
3291   unsigned Width1 = MIb.memoperands().front()->getSize();
3292   return offsetsDoNotOverlap(Width0, Offset0, Width1, Offset1);
3293 }
3294 
3295 bool SIInstrInfo::areMemAccessesTriviallyDisjoint(const MachineInstr &MIa,
3296                                                   const MachineInstr &MIb) const {
3297   assert(MIa.mayLoadOrStore() &&
3298          "MIa must load from or modify a memory location");
3299   assert(MIb.mayLoadOrStore() &&
3300          "MIb must load from or modify a memory location");
3301 
3302   if (MIa.hasUnmodeledSideEffects() || MIb.hasUnmodeledSideEffects())
3303     return false;
3304 
3305   // XXX - Can we relax this between address spaces?
3306   if (MIa.hasOrderedMemoryRef() || MIb.hasOrderedMemoryRef())
3307     return false;
3308 
3309   // TODO: Should we check the address space from the MachineMemOperand? That
3310   // would allow us to distinguish objects we know don't alias based on the
3311   // underlying address space, even if it was lowered to a different one,
3312   // e.g. private accesses lowered to use MUBUF instructions on a scratch
3313   // buffer.
3314   if (isDS(MIa)) {
3315     if (isDS(MIb))
3316       return checkInstOffsetsDoNotOverlap(MIa, MIb);
3317 
3318     return !isFLAT(MIb) || isSegmentSpecificFLAT(MIb);
3319   }
3320 
3321   if (isMUBUF(MIa) || isMTBUF(MIa)) {
3322     if (isMUBUF(MIb) || isMTBUF(MIb))
3323       return checkInstOffsetsDoNotOverlap(MIa, MIb);
3324 
3325     return !isFLAT(MIb) && !isSMRD(MIb);
3326   }
3327 
3328   if (isSMRD(MIa)) {
3329     if (isSMRD(MIb))
3330       return checkInstOffsetsDoNotOverlap(MIa, MIb);
3331 
3332     return !isFLAT(MIb) && !isMUBUF(MIb) && !isMTBUF(MIb);
3333   }
3334 
3335   if (isFLAT(MIa)) {
3336     if (isFLAT(MIb))
3337       return checkInstOffsetsDoNotOverlap(MIa, MIb);
3338 
3339     return false;
3340   }
3341 
3342   return false;
3343 }
3344 
3345 static bool getFoldableImm(Register Reg, const MachineRegisterInfo &MRI,
3346                            int64_t &Imm, MachineInstr **DefMI = nullptr) {
3347   if (Reg.isPhysical())
3348     return false;
3349   auto *Def = MRI.getUniqueVRegDef(Reg);
3350   if (Def && SIInstrInfo::isFoldableCopy(*Def) && Def->getOperand(1).isImm()) {
3351     Imm = Def->getOperand(1).getImm();
3352     if (DefMI)
3353       *DefMI = Def;
3354     return true;
3355   }
3356   return false;
3357 }
3358 
3359 static bool getFoldableImm(const MachineOperand *MO, int64_t &Imm,
3360                            MachineInstr **DefMI = nullptr) {
3361   if (!MO->isReg())
3362     return false;
3363   const MachineFunction *MF = MO->getParent()->getParent()->getParent();
3364   const MachineRegisterInfo &MRI = MF->getRegInfo();
3365   return getFoldableImm(MO->getReg(), MRI, Imm, DefMI);
3366 }
3367 
3368 static void updateLiveVariables(LiveVariables *LV, MachineInstr &MI,
3369                                 MachineInstr &NewMI) {
3370   if (LV) {
3371     unsigned NumOps = MI.getNumOperands();
3372     for (unsigned I = 1; I < NumOps; ++I) {
3373       MachineOperand &Op = MI.getOperand(I);
3374       if (Op.isReg() && Op.isKill())
3375         LV->replaceKillInstruction(Op.getReg(), MI, NewMI);
3376     }
3377   }
3378 }
3379 
3380 MachineInstr *SIInstrInfo::convertToThreeAddress(MachineInstr &MI,
3381                                                  LiveVariables *LV,
3382                                                  LiveIntervals *LIS) const {
3383   MachineBasicBlock &MBB = *MI.getParent();
3384   unsigned Opc = MI.getOpcode();
3385 
3386   // Handle MFMA.
3387   int NewMFMAOpc = AMDGPU::getMFMAEarlyClobberOp(Opc);
3388   if (NewMFMAOpc != -1) {
3389     MachineInstrBuilder MIB =
3390         BuildMI(MBB, MI, MI.getDebugLoc(), get(NewMFMAOpc));
3391     for (unsigned I = 0, E = MI.getNumOperands(); I != E; ++I)
3392       MIB.add(MI.getOperand(I));
3393     updateLiveVariables(LV, MI, *MIB);
3394     if (LIS)
3395       LIS->ReplaceMachineInstrInMaps(MI, *MIB);
3396     return MIB;
3397   }
3398 
3399   if (SIInstrInfo::isWMMA(MI)) {
3400     unsigned NewOpc = AMDGPU::mapWMMA2AddrTo3AddrOpcode(MI.getOpcode());
3401     MachineInstrBuilder MIB = BuildMI(MBB, MI, MI.getDebugLoc(), get(NewOpc))
3402                                   .setMIFlags(MI.getFlags());
3403     for (unsigned I = 0, E = MI.getNumOperands(); I != E; ++I)
3404       MIB->addOperand(MI.getOperand(I));
3405 
3406     updateLiveVariables(LV, MI, *MIB);
3407     if (LIS)
3408       LIS->ReplaceMachineInstrInMaps(MI, *MIB);
3409 
3410     return MIB;
3411   }
3412 
3413   assert(Opc != AMDGPU::V_FMAC_F16_t16_e32 &&
3414          "V_FMAC_F16_t16_e32 is not supported and not expected to be present "
3415          "pre-RA");
3416 
3417   // Handle MAC/FMAC.
3418   bool IsF16 = Opc == AMDGPU::V_MAC_F16_e32 || Opc == AMDGPU::V_MAC_F16_e64 ||
3419                Opc == AMDGPU::V_FMAC_F16_e32 || Opc == AMDGPU::V_FMAC_F16_e64 ||
3420                Opc == AMDGPU::V_FMAC_F16_t16_e64;
3421   bool IsFMA = Opc == AMDGPU::V_FMAC_F32_e32 || Opc == AMDGPU::V_FMAC_F32_e64 ||
3422                Opc == AMDGPU::V_FMAC_LEGACY_F32_e32 ||
3423                Opc == AMDGPU::V_FMAC_LEGACY_F32_e64 ||
3424                Opc == AMDGPU::V_FMAC_F16_e32 || Opc == AMDGPU::V_FMAC_F16_e64 ||
3425                Opc == AMDGPU::V_FMAC_F16_t16_e64 ||
3426                Opc == AMDGPU::V_FMAC_F64_e32 || Opc == AMDGPU::V_FMAC_F64_e64;
3427   bool IsF64 = Opc == AMDGPU::V_FMAC_F64_e32 || Opc == AMDGPU::V_FMAC_F64_e64;
3428   bool IsLegacy = Opc == AMDGPU::V_MAC_LEGACY_F32_e32 ||
3429                   Opc == AMDGPU::V_MAC_LEGACY_F32_e64 ||
3430                   Opc == AMDGPU::V_FMAC_LEGACY_F32_e32 ||
3431                   Opc == AMDGPU::V_FMAC_LEGACY_F32_e64;
3432   bool Src0Literal = false;
3433 
3434   switch (Opc) {
3435   default:
3436     return nullptr;
3437   case AMDGPU::V_MAC_F16_e64:
3438   case AMDGPU::V_FMAC_F16_e64:
3439   case AMDGPU::V_FMAC_F16_t16_e64:
3440   case AMDGPU::V_MAC_F32_e64:
3441   case AMDGPU::V_MAC_LEGACY_F32_e64:
3442   case AMDGPU::V_FMAC_F32_e64:
3443   case AMDGPU::V_FMAC_LEGACY_F32_e64:
3444   case AMDGPU::V_FMAC_F64_e64:
3445     break;
3446   case AMDGPU::V_MAC_F16_e32:
3447   case AMDGPU::V_FMAC_F16_e32:
3448   case AMDGPU::V_MAC_F32_e32:
3449   case AMDGPU::V_MAC_LEGACY_F32_e32:
3450   case AMDGPU::V_FMAC_F32_e32:
3451   case AMDGPU::V_FMAC_LEGACY_F32_e32:
3452   case AMDGPU::V_FMAC_F64_e32: {
3453     int Src0Idx = AMDGPU::getNamedOperandIdx(MI.getOpcode(),
3454                                              AMDGPU::OpName::src0);
3455     const MachineOperand *Src0 = &MI.getOperand(Src0Idx);
3456     if (!Src0->isReg() && !Src0->isImm())
3457       return nullptr;
3458 
3459     if (Src0->isImm() && !isInlineConstant(MI, Src0Idx, *Src0))
3460       Src0Literal = true;
3461 
3462     break;
3463   }
3464   }
3465 
3466   MachineInstrBuilder MIB;
3467   const MachineOperand *Dst = getNamedOperand(MI, AMDGPU::OpName::vdst);
3468   const MachineOperand *Src0 = getNamedOperand(MI, AMDGPU::OpName::src0);
3469   const MachineOperand *Src0Mods =
3470     getNamedOperand(MI, AMDGPU::OpName::src0_modifiers);
3471   const MachineOperand *Src1 = getNamedOperand(MI, AMDGPU::OpName::src1);
3472   const MachineOperand *Src1Mods =
3473     getNamedOperand(MI, AMDGPU::OpName::src1_modifiers);
3474   const MachineOperand *Src2 = getNamedOperand(MI, AMDGPU::OpName::src2);
3475   const MachineOperand *Src2Mods =
3476       getNamedOperand(MI, AMDGPU::OpName::src2_modifiers);
3477   const MachineOperand *Clamp = getNamedOperand(MI, AMDGPU::OpName::clamp);
3478   const MachineOperand *Omod = getNamedOperand(MI, AMDGPU::OpName::omod);
3479   const MachineOperand *OpSel = getNamedOperand(MI, AMDGPU::OpName::op_sel);
3480 
3481   if (!Src0Mods && !Src1Mods && !Src2Mods && !Clamp && !Omod && !IsF64 &&
3482       !IsLegacy &&
3483       // If we have an SGPR input, we will violate the constant bus restriction.
3484       (ST.getConstantBusLimit(Opc) > 1 || !Src0->isReg() ||
3485        !RI.isSGPRReg(MBB.getParent()->getRegInfo(), Src0->getReg()))) {
3486     MachineInstr *DefMI;
3487     const auto killDef = [&]() -> void {
3488       const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
3489       // The only user is the instruction which will be killed.
3490       Register DefReg = DefMI->getOperand(0).getReg();
3491       if (!MRI.hasOneNonDBGUse(DefReg))
3492         return;
3493       // We cannot just remove the DefMI here, calling pass will crash.
3494       DefMI->setDesc(get(AMDGPU::IMPLICIT_DEF));
3495       for (unsigned I = DefMI->getNumOperands() - 1; I != 0; --I)
3496         DefMI->removeOperand(I);
3497       if (LV)
3498         LV->getVarInfo(DefReg).AliveBlocks.clear();
3499     };
3500 
3501     int64_t Imm;
3502     if (!Src0Literal && getFoldableImm(Src2, Imm, &DefMI)) {
3503       unsigned NewOpc =
3504           IsFMA ? (IsF16 ? (ST.hasTrue16BitInsts() ? AMDGPU::V_FMAAK_F16_t16
3505                                                    : AMDGPU::V_FMAAK_F16)
3506                          : AMDGPU::V_FMAAK_F32)
3507                 : (IsF16 ? AMDGPU::V_MADAK_F16 : AMDGPU::V_MADAK_F32);
3508       if (pseudoToMCOpcode(NewOpc) != -1) {
3509         MIB = BuildMI(MBB, MI, MI.getDebugLoc(), get(NewOpc))
3510                   .add(*Dst)
3511                   .add(*Src0)
3512                   .add(*Src1)
3513                   .addImm(Imm);
3514         updateLiveVariables(LV, MI, *MIB);
3515         if (LIS)
3516           LIS->ReplaceMachineInstrInMaps(MI, *MIB);
3517         killDef();
3518         return MIB;
3519       }
3520     }
3521     unsigned NewOpc =
3522         IsFMA ? (IsF16 ? (ST.hasTrue16BitInsts() ? AMDGPU::V_FMAMK_F16_t16
3523                                                  : AMDGPU::V_FMAMK_F16)
3524                        : AMDGPU::V_FMAMK_F32)
3525               : (IsF16 ? AMDGPU::V_MADMK_F16 : AMDGPU::V_MADMK_F32);
3526     if (!Src0Literal && getFoldableImm(Src1, Imm, &DefMI)) {
3527       if (pseudoToMCOpcode(NewOpc) != -1) {
3528         MIB = BuildMI(MBB, MI, MI.getDebugLoc(), get(NewOpc))
3529                   .add(*Dst)
3530                   .add(*Src0)
3531                   .addImm(Imm)
3532                   .add(*Src2);
3533         updateLiveVariables(LV, MI, *MIB);
3534         if (LIS)
3535           LIS->ReplaceMachineInstrInMaps(MI, *MIB);
3536         killDef();
3537         return MIB;
3538       }
3539     }
3540     if (Src0Literal || getFoldableImm(Src0, Imm, &DefMI)) {
3541       if (Src0Literal) {
3542         Imm = Src0->getImm();
3543         DefMI = nullptr;
3544       }
3545       if (pseudoToMCOpcode(NewOpc) != -1 &&
3546           isOperandLegal(
3547               MI, AMDGPU::getNamedOperandIdx(NewOpc, AMDGPU::OpName::src0),
3548               Src1)) {
3549         MIB = BuildMI(MBB, MI, MI.getDebugLoc(), get(NewOpc))
3550                   .add(*Dst)
3551                   .add(*Src1)
3552                   .addImm(Imm)
3553                   .add(*Src2);
3554         updateLiveVariables(LV, MI, *MIB);
3555         if (LIS)
3556           LIS->ReplaceMachineInstrInMaps(MI, *MIB);
3557         if (DefMI)
3558           killDef();
3559         return MIB;
3560       }
3561     }
3562   }
3563 
3564   // VOP2 mac/fmac with a literal operand cannot be converted to VOP3 mad/fma
3565   // if VOP3 does not allow a literal operand.
3566   if (Src0Literal && !ST.hasVOP3Literal())
3567     return nullptr;
3568 
3569   unsigned NewOpc = IsFMA ? IsF16 ? AMDGPU::V_FMA_F16_gfx9_e64
3570                                   : IsF64 ? AMDGPU::V_FMA_F64_e64
3571                                           : IsLegacy
3572                                                 ? AMDGPU::V_FMA_LEGACY_F32_e64
3573                                                 : AMDGPU::V_FMA_F32_e64
3574                           : IsF16 ? AMDGPU::V_MAD_F16_e64
3575                                   : IsLegacy ? AMDGPU::V_MAD_LEGACY_F32_e64
3576                                              : AMDGPU::V_MAD_F32_e64;
3577   if (pseudoToMCOpcode(NewOpc) == -1)
3578     return nullptr;
3579 
3580   MIB = BuildMI(MBB, MI, MI.getDebugLoc(), get(NewOpc))
3581             .add(*Dst)
3582             .addImm(Src0Mods ? Src0Mods->getImm() : 0)
3583             .add(*Src0)
3584             .addImm(Src1Mods ? Src1Mods->getImm() : 0)
3585             .add(*Src1)
3586             .addImm(Src2Mods ? Src2Mods->getImm() : 0)
3587             .add(*Src2)
3588             .addImm(Clamp ? Clamp->getImm() : 0)
3589             .addImm(Omod ? Omod->getImm() : 0);
3590   if (AMDGPU::hasNamedOperand(NewOpc, AMDGPU::OpName::op_sel))
3591     MIB.addImm(OpSel ? OpSel->getImm() : 0);
3592   updateLiveVariables(LV, MI, *MIB);
3593   if (LIS)
3594     LIS->ReplaceMachineInstrInMaps(MI, *MIB);
3595   return MIB;
3596 }
3597 
3598 // It's not generally safe to move VALU instructions across these since it will
3599 // start using the register as a base index rather than directly.
3600 // XXX - Why isn't hasSideEffects sufficient for these?
3601 static bool changesVGPRIndexingMode(const MachineInstr &MI) {
3602   switch (MI.getOpcode()) {
3603   case AMDGPU::S_SET_GPR_IDX_ON:
3604   case AMDGPU::S_SET_GPR_IDX_MODE:
3605   case AMDGPU::S_SET_GPR_IDX_OFF:
3606     return true;
3607   default:
3608     return false;
3609   }
3610 }
3611 
3612 bool SIInstrInfo::isSchedulingBoundary(const MachineInstr &MI,
3613                                        const MachineBasicBlock *MBB,
3614                                        const MachineFunction &MF) const {
3615   // Skipping the check for SP writes in the base implementation. The reason it
3616   // was added was apparently due to compile time concerns.
3617   //
3618   // TODO: Do we really want this barrier? It triggers unnecessary hazard nops
3619   // but is probably avoidable.
3620 
3621   // Copied from base implementation.
3622   // Terminators and labels can't be scheduled around.
3623   if (MI.isTerminator() || MI.isPosition())
3624     return true;
3625 
3626   // INLINEASM_BR can jump to another block
3627   if (MI.getOpcode() == TargetOpcode::INLINEASM_BR)
3628     return true;
3629 
3630   if (MI.getOpcode() == AMDGPU::SCHED_BARRIER && MI.getOperand(0).getImm() == 0)
3631     return true;
3632 
3633   // Target-independent instructions do not have an implicit-use of EXEC, even
3634   // when they operate on VGPRs. Treating EXEC modifications as scheduling
3635   // boundaries prevents incorrect movements of such instructions.
3636   return MI.modifiesRegister(AMDGPU::EXEC, &RI) ||
3637          MI.getOpcode() == AMDGPU::S_SETREG_IMM32_B32 ||
3638          MI.getOpcode() == AMDGPU::S_SETREG_B32 ||
3639          MI.getOpcode() == AMDGPU::S_SETPRIO ||
3640          changesVGPRIndexingMode(MI);
3641 }
3642 
3643 bool SIInstrInfo::isAlwaysGDS(uint16_t Opcode) const {
3644   return Opcode == AMDGPU::DS_ORDERED_COUNT ||
3645          Opcode == AMDGPU::DS_GWS_INIT ||
3646          Opcode == AMDGPU::DS_GWS_SEMA_V ||
3647          Opcode == AMDGPU::DS_GWS_SEMA_BR ||
3648          Opcode == AMDGPU::DS_GWS_SEMA_P ||
3649          Opcode == AMDGPU::DS_GWS_SEMA_RELEASE_ALL ||
3650          Opcode == AMDGPU::DS_GWS_BARRIER;
3651 }
3652 
3653 bool SIInstrInfo::modifiesModeRegister(const MachineInstr &MI) {
3654   // Skip the full operand and register alias search modifiesRegister
3655   // does. There's only a handful of instructions that touch this, it's only an
3656   // implicit def, and doesn't alias any other registers.
3657   return is_contained(MI.getDesc().implicit_defs(), AMDGPU::MODE);
3658 }
3659 
3660 bool SIInstrInfo::hasUnwantedEffectsWhenEXECEmpty(const MachineInstr &MI) const {
3661   unsigned Opcode = MI.getOpcode();
3662 
3663   if (MI.mayStore() && isSMRD(MI))
3664     return true; // scalar store or atomic
3665 
3666   // This will terminate the function when other lanes may need to continue.
3667   if (MI.isReturn())
3668     return true;
3669 
3670   // These instructions cause shader I/O that may cause hardware lockups
3671   // when executed with an empty EXEC mask.
3672   //
3673   // Note: exp with VM = DONE = 0 is automatically skipped by hardware when
3674   //       EXEC = 0, but checking for that case here seems not worth it
3675   //       given the typical code patterns.
3676   if (Opcode == AMDGPU::S_SENDMSG || Opcode == AMDGPU::S_SENDMSGHALT ||
3677       isEXP(Opcode) ||
3678       Opcode == AMDGPU::DS_ORDERED_COUNT || Opcode == AMDGPU::S_TRAP ||
3679       Opcode == AMDGPU::DS_GWS_INIT || Opcode == AMDGPU::DS_GWS_BARRIER)
3680     return true;
3681 
3682   if (MI.isCall() || MI.isInlineAsm())
3683     return true; // conservative assumption
3684 
3685   // A mode change is a scalar operation that influences vector instructions.
3686   if (modifiesModeRegister(MI))
3687     return true;
3688 
3689   // These are like SALU instructions in terms of effects, so it's questionable
3690   // whether we should return true for those.
3691   //
3692   // However, executing them with EXEC = 0 causes them to operate on undefined
3693   // data, which we avoid by returning true here.
3694   if (Opcode == AMDGPU::V_READFIRSTLANE_B32 ||
3695       Opcode == AMDGPU::V_READLANE_B32 || Opcode == AMDGPU::V_WRITELANE_B32)
3696     return true;
3697 
3698   return false;
3699 }
3700 
3701 bool SIInstrInfo::mayReadEXEC(const MachineRegisterInfo &MRI,
3702                               const MachineInstr &MI) const {
3703   if (MI.isMetaInstruction())
3704     return false;
3705 
3706   // This won't read exec if this is an SGPR->SGPR copy.
3707   if (MI.isCopyLike()) {
3708     if (!RI.isSGPRReg(MRI, MI.getOperand(0).getReg()))
3709       return true;
3710 
3711     // Make sure this isn't copying exec as a normal operand
3712     return MI.readsRegister(AMDGPU::EXEC, &RI);
3713   }
3714 
3715   // Make a conservative assumption about the callee.
3716   if (MI.isCall())
3717     return true;
3718 
3719   // Be conservative with any unhandled generic opcodes.
3720   if (!isTargetSpecificOpcode(MI.getOpcode()))
3721     return true;
3722 
3723   return !isSALU(MI) || MI.readsRegister(AMDGPU::EXEC, &RI);
3724 }
3725 
3726 bool SIInstrInfo::isInlineConstant(const APInt &Imm) const {
3727   switch (Imm.getBitWidth()) {
3728   case 1: // This likely will be a condition code mask.
3729     return true;
3730 
3731   case 32:
3732     return AMDGPU::isInlinableLiteral32(Imm.getSExtValue(),
3733                                         ST.hasInv2PiInlineImm());
3734   case 64:
3735     return AMDGPU::isInlinableLiteral64(Imm.getSExtValue(),
3736                                         ST.hasInv2PiInlineImm());
3737   case 16:
3738     return ST.has16BitInsts() &&
3739            AMDGPU::isInlinableLiteral16(Imm.getSExtValue(),
3740                                         ST.hasInv2PiInlineImm());
3741   default:
3742     llvm_unreachable("invalid bitwidth");
3743   }
3744 }
3745 
3746 bool SIInstrInfo::isInlineConstant(const MachineOperand &MO,
3747                                    uint8_t OperandType) const {
3748   assert(!MO.isReg() && "isInlineConstant called on register operand!");
3749   if (!MO.isImm() ||
3750       OperandType < AMDGPU::OPERAND_SRC_FIRST ||
3751       OperandType > AMDGPU::OPERAND_SRC_LAST)
3752     return false;
3753 
3754   // MachineOperand provides no way to tell the true operand size, since it only
3755   // records a 64-bit value. We need to know the size to determine if a 32-bit
3756   // floating point immediate bit pattern is legal for an integer immediate. It
3757   // would be for any 32-bit integer operand, but would not be for a 64-bit one.
3758 
3759   int64_t Imm = MO.getImm();
3760   switch (OperandType) {
3761   case AMDGPU::OPERAND_REG_IMM_INT32:
3762   case AMDGPU::OPERAND_REG_IMM_FP32:
3763   case AMDGPU::OPERAND_REG_IMM_FP32_DEFERRED:
3764   case AMDGPU::OPERAND_REG_INLINE_C_INT32:
3765   case AMDGPU::OPERAND_REG_INLINE_C_FP32:
3766   case AMDGPU::OPERAND_REG_IMM_V2FP32:
3767   case AMDGPU::OPERAND_REG_INLINE_C_V2FP32:
3768   case AMDGPU::OPERAND_REG_IMM_V2INT32:
3769   case AMDGPU::OPERAND_REG_INLINE_C_V2INT32:
3770   case AMDGPU::OPERAND_REG_INLINE_AC_INT32:
3771   case AMDGPU::OPERAND_REG_INLINE_AC_FP32: {
3772     int32_t Trunc = static_cast<int32_t>(Imm);
3773     return AMDGPU::isInlinableLiteral32(Trunc, ST.hasInv2PiInlineImm());
3774   }
3775   case AMDGPU::OPERAND_REG_IMM_INT64:
3776   case AMDGPU::OPERAND_REG_IMM_FP64:
3777   case AMDGPU::OPERAND_REG_INLINE_C_INT64:
3778   case AMDGPU::OPERAND_REG_INLINE_C_FP64:
3779   case AMDGPU::OPERAND_REG_INLINE_AC_FP64:
3780     return AMDGPU::isInlinableLiteral64(MO.getImm(),
3781                                         ST.hasInv2PiInlineImm());
3782   case AMDGPU::OPERAND_REG_IMM_INT16:
3783   case AMDGPU::OPERAND_REG_INLINE_C_INT16:
3784   case AMDGPU::OPERAND_REG_INLINE_AC_INT16:
3785     // We would expect inline immediates to not be concerned with an integer/fp
3786     // distinction. However, in the case of 16-bit integer operations, the
3787     // "floating point" values appear to not work. It seems read the low 16-bits
3788     // of 32-bit immediates, which happens to always work for the integer
3789     // values.
3790     //
3791     // See llvm bugzilla 46302.
3792     //
3793     // TODO: Theoretically we could use op-sel to use the high bits of the
3794     // 32-bit FP values.
3795     return AMDGPU::isInlinableIntLiteral(Imm);
3796   case AMDGPU::OPERAND_REG_IMM_V2INT16:
3797   case AMDGPU::OPERAND_REG_INLINE_C_V2INT16:
3798   case AMDGPU::OPERAND_REG_INLINE_AC_V2INT16:
3799     // This suffers the same problem as the scalar 16-bit cases.
3800     return AMDGPU::isInlinableIntLiteralV216(Imm);
3801   case AMDGPU::OPERAND_REG_IMM_FP16:
3802   case AMDGPU::OPERAND_REG_IMM_FP16_DEFERRED:
3803   case AMDGPU::OPERAND_REG_INLINE_C_FP16:
3804   case AMDGPU::OPERAND_REG_INLINE_AC_FP16: {
3805     if (isInt<16>(Imm) || isUInt<16>(Imm)) {
3806       // A few special case instructions have 16-bit operands on subtargets
3807       // where 16-bit instructions are not legal.
3808       // TODO: Do the 32-bit immediates work? We shouldn't really need to handle
3809       // constants in these cases
3810       int16_t Trunc = static_cast<int16_t>(Imm);
3811       return ST.has16BitInsts() &&
3812              AMDGPU::isInlinableLiteral16(Trunc, ST.hasInv2PiInlineImm());
3813     }
3814 
3815     return false;
3816   }
3817   case AMDGPU::OPERAND_REG_IMM_V2FP16:
3818   case AMDGPU::OPERAND_REG_INLINE_C_V2FP16:
3819   case AMDGPU::OPERAND_REG_INLINE_AC_V2FP16: {
3820     uint32_t Trunc = static_cast<uint32_t>(Imm);
3821     return AMDGPU::isInlinableLiteralV216(Trunc, ST.hasInv2PiInlineImm());
3822   }
3823   case AMDGPU::OPERAND_KIMM32:
3824   case AMDGPU::OPERAND_KIMM16:
3825     return false;
3826   default:
3827     llvm_unreachable("invalid bitwidth");
3828   }
3829 }
3830 
3831 static bool compareMachineOp(const MachineOperand &Op0,
3832                              const MachineOperand &Op1) {
3833   if (Op0.getType() != Op1.getType())
3834     return false;
3835 
3836   switch (Op0.getType()) {
3837   case MachineOperand::MO_Register:
3838     return Op0.getReg() == Op1.getReg();
3839   case MachineOperand::MO_Immediate:
3840     return Op0.getImm() == Op1.getImm();
3841   default:
3842     llvm_unreachable("Didn't expect to be comparing these operand types");
3843   }
3844 }
3845 
3846 bool SIInstrInfo::isImmOperandLegal(const MachineInstr &MI, unsigned OpNo,
3847                                     const MachineOperand &MO) const {
3848   const MCInstrDesc &InstDesc = MI.getDesc();
3849   const MCOperandInfo &OpInfo = InstDesc.operands()[OpNo];
3850 
3851   assert(MO.isImm() || MO.isTargetIndex() || MO.isFI() || MO.isGlobal());
3852 
3853   if (OpInfo.OperandType == MCOI::OPERAND_IMMEDIATE)
3854     return true;
3855 
3856   if (OpInfo.RegClass < 0)
3857     return false;
3858 
3859   if (MO.isImm() && isInlineConstant(MO, OpInfo)) {
3860     if (isMAI(MI) && ST.hasMFMAInlineLiteralBug() &&
3861         OpNo ==(unsigned)AMDGPU::getNamedOperandIdx(MI.getOpcode(),
3862                                                     AMDGPU::OpName::src2))
3863       return false;
3864     return RI.opCanUseInlineConstant(OpInfo.OperandType);
3865   }
3866 
3867   if (!RI.opCanUseLiteralConstant(OpInfo.OperandType))
3868     return false;
3869 
3870   if (!isVOP3(MI) || !AMDGPU::isSISrcOperand(InstDesc, OpNo))
3871     return true;
3872 
3873   return ST.hasVOP3Literal();
3874 }
3875 
3876 bool SIInstrInfo::hasVALU32BitEncoding(unsigned Opcode) const {
3877   // GFX90A does not have V_MUL_LEGACY_F32_e32.
3878   if (Opcode == AMDGPU::V_MUL_LEGACY_F32_e64 && ST.hasGFX90AInsts())
3879     return false;
3880 
3881   int Op32 = AMDGPU::getVOPe32(Opcode);
3882   if (Op32 == -1)
3883     return false;
3884 
3885   return pseudoToMCOpcode(Op32) != -1;
3886 }
3887 
3888 bool SIInstrInfo::hasModifiers(unsigned Opcode) const {
3889   // The src0_modifier operand is present on all instructions
3890   // that have modifiers.
3891 
3892   return AMDGPU::hasNamedOperand(Opcode, AMDGPU::OpName::src0_modifiers);
3893 }
3894 
3895 bool SIInstrInfo::hasModifiersSet(const MachineInstr &MI,
3896                                   unsigned OpName) const {
3897   const MachineOperand *Mods = getNamedOperand(MI, OpName);
3898   return Mods && Mods->getImm();
3899 }
3900 
3901 bool SIInstrInfo::hasAnyModifiersSet(const MachineInstr &MI) const {
3902   return any_of(ModifierOpNames,
3903                 [&](unsigned Name) { return hasModifiersSet(MI, Name); });
3904 }
3905 
3906 bool SIInstrInfo::canShrink(const MachineInstr &MI,
3907                             const MachineRegisterInfo &MRI) const {
3908   const MachineOperand *Src2 = getNamedOperand(MI, AMDGPU::OpName::src2);
3909   // Can't shrink instruction with three operands.
3910   if (Src2) {
3911     switch (MI.getOpcode()) {
3912       default: return false;
3913 
3914       case AMDGPU::V_ADDC_U32_e64:
3915       case AMDGPU::V_SUBB_U32_e64:
3916       case AMDGPU::V_SUBBREV_U32_e64: {
3917         const MachineOperand *Src1
3918           = getNamedOperand(MI, AMDGPU::OpName::src1);
3919         if (!Src1->isReg() || !RI.isVGPR(MRI, Src1->getReg()))
3920           return false;
3921         // Additional verification is needed for sdst/src2.
3922         return true;
3923       }
3924       case AMDGPU::V_MAC_F16_e64:
3925       case AMDGPU::V_MAC_F32_e64:
3926       case AMDGPU::V_MAC_LEGACY_F32_e64:
3927       case AMDGPU::V_FMAC_F16_e64:
3928       case AMDGPU::V_FMAC_F16_t16_e64:
3929       case AMDGPU::V_FMAC_F32_e64:
3930       case AMDGPU::V_FMAC_F64_e64:
3931       case AMDGPU::V_FMAC_LEGACY_F32_e64:
3932         if (!Src2->isReg() || !RI.isVGPR(MRI, Src2->getReg()) ||
3933             hasModifiersSet(MI, AMDGPU::OpName::src2_modifiers))
3934           return false;
3935         break;
3936 
3937       case AMDGPU::V_CNDMASK_B32_e64:
3938         break;
3939     }
3940   }
3941 
3942   const MachineOperand *Src1 = getNamedOperand(MI, AMDGPU::OpName::src1);
3943   if (Src1 && (!Src1->isReg() || !RI.isVGPR(MRI, Src1->getReg()) ||
3944                hasModifiersSet(MI, AMDGPU::OpName::src1_modifiers)))
3945     return false;
3946 
3947   // We don't need to check src0, all input types are legal, so just make sure
3948   // src0 isn't using any modifiers.
3949   if (hasModifiersSet(MI, AMDGPU::OpName::src0_modifiers))
3950     return false;
3951 
3952   // Can it be shrunk to a valid 32 bit opcode?
3953   if (!hasVALU32BitEncoding(MI.getOpcode()))
3954     return false;
3955 
3956   // Check output modifiers
3957   return !hasModifiersSet(MI, AMDGPU::OpName::omod) &&
3958          !hasModifiersSet(MI, AMDGPU::OpName::clamp);
3959 }
3960 
3961 // Set VCC operand with all flags from \p Orig, except for setting it as
3962 // implicit.
3963 static void copyFlagsToImplicitVCC(MachineInstr &MI,
3964                                    const MachineOperand &Orig) {
3965 
3966   for (MachineOperand &Use : MI.implicit_operands()) {
3967     if (Use.isUse() &&
3968         (Use.getReg() == AMDGPU::VCC || Use.getReg() == AMDGPU::VCC_LO)) {
3969       Use.setIsUndef(Orig.isUndef());
3970       Use.setIsKill(Orig.isKill());
3971       return;
3972     }
3973   }
3974 }
3975 
3976 MachineInstr *SIInstrInfo::buildShrunkInst(MachineInstr &MI,
3977                                            unsigned Op32) const {
3978   MachineBasicBlock *MBB = MI.getParent();
3979   MachineInstrBuilder Inst32 =
3980     BuildMI(*MBB, MI, MI.getDebugLoc(), get(Op32))
3981     .setMIFlags(MI.getFlags());
3982 
3983   // Add the dst operand if the 32-bit encoding also has an explicit $vdst.
3984   // For VOPC instructions, this is replaced by an implicit def of vcc.
3985   if (AMDGPU::hasNamedOperand(Op32, AMDGPU::OpName::vdst)) {
3986     // dst
3987     Inst32.add(MI.getOperand(0));
3988   } else if (AMDGPU::hasNamedOperand(Op32, AMDGPU::OpName::sdst)) {
3989     // VOPCX instructions won't be writing to an explicit dst, so this should
3990     // not fail for these instructions.
3991     assert(((MI.getOperand(0).getReg() == AMDGPU::VCC) ||
3992             (MI.getOperand(0).getReg() == AMDGPU::VCC_LO)) &&
3993            "Unexpected case");
3994   }
3995 
3996   Inst32.add(*getNamedOperand(MI, AMDGPU::OpName::src0));
3997 
3998   const MachineOperand *Src1 = getNamedOperand(MI, AMDGPU::OpName::src1);
3999   if (Src1)
4000     Inst32.add(*Src1);
4001 
4002   const MachineOperand *Src2 = getNamedOperand(MI, AMDGPU::OpName::src2);
4003 
4004   if (Src2) {
4005     int Op32Src2Idx = AMDGPU::getNamedOperandIdx(Op32, AMDGPU::OpName::src2);
4006     if (Op32Src2Idx != -1) {
4007       Inst32.add(*Src2);
4008     } else {
4009       // In the case of V_CNDMASK_B32_e32, the explicit operand src2 is
4010       // replaced with an implicit read of vcc or vcc_lo. The implicit read
4011       // of vcc was already added during the initial BuildMI, but we
4012       // 1) may need to change vcc to vcc_lo to preserve the original register
4013       // 2) have to preserve the original flags.
4014       fixImplicitOperands(*Inst32);
4015       copyFlagsToImplicitVCC(*Inst32, *Src2);
4016     }
4017   }
4018 
4019   return Inst32;
4020 }
4021 
4022 bool SIInstrInfo::usesConstantBus(const MachineRegisterInfo &MRI,
4023                                   const MachineOperand &MO,
4024                                   const MCOperandInfo &OpInfo) const {
4025   // Literal constants use the constant bus.
4026   if (!MO.isReg())
4027     return !isInlineConstant(MO, OpInfo);
4028 
4029   if (!MO.isUse())
4030     return false;
4031 
4032   if (MO.getReg().isVirtual())
4033     return RI.isSGPRClass(MRI.getRegClass(MO.getReg()));
4034 
4035   // Null is free
4036   if (MO.getReg() == AMDGPU::SGPR_NULL || MO.getReg() == AMDGPU::SGPR_NULL64)
4037     return false;
4038 
4039   // SGPRs use the constant bus
4040   if (MO.isImplicit()) {
4041     return MO.getReg() == AMDGPU::M0 ||
4042            MO.getReg() == AMDGPU::VCC ||
4043            MO.getReg() == AMDGPU::VCC_LO;
4044   } else {
4045     return AMDGPU::SReg_32RegClass.contains(MO.getReg()) ||
4046            AMDGPU::SReg_64RegClass.contains(MO.getReg());
4047   }
4048 }
4049 
4050 static Register findImplicitSGPRRead(const MachineInstr &MI) {
4051   for (const MachineOperand &MO : MI.implicit_operands()) {
4052     // We only care about reads.
4053     if (MO.isDef())
4054       continue;
4055 
4056     switch (MO.getReg()) {
4057     case AMDGPU::VCC:
4058     case AMDGPU::VCC_LO:
4059     case AMDGPU::VCC_HI:
4060     case AMDGPU::M0:
4061     case AMDGPU::FLAT_SCR:
4062       return MO.getReg();
4063 
4064     default:
4065       break;
4066     }
4067   }
4068 
4069   return Register();
4070 }
4071 
4072 static bool shouldReadExec(const MachineInstr &MI) {
4073   if (SIInstrInfo::isVALU(MI)) {
4074     switch (MI.getOpcode()) {
4075     case AMDGPU::V_READLANE_B32:
4076     case AMDGPU::V_WRITELANE_B32:
4077       return false;
4078     }
4079 
4080     return true;
4081   }
4082 
4083   if (MI.isPreISelOpcode() ||
4084       SIInstrInfo::isGenericOpcode(MI.getOpcode()) ||
4085       SIInstrInfo::isSALU(MI) ||
4086       SIInstrInfo::isSMRD(MI))
4087     return false;
4088 
4089   return true;
4090 }
4091 
4092 static bool isSubRegOf(const SIRegisterInfo &TRI,
4093                        const MachineOperand &SuperVec,
4094                        const MachineOperand &SubReg) {
4095   if (SubReg.getReg().isPhysical())
4096     return TRI.isSubRegister(SuperVec.getReg(), SubReg.getReg());
4097 
4098   return SubReg.getSubReg() != AMDGPU::NoSubRegister &&
4099          SubReg.getReg() == SuperVec.getReg();
4100 }
4101 
4102 bool SIInstrInfo::verifyInstruction(const MachineInstr &MI,
4103                                     StringRef &ErrInfo) const {
4104   uint16_t Opcode = MI.getOpcode();
4105   if (SIInstrInfo::isGenericOpcode(MI.getOpcode()))
4106     return true;
4107 
4108   const MachineFunction *MF = MI.getParent()->getParent();
4109   const MachineRegisterInfo &MRI = MF->getRegInfo();
4110 
4111   int Src0Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src0);
4112   int Src1Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src1);
4113   int Src2Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src2);
4114   int Src3Idx = -1;
4115   if (Src0Idx == -1) {
4116     // VOPD V_DUAL_* instructions use different operand names.
4117     Src0Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src0X);
4118     Src1Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::vsrc1X);
4119     Src2Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src0Y);
4120     Src3Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::vsrc1Y);
4121   }
4122 
4123   // Make sure the number of operands is correct.
4124   const MCInstrDesc &Desc = get(Opcode);
4125   if (!Desc.isVariadic() &&
4126       Desc.getNumOperands() != MI.getNumExplicitOperands()) {
4127     ErrInfo = "Instruction has wrong number of operands.";
4128     return false;
4129   }
4130 
4131   if (MI.isInlineAsm()) {
4132     // Verify register classes for inlineasm constraints.
4133     for (unsigned I = InlineAsm::MIOp_FirstOperand, E = MI.getNumOperands();
4134          I != E; ++I) {
4135       const TargetRegisterClass *RC = MI.getRegClassConstraint(I, this, &RI);
4136       if (!RC)
4137         continue;
4138 
4139       const MachineOperand &Op = MI.getOperand(I);
4140       if (!Op.isReg())
4141         continue;
4142 
4143       Register Reg = Op.getReg();
4144       if (!Reg.isVirtual() && !RC->contains(Reg)) {
4145         ErrInfo = "inlineasm operand has incorrect register class.";
4146         return false;
4147       }
4148     }
4149 
4150     return true;
4151   }
4152 
4153   if (isMIMG(MI) && MI.memoperands_empty() && MI.mayLoadOrStore()) {
4154     ErrInfo = "missing memory operand from MIMG instruction.";
4155     return false;
4156   }
4157 
4158   // Make sure the register classes are correct.
4159   for (int i = 0, e = Desc.getNumOperands(); i != e; ++i) {
4160     const MachineOperand &MO = MI.getOperand(i);
4161     if (MO.isFPImm()) {
4162       ErrInfo = "FPImm Machine Operands are not supported. ISel should bitcast "
4163                 "all fp values to integers.";
4164       return false;
4165     }
4166 
4167     int RegClass = Desc.operands()[i].RegClass;
4168 
4169     switch (Desc.operands()[i].OperandType) {
4170     case MCOI::OPERAND_REGISTER:
4171       if (MI.getOperand(i).isImm() || MI.getOperand(i).isGlobal()) {
4172         ErrInfo = "Illegal immediate value for operand.";
4173         return false;
4174       }
4175       break;
4176     case AMDGPU::OPERAND_REG_IMM_INT32:
4177     case AMDGPU::OPERAND_REG_IMM_FP32:
4178     case AMDGPU::OPERAND_REG_IMM_FP32_DEFERRED:
4179     case AMDGPU::OPERAND_REG_IMM_V2FP32:
4180       break;
4181     case AMDGPU::OPERAND_REG_INLINE_C_INT32:
4182     case AMDGPU::OPERAND_REG_INLINE_C_FP32:
4183     case AMDGPU::OPERAND_REG_INLINE_C_INT64:
4184     case AMDGPU::OPERAND_REG_INLINE_C_FP64:
4185     case AMDGPU::OPERAND_REG_INLINE_C_INT16:
4186     case AMDGPU::OPERAND_REG_INLINE_C_FP16:
4187     case AMDGPU::OPERAND_REG_INLINE_AC_INT32:
4188     case AMDGPU::OPERAND_REG_INLINE_AC_FP32:
4189     case AMDGPU::OPERAND_REG_INLINE_AC_INT16:
4190     case AMDGPU::OPERAND_REG_INLINE_AC_FP16:
4191     case AMDGPU::OPERAND_REG_INLINE_AC_FP64: {
4192       if (!MO.isReg() && (!MO.isImm() || !isInlineConstant(MI, i))) {
4193         ErrInfo = "Illegal immediate value for operand.";
4194         return false;
4195       }
4196       break;
4197     }
4198     case MCOI::OPERAND_IMMEDIATE:
4199     case AMDGPU::OPERAND_KIMM32:
4200       // Check if this operand is an immediate.
4201       // FrameIndex operands will be replaced by immediates, so they are
4202       // allowed.
4203       if (!MI.getOperand(i).isImm() && !MI.getOperand(i).isFI()) {
4204         ErrInfo = "Expected immediate, but got non-immediate";
4205         return false;
4206       }
4207       [[fallthrough]];
4208     default:
4209       continue;
4210     }
4211 
4212     if (!MO.isReg())
4213       continue;
4214     Register Reg = MO.getReg();
4215     if (!Reg)
4216       continue;
4217 
4218     // FIXME: Ideally we would have separate instruction definitions with the
4219     // aligned register constraint.
4220     // FIXME: We do not verify inline asm operands, but custom inline asm
4221     // verification is broken anyway
4222     if (ST.needsAlignedVGPRs()) {
4223       const TargetRegisterClass *RC = RI.getRegClassForReg(MRI, Reg);
4224       if (RI.hasVectorRegisters(RC) && MO.getSubReg()) {
4225         const TargetRegisterClass *SubRC =
4226             RI.getSubRegisterClass(RC, MO.getSubReg());
4227         RC = RI.getCompatibleSubRegClass(RC, SubRC, MO.getSubReg());
4228         if (RC)
4229           RC = SubRC;
4230       }
4231 
4232       // Check that this is the aligned version of the class.
4233       if (!RC || !RI.isProperlyAlignedRC(*RC)) {
4234         ErrInfo = "Subtarget requires even aligned vector registers";
4235         return false;
4236       }
4237     }
4238 
4239     if (RegClass != -1) {
4240       if (Reg.isVirtual())
4241         continue;
4242 
4243       const TargetRegisterClass *RC = RI.getRegClass(RegClass);
4244       if (!RC->contains(Reg)) {
4245         ErrInfo = "Operand has incorrect register class.";
4246         return false;
4247       }
4248     }
4249   }
4250 
4251   // Verify SDWA
4252   if (isSDWA(MI)) {
4253     if (!ST.hasSDWA()) {
4254       ErrInfo = "SDWA is not supported on this target";
4255       return false;
4256     }
4257 
4258     int DstIdx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::vdst);
4259 
4260     for (int OpIdx : {DstIdx, Src0Idx, Src1Idx, Src2Idx}) {
4261       if (OpIdx == -1)
4262         continue;
4263       const MachineOperand &MO = MI.getOperand(OpIdx);
4264 
4265       if (!ST.hasSDWAScalar()) {
4266         // Only VGPRS on VI
4267         if (!MO.isReg() || !RI.hasVGPRs(RI.getRegClassForReg(MRI, MO.getReg()))) {
4268           ErrInfo = "Only VGPRs allowed as operands in SDWA instructions on VI";
4269           return false;
4270         }
4271       } else {
4272         // No immediates on GFX9
4273         if (!MO.isReg()) {
4274           ErrInfo =
4275             "Only reg allowed as operands in SDWA instructions on GFX9+";
4276           return false;
4277         }
4278       }
4279     }
4280 
4281     if (!ST.hasSDWAOmod()) {
4282       // No omod allowed on VI
4283       const MachineOperand *OMod = getNamedOperand(MI, AMDGPU::OpName::omod);
4284       if (OMod != nullptr &&
4285         (!OMod->isImm() || OMod->getImm() != 0)) {
4286         ErrInfo = "OMod not allowed in SDWA instructions on VI";
4287         return false;
4288       }
4289     }
4290 
4291     uint16_t BasicOpcode = AMDGPU::getBasicFromSDWAOp(Opcode);
4292     if (isVOPC(BasicOpcode)) {
4293       if (!ST.hasSDWASdst() && DstIdx != -1) {
4294         // Only vcc allowed as dst on VI for VOPC
4295         const MachineOperand &Dst = MI.getOperand(DstIdx);
4296         if (!Dst.isReg() || Dst.getReg() != AMDGPU::VCC) {
4297           ErrInfo = "Only VCC allowed as dst in SDWA instructions on VI";
4298           return false;
4299         }
4300       } else if (!ST.hasSDWAOutModsVOPC()) {
4301         // No clamp allowed on GFX9 for VOPC
4302         const MachineOperand *Clamp = getNamedOperand(MI, AMDGPU::OpName::clamp);
4303         if (Clamp && (!Clamp->isImm() || Clamp->getImm() != 0)) {
4304           ErrInfo = "Clamp not allowed in VOPC SDWA instructions on VI";
4305           return false;
4306         }
4307 
4308         // No omod allowed on GFX9 for VOPC
4309         const MachineOperand *OMod = getNamedOperand(MI, AMDGPU::OpName::omod);
4310         if (OMod && (!OMod->isImm() || OMod->getImm() != 0)) {
4311           ErrInfo = "OMod not allowed in VOPC SDWA instructions on VI";
4312           return false;
4313         }
4314       }
4315     }
4316 
4317     const MachineOperand *DstUnused = getNamedOperand(MI, AMDGPU::OpName::dst_unused);
4318     if (DstUnused && DstUnused->isImm() &&
4319         DstUnused->getImm() == AMDGPU::SDWA::UNUSED_PRESERVE) {
4320       const MachineOperand &Dst = MI.getOperand(DstIdx);
4321       if (!Dst.isReg() || !Dst.isTied()) {
4322         ErrInfo = "Dst register should have tied register";
4323         return false;
4324       }
4325 
4326       const MachineOperand &TiedMO =
4327           MI.getOperand(MI.findTiedOperandIdx(DstIdx));
4328       if (!TiedMO.isReg() || !TiedMO.isImplicit() || !TiedMO.isUse()) {
4329         ErrInfo =
4330             "Dst register should be tied to implicit use of preserved register";
4331         return false;
4332       } else if (TiedMO.getReg().isPhysical() &&
4333                  Dst.getReg() != TiedMO.getReg()) {
4334         ErrInfo = "Dst register should use same physical register as preserved";
4335         return false;
4336       }
4337     }
4338   }
4339 
4340   // Verify MIMG
4341   if (isMIMG(MI.getOpcode()) && !MI.mayStore()) {
4342     // Ensure that the return type used is large enough for all the options
4343     // being used TFE/LWE require an extra result register.
4344     const MachineOperand *DMask = getNamedOperand(MI, AMDGPU::OpName::dmask);
4345     if (DMask) {
4346       uint64_t DMaskImm = DMask->getImm();
4347       uint32_t RegCount =
4348           isGather4(MI.getOpcode()) ? 4 : llvm::popcount(DMaskImm);
4349       const MachineOperand *TFE = getNamedOperand(MI, AMDGPU::OpName::tfe);
4350       const MachineOperand *LWE = getNamedOperand(MI, AMDGPU::OpName::lwe);
4351       const MachineOperand *D16 = getNamedOperand(MI, AMDGPU::OpName::d16);
4352 
4353       // Adjust for packed 16 bit values
4354       if (D16 && D16->getImm() && !ST.hasUnpackedD16VMem())
4355         RegCount >>= 1;
4356 
4357       // Adjust if using LWE or TFE
4358       if ((LWE && LWE->getImm()) || (TFE && TFE->getImm()))
4359         RegCount += 1;
4360 
4361       const uint32_t DstIdx =
4362           AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::vdata);
4363       const MachineOperand &Dst = MI.getOperand(DstIdx);
4364       if (Dst.isReg()) {
4365         const TargetRegisterClass *DstRC = getOpRegClass(MI, DstIdx);
4366         uint32_t DstSize = RI.getRegSizeInBits(*DstRC) / 32;
4367         if (RegCount > DstSize) {
4368           ErrInfo = "MIMG instruction returns too many registers for dst "
4369                     "register class";
4370           return false;
4371         }
4372       }
4373     }
4374   }
4375 
4376   // Verify VOP*. Ignore multiple sgpr operands on writelane.
4377   if (isVALU(MI) && Desc.getOpcode() != AMDGPU::V_WRITELANE_B32) {
4378     unsigned ConstantBusCount = 0;
4379     bool UsesLiteral = false;
4380     const MachineOperand *LiteralVal = nullptr;
4381 
4382     int ImmIdx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::imm);
4383     if (ImmIdx != -1) {
4384       ++ConstantBusCount;
4385       UsesLiteral = true;
4386       LiteralVal = &MI.getOperand(ImmIdx);
4387     }
4388 
4389     SmallVector<Register, 2> SGPRsUsed;
4390     Register SGPRUsed;
4391 
4392     // Only look at the true operands. Only a real operand can use the constant
4393     // bus, and we don't want to check pseudo-operands like the source modifier
4394     // flags.
4395     for (int OpIdx : {Src0Idx, Src1Idx, Src2Idx, Src3Idx}) {
4396       if (OpIdx == -1)
4397         continue;
4398       const MachineOperand &MO = MI.getOperand(OpIdx);
4399       if (usesConstantBus(MRI, MO, MI.getDesc().operands()[OpIdx])) {
4400         if (MO.isReg()) {
4401           SGPRUsed = MO.getReg();
4402           if (!llvm::is_contained(SGPRsUsed, SGPRUsed)) {
4403             ++ConstantBusCount;
4404             SGPRsUsed.push_back(SGPRUsed);
4405           }
4406         } else {
4407           if (!UsesLiteral) {
4408             ++ConstantBusCount;
4409             UsesLiteral = true;
4410             LiteralVal = &MO;
4411           } else if (!MO.isIdenticalTo(*LiteralVal)) {
4412             assert(isVOP2(MI) || isVOP3(MI));
4413             ErrInfo = "VOP2/VOP3 instruction uses more than one literal";
4414             return false;
4415           }
4416         }
4417       }
4418     }
4419 
4420     SGPRUsed = findImplicitSGPRRead(MI);
4421     if (SGPRUsed) {
4422       // Implicit uses may safely overlap true operands
4423       if (llvm::all_of(SGPRsUsed, [this, SGPRUsed](unsigned SGPR) {
4424             return !RI.regsOverlap(SGPRUsed, SGPR);
4425           })) {
4426         ++ConstantBusCount;
4427         SGPRsUsed.push_back(SGPRUsed);
4428       }
4429     }
4430 
4431     // v_writelane_b32 is an exception from constant bus restriction:
4432     // vsrc0 can be sgpr, const or m0 and lane select sgpr, m0 or inline-const
4433     if (ConstantBusCount > ST.getConstantBusLimit(Opcode) &&
4434         Opcode != AMDGPU::V_WRITELANE_B32) {
4435       ErrInfo = "VOP* instruction violates constant bus restriction";
4436       return false;
4437     }
4438 
4439     if (isVOP3(MI) && UsesLiteral && !ST.hasVOP3Literal()) {
4440       ErrInfo = "VOP3 instruction uses literal";
4441       return false;
4442     }
4443   }
4444 
4445   // Special case for writelane - this can break the multiple constant bus rule,
4446   // but still can't use more than one SGPR register
4447   if (Desc.getOpcode() == AMDGPU::V_WRITELANE_B32) {
4448     unsigned SGPRCount = 0;
4449     Register SGPRUsed;
4450 
4451     for (int OpIdx : {Src0Idx, Src1Idx}) {
4452       if (OpIdx == -1)
4453         break;
4454 
4455       const MachineOperand &MO = MI.getOperand(OpIdx);
4456 
4457       if (usesConstantBus(MRI, MO, MI.getDesc().operands()[OpIdx])) {
4458         if (MO.isReg() && MO.getReg() != AMDGPU::M0) {
4459           if (MO.getReg() != SGPRUsed)
4460             ++SGPRCount;
4461           SGPRUsed = MO.getReg();
4462         }
4463       }
4464       if (SGPRCount > ST.getConstantBusLimit(Opcode)) {
4465         ErrInfo = "WRITELANE instruction violates constant bus restriction";
4466         return false;
4467       }
4468     }
4469   }
4470 
4471   // Verify misc. restrictions on specific instructions.
4472   if (Desc.getOpcode() == AMDGPU::V_DIV_SCALE_F32_e64 ||
4473       Desc.getOpcode() == AMDGPU::V_DIV_SCALE_F64_e64) {
4474     const MachineOperand &Src0 = MI.getOperand(Src0Idx);
4475     const MachineOperand &Src1 = MI.getOperand(Src1Idx);
4476     const MachineOperand &Src2 = MI.getOperand(Src2Idx);
4477     if (Src0.isReg() && Src1.isReg() && Src2.isReg()) {
4478       if (!compareMachineOp(Src0, Src1) &&
4479           !compareMachineOp(Src0, Src2)) {
4480         ErrInfo = "v_div_scale_{f32|f64} require src0 = src1 or src2";
4481         return false;
4482       }
4483     }
4484     if ((getNamedOperand(MI, AMDGPU::OpName::src0_modifiers)->getImm() &
4485          SISrcMods::ABS) ||
4486         (getNamedOperand(MI, AMDGPU::OpName::src1_modifiers)->getImm() &
4487          SISrcMods::ABS) ||
4488         (getNamedOperand(MI, AMDGPU::OpName::src2_modifiers)->getImm() &
4489          SISrcMods::ABS)) {
4490       ErrInfo = "ABS not allowed in VOP3B instructions";
4491       return false;
4492     }
4493   }
4494 
4495   if (isSOP2(MI) || isSOPC(MI)) {
4496     const MachineOperand &Src0 = MI.getOperand(Src0Idx);
4497     const MachineOperand &Src1 = MI.getOperand(Src1Idx);
4498 
4499     if (!Src0.isReg() && !Src1.isReg() &&
4500         !isInlineConstant(Src0, Desc.operands()[Src0Idx]) &&
4501         !isInlineConstant(Src1, Desc.operands()[Src1Idx]) &&
4502         !Src0.isIdenticalTo(Src1)) {
4503       ErrInfo = "SOP2/SOPC instruction requires too many immediate constants";
4504       return false;
4505     }
4506   }
4507 
4508   if (isSOPK(MI)) {
4509     auto Op = getNamedOperand(MI, AMDGPU::OpName::simm16);
4510     if (Desc.isBranch()) {
4511       if (!Op->isMBB()) {
4512         ErrInfo = "invalid branch target for SOPK instruction";
4513         return false;
4514       }
4515     } else {
4516       uint64_t Imm = Op->getImm();
4517       if (sopkIsZext(MI)) {
4518         if (!isUInt<16>(Imm)) {
4519           ErrInfo = "invalid immediate for SOPK instruction";
4520           return false;
4521         }
4522       } else {
4523         if (!isInt<16>(Imm)) {
4524           ErrInfo = "invalid immediate for SOPK instruction";
4525           return false;
4526         }
4527       }
4528     }
4529   }
4530 
4531   if (Desc.getOpcode() == AMDGPU::V_MOVRELS_B32_e32 ||
4532       Desc.getOpcode() == AMDGPU::V_MOVRELS_B32_e64 ||
4533       Desc.getOpcode() == AMDGPU::V_MOVRELD_B32_e32 ||
4534       Desc.getOpcode() == AMDGPU::V_MOVRELD_B32_e64) {
4535     const bool IsDst = Desc.getOpcode() == AMDGPU::V_MOVRELD_B32_e32 ||
4536                        Desc.getOpcode() == AMDGPU::V_MOVRELD_B32_e64;
4537 
4538     const unsigned StaticNumOps =
4539         Desc.getNumOperands() + Desc.implicit_uses().size();
4540     const unsigned NumImplicitOps = IsDst ? 2 : 1;
4541 
4542     // Allow additional implicit operands. This allows a fixup done by the post
4543     // RA scheduler where the main implicit operand is killed and implicit-defs
4544     // are added for sub-registers that remain live after this instruction.
4545     if (MI.getNumOperands() < StaticNumOps + NumImplicitOps) {
4546       ErrInfo = "missing implicit register operands";
4547       return false;
4548     }
4549 
4550     const MachineOperand *Dst = getNamedOperand(MI, AMDGPU::OpName::vdst);
4551     if (IsDst) {
4552       if (!Dst->isUse()) {
4553         ErrInfo = "v_movreld_b32 vdst should be a use operand";
4554         return false;
4555       }
4556 
4557       unsigned UseOpIdx;
4558       if (!MI.isRegTiedToUseOperand(StaticNumOps, &UseOpIdx) ||
4559           UseOpIdx != StaticNumOps + 1) {
4560         ErrInfo = "movrel implicit operands should be tied";
4561         return false;
4562       }
4563     }
4564 
4565     const MachineOperand &Src0 = MI.getOperand(Src0Idx);
4566     const MachineOperand &ImpUse
4567       = MI.getOperand(StaticNumOps + NumImplicitOps - 1);
4568     if (!ImpUse.isReg() || !ImpUse.isUse() ||
4569         !isSubRegOf(RI, ImpUse, IsDst ? *Dst : Src0)) {
4570       ErrInfo = "src0 should be subreg of implicit vector use";
4571       return false;
4572     }
4573   }
4574 
4575   // Make sure we aren't losing exec uses in the td files. This mostly requires
4576   // being careful when using let Uses to try to add other use registers.
4577   if (shouldReadExec(MI)) {
4578     if (!MI.hasRegisterImplicitUseOperand(AMDGPU::EXEC)) {
4579       ErrInfo = "VALU instruction does not implicitly read exec mask";
4580       return false;
4581     }
4582   }
4583 
4584   if (isSMRD(MI)) {
4585     if (MI.mayStore() &&
4586         ST.getGeneration() == AMDGPUSubtarget::VOLCANIC_ISLANDS) {
4587       // The register offset form of scalar stores may only use m0 as the
4588       // soffset register.
4589       const MachineOperand *Soff = getNamedOperand(MI, AMDGPU::OpName::soffset);
4590       if (Soff && Soff->getReg() != AMDGPU::M0) {
4591         ErrInfo = "scalar stores must use m0 as offset register";
4592         return false;
4593       }
4594     }
4595   }
4596 
4597   if (isFLAT(MI) && !ST.hasFlatInstOffsets()) {
4598     const MachineOperand *Offset = getNamedOperand(MI, AMDGPU::OpName::offset);
4599     if (Offset->getImm() != 0) {
4600       ErrInfo = "subtarget does not support offsets in flat instructions";
4601       return false;
4602     }
4603   }
4604 
4605   if (isMIMG(MI)) {
4606     const MachineOperand *DimOp = getNamedOperand(MI, AMDGPU::OpName::dim);
4607     if (DimOp) {
4608       int VAddr0Idx = AMDGPU::getNamedOperandIdx(Opcode,
4609                                                  AMDGPU::OpName::vaddr0);
4610       int SRsrcIdx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::srsrc);
4611       const AMDGPU::MIMGInfo *Info = AMDGPU::getMIMGInfo(Opcode);
4612       const AMDGPU::MIMGBaseOpcodeInfo *BaseOpcode =
4613           AMDGPU::getMIMGBaseOpcodeInfo(Info->BaseOpcode);
4614       const AMDGPU::MIMGDimInfo *Dim =
4615           AMDGPU::getMIMGDimInfoByEncoding(DimOp->getImm());
4616 
4617       if (!Dim) {
4618         ErrInfo = "dim is out of range";
4619         return false;
4620       }
4621 
4622       bool IsA16 = false;
4623       if (ST.hasR128A16()) {
4624         const MachineOperand *R128A16 = getNamedOperand(MI, AMDGPU::OpName::r128);
4625         IsA16 = R128A16->getImm() != 0;
4626       } else if (ST.hasA16()) {
4627         const MachineOperand *A16 = getNamedOperand(MI, AMDGPU::OpName::a16);
4628         IsA16 = A16->getImm() != 0;
4629       }
4630 
4631       bool IsNSA = SRsrcIdx - VAddr0Idx > 1;
4632 
4633       unsigned AddrWords =
4634           AMDGPU::getAddrSizeMIMGOp(BaseOpcode, Dim, IsA16, ST.hasG16());
4635 
4636       unsigned VAddrWords;
4637       if (IsNSA) {
4638         VAddrWords = SRsrcIdx - VAddr0Idx;
4639       } else {
4640         const TargetRegisterClass *RC = getOpRegClass(MI, VAddr0Idx);
4641         VAddrWords = MRI.getTargetRegisterInfo()->getRegSizeInBits(*RC) / 32;
4642         if (AddrWords > 12)
4643           AddrWords = 16;
4644       }
4645 
4646       if (VAddrWords != AddrWords) {
4647         LLVM_DEBUG(dbgs() << "bad vaddr size, expected " << AddrWords
4648                           << " but got " << VAddrWords << "\n");
4649         ErrInfo = "bad vaddr size";
4650         return false;
4651       }
4652     }
4653   }
4654 
4655   const MachineOperand *DppCt = getNamedOperand(MI, AMDGPU::OpName::dpp_ctrl);
4656   if (DppCt) {
4657     using namespace AMDGPU::DPP;
4658 
4659     unsigned DC = DppCt->getImm();
4660     if (DC == DppCtrl::DPP_UNUSED1 || DC == DppCtrl::DPP_UNUSED2 ||
4661         DC == DppCtrl::DPP_UNUSED3 || DC > DppCtrl::DPP_LAST ||
4662         (DC >= DppCtrl::DPP_UNUSED4_FIRST && DC <= DppCtrl::DPP_UNUSED4_LAST) ||
4663         (DC >= DppCtrl::DPP_UNUSED5_FIRST && DC <= DppCtrl::DPP_UNUSED5_LAST) ||
4664         (DC >= DppCtrl::DPP_UNUSED6_FIRST && DC <= DppCtrl::DPP_UNUSED6_LAST) ||
4665         (DC >= DppCtrl::DPP_UNUSED7_FIRST && DC <= DppCtrl::DPP_UNUSED7_LAST) ||
4666         (DC >= DppCtrl::DPP_UNUSED8_FIRST && DC <= DppCtrl::DPP_UNUSED8_LAST)) {
4667       ErrInfo = "Invalid dpp_ctrl value";
4668       return false;
4669     }
4670     if (DC >= DppCtrl::WAVE_SHL1 && DC <= DppCtrl::WAVE_ROR1 &&
4671         ST.getGeneration() >= AMDGPUSubtarget::GFX10) {
4672       ErrInfo = "Invalid dpp_ctrl value: "
4673                 "wavefront shifts are not supported on GFX10+";
4674       return false;
4675     }
4676     if (DC >= DppCtrl::BCAST15 && DC <= DppCtrl::BCAST31 &&
4677         ST.getGeneration() >= AMDGPUSubtarget::GFX10) {
4678       ErrInfo = "Invalid dpp_ctrl value: "
4679                 "broadcasts are not supported on GFX10+";
4680       return false;
4681     }
4682     if (DC >= DppCtrl::ROW_SHARE_FIRST && DC <= DppCtrl::ROW_XMASK_LAST &&
4683         ST.getGeneration() < AMDGPUSubtarget::GFX10) {
4684       if (DC >= DppCtrl::ROW_NEWBCAST_FIRST &&
4685           DC <= DppCtrl::ROW_NEWBCAST_LAST &&
4686           !ST.hasGFX90AInsts()) {
4687         ErrInfo = "Invalid dpp_ctrl value: "
4688                   "row_newbroadcast/row_share is not supported before "
4689                   "GFX90A/GFX10";
4690         return false;
4691       } else if (DC > DppCtrl::ROW_NEWBCAST_LAST || !ST.hasGFX90AInsts()) {
4692         ErrInfo = "Invalid dpp_ctrl value: "
4693                   "row_share and row_xmask are not supported before GFX10";
4694         return false;
4695       }
4696     }
4697 
4698     int DstIdx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::vdst);
4699 
4700     if (Opcode != AMDGPU::V_MOV_B64_DPP_PSEUDO &&
4701         ((DstIdx >= 0 &&
4702           (Desc.operands()[DstIdx].RegClass == AMDGPU::VReg_64RegClassID ||
4703            Desc.operands()[DstIdx].RegClass ==
4704                AMDGPU::VReg_64_Align2RegClassID)) ||
4705          ((Src0Idx >= 0 &&
4706            (Desc.operands()[Src0Idx].RegClass == AMDGPU::VReg_64RegClassID ||
4707             Desc.operands()[Src0Idx].RegClass ==
4708                 AMDGPU::VReg_64_Align2RegClassID)))) &&
4709         !AMDGPU::isLegal64BitDPPControl(DC)) {
4710       ErrInfo = "Invalid dpp_ctrl value: "
4711                 "64 bit dpp only support row_newbcast";
4712       return false;
4713     }
4714   }
4715 
4716   if ((MI.mayStore() || MI.mayLoad()) && !isVGPRSpill(MI)) {
4717     const MachineOperand *Dst = getNamedOperand(MI, AMDGPU::OpName::vdst);
4718     uint16_t DataNameIdx = isDS(Opcode) ? AMDGPU::OpName::data0
4719                                         : AMDGPU::OpName::vdata;
4720     const MachineOperand *Data = getNamedOperand(MI, DataNameIdx);
4721     const MachineOperand *Data2 = getNamedOperand(MI, AMDGPU::OpName::data1);
4722     if (Data && !Data->isReg())
4723       Data = nullptr;
4724 
4725     if (ST.hasGFX90AInsts()) {
4726       if (Dst && Data &&
4727           (RI.isAGPR(MRI, Dst->getReg()) != RI.isAGPR(MRI, Data->getReg()))) {
4728         ErrInfo = "Invalid register class: "
4729                   "vdata and vdst should be both VGPR or AGPR";
4730         return false;
4731       }
4732       if (Data && Data2 &&
4733           (RI.isAGPR(MRI, Data->getReg()) != RI.isAGPR(MRI, Data2->getReg()))) {
4734         ErrInfo = "Invalid register class: "
4735                   "both data operands should be VGPR or AGPR";
4736         return false;
4737       }
4738     } else {
4739       if ((Dst && RI.isAGPR(MRI, Dst->getReg())) ||
4740           (Data && RI.isAGPR(MRI, Data->getReg())) ||
4741           (Data2 && RI.isAGPR(MRI, Data2->getReg()))) {
4742         ErrInfo = "Invalid register class: "
4743                   "agpr loads and stores not supported on this GPU";
4744         return false;
4745       }
4746     }
4747   }
4748 
4749   if (ST.needsAlignedVGPRs()) {
4750     const auto isAlignedReg = [&MI, &MRI, this](unsigned OpName) -> bool {
4751       const MachineOperand *Op = getNamedOperand(MI, OpName);
4752       if (!Op)
4753         return true;
4754       Register Reg = Op->getReg();
4755       if (Reg.isPhysical())
4756         return !(RI.getHWRegIndex(Reg) & 1);
4757       const TargetRegisterClass &RC = *MRI.getRegClass(Reg);
4758       return RI.getRegSizeInBits(RC) > 32 && RI.isProperlyAlignedRC(RC) &&
4759              !(RI.getChannelFromSubReg(Op->getSubReg()) & 1);
4760     };
4761 
4762     if (MI.getOpcode() == AMDGPU::DS_GWS_INIT ||
4763         MI.getOpcode() == AMDGPU::DS_GWS_SEMA_BR ||
4764         MI.getOpcode() == AMDGPU::DS_GWS_BARRIER) {
4765 
4766       if (!isAlignedReg(AMDGPU::OpName::data0)) {
4767         ErrInfo = "Subtarget requires even aligned vector registers "
4768                   "for DS_GWS instructions";
4769         return false;
4770       }
4771     }
4772 
4773     if (isMIMG(MI)) {
4774       if (!isAlignedReg(AMDGPU::OpName::vaddr)) {
4775         ErrInfo = "Subtarget requires even aligned vector registers "
4776                   "for vaddr operand of image instructions";
4777         return false;
4778       }
4779     }
4780   }
4781 
4782   if (MI.getOpcode() == AMDGPU::V_ACCVGPR_WRITE_B32_e64 &&
4783       !ST.hasGFX90AInsts()) {
4784     const MachineOperand *Src = getNamedOperand(MI, AMDGPU::OpName::src0);
4785     if (Src->isReg() && RI.isSGPRReg(MRI, Src->getReg())) {
4786       ErrInfo = "Invalid register class: "
4787                 "v_accvgpr_write with an SGPR is not supported on this GPU";
4788       return false;
4789     }
4790   }
4791 
4792   if (Desc.getOpcode() == AMDGPU::G_AMDGPU_WAVE_ADDRESS) {
4793     const MachineOperand &SrcOp = MI.getOperand(1);
4794     if (!SrcOp.isReg() || SrcOp.getReg().isVirtual()) {
4795       ErrInfo = "pseudo expects only physical SGPRs";
4796       return false;
4797     }
4798   }
4799 
4800   return true;
4801 }
4802 
4803 unsigned SIInstrInfo::getVALUOp(const MachineInstr &MI) const {
4804   switch (MI.getOpcode()) {
4805   default: return AMDGPU::INSTRUCTION_LIST_END;
4806   case AMDGPU::REG_SEQUENCE: return AMDGPU::REG_SEQUENCE;
4807   case AMDGPU::COPY: return AMDGPU::COPY;
4808   case AMDGPU::PHI: return AMDGPU::PHI;
4809   case AMDGPU::INSERT_SUBREG: return AMDGPU::INSERT_SUBREG;
4810   case AMDGPU::WQM: return AMDGPU::WQM;
4811   case AMDGPU::SOFT_WQM: return AMDGPU::SOFT_WQM;
4812   case AMDGPU::STRICT_WWM: return AMDGPU::STRICT_WWM;
4813   case AMDGPU::STRICT_WQM: return AMDGPU::STRICT_WQM;
4814   case AMDGPU::S_MOV_B32: {
4815     const MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
4816     return MI.getOperand(1).isReg() ||
4817            RI.isAGPR(MRI, MI.getOperand(0).getReg()) ?
4818            AMDGPU::COPY : AMDGPU::V_MOV_B32_e32;
4819   }
4820   case AMDGPU::S_ADD_I32:
4821     return ST.hasAddNoCarry() ? AMDGPU::V_ADD_U32_e64 : AMDGPU::V_ADD_CO_U32_e32;
4822   case AMDGPU::S_ADDC_U32:
4823     return AMDGPU::V_ADDC_U32_e32;
4824   case AMDGPU::S_SUB_I32:
4825     return ST.hasAddNoCarry() ? AMDGPU::V_SUB_U32_e64 : AMDGPU::V_SUB_CO_U32_e32;
4826     // FIXME: These are not consistently handled, and selected when the carry is
4827     // used.
4828   case AMDGPU::S_ADD_U32:
4829     return AMDGPU::V_ADD_CO_U32_e32;
4830   case AMDGPU::S_SUB_U32:
4831     return AMDGPU::V_SUB_CO_U32_e32;
4832   case AMDGPU::S_SUBB_U32: return AMDGPU::V_SUBB_U32_e32;
4833   case AMDGPU::S_MUL_I32: return AMDGPU::V_MUL_LO_U32_e64;
4834   case AMDGPU::S_MUL_HI_U32: return AMDGPU::V_MUL_HI_U32_e64;
4835   case AMDGPU::S_MUL_HI_I32: return AMDGPU::V_MUL_HI_I32_e64;
4836   case AMDGPU::S_AND_B32: return AMDGPU::V_AND_B32_e64;
4837   case AMDGPU::S_OR_B32: return AMDGPU::V_OR_B32_e64;
4838   case AMDGPU::S_XOR_B32: return AMDGPU::V_XOR_B32_e64;
4839   case AMDGPU::S_XNOR_B32:
4840     return ST.hasDLInsts() ? AMDGPU::V_XNOR_B32_e64 : AMDGPU::INSTRUCTION_LIST_END;
4841   case AMDGPU::S_MIN_I32: return AMDGPU::V_MIN_I32_e64;
4842   case AMDGPU::S_MIN_U32: return AMDGPU::V_MIN_U32_e64;
4843   case AMDGPU::S_MAX_I32: return AMDGPU::V_MAX_I32_e64;
4844   case AMDGPU::S_MAX_U32: return AMDGPU::V_MAX_U32_e64;
4845   case AMDGPU::S_ASHR_I32: return AMDGPU::V_ASHR_I32_e32;
4846   case AMDGPU::S_ASHR_I64: return AMDGPU::V_ASHR_I64_e64;
4847   case AMDGPU::S_LSHL_B32: return AMDGPU::V_LSHL_B32_e32;
4848   case AMDGPU::S_LSHL_B64: return AMDGPU::V_LSHL_B64_e64;
4849   case AMDGPU::S_LSHR_B32: return AMDGPU::V_LSHR_B32_e32;
4850   case AMDGPU::S_LSHR_B64: return AMDGPU::V_LSHR_B64_e64;
4851   case AMDGPU::S_SEXT_I32_I8: return AMDGPU::V_BFE_I32_e64;
4852   case AMDGPU::S_SEXT_I32_I16: return AMDGPU::V_BFE_I32_e64;
4853   case AMDGPU::S_BFE_U32: return AMDGPU::V_BFE_U32_e64;
4854   case AMDGPU::S_BFE_I32: return AMDGPU::V_BFE_I32_e64;
4855   case AMDGPU::S_BFM_B32: return AMDGPU::V_BFM_B32_e64;
4856   case AMDGPU::S_BREV_B32: return AMDGPU::V_BFREV_B32_e32;
4857   case AMDGPU::S_NOT_B32: return AMDGPU::V_NOT_B32_e32;
4858   case AMDGPU::S_NOT_B64: return AMDGPU::V_NOT_B32_e32;
4859   case AMDGPU::S_CMP_EQ_I32: return AMDGPU::V_CMP_EQ_I32_e64;
4860   case AMDGPU::S_CMP_LG_I32: return AMDGPU::V_CMP_NE_I32_e64;
4861   case AMDGPU::S_CMP_GT_I32: return AMDGPU::V_CMP_GT_I32_e64;
4862   case AMDGPU::S_CMP_GE_I32: return AMDGPU::V_CMP_GE_I32_e64;
4863   case AMDGPU::S_CMP_LT_I32: return AMDGPU::V_CMP_LT_I32_e64;
4864   case AMDGPU::S_CMP_LE_I32: return AMDGPU::V_CMP_LE_I32_e64;
4865   case AMDGPU::S_CMP_EQ_U32: return AMDGPU::V_CMP_EQ_U32_e64;
4866   case AMDGPU::S_CMP_LG_U32: return AMDGPU::V_CMP_NE_U32_e64;
4867   case AMDGPU::S_CMP_GT_U32: return AMDGPU::V_CMP_GT_U32_e64;
4868   case AMDGPU::S_CMP_GE_U32: return AMDGPU::V_CMP_GE_U32_e64;
4869   case AMDGPU::S_CMP_LT_U32: return AMDGPU::V_CMP_LT_U32_e64;
4870   case AMDGPU::S_CMP_LE_U32: return AMDGPU::V_CMP_LE_U32_e64;
4871   case AMDGPU::S_CMP_EQ_U64: return AMDGPU::V_CMP_EQ_U64_e64;
4872   case AMDGPU::S_CMP_LG_U64: return AMDGPU::V_CMP_NE_U64_e64;
4873   case AMDGPU::S_BCNT1_I32_B32: return AMDGPU::V_BCNT_U32_B32_e64;
4874   case AMDGPU::S_FF1_I32_B32: return AMDGPU::V_FFBL_B32_e32;
4875   case AMDGPU::S_FLBIT_I32_B32: return AMDGPU::V_FFBH_U32_e32;
4876   case AMDGPU::S_FLBIT_I32: return AMDGPU::V_FFBH_I32_e64;
4877   case AMDGPU::S_CBRANCH_SCC0: return AMDGPU::S_CBRANCH_VCCZ;
4878   case AMDGPU::S_CBRANCH_SCC1: return AMDGPU::S_CBRANCH_VCCNZ;
4879   }
4880   llvm_unreachable(
4881       "Unexpected scalar opcode without corresponding vector one!");
4882 }
4883 
4884 static const TargetRegisterClass *
4885 adjustAllocatableRegClass(const GCNSubtarget &ST, const SIRegisterInfo &RI,
4886                           const MachineRegisterInfo &MRI,
4887                           const MCInstrDesc &TID, unsigned RCID,
4888                           bool IsAllocatable) {
4889   if ((IsAllocatable || !ST.hasGFX90AInsts() || !MRI.reservedRegsFrozen()) &&
4890       (((TID.mayLoad() || TID.mayStore()) &&
4891         !(TID.TSFlags & SIInstrFlags::VGPRSpill)) ||
4892        (TID.TSFlags & (SIInstrFlags::DS | SIInstrFlags::MIMG)))) {
4893     switch (RCID) {
4894     case AMDGPU::AV_32RegClassID:
4895       RCID = AMDGPU::VGPR_32RegClassID;
4896       break;
4897     case AMDGPU::AV_64RegClassID:
4898       RCID = AMDGPU::VReg_64RegClassID;
4899       break;
4900     case AMDGPU::AV_96RegClassID:
4901       RCID = AMDGPU::VReg_96RegClassID;
4902       break;
4903     case AMDGPU::AV_128RegClassID:
4904       RCID = AMDGPU::VReg_128RegClassID;
4905       break;
4906     case AMDGPU::AV_160RegClassID:
4907       RCID = AMDGPU::VReg_160RegClassID;
4908       break;
4909     case AMDGPU::AV_512RegClassID:
4910       RCID = AMDGPU::VReg_512RegClassID;
4911       break;
4912     default:
4913       break;
4914     }
4915   }
4916 
4917   return RI.getProperlyAlignedRC(RI.getRegClass(RCID));
4918 }
4919 
4920 const TargetRegisterClass *SIInstrInfo::getRegClass(const MCInstrDesc &TID,
4921     unsigned OpNum, const TargetRegisterInfo *TRI,
4922     const MachineFunction &MF)
4923   const {
4924   if (OpNum >= TID.getNumOperands())
4925     return nullptr;
4926   auto RegClass = TID.operands()[OpNum].RegClass;
4927   bool IsAllocatable = false;
4928   if (TID.TSFlags & (SIInstrFlags::DS | SIInstrFlags::FLAT)) {
4929     // vdst and vdata should be both VGPR or AGPR, same for the DS instructions
4930     // with two data operands. Request register class constrained to VGPR only
4931     // of both operands present as Machine Copy Propagation can not check this
4932     // constraint and possibly other passes too.
4933     //
4934     // The check is limited to FLAT and DS because atomics in non-flat encoding
4935     // have their vdst and vdata tied to be the same register.
4936     const int VDstIdx = AMDGPU::getNamedOperandIdx(TID.Opcode,
4937                                                    AMDGPU::OpName::vdst);
4938     const int DataIdx = AMDGPU::getNamedOperandIdx(TID.Opcode,
4939         (TID.TSFlags & SIInstrFlags::DS) ? AMDGPU::OpName::data0
4940                                          : AMDGPU::OpName::vdata);
4941     if (DataIdx != -1) {
4942       IsAllocatable = VDstIdx != -1 || AMDGPU::hasNamedOperand(
4943                                            TID.Opcode, AMDGPU::OpName::data1);
4944     }
4945   }
4946   return adjustAllocatableRegClass(ST, RI, MF.getRegInfo(), TID, RegClass,
4947                                    IsAllocatable);
4948 }
4949 
4950 const TargetRegisterClass *SIInstrInfo::getOpRegClass(const MachineInstr &MI,
4951                                                       unsigned OpNo) const {
4952   const MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
4953   const MCInstrDesc &Desc = get(MI.getOpcode());
4954   if (MI.isVariadic() || OpNo >= Desc.getNumOperands() ||
4955       Desc.operands()[OpNo].RegClass == -1) {
4956     Register Reg = MI.getOperand(OpNo).getReg();
4957 
4958     if (Reg.isVirtual())
4959       return MRI.getRegClass(Reg);
4960     return RI.getPhysRegBaseClass(Reg);
4961   }
4962 
4963   unsigned RCID = Desc.operands()[OpNo].RegClass;
4964   return adjustAllocatableRegClass(ST, RI, MRI, Desc, RCID, true);
4965 }
4966 
4967 void SIInstrInfo::legalizeOpWithMove(MachineInstr &MI, unsigned OpIdx) const {
4968   MachineBasicBlock::iterator I = MI;
4969   MachineBasicBlock *MBB = MI.getParent();
4970   MachineOperand &MO = MI.getOperand(OpIdx);
4971   MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
4972   unsigned RCID = get(MI.getOpcode()).operands()[OpIdx].RegClass;
4973   const TargetRegisterClass *RC = RI.getRegClass(RCID);
4974   unsigned Size = RI.getRegSizeInBits(*RC);
4975   unsigned Opcode = (Size == 64) ? AMDGPU::V_MOV_B64_PSEUDO : AMDGPU::V_MOV_B32_e32;
4976   if (MO.isReg())
4977     Opcode = AMDGPU::COPY;
4978   else if (RI.isSGPRClass(RC))
4979     Opcode = (Size == 64) ? AMDGPU::S_MOV_B64 : AMDGPU::S_MOV_B32;
4980 
4981   const TargetRegisterClass *VRC = RI.getEquivalentVGPRClass(RC);
4982   const TargetRegisterClass *VRC64 = RI.getVGPR64Class();
4983   if (RI.getCommonSubClass(VRC64, VRC))
4984     VRC = VRC64;
4985   else
4986     VRC = &AMDGPU::VGPR_32RegClass;
4987 
4988   Register Reg = MRI.createVirtualRegister(VRC);
4989   DebugLoc DL = MBB->findDebugLoc(I);
4990   BuildMI(*MI.getParent(), I, DL, get(Opcode), Reg).add(MO);
4991   MO.ChangeToRegister(Reg, false);
4992 }
4993 
4994 unsigned SIInstrInfo::buildExtractSubReg(MachineBasicBlock::iterator MI,
4995                                          MachineRegisterInfo &MRI,
4996                                          MachineOperand &SuperReg,
4997                                          const TargetRegisterClass *SuperRC,
4998                                          unsigned SubIdx,
4999                                          const TargetRegisterClass *SubRC)
5000                                          const {
5001   MachineBasicBlock *MBB = MI->getParent();
5002   DebugLoc DL = MI->getDebugLoc();
5003   Register SubReg = MRI.createVirtualRegister(SubRC);
5004 
5005   if (SuperReg.getSubReg() == AMDGPU::NoSubRegister) {
5006     BuildMI(*MBB, MI, DL, get(TargetOpcode::COPY), SubReg)
5007       .addReg(SuperReg.getReg(), 0, SubIdx);
5008     return SubReg;
5009   }
5010 
5011   // Just in case the super register is itself a sub-register, copy it to a new
5012   // value so we don't need to worry about merging its subreg index with the
5013   // SubIdx passed to this function. The register coalescer should be able to
5014   // eliminate this extra copy.
5015   Register NewSuperReg = MRI.createVirtualRegister(SuperRC);
5016 
5017   BuildMI(*MBB, MI, DL, get(TargetOpcode::COPY), NewSuperReg)
5018     .addReg(SuperReg.getReg(), 0, SuperReg.getSubReg());
5019 
5020   BuildMI(*MBB, MI, DL, get(TargetOpcode::COPY), SubReg)
5021     .addReg(NewSuperReg, 0, SubIdx);
5022 
5023   return SubReg;
5024 }
5025 
5026 MachineOperand SIInstrInfo::buildExtractSubRegOrImm(
5027   MachineBasicBlock::iterator MII,
5028   MachineRegisterInfo &MRI,
5029   MachineOperand &Op,
5030   const TargetRegisterClass *SuperRC,
5031   unsigned SubIdx,
5032   const TargetRegisterClass *SubRC) const {
5033   if (Op.isImm()) {
5034     if (SubIdx == AMDGPU::sub0)
5035       return MachineOperand::CreateImm(static_cast<int32_t>(Op.getImm()));
5036     if (SubIdx == AMDGPU::sub1)
5037       return MachineOperand::CreateImm(static_cast<int32_t>(Op.getImm() >> 32));
5038 
5039     llvm_unreachable("Unhandled register index for immediate");
5040   }
5041 
5042   unsigned SubReg = buildExtractSubReg(MII, MRI, Op, SuperRC,
5043                                        SubIdx, SubRC);
5044   return MachineOperand::CreateReg(SubReg, false);
5045 }
5046 
5047 // Change the order of operands from (0, 1, 2) to (0, 2, 1)
5048 void SIInstrInfo::swapOperands(MachineInstr &Inst) const {
5049   assert(Inst.getNumExplicitOperands() == 3);
5050   MachineOperand Op1 = Inst.getOperand(1);
5051   Inst.removeOperand(1);
5052   Inst.addOperand(Op1);
5053 }
5054 
5055 bool SIInstrInfo::isLegalRegOperand(const MachineRegisterInfo &MRI,
5056                                     const MCOperandInfo &OpInfo,
5057                                     const MachineOperand &MO) const {
5058   if (!MO.isReg())
5059     return false;
5060 
5061   Register Reg = MO.getReg();
5062 
5063   const TargetRegisterClass *DRC = RI.getRegClass(OpInfo.RegClass);
5064   if (Reg.isPhysical())
5065     return DRC->contains(Reg);
5066 
5067   const TargetRegisterClass *RC = MRI.getRegClass(Reg);
5068 
5069   if (MO.getSubReg()) {
5070     const MachineFunction *MF = MO.getParent()->getParent()->getParent();
5071     const TargetRegisterClass *SuperRC = RI.getLargestLegalSuperClass(RC, *MF);
5072     if (!SuperRC)
5073       return false;
5074 
5075     DRC = RI.getMatchingSuperRegClass(SuperRC, DRC, MO.getSubReg());
5076     if (!DRC)
5077       return false;
5078   }
5079   return RC->hasSuperClassEq(DRC);
5080 }
5081 
5082 bool SIInstrInfo::isLegalVSrcOperand(const MachineRegisterInfo &MRI,
5083                                      const MCOperandInfo &OpInfo,
5084                                      const MachineOperand &MO) const {
5085   if (MO.isReg())
5086     return isLegalRegOperand(MRI, OpInfo, MO);
5087 
5088   // Handle non-register types that are treated like immediates.
5089   assert(MO.isImm() || MO.isTargetIndex() || MO.isFI() || MO.isGlobal());
5090   return true;
5091 }
5092 
5093 bool SIInstrInfo::isOperandLegal(const MachineInstr &MI, unsigned OpIdx,
5094                                  const MachineOperand *MO) const {
5095   const MachineFunction &MF = *MI.getParent()->getParent();
5096   const MachineRegisterInfo &MRI = MF.getRegInfo();
5097   const MCInstrDesc &InstDesc = MI.getDesc();
5098   const MCOperandInfo &OpInfo = InstDesc.operands()[OpIdx];
5099   const TargetRegisterClass *DefinedRC =
5100       OpInfo.RegClass != -1 ? RI.getRegClass(OpInfo.RegClass) : nullptr;
5101   if (!MO)
5102     MO = &MI.getOperand(OpIdx);
5103 
5104   int ConstantBusLimit = ST.getConstantBusLimit(MI.getOpcode());
5105   int LiteralLimit = !isVOP3(MI) || ST.hasVOP3Literal() ? 1 : 0;
5106   if (isVALU(MI) && usesConstantBus(MRI, *MO, OpInfo)) {
5107     if (!MO->isReg() && !isInlineConstant(*MO, OpInfo) && !LiteralLimit--)
5108       return false;
5109 
5110     SmallDenseSet<RegSubRegPair> SGPRsUsed;
5111     if (MO->isReg())
5112       SGPRsUsed.insert(RegSubRegPair(MO->getReg(), MO->getSubReg()));
5113 
5114     for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
5115       if (i == OpIdx)
5116         continue;
5117       const MachineOperand &Op = MI.getOperand(i);
5118       if (Op.isReg()) {
5119         RegSubRegPair SGPR(Op.getReg(), Op.getSubReg());
5120         if (!SGPRsUsed.count(SGPR) &&
5121             // FIXME: This can access off the end of the operands() array.
5122             usesConstantBus(MRI, Op, InstDesc.operands().begin()[i])) {
5123           if (--ConstantBusLimit <= 0)
5124             return false;
5125           SGPRsUsed.insert(SGPR);
5126         }
5127       } else if (InstDesc.operands()[i].OperandType == AMDGPU::OPERAND_KIMM32 ||
5128                  (AMDGPU::isSISrcOperand(InstDesc, i) &&
5129                   !isInlineConstant(Op, InstDesc.operands()[i]))) {
5130         if (!LiteralLimit--)
5131           return false;
5132         if (--ConstantBusLimit <= 0)
5133           return false;
5134       }
5135     }
5136   }
5137 
5138   if (MO->isReg()) {
5139     if (!DefinedRC)
5140       return OpInfo.OperandType == MCOI::OPERAND_UNKNOWN;
5141     if (!isLegalRegOperand(MRI, OpInfo, *MO))
5142       return false;
5143     bool IsAGPR = RI.isAGPR(MRI, MO->getReg());
5144     if (IsAGPR && !ST.hasMAIInsts())
5145       return false;
5146     unsigned Opc = MI.getOpcode();
5147     if (IsAGPR &&
5148         (!ST.hasGFX90AInsts() || !MRI.reservedRegsFrozen()) &&
5149         (MI.mayLoad() || MI.mayStore() || isDS(Opc) || isMIMG(Opc)))
5150       return false;
5151     // Atomics should have both vdst and vdata either vgpr or agpr.
5152     const int VDstIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::vdst);
5153     const int DataIdx = AMDGPU::getNamedOperandIdx(Opc,
5154         isDS(Opc) ? AMDGPU::OpName::data0 : AMDGPU::OpName::vdata);
5155     if ((int)OpIdx == VDstIdx && DataIdx != -1 &&
5156         MI.getOperand(DataIdx).isReg() &&
5157         RI.isAGPR(MRI, MI.getOperand(DataIdx).getReg()) != IsAGPR)
5158       return false;
5159     if ((int)OpIdx == DataIdx) {
5160       if (VDstIdx != -1 &&
5161           RI.isAGPR(MRI, MI.getOperand(VDstIdx).getReg()) != IsAGPR)
5162         return false;
5163       // DS instructions with 2 src operands also must have tied RC.
5164       const int Data1Idx = AMDGPU::getNamedOperandIdx(Opc,
5165                                                       AMDGPU::OpName::data1);
5166       if (Data1Idx != -1 && MI.getOperand(Data1Idx).isReg() &&
5167           RI.isAGPR(MRI, MI.getOperand(Data1Idx).getReg()) != IsAGPR)
5168         return false;
5169     }
5170     if (Opc == AMDGPU::V_ACCVGPR_WRITE_B32_e64 && !ST.hasGFX90AInsts() &&
5171         (int)OpIdx == AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0) &&
5172         RI.isSGPRReg(MRI, MO->getReg()))
5173       return false;
5174     return true;
5175   }
5176 
5177   // Handle non-register types that are treated like immediates.
5178   assert(MO->isImm() || MO->isTargetIndex() || MO->isFI() || MO->isGlobal());
5179 
5180   if (!DefinedRC) {
5181     // This operand expects an immediate.
5182     return true;
5183   }
5184 
5185   return isImmOperandLegal(MI, OpIdx, *MO);
5186 }
5187 
5188 void SIInstrInfo::legalizeOperandsVOP2(MachineRegisterInfo &MRI,
5189                                        MachineInstr &MI) const {
5190   unsigned Opc = MI.getOpcode();
5191   const MCInstrDesc &InstrDesc = get(Opc);
5192 
5193   int Src0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0);
5194   MachineOperand &Src0 = MI.getOperand(Src0Idx);
5195 
5196   int Src1Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src1);
5197   MachineOperand &Src1 = MI.getOperand(Src1Idx);
5198 
5199   // If there is an implicit SGPR use such as VCC use for v_addc_u32/v_subb_u32
5200   // we need to only have one constant bus use before GFX10.
5201   bool HasImplicitSGPR = findImplicitSGPRRead(MI);
5202   if (HasImplicitSGPR && ST.getConstantBusLimit(Opc) <= 1 && Src0.isReg() &&
5203       RI.isSGPRReg(MRI, Src0.getReg()))
5204     legalizeOpWithMove(MI, Src0Idx);
5205 
5206   // Special case: V_WRITELANE_B32 accepts only immediate or SGPR operands for
5207   // both the value to write (src0) and lane select (src1).  Fix up non-SGPR
5208   // src0/src1 with V_READFIRSTLANE.
5209   if (Opc == AMDGPU::V_WRITELANE_B32) {
5210     const DebugLoc &DL = MI.getDebugLoc();
5211     if (Src0.isReg() && RI.isVGPR(MRI, Src0.getReg())) {
5212       Register Reg = MRI.createVirtualRegister(&AMDGPU::SReg_32_XM0RegClass);
5213       BuildMI(*MI.getParent(), MI, DL, get(AMDGPU::V_READFIRSTLANE_B32), Reg)
5214           .add(Src0);
5215       Src0.ChangeToRegister(Reg, false);
5216     }
5217     if (Src1.isReg() && RI.isVGPR(MRI, Src1.getReg())) {
5218       Register Reg = MRI.createVirtualRegister(&AMDGPU::SReg_32_XM0RegClass);
5219       const DebugLoc &DL = MI.getDebugLoc();
5220       BuildMI(*MI.getParent(), MI, DL, get(AMDGPU::V_READFIRSTLANE_B32), Reg)
5221           .add(Src1);
5222       Src1.ChangeToRegister(Reg, false);
5223     }
5224     return;
5225   }
5226 
5227   // No VOP2 instructions support AGPRs.
5228   if (Src0.isReg() && RI.isAGPR(MRI, Src0.getReg()))
5229     legalizeOpWithMove(MI, Src0Idx);
5230 
5231   if (Src1.isReg() && RI.isAGPR(MRI, Src1.getReg()))
5232     legalizeOpWithMove(MI, Src1Idx);
5233 
5234   // VOP2 src0 instructions support all operand types, so we don't need to check
5235   // their legality. If src1 is already legal, we don't need to do anything.
5236   if (isLegalRegOperand(MRI, InstrDesc.operands()[Src1Idx], Src1))
5237     return;
5238 
5239   // Special case: V_READLANE_B32 accepts only immediate or SGPR operands for
5240   // lane select. Fix up using V_READFIRSTLANE, since we assume that the lane
5241   // select is uniform.
5242   if (Opc == AMDGPU::V_READLANE_B32 && Src1.isReg() &&
5243       RI.isVGPR(MRI, Src1.getReg())) {
5244     Register Reg = MRI.createVirtualRegister(&AMDGPU::SReg_32_XM0RegClass);
5245     const DebugLoc &DL = MI.getDebugLoc();
5246     BuildMI(*MI.getParent(), MI, DL, get(AMDGPU::V_READFIRSTLANE_B32), Reg)
5247         .add(Src1);
5248     Src1.ChangeToRegister(Reg, false);
5249     return;
5250   }
5251 
5252   // We do not use commuteInstruction here because it is too aggressive and will
5253   // commute if it is possible. We only want to commute here if it improves
5254   // legality. This can be called a fairly large number of times so don't waste
5255   // compile time pointlessly swapping and checking legality again.
5256   if (HasImplicitSGPR || !MI.isCommutable()) {
5257     legalizeOpWithMove(MI, Src1Idx);
5258     return;
5259   }
5260 
5261   // If src0 can be used as src1, commuting will make the operands legal.
5262   // Otherwise we have to give up and insert a move.
5263   //
5264   // TODO: Other immediate-like operand kinds could be commuted if there was a
5265   // MachineOperand::ChangeTo* for them.
5266   if ((!Src1.isImm() && !Src1.isReg()) ||
5267       !isLegalRegOperand(MRI, InstrDesc.operands()[Src1Idx], Src0)) {
5268     legalizeOpWithMove(MI, Src1Idx);
5269     return;
5270   }
5271 
5272   int CommutedOpc = commuteOpcode(MI);
5273   if (CommutedOpc == -1) {
5274     legalizeOpWithMove(MI, Src1Idx);
5275     return;
5276   }
5277 
5278   MI.setDesc(get(CommutedOpc));
5279 
5280   Register Src0Reg = Src0.getReg();
5281   unsigned Src0SubReg = Src0.getSubReg();
5282   bool Src0Kill = Src0.isKill();
5283 
5284   if (Src1.isImm())
5285     Src0.ChangeToImmediate(Src1.getImm());
5286   else if (Src1.isReg()) {
5287     Src0.ChangeToRegister(Src1.getReg(), false, false, Src1.isKill());
5288     Src0.setSubReg(Src1.getSubReg());
5289   } else
5290     llvm_unreachable("Should only have register or immediate operands");
5291 
5292   Src1.ChangeToRegister(Src0Reg, false, false, Src0Kill);
5293   Src1.setSubReg(Src0SubReg);
5294   fixImplicitOperands(MI);
5295 }
5296 
5297 // Legalize VOP3 operands. All operand types are supported for any operand
5298 // but only one literal constant and only starting from GFX10.
5299 void SIInstrInfo::legalizeOperandsVOP3(MachineRegisterInfo &MRI,
5300                                        MachineInstr &MI) const {
5301   unsigned Opc = MI.getOpcode();
5302 
5303   int VOP3Idx[3] = {
5304     AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0),
5305     AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src1),
5306     AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src2)
5307   };
5308 
5309   if (Opc == AMDGPU::V_PERMLANE16_B32_e64 ||
5310       Opc == AMDGPU::V_PERMLANEX16_B32_e64) {
5311     // src1 and src2 must be scalar
5312     MachineOperand &Src1 = MI.getOperand(VOP3Idx[1]);
5313     MachineOperand &Src2 = MI.getOperand(VOP3Idx[2]);
5314     const DebugLoc &DL = MI.getDebugLoc();
5315     if (Src1.isReg() && !RI.isSGPRClass(MRI.getRegClass(Src1.getReg()))) {
5316       Register Reg = MRI.createVirtualRegister(&AMDGPU::SReg_32_XM0RegClass);
5317       BuildMI(*MI.getParent(), MI, DL, get(AMDGPU::V_READFIRSTLANE_B32), Reg)
5318         .add(Src1);
5319       Src1.ChangeToRegister(Reg, false);
5320     }
5321     if (Src2.isReg() && !RI.isSGPRClass(MRI.getRegClass(Src2.getReg()))) {
5322       Register Reg = MRI.createVirtualRegister(&AMDGPU::SReg_32_XM0RegClass);
5323       BuildMI(*MI.getParent(), MI, DL, get(AMDGPU::V_READFIRSTLANE_B32), Reg)
5324         .add(Src2);
5325       Src2.ChangeToRegister(Reg, false);
5326     }
5327   }
5328 
5329   // Find the one SGPR operand we are allowed to use.
5330   int ConstantBusLimit = ST.getConstantBusLimit(Opc);
5331   int LiteralLimit = ST.hasVOP3Literal() ? 1 : 0;
5332   SmallDenseSet<unsigned> SGPRsUsed;
5333   Register SGPRReg = findUsedSGPR(MI, VOP3Idx);
5334   if (SGPRReg) {
5335     SGPRsUsed.insert(SGPRReg);
5336     --ConstantBusLimit;
5337   }
5338 
5339   for (int Idx : VOP3Idx) {
5340     if (Idx == -1)
5341       break;
5342     MachineOperand &MO = MI.getOperand(Idx);
5343 
5344     if (!MO.isReg()) {
5345       if (isInlineConstant(MO, get(Opc).operands()[Idx]))
5346         continue;
5347 
5348       if (LiteralLimit > 0 && ConstantBusLimit > 0) {
5349         --LiteralLimit;
5350         --ConstantBusLimit;
5351         continue;
5352       }
5353 
5354       --LiteralLimit;
5355       --ConstantBusLimit;
5356       legalizeOpWithMove(MI, Idx);
5357       continue;
5358     }
5359 
5360     if (RI.hasAGPRs(RI.getRegClassForReg(MRI, MO.getReg())) &&
5361         !isOperandLegal(MI, Idx, &MO)) {
5362       legalizeOpWithMove(MI, Idx);
5363       continue;
5364     }
5365 
5366     if (!RI.isSGPRClass(RI.getRegClassForReg(MRI, MO.getReg())))
5367       continue; // VGPRs are legal
5368 
5369     // We can use one SGPR in each VOP3 instruction prior to GFX10
5370     // and two starting from GFX10.
5371     if (SGPRsUsed.count(MO.getReg()))
5372       continue;
5373     if (ConstantBusLimit > 0) {
5374       SGPRsUsed.insert(MO.getReg());
5375       --ConstantBusLimit;
5376       continue;
5377     }
5378 
5379     // If we make it this far, then the operand is not legal and we must
5380     // legalize it.
5381     legalizeOpWithMove(MI, Idx);
5382   }
5383 }
5384 
5385 Register SIInstrInfo::readlaneVGPRToSGPR(Register SrcReg, MachineInstr &UseMI,
5386                                          MachineRegisterInfo &MRI) const {
5387   const TargetRegisterClass *VRC = MRI.getRegClass(SrcReg);
5388   const TargetRegisterClass *SRC = RI.getEquivalentSGPRClass(VRC);
5389   Register DstReg = MRI.createVirtualRegister(SRC);
5390   unsigned SubRegs = RI.getRegSizeInBits(*VRC) / 32;
5391 
5392   if (RI.hasAGPRs(VRC)) {
5393     VRC = RI.getEquivalentVGPRClass(VRC);
5394     Register NewSrcReg = MRI.createVirtualRegister(VRC);
5395     BuildMI(*UseMI.getParent(), UseMI, UseMI.getDebugLoc(),
5396             get(TargetOpcode::COPY), NewSrcReg)
5397         .addReg(SrcReg);
5398     SrcReg = NewSrcReg;
5399   }
5400 
5401   if (SubRegs == 1) {
5402     BuildMI(*UseMI.getParent(), UseMI, UseMI.getDebugLoc(),
5403             get(AMDGPU::V_READFIRSTLANE_B32), DstReg)
5404         .addReg(SrcReg);
5405     return DstReg;
5406   }
5407 
5408   SmallVector<Register, 8> SRegs;
5409   for (unsigned i = 0; i < SubRegs; ++i) {
5410     Register SGPR = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
5411     BuildMI(*UseMI.getParent(), UseMI, UseMI.getDebugLoc(),
5412             get(AMDGPU::V_READFIRSTLANE_B32), SGPR)
5413         .addReg(SrcReg, 0, RI.getSubRegFromChannel(i));
5414     SRegs.push_back(SGPR);
5415   }
5416 
5417   MachineInstrBuilder MIB =
5418       BuildMI(*UseMI.getParent(), UseMI, UseMI.getDebugLoc(),
5419               get(AMDGPU::REG_SEQUENCE), DstReg);
5420   for (unsigned i = 0; i < SubRegs; ++i) {
5421     MIB.addReg(SRegs[i]);
5422     MIB.addImm(RI.getSubRegFromChannel(i));
5423   }
5424   return DstReg;
5425 }
5426 
5427 void SIInstrInfo::legalizeOperandsSMRD(MachineRegisterInfo &MRI,
5428                                        MachineInstr &MI) const {
5429 
5430   // If the pointer is store in VGPRs, then we need to move them to
5431   // SGPRs using v_readfirstlane.  This is safe because we only select
5432   // loads with uniform pointers to SMRD instruction so we know the
5433   // pointer value is uniform.
5434   MachineOperand *SBase = getNamedOperand(MI, AMDGPU::OpName::sbase);
5435   if (SBase && !RI.isSGPRClass(MRI.getRegClass(SBase->getReg()))) {
5436     Register SGPR = readlaneVGPRToSGPR(SBase->getReg(), MI, MRI);
5437     SBase->setReg(SGPR);
5438   }
5439   MachineOperand *SOff = getNamedOperand(MI, AMDGPU::OpName::soffset);
5440   if (SOff && !RI.isSGPRClass(MRI.getRegClass(SOff->getReg()))) {
5441     Register SGPR = readlaneVGPRToSGPR(SOff->getReg(), MI, MRI);
5442     SOff->setReg(SGPR);
5443   }
5444 }
5445 
5446 bool SIInstrInfo::moveFlatAddrToVGPR(MachineInstr &Inst) const {
5447   unsigned Opc = Inst.getOpcode();
5448   int OldSAddrIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::saddr);
5449   if (OldSAddrIdx < 0)
5450     return false;
5451 
5452   assert(isSegmentSpecificFLAT(Inst));
5453 
5454   int NewOpc = AMDGPU::getGlobalVaddrOp(Opc);
5455   if (NewOpc < 0)
5456     NewOpc = AMDGPU::getFlatScratchInstSVfromSS(Opc);
5457   if (NewOpc < 0)
5458     return false;
5459 
5460   MachineRegisterInfo &MRI = Inst.getMF()->getRegInfo();
5461   MachineOperand &SAddr = Inst.getOperand(OldSAddrIdx);
5462   if (RI.isSGPRReg(MRI, SAddr.getReg()))
5463     return false;
5464 
5465   int NewVAddrIdx = AMDGPU::getNamedOperandIdx(NewOpc, AMDGPU::OpName::vaddr);
5466   if (NewVAddrIdx < 0)
5467     return false;
5468 
5469   int OldVAddrIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::vaddr);
5470 
5471   // Check vaddr, it shall be zero or absent.
5472   MachineInstr *VAddrDef = nullptr;
5473   if (OldVAddrIdx >= 0) {
5474     MachineOperand &VAddr = Inst.getOperand(OldVAddrIdx);
5475     VAddrDef = MRI.getUniqueVRegDef(VAddr.getReg());
5476     if (!VAddrDef || VAddrDef->getOpcode() != AMDGPU::V_MOV_B32_e32 ||
5477         !VAddrDef->getOperand(1).isImm() ||
5478         VAddrDef->getOperand(1).getImm() != 0)
5479       return false;
5480   }
5481 
5482   const MCInstrDesc &NewDesc = get(NewOpc);
5483   Inst.setDesc(NewDesc);
5484 
5485   // Callers expect iterator to be valid after this call, so modify the
5486   // instruction in place.
5487   if (OldVAddrIdx == NewVAddrIdx) {
5488     MachineOperand &NewVAddr = Inst.getOperand(NewVAddrIdx);
5489     // Clear use list from the old vaddr holding a zero register.
5490     MRI.removeRegOperandFromUseList(&NewVAddr);
5491     MRI.moveOperands(&NewVAddr, &SAddr, 1);
5492     Inst.removeOperand(OldSAddrIdx);
5493     // Update the use list with the pointer we have just moved from vaddr to
5494     // saddr position. Otherwise new vaddr will be missing from the use list.
5495     MRI.removeRegOperandFromUseList(&NewVAddr);
5496     MRI.addRegOperandToUseList(&NewVAddr);
5497   } else {
5498     assert(OldSAddrIdx == NewVAddrIdx);
5499 
5500     if (OldVAddrIdx >= 0) {
5501       int NewVDstIn = AMDGPU::getNamedOperandIdx(NewOpc,
5502                                                  AMDGPU::OpName::vdst_in);
5503 
5504       // removeOperand doesn't try to fixup tied operand indexes at it goes, so
5505       // it asserts. Untie the operands for now and retie them afterwards.
5506       if (NewVDstIn != -1) {
5507         int OldVDstIn = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::vdst_in);
5508         Inst.untieRegOperand(OldVDstIn);
5509       }
5510 
5511       Inst.removeOperand(OldVAddrIdx);
5512 
5513       if (NewVDstIn != -1) {
5514         int NewVDst = AMDGPU::getNamedOperandIdx(NewOpc, AMDGPU::OpName::vdst);
5515         Inst.tieOperands(NewVDst, NewVDstIn);
5516       }
5517     }
5518   }
5519 
5520   if (VAddrDef && MRI.use_nodbg_empty(VAddrDef->getOperand(0).getReg()))
5521     VAddrDef->eraseFromParent();
5522 
5523   return true;
5524 }
5525 
5526 // FIXME: Remove this when SelectionDAG is obsoleted.
5527 void SIInstrInfo::legalizeOperandsFLAT(MachineRegisterInfo &MRI,
5528                                        MachineInstr &MI) const {
5529   if (!isSegmentSpecificFLAT(MI))
5530     return;
5531 
5532   // Fixup SGPR operands in VGPRs. We only select these when the DAG divergence
5533   // thinks they are uniform, so a readfirstlane should be valid.
5534   MachineOperand *SAddr = getNamedOperand(MI, AMDGPU::OpName::saddr);
5535   if (!SAddr || RI.isSGPRClass(MRI.getRegClass(SAddr->getReg())))
5536     return;
5537 
5538   if (moveFlatAddrToVGPR(MI))
5539     return;
5540 
5541   Register ToSGPR = readlaneVGPRToSGPR(SAddr->getReg(), MI, MRI);
5542   SAddr->setReg(ToSGPR);
5543 }
5544 
5545 void SIInstrInfo::legalizeGenericOperand(MachineBasicBlock &InsertMBB,
5546                                          MachineBasicBlock::iterator I,
5547                                          const TargetRegisterClass *DstRC,
5548                                          MachineOperand &Op,
5549                                          MachineRegisterInfo &MRI,
5550                                          const DebugLoc &DL) const {
5551   Register OpReg = Op.getReg();
5552   unsigned OpSubReg = Op.getSubReg();
5553 
5554   const TargetRegisterClass *OpRC = RI.getSubClassWithSubReg(
5555       RI.getRegClassForReg(MRI, OpReg), OpSubReg);
5556 
5557   // Check if operand is already the correct register class.
5558   if (DstRC == OpRC)
5559     return;
5560 
5561   Register DstReg = MRI.createVirtualRegister(DstRC);
5562   auto Copy = BuildMI(InsertMBB, I, DL, get(AMDGPU::COPY), DstReg).add(Op);
5563 
5564   Op.setReg(DstReg);
5565   Op.setSubReg(0);
5566 
5567   MachineInstr *Def = MRI.getVRegDef(OpReg);
5568   if (!Def)
5569     return;
5570 
5571   // Try to eliminate the copy if it is copying an immediate value.
5572   if (Def->isMoveImmediate() && DstRC != &AMDGPU::VReg_1RegClass)
5573     FoldImmediate(*Copy, *Def, OpReg, &MRI);
5574 
5575   bool ImpDef = Def->isImplicitDef();
5576   while (!ImpDef && Def && Def->isCopy()) {
5577     if (Def->getOperand(1).getReg().isPhysical())
5578       break;
5579     Def = MRI.getUniqueVRegDef(Def->getOperand(1).getReg());
5580     ImpDef = Def && Def->isImplicitDef();
5581   }
5582   if (!RI.isSGPRClass(DstRC) && !Copy->readsRegister(AMDGPU::EXEC, &RI) &&
5583       !ImpDef)
5584     Copy.addReg(AMDGPU::EXEC, RegState::Implicit);
5585 }
5586 
5587 // Emit the actual waterfall loop, executing the wrapped instruction for each
5588 // unique value of \p Rsrc across all lanes. In the best case we execute 1
5589 // iteration, in the worst case we execute 64 (once per lane).
5590 static void
5591 emitLoadSRsrcFromVGPRLoop(const SIInstrInfo &TII, MachineRegisterInfo &MRI,
5592                           MachineBasicBlock &OrigBB, MachineBasicBlock &LoopBB,
5593                           MachineBasicBlock &BodyBB, const DebugLoc &DL,
5594                           MachineOperand &Rsrc) {
5595   MachineFunction &MF = *OrigBB.getParent();
5596   const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
5597   const SIRegisterInfo *TRI = ST.getRegisterInfo();
5598   unsigned Exec = ST.isWave32() ? AMDGPU::EXEC_LO : AMDGPU::EXEC;
5599   unsigned SaveExecOpc =
5600       ST.isWave32() ? AMDGPU::S_AND_SAVEEXEC_B32 : AMDGPU::S_AND_SAVEEXEC_B64;
5601   unsigned XorTermOpc =
5602       ST.isWave32() ? AMDGPU::S_XOR_B32_term : AMDGPU::S_XOR_B64_term;
5603   unsigned AndOpc =
5604       ST.isWave32() ? AMDGPU::S_AND_B32 : AMDGPU::S_AND_B64;
5605   const auto *BoolXExecRC = TRI->getRegClass(AMDGPU::SReg_1_XEXECRegClassID);
5606 
5607   MachineBasicBlock::iterator I = LoopBB.begin();
5608 
5609   SmallVector<Register, 8> ReadlanePieces;
5610   Register CondReg;
5611 
5612   Register VRsrc = Rsrc.getReg();
5613   unsigned VRsrcUndef = getUndefRegState(Rsrc.isUndef());
5614 
5615   unsigned RegSize = TRI->getRegSizeInBits(Rsrc.getReg(), MRI);
5616   unsigned NumSubRegs =  RegSize / 32;
5617   assert(NumSubRegs % 2 == 0 && NumSubRegs <= 32 && "Unhandled register size");
5618 
5619   for (unsigned Idx = 0; Idx < NumSubRegs; Idx += 2) {
5620 
5621     Register CurRegLo = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
5622     Register CurRegHi = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
5623 
5624     // Read the next variant <- also loop target.
5625     BuildMI(LoopBB, I, DL, TII.get(AMDGPU::V_READFIRSTLANE_B32), CurRegLo)
5626             .addReg(VRsrc, VRsrcUndef, TRI->getSubRegFromChannel(Idx));
5627 
5628     // Read the next variant <- also loop target.
5629     BuildMI(LoopBB, I, DL, TII.get(AMDGPU::V_READFIRSTLANE_B32), CurRegHi)
5630             .addReg(VRsrc, VRsrcUndef, TRI->getSubRegFromChannel(Idx + 1));
5631 
5632     ReadlanePieces.push_back(CurRegLo);
5633     ReadlanePieces.push_back(CurRegHi);
5634 
5635     // Comparison is to be done as 64-bit.
5636     Register CurReg = MRI.createVirtualRegister(&AMDGPU::SGPR_64RegClass);
5637     BuildMI(LoopBB, I, DL, TII.get(AMDGPU::REG_SEQUENCE), CurReg)
5638             .addReg(CurRegLo)
5639             .addImm(AMDGPU::sub0)
5640             .addReg(CurRegHi)
5641             .addImm(AMDGPU::sub1);
5642 
5643     Register NewCondReg = MRI.createVirtualRegister(BoolXExecRC);
5644     auto Cmp =
5645         BuildMI(LoopBB, I, DL, TII.get(AMDGPU::V_CMP_EQ_U64_e64), NewCondReg)
5646             .addReg(CurReg);
5647     if (NumSubRegs <= 2)
5648       Cmp.addReg(VRsrc);
5649     else
5650       Cmp.addReg(VRsrc, VRsrcUndef, TRI->getSubRegFromChannel(Idx, 2));
5651 
5652     // Combine the comparison results with AND.
5653     if (!CondReg) // First.
5654       CondReg = NewCondReg;
5655     else { // If not the first, we create an AND.
5656       Register AndReg = MRI.createVirtualRegister(BoolXExecRC);
5657       BuildMI(LoopBB, I, DL, TII.get(AndOpc), AndReg)
5658               .addReg(CondReg)
5659               .addReg(NewCondReg);
5660       CondReg = AndReg;
5661     }
5662   } // End for loop.
5663 
5664   auto SRsrcRC = TRI->getEquivalentSGPRClass(MRI.getRegClass(VRsrc));
5665   Register SRsrc = MRI.createVirtualRegister(SRsrcRC);
5666 
5667   // Build scalar Rsrc.
5668   auto Merge = BuildMI(LoopBB, I, DL, TII.get(AMDGPU::REG_SEQUENCE), SRsrc);
5669   unsigned Channel = 0;
5670   for (Register Piece : ReadlanePieces) {
5671     Merge.addReg(Piece)
5672          .addImm(TRI->getSubRegFromChannel(Channel++));
5673   }
5674 
5675   // Update Rsrc operand to use the SGPR Rsrc.
5676   Rsrc.setReg(SRsrc);
5677   Rsrc.setIsKill();
5678 
5679   Register SaveExec = MRI.createVirtualRegister(BoolXExecRC);
5680   MRI.setSimpleHint(SaveExec, CondReg);
5681 
5682   // Update EXEC to matching lanes, saving original to SaveExec.
5683   BuildMI(LoopBB, I, DL, TII.get(SaveExecOpc), SaveExec)
5684       .addReg(CondReg, RegState::Kill);
5685 
5686   // The original instruction is here; we insert the terminators after it.
5687   I = BodyBB.end();
5688 
5689   // Update EXEC, switch all done bits to 0 and all todo bits to 1.
5690   BuildMI(BodyBB, I, DL, TII.get(XorTermOpc), Exec)
5691       .addReg(Exec)
5692       .addReg(SaveExec);
5693 
5694   BuildMI(BodyBB, I, DL, TII.get(AMDGPU::SI_WATERFALL_LOOP)).addMBB(&LoopBB);
5695 }
5696 
5697 // Build a waterfall loop around \p MI, replacing the VGPR \p Rsrc register
5698 // with SGPRs by iterating over all unique values across all lanes.
5699 // Returns the loop basic block that now contains \p MI.
5700 static MachineBasicBlock *
5701 loadSRsrcFromVGPR(const SIInstrInfo &TII, MachineInstr &MI,
5702                   MachineOperand &Rsrc, MachineDominatorTree *MDT,
5703                   MachineBasicBlock::iterator Begin = nullptr,
5704                   MachineBasicBlock::iterator End = nullptr) {
5705   MachineBasicBlock &MBB = *MI.getParent();
5706   MachineFunction &MF = *MBB.getParent();
5707   const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
5708   const SIRegisterInfo *TRI = ST.getRegisterInfo();
5709   MachineRegisterInfo &MRI = MF.getRegInfo();
5710   if (!Begin.isValid())
5711     Begin = &MI;
5712   if (!End.isValid()) {
5713     End = &MI;
5714     ++End;
5715   }
5716   const DebugLoc &DL = MI.getDebugLoc();
5717   unsigned Exec = ST.isWave32() ? AMDGPU::EXEC_LO : AMDGPU::EXEC;
5718   unsigned MovExecOpc = ST.isWave32() ? AMDGPU::S_MOV_B32 : AMDGPU::S_MOV_B64;
5719   const auto *BoolXExecRC = TRI->getRegClass(AMDGPU::SReg_1_XEXECRegClassID);
5720 
5721   Register SaveExec = MRI.createVirtualRegister(BoolXExecRC);
5722 
5723   // Save the EXEC mask
5724   BuildMI(MBB, Begin, DL, TII.get(MovExecOpc), SaveExec).addReg(Exec);
5725 
5726   // Killed uses in the instruction we are waterfalling around will be
5727   // incorrect due to the added control-flow.
5728   MachineBasicBlock::iterator AfterMI = MI;
5729   ++AfterMI;
5730   for (auto I = Begin; I != AfterMI; I++) {
5731     for (auto &MO : I->uses()) {
5732       if (MO.isReg() && MO.isUse()) {
5733         MRI.clearKillFlags(MO.getReg());
5734       }
5735     }
5736   }
5737 
5738   // To insert the loop we need to split the block. Move everything after this
5739   // point to a new block, and insert a new empty block between the two.
5740   MachineBasicBlock *LoopBB = MF.CreateMachineBasicBlock();
5741   MachineBasicBlock *BodyBB = MF.CreateMachineBasicBlock();
5742   MachineBasicBlock *RemainderBB = MF.CreateMachineBasicBlock();
5743   MachineFunction::iterator MBBI(MBB);
5744   ++MBBI;
5745 
5746   MF.insert(MBBI, LoopBB);
5747   MF.insert(MBBI, BodyBB);
5748   MF.insert(MBBI, RemainderBB);
5749 
5750   LoopBB->addSuccessor(BodyBB);
5751   BodyBB->addSuccessor(LoopBB);
5752   BodyBB->addSuccessor(RemainderBB);
5753 
5754   // Move Begin to MI to the BodyBB, and the remainder of the block to
5755   // RemainderBB.
5756   RemainderBB->transferSuccessorsAndUpdatePHIs(&MBB);
5757   RemainderBB->splice(RemainderBB->begin(), &MBB, End, MBB.end());
5758   BodyBB->splice(BodyBB->begin(), &MBB, Begin, MBB.end());
5759 
5760   MBB.addSuccessor(LoopBB);
5761 
5762   // Update dominators. We know that MBB immediately dominates LoopBB, that
5763   // LoopBB immediately dominates BodyBB, and BodyBB immediately dominates
5764   // RemainderBB. RemainderBB immediately dominates all of the successors
5765   // transferred to it from MBB that MBB used to properly dominate.
5766   if (MDT) {
5767     MDT->addNewBlock(LoopBB, &MBB);
5768     MDT->addNewBlock(BodyBB, LoopBB);
5769     MDT->addNewBlock(RemainderBB, BodyBB);
5770     for (auto &Succ : RemainderBB->successors()) {
5771       if (MDT->properlyDominates(&MBB, Succ)) {
5772         MDT->changeImmediateDominator(Succ, RemainderBB);
5773       }
5774     }
5775   }
5776 
5777   emitLoadSRsrcFromVGPRLoop(TII, MRI, MBB, *LoopBB, *BodyBB, DL, Rsrc);
5778 
5779   // Restore the EXEC mask
5780   MachineBasicBlock::iterator First = RemainderBB->begin();
5781   BuildMI(*RemainderBB, First, DL, TII.get(MovExecOpc), Exec).addReg(SaveExec);
5782   return BodyBB;
5783 }
5784 
5785 // Extract pointer from Rsrc and return a zero-value Rsrc replacement.
5786 static std::tuple<unsigned, unsigned>
5787 extractRsrcPtr(const SIInstrInfo &TII, MachineInstr &MI, MachineOperand &Rsrc) {
5788   MachineBasicBlock &MBB = *MI.getParent();
5789   MachineFunction &MF = *MBB.getParent();
5790   MachineRegisterInfo &MRI = MF.getRegInfo();
5791 
5792   // Extract the ptr from the resource descriptor.
5793   unsigned RsrcPtr =
5794       TII.buildExtractSubReg(MI, MRI, Rsrc, &AMDGPU::VReg_128RegClass,
5795                              AMDGPU::sub0_sub1, &AMDGPU::VReg_64RegClass);
5796 
5797   // Create an empty resource descriptor
5798   Register Zero64 = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
5799   Register SRsrcFormatLo = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
5800   Register SRsrcFormatHi = MRI.createVirtualRegister(&AMDGPU::SGPR_32RegClass);
5801   Register NewSRsrc = MRI.createVirtualRegister(&AMDGPU::SGPR_128RegClass);
5802   uint64_t RsrcDataFormat = TII.getDefaultRsrcDataFormat();
5803 
5804   // Zero64 = 0
5805   BuildMI(MBB, MI, MI.getDebugLoc(), TII.get(AMDGPU::S_MOV_B64), Zero64)
5806       .addImm(0);
5807 
5808   // SRsrcFormatLo = RSRC_DATA_FORMAT{31-0}
5809   BuildMI(MBB, MI, MI.getDebugLoc(), TII.get(AMDGPU::S_MOV_B32), SRsrcFormatLo)
5810       .addImm(RsrcDataFormat & 0xFFFFFFFF);
5811 
5812   // SRsrcFormatHi = RSRC_DATA_FORMAT{63-32}
5813   BuildMI(MBB, MI, MI.getDebugLoc(), TII.get(AMDGPU::S_MOV_B32), SRsrcFormatHi)
5814       .addImm(RsrcDataFormat >> 32);
5815 
5816   // NewSRsrc = {Zero64, SRsrcFormat}
5817   BuildMI(MBB, MI, MI.getDebugLoc(), TII.get(AMDGPU::REG_SEQUENCE), NewSRsrc)
5818       .addReg(Zero64)
5819       .addImm(AMDGPU::sub0_sub1)
5820       .addReg(SRsrcFormatLo)
5821       .addImm(AMDGPU::sub2)
5822       .addReg(SRsrcFormatHi)
5823       .addImm(AMDGPU::sub3);
5824 
5825   return std::tuple(RsrcPtr, NewSRsrc);
5826 }
5827 
5828 MachineBasicBlock *
5829 SIInstrInfo::legalizeOperands(MachineInstr &MI,
5830                               MachineDominatorTree *MDT) const {
5831   MachineFunction &MF = *MI.getParent()->getParent();
5832   MachineRegisterInfo &MRI = MF.getRegInfo();
5833   MachineBasicBlock *CreatedBB = nullptr;
5834 
5835   // Legalize VOP2
5836   if (isVOP2(MI) || isVOPC(MI)) {
5837     legalizeOperandsVOP2(MRI, MI);
5838     return CreatedBB;
5839   }
5840 
5841   // Legalize VOP3
5842   if (isVOP3(MI)) {
5843     legalizeOperandsVOP3(MRI, MI);
5844     return CreatedBB;
5845   }
5846 
5847   // Legalize SMRD
5848   if (isSMRD(MI)) {
5849     legalizeOperandsSMRD(MRI, MI);
5850     return CreatedBB;
5851   }
5852 
5853   // Legalize FLAT
5854   if (isFLAT(MI)) {
5855     legalizeOperandsFLAT(MRI, MI);
5856     return CreatedBB;
5857   }
5858 
5859   // Legalize REG_SEQUENCE and PHI
5860   // The register class of the operands much be the same type as the register
5861   // class of the output.
5862   if (MI.getOpcode() == AMDGPU::PHI) {
5863     const TargetRegisterClass *RC = nullptr, *SRC = nullptr, *VRC = nullptr;
5864     for (unsigned i = 1, e = MI.getNumOperands(); i != e; i += 2) {
5865       if (!MI.getOperand(i).isReg() || !MI.getOperand(i).getReg().isVirtual())
5866         continue;
5867       const TargetRegisterClass *OpRC =
5868           MRI.getRegClass(MI.getOperand(i).getReg());
5869       if (RI.hasVectorRegisters(OpRC)) {
5870         VRC = OpRC;
5871       } else {
5872         SRC = OpRC;
5873       }
5874     }
5875 
5876     // If any of the operands are VGPR registers, then they all most be
5877     // otherwise we will create illegal VGPR->SGPR copies when legalizing
5878     // them.
5879     if (VRC || !RI.isSGPRClass(getOpRegClass(MI, 0))) {
5880       if (!VRC) {
5881         assert(SRC);
5882         if (getOpRegClass(MI, 0) == &AMDGPU::VReg_1RegClass) {
5883           VRC = &AMDGPU::VReg_1RegClass;
5884         } else
5885           VRC = RI.isAGPRClass(getOpRegClass(MI, 0))
5886                     ? RI.getEquivalentAGPRClass(SRC)
5887                     : RI.getEquivalentVGPRClass(SRC);
5888       } else {
5889         VRC = RI.isAGPRClass(getOpRegClass(MI, 0))
5890                   ? RI.getEquivalentAGPRClass(VRC)
5891                   : RI.getEquivalentVGPRClass(VRC);
5892       }
5893       RC = VRC;
5894     } else {
5895       RC = SRC;
5896     }
5897 
5898     // Update all the operands so they have the same type.
5899     for (unsigned I = 1, E = MI.getNumOperands(); I != E; I += 2) {
5900       MachineOperand &Op = MI.getOperand(I);
5901       if (!Op.isReg() || !Op.getReg().isVirtual())
5902         continue;
5903 
5904       // MI is a PHI instruction.
5905       MachineBasicBlock *InsertBB = MI.getOperand(I + 1).getMBB();
5906       MachineBasicBlock::iterator Insert = InsertBB->getFirstTerminator();
5907 
5908       // Avoid creating no-op copies with the same src and dst reg class.  These
5909       // confuse some of the machine passes.
5910       legalizeGenericOperand(*InsertBB, Insert, RC, Op, MRI, MI.getDebugLoc());
5911     }
5912   }
5913 
5914   // REG_SEQUENCE doesn't really require operand legalization, but if one has a
5915   // VGPR dest type and SGPR sources, insert copies so all operands are
5916   // VGPRs. This seems to help operand folding / the register coalescer.
5917   if (MI.getOpcode() == AMDGPU::REG_SEQUENCE) {
5918     MachineBasicBlock *MBB = MI.getParent();
5919     const TargetRegisterClass *DstRC = getOpRegClass(MI, 0);
5920     if (RI.hasVGPRs(DstRC)) {
5921       // Update all the operands so they are VGPR register classes. These may
5922       // not be the same register class because REG_SEQUENCE supports mixing
5923       // subregister index types e.g. sub0_sub1 + sub2 + sub3
5924       for (unsigned I = 1, E = MI.getNumOperands(); I != E; I += 2) {
5925         MachineOperand &Op = MI.getOperand(I);
5926         if (!Op.isReg() || !Op.getReg().isVirtual())
5927           continue;
5928 
5929         const TargetRegisterClass *OpRC = MRI.getRegClass(Op.getReg());
5930         const TargetRegisterClass *VRC = RI.getEquivalentVGPRClass(OpRC);
5931         if (VRC == OpRC)
5932           continue;
5933 
5934         legalizeGenericOperand(*MBB, MI, VRC, Op, MRI, MI.getDebugLoc());
5935         Op.setIsKill();
5936       }
5937     }
5938 
5939     return CreatedBB;
5940   }
5941 
5942   // Legalize INSERT_SUBREG
5943   // src0 must have the same register class as dst
5944   if (MI.getOpcode() == AMDGPU::INSERT_SUBREG) {
5945     Register Dst = MI.getOperand(0).getReg();
5946     Register Src0 = MI.getOperand(1).getReg();
5947     const TargetRegisterClass *DstRC = MRI.getRegClass(Dst);
5948     const TargetRegisterClass *Src0RC = MRI.getRegClass(Src0);
5949     if (DstRC != Src0RC) {
5950       MachineBasicBlock *MBB = MI.getParent();
5951       MachineOperand &Op = MI.getOperand(1);
5952       legalizeGenericOperand(*MBB, MI, DstRC, Op, MRI, MI.getDebugLoc());
5953     }
5954     return CreatedBB;
5955   }
5956 
5957   // Legalize SI_INIT_M0
5958   if (MI.getOpcode() == AMDGPU::SI_INIT_M0) {
5959     MachineOperand &Src = MI.getOperand(0);
5960     if (Src.isReg() && RI.hasVectorRegisters(MRI.getRegClass(Src.getReg())))
5961       Src.setReg(readlaneVGPRToSGPR(Src.getReg(), MI, MRI));
5962     return CreatedBB;
5963   }
5964 
5965   // Legalize MIMG and MUBUF/MTBUF for shaders.
5966   //
5967   // Shaders only generate MUBUF/MTBUF instructions via intrinsics or via
5968   // scratch memory access. In both cases, the legalization never involves
5969   // conversion to the addr64 form.
5970   if (isMIMG(MI) || (AMDGPU::isGraphics(MF.getFunction().getCallingConv()) &&
5971                      (isMUBUF(MI) || isMTBUF(MI)))) {
5972     MachineOperand *SRsrc = getNamedOperand(MI, AMDGPU::OpName::srsrc);
5973     if (SRsrc && !RI.isSGPRClass(MRI.getRegClass(SRsrc->getReg())))
5974       CreatedBB = loadSRsrcFromVGPR(*this, MI, *SRsrc, MDT);
5975 
5976     MachineOperand *SSamp = getNamedOperand(MI, AMDGPU::OpName::ssamp);
5977     if (SSamp && !RI.isSGPRClass(MRI.getRegClass(SSamp->getReg())))
5978       CreatedBB = loadSRsrcFromVGPR(*this, MI, *SSamp, MDT);
5979 
5980     return CreatedBB;
5981   }
5982 
5983   // Legalize SI_CALL
5984   if (MI.getOpcode() == AMDGPU::SI_CALL_ISEL) {
5985     MachineOperand *Dest = &MI.getOperand(0);
5986     if (!RI.isSGPRClass(MRI.getRegClass(Dest->getReg()))) {
5987       // Move everything between ADJCALLSTACKUP and ADJCALLSTACKDOWN and
5988       // following copies, we also need to move copies from and to physical
5989       // registers into the loop block.
5990       unsigned FrameSetupOpcode = getCallFrameSetupOpcode();
5991       unsigned FrameDestroyOpcode = getCallFrameDestroyOpcode();
5992 
5993       // Also move the copies to physical registers into the loop block
5994       MachineBasicBlock &MBB = *MI.getParent();
5995       MachineBasicBlock::iterator Start(&MI);
5996       while (Start->getOpcode() != FrameSetupOpcode)
5997         --Start;
5998       MachineBasicBlock::iterator End(&MI);
5999       while (End->getOpcode() != FrameDestroyOpcode)
6000         ++End;
6001       // Also include following copies of the return value
6002       ++End;
6003       while (End != MBB.end() && End->isCopy() && End->getOperand(1).isReg() &&
6004              MI.definesRegister(End->getOperand(1).getReg()))
6005         ++End;
6006       CreatedBB = loadSRsrcFromVGPR(*this, MI, *Dest, MDT, Start, End);
6007     }
6008   }
6009 
6010   // Legalize MUBUF* instructions.
6011   int RsrcIdx =
6012       AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::srsrc);
6013   if (RsrcIdx != -1) {
6014     // We have an MUBUF instruction
6015     MachineOperand *Rsrc = &MI.getOperand(RsrcIdx);
6016     unsigned RsrcRC = get(MI.getOpcode()).operands()[RsrcIdx].RegClass;
6017     if (RI.getCommonSubClass(MRI.getRegClass(Rsrc->getReg()),
6018                              RI.getRegClass(RsrcRC))) {
6019       // The operands are legal.
6020       // FIXME: We may need to legalize operands besides srsrc.
6021       return CreatedBB;
6022     }
6023 
6024     // Legalize a VGPR Rsrc.
6025     //
6026     // If the instruction is _ADDR64, we can avoid a waterfall by extracting
6027     // the base pointer from the VGPR Rsrc, adding it to the VAddr, then using
6028     // a zero-value SRsrc.
6029     //
6030     // If the instruction is _OFFSET (both idxen and offen disabled), and we
6031     // support ADDR64 instructions, we can convert to ADDR64 and do the same as
6032     // above.
6033     //
6034     // Otherwise we are on non-ADDR64 hardware, and/or we have
6035     // idxen/offen/bothen and we fall back to a waterfall loop.
6036 
6037     MachineBasicBlock &MBB = *MI.getParent();
6038 
6039     MachineOperand *VAddr = getNamedOperand(MI, AMDGPU::OpName::vaddr);
6040     if (VAddr && AMDGPU::getIfAddr64Inst(MI.getOpcode()) != -1) {
6041       // This is already an ADDR64 instruction so we need to add the pointer
6042       // extracted from the resource descriptor to the current value of VAddr.
6043       Register NewVAddrLo = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
6044       Register NewVAddrHi = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
6045       Register NewVAddr = MRI.createVirtualRegister(&AMDGPU::VReg_64RegClass);
6046 
6047       const auto *BoolXExecRC = RI.getRegClass(AMDGPU::SReg_1_XEXECRegClassID);
6048       Register CondReg0 = MRI.createVirtualRegister(BoolXExecRC);
6049       Register CondReg1 = MRI.createVirtualRegister(BoolXExecRC);
6050 
6051       unsigned RsrcPtr, NewSRsrc;
6052       std::tie(RsrcPtr, NewSRsrc) = extractRsrcPtr(*this, MI, *Rsrc);
6053 
6054       // NewVaddrLo = RsrcPtr:sub0 + VAddr:sub0
6055       const DebugLoc &DL = MI.getDebugLoc();
6056       BuildMI(MBB, MI, DL, get(AMDGPU::V_ADD_CO_U32_e64), NewVAddrLo)
6057         .addDef(CondReg0)
6058         .addReg(RsrcPtr, 0, AMDGPU::sub0)
6059         .addReg(VAddr->getReg(), 0, AMDGPU::sub0)
6060         .addImm(0);
6061 
6062       // NewVaddrHi = RsrcPtr:sub1 + VAddr:sub1
6063       BuildMI(MBB, MI, DL, get(AMDGPU::V_ADDC_U32_e64), NewVAddrHi)
6064         .addDef(CondReg1, RegState::Dead)
6065         .addReg(RsrcPtr, 0, AMDGPU::sub1)
6066         .addReg(VAddr->getReg(), 0, AMDGPU::sub1)
6067         .addReg(CondReg0, RegState::Kill)
6068         .addImm(0);
6069 
6070       // NewVaddr = {NewVaddrHi, NewVaddrLo}
6071       BuildMI(MBB, MI, MI.getDebugLoc(), get(AMDGPU::REG_SEQUENCE), NewVAddr)
6072           .addReg(NewVAddrLo)
6073           .addImm(AMDGPU::sub0)
6074           .addReg(NewVAddrHi)
6075           .addImm(AMDGPU::sub1);
6076 
6077       VAddr->setReg(NewVAddr);
6078       Rsrc->setReg(NewSRsrc);
6079     } else if (!VAddr && ST.hasAddr64()) {
6080       // This instructions is the _OFFSET variant, so we need to convert it to
6081       // ADDR64.
6082       assert(ST.getGeneration() < AMDGPUSubtarget::VOLCANIC_ISLANDS &&
6083              "FIXME: Need to emit flat atomics here");
6084 
6085       unsigned RsrcPtr, NewSRsrc;
6086       std::tie(RsrcPtr, NewSRsrc) = extractRsrcPtr(*this, MI, *Rsrc);
6087 
6088       Register NewVAddr = MRI.createVirtualRegister(&AMDGPU::VReg_64RegClass);
6089       MachineOperand *VData = getNamedOperand(MI, AMDGPU::OpName::vdata);
6090       MachineOperand *Offset = getNamedOperand(MI, AMDGPU::OpName::offset);
6091       MachineOperand *SOffset = getNamedOperand(MI, AMDGPU::OpName::soffset);
6092       unsigned Addr64Opcode = AMDGPU::getAddr64Inst(MI.getOpcode());
6093 
6094       // Atomics with return have an additional tied operand and are
6095       // missing some of the special bits.
6096       MachineOperand *VDataIn = getNamedOperand(MI, AMDGPU::OpName::vdata_in);
6097       MachineInstr *Addr64;
6098 
6099       if (!VDataIn) {
6100         // Regular buffer load / store.
6101         MachineInstrBuilder MIB =
6102             BuildMI(MBB, MI, MI.getDebugLoc(), get(Addr64Opcode))
6103                 .add(*VData)
6104                 .addReg(NewVAddr)
6105                 .addReg(NewSRsrc)
6106                 .add(*SOffset)
6107                 .add(*Offset);
6108 
6109         if (const MachineOperand *CPol =
6110                 getNamedOperand(MI, AMDGPU::OpName::cpol)) {
6111           MIB.addImm(CPol->getImm());
6112         }
6113 
6114         if (const MachineOperand *TFE =
6115                 getNamedOperand(MI, AMDGPU::OpName::tfe)) {
6116           MIB.addImm(TFE->getImm());
6117         }
6118 
6119         MIB.addImm(getNamedImmOperand(MI, AMDGPU::OpName::swz));
6120 
6121         MIB.cloneMemRefs(MI);
6122         Addr64 = MIB;
6123       } else {
6124         // Atomics with return.
6125         Addr64 = BuildMI(MBB, MI, MI.getDebugLoc(), get(Addr64Opcode))
6126                      .add(*VData)
6127                      .add(*VDataIn)
6128                      .addReg(NewVAddr)
6129                      .addReg(NewSRsrc)
6130                      .add(*SOffset)
6131                      .add(*Offset)
6132                      .addImm(getNamedImmOperand(MI, AMDGPU::OpName::cpol))
6133                      .cloneMemRefs(MI);
6134       }
6135 
6136       MI.removeFromParent();
6137 
6138       // NewVaddr = {NewVaddrHi, NewVaddrLo}
6139       BuildMI(MBB, Addr64, Addr64->getDebugLoc(), get(AMDGPU::REG_SEQUENCE),
6140               NewVAddr)
6141           .addReg(RsrcPtr, 0, AMDGPU::sub0)
6142           .addImm(AMDGPU::sub0)
6143           .addReg(RsrcPtr, 0, AMDGPU::sub1)
6144           .addImm(AMDGPU::sub1);
6145     } else {
6146       // This is another variant; legalize Rsrc with waterfall loop from VGPRs
6147       // to SGPRs.
6148       CreatedBB = loadSRsrcFromVGPR(*this, MI, *Rsrc, MDT);
6149       return CreatedBB;
6150     }
6151   }
6152   return CreatedBB;
6153 }
6154 
6155 MachineBasicBlock *SIInstrInfo::moveToVALU(MachineInstr &TopInst,
6156                                            MachineDominatorTree *MDT) const {
6157   SetVectorType Worklist;
6158   Worklist.insert(&TopInst);
6159   MachineBasicBlock *CreatedBB = nullptr;
6160   MachineBasicBlock *CreatedBBTmp = nullptr;
6161 
6162   while (!Worklist.empty()) {
6163     MachineInstr &Inst = *Worklist.pop_back_val();
6164     MachineBasicBlock *MBB = Inst.getParent();
6165     MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
6166 
6167     unsigned Opcode = Inst.getOpcode();
6168     unsigned NewOpcode = getVALUOp(Inst);
6169 
6170     // Handle some special cases
6171     switch (Opcode) {
6172     default:
6173       break;
6174     case AMDGPU::S_ADD_U64_PSEUDO:
6175     case AMDGPU::S_SUB_U64_PSEUDO:
6176       splitScalar64BitAddSub(Worklist, Inst, MDT);
6177       Inst.eraseFromParent();
6178       continue;
6179     case AMDGPU::S_ADD_I32:
6180     case AMDGPU::S_SUB_I32: {
6181       // FIXME: The u32 versions currently selected use the carry.
6182       bool Changed;
6183       std::tie(Changed, CreatedBBTmp) = moveScalarAddSub(Worklist, Inst, MDT);
6184       if (CreatedBBTmp && TopInst.getParent() == CreatedBBTmp)
6185         CreatedBB = CreatedBBTmp;
6186       if (Changed)
6187         continue;
6188 
6189       // Default handling
6190       break;
6191     }
6192     case AMDGPU::S_AND_B64:
6193       splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_AND_B32, MDT);
6194       Inst.eraseFromParent();
6195       continue;
6196 
6197     case AMDGPU::S_OR_B64:
6198       splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_OR_B32, MDT);
6199       Inst.eraseFromParent();
6200       continue;
6201 
6202     case AMDGPU::S_XOR_B64:
6203       splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_XOR_B32, MDT);
6204       Inst.eraseFromParent();
6205       continue;
6206 
6207     case AMDGPU::S_NAND_B64:
6208       splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_NAND_B32, MDT);
6209       Inst.eraseFromParent();
6210       continue;
6211 
6212     case AMDGPU::S_NOR_B64:
6213       splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_NOR_B32, MDT);
6214       Inst.eraseFromParent();
6215       continue;
6216 
6217     case AMDGPU::S_XNOR_B64:
6218       if (ST.hasDLInsts())
6219         splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_XNOR_B32, MDT);
6220       else
6221         splitScalar64BitXnor(Worklist, Inst, MDT);
6222       Inst.eraseFromParent();
6223       continue;
6224 
6225     case AMDGPU::S_ANDN2_B64:
6226       splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_ANDN2_B32, MDT);
6227       Inst.eraseFromParent();
6228       continue;
6229 
6230     case AMDGPU::S_ORN2_B64:
6231       splitScalar64BitBinaryOp(Worklist, Inst, AMDGPU::S_ORN2_B32, MDT);
6232       Inst.eraseFromParent();
6233       continue;
6234 
6235     case AMDGPU::S_BREV_B64:
6236       splitScalar64BitUnaryOp(Worklist, Inst, AMDGPU::S_BREV_B32, true);
6237       Inst.eraseFromParent();
6238       continue;
6239 
6240     case AMDGPU::S_NOT_B64:
6241       splitScalar64BitUnaryOp(Worklist, Inst, AMDGPU::S_NOT_B32);
6242       Inst.eraseFromParent();
6243       continue;
6244 
6245     case AMDGPU::S_BCNT1_I32_B64:
6246       splitScalar64BitBCNT(Worklist, Inst);
6247       Inst.eraseFromParent();
6248       continue;
6249 
6250     case AMDGPU::S_BFE_I64:
6251       splitScalar64BitBFE(Worklist, Inst);
6252       Inst.eraseFromParent();
6253       continue;
6254 
6255     case AMDGPU::S_LSHL_B32:
6256       if (ST.hasOnlyRevVALUShifts()) {
6257         NewOpcode = AMDGPU::V_LSHLREV_B32_e64;
6258         swapOperands(Inst);
6259       }
6260       break;
6261     case AMDGPU::S_ASHR_I32:
6262       if (ST.hasOnlyRevVALUShifts()) {
6263         NewOpcode = AMDGPU::V_ASHRREV_I32_e64;
6264         swapOperands(Inst);
6265       }
6266       break;
6267     case AMDGPU::S_LSHR_B32:
6268       if (ST.hasOnlyRevVALUShifts()) {
6269         NewOpcode = AMDGPU::V_LSHRREV_B32_e64;
6270         swapOperands(Inst);
6271       }
6272       break;
6273     case AMDGPU::S_LSHL_B64:
6274       if (ST.hasOnlyRevVALUShifts()) {
6275         NewOpcode = AMDGPU::V_LSHLREV_B64_e64;
6276         swapOperands(Inst);
6277       }
6278       break;
6279     case AMDGPU::S_ASHR_I64:
6280       if (ST.hasOnlyRevVALUShifts()) {
6281         NewOpcode = AMDGPU::V_ASHRREV_I64_e64;
6282         swapOperands(Inst);
6283       }
6284       break;
6285     case AMDGPU::S_LSHR_B64:
6286       if (ST.hasOnlyRevVALUShifts()) {
6287         NewOpcode = AMDGPU::V_LSHRREV_B64_e64;
6288         swapOperands(Inst);
6289       }
6290       break;
6291 
6292     case AMDGPU::S_ABS_I32:
6293       lowerScalarAbs(Worklist, Inst);
6294       Inst.eraseFromParent();
6295       continue;
6296 
6297     case AMDGPU::S_CBRANCH_SCC0:
6298     case AMDGPU::S_CBRANCH_SCC1: {
6299         // Clear unused bits of vcc
6300         Register CondReg = Inst.getOperand(1).getReg();
6301         bool IsSCC = CondReg == AMDGPU::SCC;
6302         Register VCC = RI.getVCC();
6303         Register EXEC = ST.isWave32() ? AMDGPU::EXEC_LO : AMDGPU::EXEC;
6304         unsigned Opc = ST.isWave32() ? AMDGPU::S_AND_B32 : AMDGPU::S_AND_B64;
6305         BuildMI(*MBB, Inst, Inst.getDebugLoc(), get(Opc), VCC)
6306             .addReg(EXEC)
6307             .addReg(IsSCC ? VCC : CondReg);
6308         Inst.removeOperand(1);
6309       }
6310       break;
6311 
6312     case AMDGPU::S_BFE_U64:
6313     case AMDGPU::S_BFM_B64:
6314       llvm_unreachable("Moving this op to VALU not implemented");
6315 
6316     case AMDGPU::S_PACK_LL_B32_B16:
6317     case AMDGPU::S_PACK_LH_B32_B16:
6318     case AMDGPU::S_PACK_HL_B32_B16:
6319     case AMDGPU::S_PACK_HH_B32_B16:
6320       movePackToVALU(Worklist, MRI, Inst);
6321       Inst.eraseFromParent();
6322       continue;
6323 
6324     case AMDGPU::S_XNOR_B32:
6325       lowerScalarXnor(Worklist, Inst);
6326       Inst.eraseFromParent();
6327       continue;
6328 
6329     case AMDGPU::S_NAND_B32:
6330       splitScalarNotBinop(Worklist, Inst, AMDGPU::S_AND_B32);
6331       Inst.eraseFromParent();
6332       continue;
6333 
6334     case AMDGPU::S_NOR_B32:
6335       splitScalarNotBinop(Worklist, Inst, AMDGPU::S_OR_B32);
6336       Inst.eraseFromParent();
6337       continue;
6338 
6339     case AMDGPU::S_ANDN2_B32:
6340       splitScalarBinOpN2(Worklist, Inst, AMDGPU::S_AND_B32);
6341       Inst.eraseFromParent();
6342       continue;
6343 
6344     case AMDGPU::S_ORN2_B32:
6345       splitScalarBinOpN2(Worklist, Inst, AMDGPU::S_OR_B32);
6346       Inst.eraseFromParent();
6347       continue;
6348 
6349     // TODO: remove as soon as everything is ready
6350     // to replace VGPR to SGPR copy with V_READFIRSTLANEs.
6351     // S_ADD/SUB_CO_PSEUDO as well as S_UADDO/USUBO_PSEUDO
6352     // can only be selected from the uniform SDNode.
6353     case AMDGPU::S_ADD_CO_PSEUDO:
6354     case AMDGPU::S_SUB_CO_PSEUDO: {
6355       unsigned Opc = (Inst.getOpcode() == AMDGPU::S_ADD_CO_PSEUDO)
6356                          ? AMDGPU::V_ADDC_U32_e64
6357                          : AMDGPU::V_SUBB_U32_e64;
6358       const auto *CarryRC = RI.getRegClass(AMDGPU::SReg_1_XEXECRegClassID);
6359 
6360       Register CarryInReg = Inst.getOperand(4).getReg();
6361       if (!MRI.constrainRegClass(CarryInReg, CarryRC)) {
6362         Register NewCarryReg = MRI.createVirtualRegister(CarryRC);
6363         BuildMI(*MBB, &Inst, Inst.getDebugLoc(), get(AMDGPU::COPY), NewCarryReg)
6364             .addReg(CarryInReg);
6365       }
6366 
6367       Register CarryOutReg = Inst.getOperand(1).getReg();
6368 
6369       Register DestReg = MRI.createVirtualRegister(RI.getEquivalentVGPRClass(
6370           MRI.getRegClass(Inst.getOperand(0).getReg())));
6371       MachineInstr *CarryOp =
6372           BuildMI(*MBB, &Inst, Inst.getDebugLoc(), get(Opc), DestReg)
6373               .addReg(CarryOutReg, RegState::Define)
6374               .add(Inst.getOperand(2))
6375               .add(Inst.getOperand(3))
6376               .addReg(CarryInReg)
6377               .addImm(0);
6378       CreatedBBTmp = legalizeOperands(*CarryOp);
6379       if (CreatedBBTmp && TopInst.getParent() == CreatedBBTmp)
6380         CreatedBB = CreatedBBTmp;
6381       MRI.replaceRegWith(Inst.getOperand(0).getReg(), DestReg);
6382       addUsersToMoveToVALUWorklist(DestReg, MRI, Worklist);
6383       Inst.eraseFromParent();
6384     }
6385       continue;
6386     case AMDGPU::S_UADDO_PSEUDO:
6387     case AMDGPU::S_USUBO_PSEUDO: {
6388       const DebugLoc &DL = Inst.getDebugLoc();
6389       MachineOperand &Dest0 = Inst.getOperand(0);
6390       MachineOperand &Dest1 = Inst.getOperand(1);
6391       MachineOperand &Src0 = Inst.getOperand(2);
6392       MachineOperand &Src1 = Inst.getOperand(3);
6393 
6394       unsigned Opc = (Inst.getOpcode() == AMDGPU::S_UADDO_PSEUDO)
6395                          ? AMDGPU::V_ADD_CO_U32_e64
6396                          : AMDGPU::V_SUB_CO_U32_e64;
6397       const TargetRegisterClass *NewRC =
6398           RI.getEquivalentVGPRClass(MRI.getRegClass(Dest0.getReg()));
6399       Register DestReg = MRI.createVirtualRegister(NewRC);
6400       MachineInstr *NewInstr = BuildMI(*MBB, &Inst, DL, get(Opc), DestReg)
6401                                    .addReg(Dest1.getReg(), RegState::Define)
6402                                    .add(Src0)
6403                                    .add(Src1)
6404                                    .addImm(0); // clamp bit
6405 
6406       CreatedBBTmp = legalizeOperands(*NewInstr, MDT);
6407       if (CreatedBBTmp && TopInst.getParent() == CreatedBBTmp)
6408         CreatedBB = CreatedBBTmp;
6409 
6410       MRI.replaceRegWith(Dest0.getReg(), DestReg);
6411       addUsersToMoveToVALUWorklist(NewInstr->getOperand(0).getReg(), MRI,
6412                                    Worklist);
6413       Inst.eraseFromParent();
6414     }
6415       continue;
6416 
6417     case AMDGPU::S_CSELECT_B32:
6418     case AMDGPU::S_CSELECT_B64:
6419       lowerSelect(Worklist, Inst, MDT);
6420       Inst.eraseFromParent();
6421       continue;
6422     case AMDGPU::S_CMP_EQ_I32:
6423     case AMDGPU::S_CMP_LG_I32:
6424     case AMDGPU::S_CMP_GT_I32:
6425     case AMDGPU::S_CMP_GE_I32:
6426     case AMDGPU::S_CMP_LT_I32:
6427     case AMDGPU::S_CMP_LE_I32:
6428     case AMDGPU::S_CMP_EQ_U32:
6429     case AMDGPU::S_CMP_LG_U32:
6430     case AMDGPU::S_CMP_GT_U32:
6431     case AMDGPU::S_CMP_GE_U32:
6432     case AMDGPU::S_CMP_LT_U32:
6433     case AMDGPU::S_CMP_LE_U32:
6434     case AMDGPU::S_CMP_EQ_U64:
6435     case AMDGPU::S_CMP_LG_U64: {
6436         const MCInstrDesc &NewDesc = get(NewOpcode);
6437         Register CondReg = MRI.createVirtualRegister(RI.getWaveMaskRegClass());
6438         MachineInstr *NewInstr =
6439             BuildMI(*MBB, Inst, Inst.getDebugLoc(), NewDesc, CondReg)
6440                 .add(Inst.getOperand(0))
6441                 .add(Inst.getOperand(1));
6442         legalizeOperands(*NewInstr, MDT);
6443         int SCCIdx = Inst.findRegisterDefOperandIdx(AMDGPU::SCC);
6444         MachineOperand SCCOp = Inst.getOperand(SCCIdx);
6445         addSCCDefUsersToVALUWorklist(SCCOp, Inst, Worklist, CondReg);
6446         Inst.eraseFromParent();
6447       }
6448       continue;
6449     }
6450 
6451     if (NewOpcode == AMDGPU::INSTRUCTION_LIST_END) {
6452       // We cannot move this instruction to the VALU, so we should try to
6453       // legalize its operands instead.
6454       CreatedBBTmp = legalizeOperands(Inst, MDT);
6455       if (CreatedBBTmp && TopInst.getParent() == CreatedBBTmp)
6456         CreatedBB = CreatedBBTmp;
6457       continue;
6458     }
6459 
6460     // Handle converting generic instructions like COPY-to-SGPR into
6461     // COPY-to-VGPR.
6462     if (NewOpcode == Opcode) {
6463       Register DstReg = Inst.getOperand(0).getReg();
6464       const TargetRegisterClass *NewDstRC = getDestEquivalentVGPRClass(Inst);
6465 
6466       if (Inst.isCopy() && Inst.getOperand(1).getReg().isVirtual() &&
6467           NewDstRC == RI.getRegClassForReg(MRI, Inst.getOperand(1).getReg())) {
6468         // Instead of creating a copy where src and dst are the same register
6469         // class, we just replace all uses of dst with src.  These kinds of
6470         // copies interfere with the heuristics MachineSink uses to decide
6471         // whether or not to split a critical edge.  Since the pass assumes
6472         // that copies will end up as machine instructions and not be
6473         // eliminated.
6474         addUsersToMoveToVALUWorklist(DstReg, MRI, Worklist);
6475         MRI.replaceRegWith(DstReg, Inst.getOperand(1).getReg());
6476         MRI.clearKillFlags(Inst.getOperand(1).getReg());
6477         Inst.getOperand(0).setReg(DstReg);
6478 
6479         // Make sure we don't leave around a dead VGPR->SGPR copy. Normally
6480         // these are deleted later, but at -O0 it would leave a suspicious
6481         // looking illegal copy of an undef register.
6482         for (unsigned I = Inst.getNumOperands() - 1; I != 0; --I)
6483           Inst.removeOperand(I);
6484         Inst.setDesc(get(AMDGPU::IMPLICIT_DEF));
6485         continue;
6486       }
6487 
6488       Register NewDstReg = MRI.createVirtualRegister(NewDstRC);
6489       MRI.replaceRegWith(DstReg, NewDstReg);
6490       legalizeOperands(Inst, MDT);
6491       addUsersToMoveToVALUWorklist(NewDstReg, MRI, Worklist);
6492       continue;
6493     }
6494 
6495     // Use the new VALU Opcode.
6496     auto NewInstr = BuildMI(*MBB, Inst, Inst.getDebugLoc(), get(NewOpcode))
6497                         .setMIFlags(Inst.getFlags());
6498     for (const MachineOperand &Op : Inst.explicit_operands())
6499       NewInstr->addOperand(Op);
6500 
6501     // Remove any references to SCC. Vector instructions can't read from it, and
6502     // We're just about to add the implicit use / defs of VCC, and we don't want
6503     // both.
6504     for (MachineOperand &Op : Inst.implicit_operands()) {
6505       if (Op.getReg() == AMDGPU::SCC) {
6506         // Only propagate through live-def of SCC.
6507         if (Op.isDef() && !Op.isDead())
6508           addSCCDefUsersToVALUWorklist(Op, Inst, Worklist);
6509         if (Op.isUse())
6510           addSCCDefsToVALUWorklist(NewInstr, Worklist);
6511       }
6512     }
6513 
6514     Inst.eraseFromParent();
6515 
6516     Register NewDstReg;
6517     if (NewInstr->getOperand(0).isReg() && NewInstr->getOperand(0).isDef()) {
6518       Register DstReg = NewInstr->getOperand(0).getReg();
6519       assert(DstReg.isVirtual());
6520 
6521       // Update the destination register class.
6522       const TargetRegisterClass *NewDstRC =
6523           getDestEquivalentVGPRClass(*NewInstr);
6524       assert(NewDstRC);
6525 
6526       NewDstReg = MRI.createVirtualRegister(NewDstRC);
6527       MRI.replaceRegWith(DstReg, NewDstReg);
6528     }
6529 
6530     if (Opcode == AMDGPU::S_SEXT_I32_I8 || Opcode == AMDGPU::S_SEXT_I32_I16) {
6531       // We are converting these to a BFE, so we need to add the missing
6532       // operands for the size and offset.
6533       unsigned Size = (Opcode == AMDGPU::S_SEXT_I32_I8) ? 8 : 16;
6534       NewInstr.addImm(0);
6535       NewInstr.addImm(Size);
6536     } else if (Opcode == AMDGPU::S_BCNT1_I32_B32) {
6537       // The VALU version adds the second operand to the result, so insert an
6538       // extra 0 operand.
6539       NewInstr.addImm(0);
6540     }
6541 
6542     if (Opcode == AMDGPU::S_BFE_I32 || Opcode == AMDGPU::S_BFE_U32) {
6543       const MachineOperand &OffsetWidthOp = NewInstr->getOperand(2);
6544       // If we need to move this to VGPRs, we need to unpack the second operand
6545       // back into the 2 separate ones for bit offset and width.
6546       assert(OffsetWidthOp.isImm() &&
6547              "Scalar BFE is only implemented for constant width and offset");
6548       uint32_t Imm = OffsetWidthOp.getImm();
6549 
6550       uint32_t Offset = Imm & 0x3f; // Extract bits [5:0].
6551       uint32_t BitWidth = (Imm & 0x7f0000) >> 16; // Extract bits [22:16].
6552       NewInstr->removeOperand(2);
6553       NewInstr.addImm(Offset);
6554       NewInstr.addImm(BitWidth);
6555     }
6556 
6557     fixImplicitOperands(*NewInstr);
6558 
6559     // Legalize the operands
6560     CreatedBBTmp = legalizeOperands(*NewInstr, MDT);
6561     if (CreatedBBTmp && TopInst.getParent() == CreatedBBTmp)
6562       CreatedBB = CreatedBBTmp;
6563 
6564     if (NewDstReg)
6565       addUsersToMoveToVALUWorklist(NewDstReg, MRI, Worklist);
6566   }
6567   return CreatedBB;
6568 }
6569 
6570 // Add/sub require special handling to deal with carry outs.
6571 std::pair<bool, MachineBasicBlock *>
6572 SIInstrInfo::moveScalarAddSub(SetVectorType &Worklist, MachineInstr &Inst,
6573                               MachineDominatorTree *MDT) const {
6574   if (ST.hasAddNoCarry()) {
6575     // Assume there is no user of scc since we don't select this in that case.
6576     // Since scc isn't used, it doesn't really matter if the i32 or u32 variant
6577     // is used.
6578 
6579     MachineBasicBlock &MBB = *Inst.getParent();
6580     MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
6581 
6582     Register OldDstReg = Inst.getOperand(0).getReg();
6583     Register ResultReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
6584 
6585     unsigned Opc = Inst.getOpcode();
6586     assert(Opc == AMDGPU::S_ADD_I32 || Opc == AMDGPU::S_SUB_I32);
6587 
6588     unsigned NewOpc = Opc == AMDGPU::S_ADD_I32 ?
6589       AMDGPU::V_ADD_U32_e64 : AMDGPU::V_SUB_U32_e64;
6590 
6591     assert(Inst.getOperand(3).getReg() == AMDGPU::SCC);
6592     Inst.removeOperand(3);
6593 
6594     Inst.setDesc(get(NewOpc));
6595     Inst.addOperand(MachineOperand::CreateImm(0)); // clamp bit
6596     Inst.addImplicitDefUseOperands(*MBB.getParent());
6597     MRI.replaceRegWith(OldDstReg, ResultReg);
6598     MachineBasicBlock *NewBB = legalizeOperands(Inst, MDT);
6599 
6600     addUsersToMoveToVALUWorklist(ResultReg, MRI, Worklist);
6601     return std::pair(true, NewBB);
6602   }
6603 
6604   return std::pair(false, nullptr);
6605 }
6606 
6607 void SIInstrInfo::lowerSelect(SetVectorType &Worklist, MachineInstr &Inst,
6608                               MachineDominatorTree *MDT) const {
6609 
6610   MachineBasicBlock &MBB = *Inst.getParent();
6611   MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
6612   MachineBasicBlock::iterator MII = Inst;
6613   DebugLoc DL = Inst.getDebugLoc();
6614 
6615   MachineOperand &Dest = Inst.getOperand(0);
6616   MachineOperand &Src0 = Inst.getOperand(1);
6617   MachineOperand &Src1 = Inst.getOperand(2);
6618   MachineOperand &Cond = Inst.getOperand(3);
6619 
6620   Register SCCSource = Cond.getReg();
6621   bool IsSCC = (SCCSource == AMDGPU::SCC);
6622 
6623   // If this is a trivial select where the condition is effectively not SCC
6624   // (SCCSource is a source of copy to SCC), then the select is semantically
6625   // equivalent to copying SCCSource. Hence, there is no need to create
6626   // V_CNDMASK, we can just use that and bail out.
6627   if (!IsSCC && Src0.isImm() && (Src0.getImm() == -1) && Src1.isImm() &&
6628       (Src1.getImm() == 0)) {
6629     MRI.replaceRegWith(Dest.getReg(), SCCSource);
6630     return;
6631   }
6632 
6633   const TargetRegisterClass *TC =
6634       RI.getRegClass(AMDGPU::SReg_1_XEXECRegClassID);
6635 
6636   Register CopySCC = MRI.createVirtualRegister(TC);
6637 
6638   if (IsSCC) {
6639     // Now look for the closest SCC def if it is a copy
6640     // replacing the SCCSource with the COPY source register
6641     bool CopyFound = false;
6642     for (MachineInstr &CandI :
6643          make_range(std::next(MachineBasicBlock::reverse_iterator(Inst)),
6644                     Inst.getParent()->rend())) {
6645       if (CandI.findRegisterDefOperandIdx(AMDGPU::SCC, false, false, &RI) !=
6646           -1) {
6647         if (CandI.isCopy() && CandI.getOperand(0).getReg() == AMDGPU::SCC) {
6648           BuildMI(MBB, MII, DL, get(AMDGPU::COPY), CopySCC)
6649               .addReg(CandI.getOperand(1).getReg());
6650           CopyFound = true;
6651         }
6652         break;
6653       }
6654     }
6655     if (!CopyFound) {
6656       // SCC def is not a copy
6657       // Insert a trivial select instead of creating a copy, because a copy from
6658       // SCC would semantically mean just copying a single bit, but we may need
6659       // the result to be a vector condition mask that needs preserving.
6660       unsigned Opcode = (ST.getWavefrontSize() == 64) ? AMDGPU::S_CSELECT_B64
6661                                                       : AMDGPU::S_CSELECT_B32;
6662       auto NewSelect =
6663           BuildMI(MBB, MII, DL, get(Opcode), CopySCC).addImm(-1).addImm(0);
6664       NewSelect->getOperand(3).setIsUndef(Cond.isUndef());
6665     }
6666   }
6667 
6668   Register ResultReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
6669 
6670   auto UpdatedInst =
6671       BuildMI(MBB, MII, DL, get(AMDGPU::V_CNDMASK_B32_e64), ResultReg)
6672           .addImm(0)
6673           .add(Src1) // False
6674           .addImm(0)
6675           .add(Src0) // True
6676           .addReg(IsSCC ? CopySCC : SCCSource);
6677 
6678   MRI.replaceRegWith(Dest.getReg(), ResultReg);
6679   legalizeOperands(*UpdatedInst, MDT);
6680   addUsersToMoveToVALUWorklist(ResultReg, MRI, Worklist);
6681 }
6682 
6683 void SIInstrInfo::lowerScalarAbs(SetVectorType &Worklist,
6684                                  MachineInstr &Inst) const {
6685   MachineBasicBlock &MBB = *Inst.getParent();
6686   MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
6687   MachineBasicBlock::iterator MII = Inst;
6688   DebugLoc DL = Inst.getDebugLoc();
6689 
6690   MachineOperand &Dest = Inst.getOperand(0);
6691   MachineOperand &Src = Inst.getOperand(1);
6692   Register TmpReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
6693   Register ResultReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
6694 
6695   unsigned SubOp = ST.hasAddNoCarry() ?
6696     AMDGPU::V_SUB_U32_e32 : AMDGPU::V_SUB_CO_U32_e32;
6697 
6698   BuildMI(MBB, MII, DL, get(SubOp), TmpReg)
6699     .addImm(0)
6700     .addReg(Src.getReg());
6701 
6702   BuildMI(MBB, MII, DL, get(AMDGPU::V_MAX_I32_e64), ResultReg)
6703     .addReg(Src.getReg())
6704     .addReg(TmpReg);
6705 
6706   MRI.replaceRegWith(Dest.getReg(), ResultReg);
6707   addUsersToMoveToVALUWorklist(ResultReg, MRI, Worklist);
6708 }
6709 
6710 void SIInstrInfo::lowerScalarXnor(SetVectorType &Worklist,
6711                                   MachineInstr &Inst) const {
6712   MachineBasicBlock &MBB = *Inst.getParent();
6713   MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
6714   MachineBasicBlock::iterator MII = Inst;
6715   const DebugLoc &DL = Inst.getDebugLoc();
6716 
6717   MachineOperand &Dest = Inst.getOperand(0);
6718   MachineOperand &Src0 = Inst.getOperand(1);
6719   MachineOperand &Src1 = Inst.getOperand(2);
6720 
6721   if (ST.hasDLInsts()) {
6722     Register NewDest = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
6723     legalizeGenericOperand(MBB, MII, &AMDGPU::VGPR_32RegClass, Src0, MRI, DL);
6724     legalizeGenericOperand(MBB, MII, &AMDGPU::VGPR_32RegClass, Src1, MRI, DL);
6725 
6726     BuildMI(MBB, MII, DL, get(AMDGPU::V_XNOR_B32_e64), NewDest)
6727       .add(Src0)
6728       .add(Src1);
6729 
6730     MRI.replaceRegWith(Dest.getReg(), NewDest);
6731     addUsersToMoveToVALUWorklist(NewDest, MRI, Worklist);
6732   } else {
6733     // Using the identity !(x ^ y) == (!x ^ y) == (x ^ !y), we can
6734     // invert either source and then perform the XOR. If either source is a
6735     // scalar register, then we can leave the inversion on the scalar unit to
6736     // achieve a better distribution of scalar and vector instructions.
6737     bool Src0IsSGPR = Src0.isReg() &&
6738                       RI.isSGPRClass(MRI.getRegClass(Src0.getReg()));
6739     bool Src1IsSGPR = Src1.isReg() &&
6740                       RI.isSGPRClass(MRI.getRegClass(Src1.getReg()));
6741     MachineInstr *Xor;
6742     Register Temp = MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
6743     Register NewDest = MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
6744 
6745     // Build a pair of scalar instructions and add them to the work list.
6746     // The next iteration over the work list will lower these to the vector
6747     // unit as necessary.
6748     if (Src0IsSGPR) {
6749       BuildMI(MBB, MII, DL, get(AMDGPU::S_NOT_B32), Temp).add(Src0);
6750       Xor = BuildMI(MBB, MII, DL, get(AMDGPU::S_XOR_B32), NewDest)
6751       .addReg(Temp)
6752       .add(Src1);
6753     } else if (Src1IsSGPR) {
6754       BuildMI(MBB, MII, DL, get(AMDGPU::S_NOT_B32), Temp).add(Src1);
6755       Xor = BuildMI(MBB, MII, DL, get(AMDGPU::S_XOR_B32), NewDest)
6756       .add(Src0)
6757       .addReg(Temp);
6758     } else {
6759       Xor = BuildMI(MBB, MII, DL, get(AMDGPU::S_XOR_B32), Temp)
6760         .add(Src0)
6761         .add(Src1);
6762       MachineInstr *Not =
6763           BuildMI(MBB, MII, DL, get(AMDGPU::S_NOT_B32), NewDest).addReg(Temp);
6764       Worklist.insert(Not);
6765     }
6766 
6767     MRI.replaceRegWith(Dest.getReg(), NewDest);
6768 
6769     Worklist.insert(Xor);
6770 
6771     addUsersToMoveToVALUWorklist(NewDest, MRI, Worklist);
6772   }
6773 }
6774 
6775 void SIInstrInfo::splitScalarNotBinop(SetVectorType &Worklist,
6776                                       MachineInstr &Inst,
6777                                       unsigned Opcode) const {
6778   MachineBasicBlock &MBB = *Inst.getParent();
6779   MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
6780   MachineBasicBlock::iterator MII = Inst;
6781   const DebugLoc &DL = Inst.getDebugLoc();
6782 
6783   MachineOperand &Dest = Inst.getOperand(0);
6784   MachineOperand &Src0 = Inst.getOperand(1);
6785   MachineOperand &Src1 = Inst.getOperand(2);
6786 
6787   Register NewDest = MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
6788   Register Interm = MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
6789 
6790   MachineInstr &Op = *BuildMI(MBB, MII, DL, get(Opcode), Interm)
6791     .add(Src0)
6792     .add(Src1);
6793 
6794   MachineInstr &Not = *BuildMI(MBB, MII, DL, get(AMDGPU::S_NOT_B32), NewDest)
6795     .addReg(Interm);
6796 
6797   Worklist.insert(&Op);
6798   Worklist.insert(&Not);
6799 
6800   MRI.replaceRegWith(Dest.getReg(), NewDest);
6801   addUsersToMoveToVALUWorklist(NewDest, MRI, Worklist);
6802 }
6803 
6804 void SIInstrInfo::splitScalarBinOpN2(SetVectorType& Worklist,
6805                                      MachineInstr &Inst,
6806                                      unsigned Opcode) const {
6807   MachineBasicBlock &MBB = *Inst.getParent();
6808   MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
6809   MachineBasicBlock::iterator MII = Inst;
6810   const DebugLoc &DL = Inst.getDebugLoc();
6811 
6812   MachineOperand &Dest = Inst.getOperand(0);
6813   MachineOperand &Src0 = Inst.getOperand(1);
6814   MachineOperand &Src1 = Inst.getOperand(2);
6815 
6816   Register NewDest = MRI.createVirtualRegister(&AMDGPU::SReg_32_XM0RegClass);
6817   Register Interm = MRI.createVirtualRegister(&AMDGPU::SReg_32_XM0RegClass);
6818 
6819   MachineInstr &Not = *BuildMI(MBB, MII, DL, get(AMDGPU::S_NOT_B32), Interm)
6820     .add(Src1);
6821 
6822   MachineInstr &Op = *BuildMI(MBB, MII, DL, get(Opcode), NewDest)
6823     .add(Src0)
6824     .addReg(Interm);
6825 
6826   Worklist.insert(&Not);
6827   Worklist.insert(&Op);
6828 
6829   MRI.replaceRegWith(Dest.getReg(), NewDest);
6830   addUsersToMoveToVALUWorklist(NewDest, MRI, Worklist);
6831 }
6832 
6833 void SIInstrInfo::splitScalar64BitUnaryOp(
6834     SetVectorType &Worklist, MachineInstr &Inst,
6835     unsigned Opcode, bool Swap) const {
6836   MachineBasicBlock &MBB = *Inst.getParent();
6837   MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
6838 
6839   MachineOperand &Dest = Inst.getOperand(0);
6840   MachineOperand &Src0 = Inst.getOperand(1);
6841   DebugLoc DL = Inst.getDebugLoc();
6842 
6843   MachineBasicBlock::iterator MII = Inst;
6844 
6845   const MCInstrDesc &InstDesc = get(Opcode);
6846   const TargetRegisterClass *Src0RC = Src0.isReg() ?
6847     MRI.getRegClass(Src0.getReg()) :
6848     &AMDGPU::SGPR_32RegClass;
6849 
6850   const TargetRegisterClass *Src0SubRC =
6851       RI.getSubRegisterClass(Src0RC, AMDGPU::sub0);
6852 
6853   MachineOperand SrcReg0Sub0 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
6854                                                        AMDGPU::sub0, Src0SubRC);
6855 
6856   const TargetRegisterClass *DestRC = MRI.getRegClass(Dest.getReg());
6857   const TargetRegisterClass *NewDestRC = RI.getEquivalentVGPRClass(DestRC);
6858   const TargetRegisterClass *NewDestSubRC =
6859       RI.getSubRegisterClass(NewDestRC, AMDGPU::sub0);
6860 
6861   Register DestSub0 = MRI.createVirtualRegister(NewDestSubRC);
6862   MachineInstr &LoHalf = *BuildMI(MBB, MII, DL, InstDesc, DestSub0).add(SrcReg0Sub0);
6863 
6864   MachineOperand SrcReg0Sub1 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
6865                                                        AMDGPU::sub1, Src0SubRC);
6866 
6867   Register DestSub1 = MRI.createVirtualRegister(NewDestSubRC);
6868   MachineInstr &HiHalf = *BuildMI(MBB, MII, DL, InstDesc, DestSub1).add(SrcReg0Sub1);
6869 
6870   if (Swap)
6871     std::swap(DestSub0, DestSub1);
6872 
6873   Register FullDestReg = MRI.createVirtualRegister(NewDestRC);
6874   BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), FullDestReg)
6875     .addReg(DestSub0)
6876     .addImm(AMDGPU::sub0)
6877     .addReg(DestSub1)
6878     .addImm(AMDGPU::sub1);
6879 
6880   MRI.replaceRegWith(Dest.getReg(), FullDestReg);
6881 
6882   Worklist.insert(&LoHalf);
6883   Worklist.insert(&HiHalf);
6884 
6885   // We don't need to legalizeOperands here because for a single operand, src0
6886   // will support any kind of input.
6887 
6888   // Move all users of this moved value.
6889   addUsersToMoveToVALUWorklist(FullDestReg, MRI, Worklist);
6890 }
6891 
6892 void SIInstrInfo::splitScalar64BitAddSub(SetVectorType &Worklist,
6893                                          MachineInstr &Inst,
6894                                          MachineDominatorTree *MDT) const {
6895   bool IsAdd = (Inst.getOpcode() == AMDGPU::S_ADD_U64_PSEUDO);
6896 
6897   MachineBasicBlock &MBB = *Inst.getParent();
6898   MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
6899   const auto *CarryRC = RI.getRegClass(AMDGPU::SReg_1_XEXECRegClassID);
6900 
6901   Register FullDestReg = MRI.createVirtualRegister(&AMDGPU::VReg_64RegClass);
6902   Register DestSub0 = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
6903   Register DestSub1 = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
6904 
6905   Register CarryReg = MRI.createVirtualRegister(CarryRC);
6906   Register DeadCarryReg = MRI.createVirtualRegister(CarryRC);
6907 
6908   MachineOperand &Dest = Inst.getOperand(0);
6909   MachineOperand &Src0 = Inst.getOperand(1);
6910   MachineOperand &Src1 = Inst.getOperand(2);
6911   const DebugLoc &DL = Inst.getDebugLoc();
6912   MachineBasicBlock::iterator MII = Inst;
6913 
6914   const TargetRegisterClass *Src0RC = MRI.getRegClass(Src0.getReg());
6915   const TargetRegisterClass *Src1RC = MRI.getRegClass(Src1.getReg());
6916   const TargetRegisterClass *Src0SubRC =
6917       RI.getSubRegisterClass(Src0RC, AMDGPU::sub0);
6918   const TargetRegisterClass *Src1SubRC =
6919       RI.getSubRegisterClass(Src1RC, AMDGPU::sub0);
6920 
6921   MachineOperand SrcReg0Sub0 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
6922                                                        AMDGPU::sub0, Src0SubRC);
6923   MachineOperand SrcReg1Sub0 = buildExtractSubRegOrImm(MII, MRI, Src1, Src1RC,
6924                                                        AMDGPU::sub0, Src1SubRC);
6925 
6926 
6927   MachineOperand SrcReg0Sub1 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
6928                                                        AMDGPU::sub1, Src0SubRC);
6929   MachineOperand SrcReg1Sub1 = buildExtractSubRegOrImm(MII, MRI, Src1, Src1RC,
6930                                                        AMDGPU::sub1, Src1SubRC);
6931 
6932   unsigned LoOpc = IsAdd ? AMDGPU::V_ADD_CO_U32_e64 : AMDGPU::V_SUB_CO_U32_e64;
6933   MachineInstr *LoHalf =
6934     BuildMI(MBB, MII, DL, get(LoOpc), DestSub0)
6935     .addReg(CarryReg, RegState::Define)
6936     .add(SrcReg0Sub0)
6937     .add(SrcReg1Sub0)
6938     .addImm(0); // clamp bit
6939 
6940   unsigned HiOpc = IsAdd ? AMDGPU::V_ADDC_U32_e64 : AMDGPU::V_SUBB_U32_e64;
6941   MachineInstr *HiHalf =
6942     BuildMI(MBB, MII, DL, get(HiOpc), DestSub1)
6943     .addReg(DeadCarryReg, RegState::Define | RegState::Dead)
6944     .add(SrcReg0Sub1)
6945     .add(SrcReg1Sub1)
6946     .addReg(CarryReg, RegState::Kill)
6947     .addImm(0); // clamp bit
6948 
6949   BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), FullDestReg)
6950     .addReg(DestSub0)
6951     .addImm(AMDGPU::sub0)
6952     .addReg(DestSub1)
6953     .addImm(AMDGPU::sub1);
6954 
6955   MRI.replaceRegWith(Dest.getReg(), FullDestReg);
6956 
6957   // Try to legalize the operands in case we need to swap the order to keep it
6958   // valid.
6959   legalizeOperands(*LoHalf, MDT);
6960   legalizeOperands(*HiHalf, MDT);
6961 
6962   // Move all users of this moved value.
6963   addUsersToMoveToVALUWorklist(FullDestReg, MRI, Worklist);
6964 }
6965 
6966 void SIInstrInfo::splitScalar64BitBinaryOp(SetVectorType &Worklist,
6967                                            MachineInstr &Inst, unsigned Opcode,
6968                                            MachineDominatorTree *MDT) const {
6969   MachineBasicBlock &MBB = *Inst.getParent();
6970   MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
6971 
6972   MachineOperand &Dest = Inst.getOperand(0);
6973   MachineOperand &Src0 = Inst.getOperand(1);
6974   MachineOperand &Src1 = Inst.getOperand(2);
6975   DebugLoc DL = Inst.getDebugLoc();
6976 
6977   MachineBasicBlock::iterator MII = Inst;
6978 
6979   const MCInstrDesc &InstDesc = get(Opcode);
6980   const TargetRegisterClass *Src0RC = Src0.isReg() ?
6981     MRI.getRegClass(Src0.getReg()) :
6982     &AMDGPU::SGPR_32RegClass;
6983 
6984   const TargetRegisterClass *Src0SubRC =
6985       RI.getSubRegisterClass(Src0RC, AMDGPU::sub0);
6986   const TargetRegisterClass *Src1RC = Src1.isReg() ?
6987     MRI.getRegClass(Src1.getReg()) :
6988     &AMDGPU::SGPR_32RegClass;
6989 
6990   const TargetRegisterClass *Src1SubRC =
6991       RI.getSubRegisterClass(Src1RC, AMDGPU::sub0);
6992 
6993   MachineOperand SrcReg0Sub0 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
6994                                                        AMDGPU::sub0, Src0SubRC);
6995   MachineOperand SrcReg1Sub0 = buildExtractSubRegOrImm(MII, MRI, Src1, Src1RC,
6996                                                        AMDGPU::sub0, Src1SubRC);
6997   MachineOperand SrcReg0Sub1 = buildExtractSubRegOrImm(MII, MRI, Src0, Src0RC,
6998                                                        AMDGPU::sub1, Src0SubRC);
6999   MachineOperand SrcReg1Sub1 = buildExtractSubRegOrImm(MII, MRI, Src1, Src1RC,
7000                                                        AMDGPU::sub1, Src1SubRC);
7001 
7002   const TargetRegisterClass *DestRC = MRI.getRegClass(Dest.getReg());
7003   const TargetRegisterClass *NewDestRC = RI.getEquivalentVGPRClass(DestRC);
7004   const TargetRegisterClass *NewDestSubRC =
7005       RI.getSubRegisterClass(NewDestRC, AMDGPU::sub0);
7006 
7007   Register DestSub0 = MRI.createVirtualRegister(NewDestSubRC);
7008   MachineInstr &LoHalf = *BuildMI(MBB, MII, DL, InstDesc, DestSub0)
7009                               .add(SrcReg0Sub0)
7010                               .add(SrcReg1Sub0);
7011 
7012   Register DestSub1 = MRI.createVirtualRegister(NewDestSubRC);
7013   MachineInstr &HiHalf = *BuildMI(MBB, MII, DL, InstDesc, DestSub1)
7014                               .add(SrcReg0Sub1)
7015                               .add(SrcReg1Sub1);
7016 
7017   Register FullDestReg = MRI.createVirtualRegister(NewDestRC);
7018   BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), FullDestReg)
7019     .addReg(DestSub0)
7020     .addImm(AMDGPU::sub0)
7021     .addReg(DestSub1)
7022     .addImm(AMDGPU::sub1);
7023 
7024   MRI.replaceRegWith(Dest.getReg(), FullDestReg);
7025 
7026   Worklist.insert(&LoHalf);
7027   Worklist.insert(&HiHalf);
7028 
7029   // Move all users of this moved value.
7030   addUsersToMoveToVALUWorklist(FullDestReg, MRI, Worklist);
7031 }
7032 
7033 void SIInstrInfo::splitScalar64BitXnor(SetVectorType &Worklist,
7034                                        MachineInstr &Inst,
7035                                        MachineDominatorTree *MDT) const {
7036   MachineBasicBlock &MBB = *Inst.getParent();
7037   MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
7038 
7039   MachineOperand &Dest = Inst.getOperand(0);
7040   MachineOperand &Src0 = Inst.getOperand(1);
7041   MachineOperand &Src1 = Inst.getOperand(2);
7042   const DebugLoc &DL = Inst.getDebugLoc();
7043 
7044   MachineBasicBlock::iterator MII = Inst;
7045 
7046   const TargetRegisterClass *DestRC = MRI.getRegClass(Dest.getReg());
7047 
7048   Register Interm = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
7049 
7050   MachineOperand* Op0;
7051   MachineOperand* Op1;
7052 
7053   if (Src0.isReg() && RI.isSGPRReg(MRI, Src0.getReg())) {
7054     Op0 = &Src0;
7055     Op1 = &Src1;
7056   } else {
7057     Op0 = &Src1;
7058     Op1 = &Src0;
7059   }
7060 
7061   BuildMI(MBB, MII, DL, get(AMDGPU::S_NOT_B64), Interm)
7062     .add(*Op0);
7063 
7064   Register NewDest = MRI.createVirtualRegister(DestRC);
7065 
7066   MachineInstr &Xor = *BuildMI(MBB, MII, DL, get(AMDGPU::S_XOR_B64), NewDest)
7067     .addReg(Interm)
7068     .add(*Op1);
7069 
7070   MRI.replaceRegWith(Dest.getReg(), NewDest);
7071 
7072   Worklist.insert(&Xor);
7073 }
7074 
7075 void SIInstrInfo::splitScalar64BitBCNT(
7076     SetVectorType &Worklist, MachineInstr &Inst) const {
7077   MachineBasicBlock &MBB = *Inst.getParent();
7078   MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
7079 
7080   MachineBasicBlock::iterator MII = Inst;
7081   const DebugLoc &DL = Inst.getDebugLoc();
7082 
7083   MachineOperand &Dest = Inst.getOperand(0);
7084   MachineOperand &Src = Inst.getOperand(1);
7085 
7086   const MCInstrDesc &InstDesc = get(AMDGPU::V_BCNT_U32_B32_e64);
7087   const TargetRegisterClass *SrcRC = Src.isReg() ?
7088     MRI.getRegClass(Src.getReg()) :
7089     &AMDGPU::SGPR_32RegClass;
7090 
7091   Register MidReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
7092   Register ResultReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
7093 
7094   const TargetRegisterClass *SrcSubRC =
7095       RI.getSubRegisterClass(SrcRC, AMDGPU::sub0);
7096 
7097   MachineOperand SrcRegSub0 = buildExtractSubRegOrImm(MII, MRI, Src, SrcRC,
7098                                                       AMDGPU::sub0, SrcSubRC);
7099   MachineOperand SrcRegSub1 = buildExtractSubRegOrImm(MII, MRI, Src, SrcRC,
7100                                                       AMDGPU::sub1, SrcSubRC);
7101 
7102   BuildMI(MBB, MII, DL, InstDesc, MidReg).add(SrcRegSub0).addImm(0);
7103 
7104   BuildMI(MBB, MII, DL, InstDesc, ResultReg).add(SrcRegSub1).addReg(MidReg);
7105 
7106   MRI.replaceRegWith(Dest.getReg(), ResultReg);
7107 
7108   // We don't need to legalize operands here. src0 for either instruction can be
7109   // an SGPR, and the second input is unused or determined here.
7110   addUsersToMoveToVALUWorklist(ResultReg, MRI, Worklist);
7111 }
7112 
7113 void SIInstrInfo::splitScalar64BitBFE(SetVectorType &Worklist,
7114                                       MachineInstr &Inst) const {
7115   MachineBasicBlock &MBB = *Inst.getParent();
7116   MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
7117   MachineBasicBlock::iterator MII = Inst;
7118   const DebugLoc &DL = Inst.getDebugLoc();
7119 
7120   MachineOperand &Dest = Inst.getOperand(0);
7121   uint32_t Imm = Inst.getOperand(2).getImm();
7122   uint32_t Offset = Imm & 0x3f; // Extract bits [5:0].
7123   uint32_t BitWidth = (Imm & 0x7f0000) >> 16; // Extract bits [22:16].
7124 
7125   (void) Offset;
7126 
7127   // Only sext_inreg cases handled.
7128   assert(Inst.getOpcode() == AMDGPU::S_BFE_I64 && BitWidth <= 32 &&
7129          Offset == 0 && "Not implemented");
7130 
7131   if (BitWidth < 32) {
7132     Register MidRegLo = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
7133     Register MidRegHi = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
7134     Register ResultReg = MRI.createVirtualRegister(&AMDGPU::VReg_64RegClass);
7135 
7136     BuildMI(MBB, MII, DL, get(AMDGPU::V_BFE_I32_e64), MidRegLo)
7137         .addReg(Inst.getOperand(1).getReg(), 0, AMDGPU::sub0)
7138         .addImm(0)
7139         .addImm(BitWidth);
7140 
7141     BuildMI(MBB, MII, DL, get(AMDGPU::V_ASHRREV_I32_e32), MidRegHi)
7142       .addImm(31)
7143       .addReg(MidRegLo);
7144 
7145     BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), ResultReg)
7146       .addReg(MidRegLo)
7147       .addImm(AMDGPU::sub0)
7148       .addReg(MidRegHi)
7149       .addImm(AMDGPU::sub1);
7150 
7151     MRI.replaceRegWith(Dest.getReg(), ResultReg);
7152     addUsersToMoveToVALUWorklist(ResultReg, MRI, Worklist);
7153     return;
7154   }
7155 
7156   MachineOperand &Src = Inst.getOperand(1);
7157   Register TmpReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
7158   Register ResultReg = MRI.createVirtualRegister(&AMDGPU::VReg_64RegClass);
7159 
7160   BuildMI(MBB, MII, DL, get(AMDGPU::V_ASHRREV_I32_e64), TmpReg)
7161     .addImm(31)
7162     .addReg(Src.getReg(), 0, AMDGPU::sub0);
7163 
7164   BuildMI(MBB, MII, DL, get(TargetOpcode::REG_SEQUENCE), ResultReg)
7165     .addReg(Src.getReg(), 0, AMDGPU::sub0)
7166     .addImm(AMDGPU::sub0)
7167     .addReg(TmpReg)
7168     .addImm(AMDGPU::sub1);
7169 
7170   MRI.replaceRegWith(Dest.getReg(), ResultReg);
7171   addUsersToMoveToVALUWorklist(ResultReg, MRI, Worklist);
7172 }
7173 
7174 void SIInstrInfo::addUsersToMoveToVALUWorklist(
7175   Register DstReg,
7176   MachineRegisterInfo &MRI,
7177   SetVectorType &Worklist) const {
7178   for (MachineRegisterInfo::use_iterator I = MRI.use_begin(DstReg),
7179          E = MRI.use_end(); I != E;) {
7180     MachineInstr &UseMI = *I->getParent();
7181 
7182     unsigned OpNo = 0;
7183 
7184     switch (UseMI.getOpcode()) {
7185     case AMDGPU::COPY:
7186     case AMDGPU::WQM:
7187     case AMDGPU::SOFT_WQM:
7188     case AMDGPU::STRICT_WWM:
7189     case AMDGPU::STRICT_WQM:
7190     case AMDGPU::REG_SEQUENCE:
7191     case AMDGPU::PHI:
7192     case AMDGPU::INSERT_SUBREG:
7193       break;
7194     default:
7195       OpNo = I.getOperandNo();
7196       break;
7197     }
7198 
7199     if (!RI.hasVectorRegisters(getOpRegClass(UseMI, OpNo))) {
7200       Worklist.insert(&UseMI);
7201 
7202       do {
7203         ++I;
7204       } while (I != E && I->getParent() == &UseMI);
7205     } else {
7206       ++I;
7207     }
7208   }
7209 }
7210 
7211 void SIInstrInfo::movePackToVALU(SetVectorType &Worklist,
7212                                  MachineRegisterInfo &MRI,
7213                                  MachineInstr &Inst) const {
7214   Register ResultReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
7215   MachineBasicBlock *MBB = Inst.getParent();
7216   MachineOperand &Src0 = Inst.getOperand(1);
7217   MachineOperand &Src1 = Inst.getOperand(2);
7218   const DebugLoc &DL = Inst.getDebugLoc();
7219 
7220   switch (Inst.getOpcode()) {
7221   case AMDGPU::S_PACK_LL_B32_B16: {
7222     Register ImmReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
7223     Register TmpReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
7224 
7225     // FIXME: Can do a lot better if we know the high bits of src0 or src1 are
7226     // 0.
7227     BuildMI(*MBB, Inst, DL, get(AMDGPU::V_MOV_B32_e32), ImmReg)
7228       .addImm(0xffff);
7229 
7230     BuildMI(*MBB, Inst, DL, get(AMDGPU::V_AND_B32_e64), TmpReg)
7231       .addReg(ImmReg, RegState::Kill)
7232       .add(Src0);
7233 
7234     BuildMI(*MBB, Inst, DL, get(AMDGPU::V_LSHL_OR_B32_e64), ResultReg)
7235       .add(Src1)
7236       .addImm(16)
7237       .addReg(TmpReg, RegState::Kill);
7238     break;
7239   }
7240   case AMDGPU::S_PACK_LH_B32_B16: {
7241     Register ImmReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
7242     BuildMI(*MBB, Inst, DL, get(AMDGPU::V_MOV_B32_e32), ImmReg)
7243       .addImm(0xffff);
7244     BuildMI(*MBB, Inst, DL, get(AMDGPU::V_BFI_B32_e64), ResultReg)
7245       .addReg(ImmReg, RegState::Kill)
7246       .add(Src0)
7247       .add(Src1);
7248     break;
7249   }
7250   case AMDGPU::S_PACK_HL_B32_B16: {
7251     Register TmpReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
7252     BuildMI(*MBB, Inst, DL, get(AMDGPU::V_LSHRREV_B32_e64), TmpReg)
7253         .addImm(16)
7254         .add(Src0);
7255     BuildMI(*MBB, Inst, DL, get(AMDGPU::V_LSHL_OR_B32_e64), ResultReg)
7256         .add(Src1)
7257         .addImm(16)
7258         .addReg(TmpReg, RegState::Kill);
7259     break;
7260   }
7261   case AMDGPU::S_PACK_HH_B32_B16: {
7262     Register ImmReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
7263     Register TmpReg = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
7264     BuildMI(*MBB, Inst, DL, get(AMDGPU::V_LSHRREV_B32_e64), TmpReg)
7265       .addImm(16)
7266       .add(Src0);
7267     BuildMI(*MBB, Inst, DL, get(AMDGPU::V_MOV_B32_e32), ImmReg)
7268       .addImm(0xffff0000);
7269     BuildMI(*MBB, Inst, DL, get(AMDGPU::V_AND_OR_B32_e64), ResultReg)
7270       .add(Src1)
7271       .addReg(ImmReg, RegState::Kill)
7272       .addReg(TmpReg, RegState::Kill);
7273     break;
7274   }
7275   default:
7276     llvm_unreachable("unhandled s_pack_* instruction");
7277   }
7278 
7279   MachineOperand &Dest = Inst.getOperand(0);
7280   MRI.replaceRegWith(Dest.getReg(), ResultReg);
7281   addUsersToMoveToVALUWorklist(ResultReg, MRI, Worklist);
7282 }
7283 
7284 void SIInstrInfo::addSCCDefUsersToVALUWorklist(MachineOperand &Op,
7285                                                MachineInstr &SCCDefInst,
7286                                                SetVectorType &Worklist,
7287                                                Register NewCond) const {
7288 
7289   // Ensure that def inst defines SCC, which is still live.
7290   assert(Op.isReg() && Op.getReg() == AMDGPU::SCC && Op.isDef() &&
7291          !Op.isDead() && Op.getParent() == &SCCDefInst);
7292   SmallVector<MachineInstr *, 4> CopyToDelete;
7293   // This assumes that all the users of SCC are in the same block
7294   // as the SCC def.
7295   for (MachineInstr &MI : // Skip the def inst itself.
7296        make_range(std::next(MachineBasicBlock::iterator(SCCDefInst)),
7297                   SCCDefInst.getParent()->end())) {
7298     // Check if SCC is used first.
7299     int SCCIdx = MI.findRegisterUseOperandIdx(AMDGPU::SCC, false, &RI);
7300     if (SCCIdx != -1) {
7301       if (MI.isCopy()) {
7302         MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
7303         Register DestReg = MI.getOperand(0).getReg();
7304 
7305         MRI.replaceRegWith(DestReg, NewCond);
7306         CopyToDelete.push_back(&MI);
7307       } else {
7308 
7309         if (NewCond.isValid())
7310           MI.getOperand(SCCIdx).setReg(NewCond);
7311 
7312         Worklist.insert(&MI);
7313       }
7314     }
7315     // Exit if we find another SCC def.
7316     if (MI.findRegisterDefOperandIdx(AMDGPU::SCC, false, false, &RI) != -1)
7317       break;
7318   }
7319   for (auto &Copy : CopyToDelete)
7320     Copy->eraseFromParent();
7321 }
7322 
7323 // Instructions that use SCC may be converted to VALU instructions. When that
7324 // happens, the SCC register is changed to VCC_LO. The instruction that defines
7325 // SCC must be changed to an instruction that defines VCC. This function makes
7326 // sure that the instruction that defines SCC is added to the moveToVALU
7327 // worklist.
7328 void SIInstrInfo::addSCCDefsToVALUWorklist(MachineInstr *SCCUseInst,
7329                                            SetVectorType &Worklist) const {
7330   // Look for a preceding instruction that either defines VCC or SCC. If VCC
7331   // then there is nothing to do because the defining instruction has been
7332   // converted to a VALU already. If SCC then that instruction needs to be
7333   // converted to a VALU.
7334   for (MachineInstr &MI :
7335        make_range(std::next(MachineBasicBlock::reverse_iterator(SCCUseInst)),
7336                   SCCUseInst->getParent()->rend())) {
7337     if (MI.modifiesRegister(AMDGPU::VCC, &RI))
7338       break;
7339     if (MI.definesRegister(AMDGPU::SCC, &RI)) {
7340       Worklist.insert(&MI);
7341       break;
7342     }
7343   }
7344 }
7345 
7346 const TargetRegisterClass *SIInstrInfo::getDestEquivalentVGPRClass(
7347   const MachineInstr &Inst) const {
7348   const TargetRegisterClass *NewDstRC = getOpRegClass(Inst, 0);
7349 
7350   switch (Inst.getOpcode()) {
7351   // For target instructions, getOpRegClass just returns the virtual register
7352   // class associated with the operand, so we need to find an equivalent VGPR
7353   // register class in order to move the instruction to the VALU.
7354   case AMDGPU::COPY:
7355   case AMDGPU::PHI:
7356   case AMDGPU::REG_SEQUENCE:
7357   case AMDGPU::INSERT_SUBREG:
7358   case AMDGPU::WQM:
7359   case AMDGPU::SOFT_WQM:
7360   case AMDGPU::STRICT_WWM:
7361   case AMDGPU::STRICT_WQM: {
7362     const TargetRegisterClass *SrcRC = getOpRegClass(Inst, 1);
7363     if (RI.isAGPRClass(SrcRC)) {
7364       if (RI.isAGPRClass(NewDstRC))
7365         return nullptr;
7366 
7367       switch (Inst.getOpcode()) {
7368       case AMDGPU::PHI:
7369       case AMDGPU::REG_SEQUENCE:
7370       case AMDGPU::INSERT_SUBREG:
7371         NewDstRC = RI.getEquivalentAGPRClass(NewDstRC);
7372         break;
7373       default:
7374         NewDstRC = RI.getEquivalentVGPRClass(NewDstRC);
7375       }
7376 
7377       if (!NewDstRC)
7378         return nullptr;
7379     } else {
7380       if (RI.isVGPRClass(NewDstRC) || NewDstRC == &AMDGPU::VReg_1RegClass)
7381         return nullptr;
7382 
7383       NewDstRC = RI.getEquivalentVGPRClass(NewDstRC);
7384       if (!NewDstRC)
7385         return nullptr;
7386     }
7387 
7388     return NewDstRC;
7389   }
7390   default:
7391     return NewDstRC;
7392   }
7393 }
7394 
7395 // Find the one SGPR operand we are allowed to use.
7396 Register SIInstrInfo::findUsedSGPR(const MachineInstr &MI,
7397                                    int OpIndices[3]) const {
7398   const MCInstrDesc &Desc = MI.getDesc();
7399 
7400   // Find the one SGPR operand we are allowed to use.
7401   //
7402   // First we need to consider the instruction's operand requirements before
7403   // legalizing. Some operands are required to be SGPRs, such as implicit uses
7404   // of VCC, but we are still bound by the constant bus requirement to only use
7405   // one.
7406   //
7407   // If the operand's class is an SGPR, we can never move it.
7408 
7409   Register SGPRReg = findImplicitSGPRRead(MI);
7410   if (SGPRReg)
7411     return SGPRReg;
7412 
7413   Register UsedSGPRs[3] = {Register()};
7414   const MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
7415 
7416   for (unsigned i = 0; i < 3; ++i) {
7417     int Idx = OpIndices[i];
7418     if (Idx == -1)
7419       break;
7420 
7421     const MachineOperand &MO = MI.getOperand(Idx);
7422     if (!MO.isReg())
7423       continue;
7424 
7425     // Is this operand statically required to be an SGPR based on the operand
7426     // constraints?
7427     const TargetRegisterClass *OpRC =
7428         RI.getRegClass(Desc.operands()[Idx].RegClass);
7429     bool IsRequiredSGPR = RI.isSGPRClass(OpRC);
7430     if (IsRequiredSGPR)
7431       return MO.getReg();
7432 
7433     // If this could be a VGPR or an SGPR, Check the dynamic register class.
7434     Register Reg = MO.getReg();
7435     const TargetRegisterClass *RegRC = MRI.getRegClass(Reg);
7436     if (RI.isSGPRClass(RegRC))
7437       UsedSGPRs[i] = Reg;
7438   }
7439 
7440   // We don't have a required SGPR operand, so we have a bit more freedom in
7441   // selecting operands to move.
7442 
7443   // Try to select the most used SGPR. If an SGPR is equal to one of the
7444   // others, we choose that.
7445   //
7446   // e.g.
7447   // V_FMA_F32 v0, s0, s0, s0 -> No moves
7448   // V_FMA_F32 v0, s0, s1, s0 -> Move s1
7449 
7450   // TODO: If some of the operands are 64-bit SGPRs and some 32, we should
7451   // prefer those.
7452 
7453   if (UsedSGPRs[0]) {
7454     if (UsedSGPRs[0] == UsedSGPRs[1] || UsedSGPRs[0] == UsedSGPRs[2])
7455       SGPRReg = UsedSGPRs[0];
7456   }
7457 
7458   if (!SGPRReg && UsedSGPRs[1]) {
7459     if (UsedSGPRs[1] == UsedSGPRs[2])
7460       SGPRReg = UsedSGPRs[1];
7461   }
7462 
7463   return SGPRReg;
7464 }
7465 
7466 MachineOperand *SIInstrInfo::getNamedOperand(MachineInstr &MI,
7467                                              unsigned OperandName) const {
7468   int Idx = AMDGPU::getNamedOperandIdx(MI.getOpcode(), OperandName);
7469   if (Idx == -1)
7470     return nullptr;
7471 
7472   return &MI.getOperand(Idx);
7473 }
7474 
7475 uint64_t SIInstrInfo::getDefaultRsrcDataFormat() const {
7476   if (ST.getGeneration() >= AMDGPUSubtarget::GFX10) {
7477     int64_t Format = ST.getGeneration() >= AMDGPUSubtarget::GFX11
7478                          ? (int64_t)AMDGPU::UfmtGFX11::UFMT_32_FLOAT
7479                          : (int64_t)AMDGPU::UfmtGFX10::UFMT_32_FLOAT;
7480     return (Format << 44) |
7481            (1ULL << 56) | // RESOURCE_LEVEL = 1
7482            (3ULL << 60); // OOB_SELECT = 3
7483   }
7484 
7485   uint64_t RsrcDataFormat = AMDGPU::RSRC_DATA_FORMAT;
7486   if (ST.isAmdHsaOS()) {
7487     // Set ATC = 1. GFX9 doesn't have this bit.
7488     if (ST.getGeneration() <= AMDGPUSubtarget::VOLCANIC_ISLANDS)
7489       RsrcDataFormat |= (1ULL << 56);
7490 
7491     // Set MTYPE = 2 (MTYPE_UC = uncached). GFX9 doesn't have this.
7492     // BTW, it disables TC L2 and therefore decreases performance.
7493     if (ST.getGeneration() == AMDGPUSubtarget::VOLCANIC_ISLANDS)
7494       RsrcDataFormat |= (2ULL << 59);
7495   }
7496 
7497   return RsrcDataFormat;
7498 }
7499 
7500 uint64_t SIInstrInfo::getScratchRsrcWords23() const {
7501   uint64_t Rsrc23 = getDefaultRsrcDataFormat() |
7502                     AMDGPU::RSRC_TID_ENABLE |
7503                     0xffffffff; // Size;
7504 
7505   // GFX9 doesn't have ELEMENT_SIZE.
7506   if (ST.getGeneration() <= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
7507     uint64_t EltSizeValue = Log2_32(ST.getMaxPrivateElementSize(true)) - 1;
7508     Rsrc23 |= EltSizeValue << AMDGPU::RSRC_ELEMENT_SIZE_SHIFT;
7509   }
7510 
7511   // IndexStride = 64 / 32.
7512   uint64_t IndexStride = ST.getWavefrontSize() == 64 ? 3 : 2;
7513   Rsrc23 |= IndexStride << AMDGPU::RSRC_INDEX_STRIDE_SHIFT;
7514 
7515   // If TID_ENABLE is set, DATA_FORMAT specifies stride bits [14:17].
7516   // Clear them unless we want a huge stride.
7517   if (ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS &&
7518       ST.getGeneration() <= AMDGPUSubtarget::GFX9)
7519     Rsrc23 &= ~AMDGPU::RSRC_DATA_FORMAT;
7520 
7521   return Rsrc23;
7522 }
7523 
7524 bool SIInstrInfo::isLowLatencyInstruction(const MachineInstr &MI) const {
7525   unsigned Opc = MI.getOpcode();
7526 
7527   return isSMRD(Opc);
7528 }
7529 
7530 bool SIInstrInfo::isHighLatencyDef(int Opc) const {
7531   return get(Opc).mayLoad() &&
7532          (isMUBUF(Opc) || isMTBUF(Opc) || isMIMG(Opc) || isFLAT(Opc));
7533 }
7534 
7535 unsigned SIInstrInfo::isStackAccess(const MachineInstr &MI,
7536                                     int &FrameIndex) const {
7537   const MachineOperand *Addr = getNamedOperand(MI, AMDGPU::OpName::vaddr);
7538   if (!Addr || !Addr->isFI())
7539     return Register();
7540 
7541   assert(!MI.memoperands_empty() &&
7542          (*MI.memoperands_begin())->getAddrSpace() == AMDGPUAS::PRIVATE_ADDRESS);
7543 
7544   FrameIndex = Addr->getIndex();
7545   return getNamedOperand(MI, AMDGPU::OpName::vdata)->getReg();
7546 }
7547 
7548 unsigned SIInstrInfo::isSGPRStackAccess(const MachineInstr &MI,
7549                                         int &FrameIndex) const {
7550   const MachineOperand *Addr = getNamedOperand(MI, AMDGPU::OpName::addr);
7551   assert(Addr && Addr->isFI());
7552   FrameIndex = Addr->getIndex();
7553   return getNamedOperand(MI, AMDGPU::OpName::data)->getReg();
7554 }
7555 
7556 unsigned SIInstrInfo::isLoadFromStackSlot(const MachineInstr &MI,
7557                                           int &FrameIndex) const {
7558   if (!MI.mayLoad())
7559     return Register();
7560 
7561   if (isMUBUF(MI) || isVGPRSpill(MI))
7562     return isStackAccess(MI, FrameIndex);
7563 
7564   if (isSGPRSpill(MI))
7565     return isSGPRStackAccess(MI, FrameIndex);
7566 
7567   return Register();
7568 }
7569 
7570 unsigned SIInstrInfo::isStoreToStackSlot(const MachineInstr &MI,
7571                                          int &FrameIndex) const {
7572   if (!MI.mayStore())
7573     return Register();
7574 
7575   if (isMUBUF(MI) || isVGPRSpill(MI))
7576     return isStackAccess(MI, FrameIndex);
7577 
7578   if (isSGPRSpill(MI))
7579     return isSGPRStackAccess(MI, FrameIndex);
7580 
7581   return Register();
7582 }
7583 
7584 unsigned SIInstrInfo::getInstBundleSize(const MachineInstr &MI) const {
7585   unsigned Size = 0;
7586   MachineBasicBlock::const_instr_iterator I = MI.getIterator();
7587   MachineBasicBlock::const_instr_iterator E = MI.getParent()->instr_end();
7588   while (++I != E && I->isInsideBundle()) {
7589     assert(!I->isBundle() && "No nested bundle!");
7590     Size += getInstSizeInBytes(*I);
7591   }
7592 
7593   return Size;
7594 }
7595 
7596 unsigned SIInstrInfo::getInstSizeInBytes(const MachineInstr &MI) const {
7597   unsigned Opc = MI.getOpcode();
7598   const MCInstrDesc &Desc = getMCOpcodeFromPseudo(Opc);
7599   unsigned DescSize = Desc.getSize();
7600 
7601   // If we have a definitive size, we can use it. Otherwise we need to inspect
7602   // the operands to know the size.
7603   if (isFixedSize(MI)) {
7604     unsigned Size = DescSize;
7605 
7606     // If we hit the buggy offset, an extra nop will be inserted in MC so
7607     // estimate the worst case.
7608     if (MI.isBranch() && ST.hasOffset3fBug())
7609       Size += 4;
7610 
7611     return Size;
7612   }
7613 
7614   // Instructions may have a 32-bit literal encoded after them. Check
7615   // operands that could ever be literals.
7616   if (isVALU(MI) || isSALU(MI)) {
7617     if (isDPP(MI))
7618       return DescSize;
7619     bool HasLiteral = false;
7620     for (int I = 0, E = MI.getNumExplicitOperands(); I != E; ++I) {
7621       const MachineOperand &Op = MI.getOperand(I);
7622       const MCOperandInfo &OpInfo = Desc.operands()[I];
7623       if (!Op.isReg() && !isInlineConstant(Op, OpInfo)) {
7624         HasLiteral = true;
7625         break;
7626       }
7627     }
7628     return HasLiteral ? DescSize + 4 : DescSize;
7629   }
7630 
7631   // Check whether we have extra NSA words.
7632   if (isMIMG(MI)) {
7633     int VAddr0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::vaddr0);
7634     if (VAddr0Idx < 0)
7635       return 8;
7636 
7637     int RSrcIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::srsrc);
7638     return 8 + 4 * ((RSrcIdx - VAddr0Idx + 2) / 4);
7639   }
7640 
7641   switch (Opc) {
7642   case TargetOpcode::BUNDLE:
7643     return getInstBundleSize(MI);
7644   case TargetOpcode::INLINEASM:
7645   case TargetOpcode::INLINEASM_BR: {
7646     const MachineFunction *MF = MI.getParent()->getParent();
7647     const char *AsmStr = MI.getOperand(0).getSymbolName();
7648     return getInlineAsmLength(AsmStr, *MF->getTarget().getMCAsmInfo(), &ST);
7649   }
7650   default:
7651     if (MI.isMetaInstruction())
7652       return 0;
7653     return DescSize;
7654   }
7655 }
7656 
7657 bool SIInstrInfo::mayAccessFlatAddressSpace(const MachineInstr &MI) const {
7658   if (!isFLAT(MI))
7659     return false;
7660 
7661   if (MI.memoperands_empty())
7662     return true;
7663 
7664   for (const MachineMemOperand *MMO : MI.memoperands()) {
7665     if (MMO->getAddrSpace() == AMDGPUAS::FLAT_ADDRESS)
7666       return true;
7667   }
7668   return false;
7669 }
7670 
7671 bool SIInstrInfo::isNonUniformBranchInstr(MachineInstr &Branch) const {
7672   return Branch.getOpcode() == AMDGPU::SI_NON_UNIFORM_BRCOND_PSEUDO;
7673 }
7674 
7675 void SIInstrInfo::convertNonUniformIfRegion(MachineBasicBlock *IfEntry,
7676                                             MachineBasicBlock *IfEnd) const {
7677   MachineBasicBlock::iterator TI = IfEntry->getFirstTerminator();
7678   assert(TI != IfEntry->end());
7679 
7680   MachineInstr *Branch = &(*TI);
7681   MachineFunction *MF = IfEntry->getParent();
7682   MachineRegisterInfo &MRI = IfEntry->getParent()->getRegInfo();
7683 
7684   if (Branch->getOpcode() == AMDGPU::SI_NON_UNIFORM_BRCOND_PSEUDO) {
7685     Register DstReg = MRI.createVirtualRegister(RI.getBoolRC());
7686     MachineInstr *SIIF =
7687         BuildMI(*MF, Branch->getDebugLoc(), get(AMDGPU::SI_IF), DstReg)
7688             .add(Branch->getOperand(0))
7689             .add(Branch->getOperand(1));
7690     MachineInstr *SIEND =
7691         BuildMI(*MF, Branch->getDebugLoc(), get(AMDGPU::SI_END_CF))
7692             .addReg(DstReg);
7693 
7694     IfEntry->erase(TI);
7695     IfEntry->insert(IfEntry->end(), SIIF);
7696     IfEnd->insert(IfEnd->getFirstNonPHI(), SIEND);
7697   }
7698 }
7699 
7700 void SIInstrInfo::convertNonUniformLoopRegion(
7701     MachineBasicBlock *LoopEntry, MachineBasicBlock *LoopEnd) const {
7702   MachineBasicBlock::iterator TI = LoopEnd->getFirstTerminator();
7703   // We expect 2 terminators, one conditional and one unconditional.
7704   assert(TI != LoopEnd->end());
7705 
7706   MachineInstr *Branch = &(*TI);
7707   MachineFunction *MF = LoopEnd->getParent();
7708   MachineRegisterInfo &MRI = LoopEnd->getParent()->getRegInfo();
7709 
7710   if (Branch->getOpcode() == AMDGPU::SI_NON_UNIFORM_BRCOND_PSEUDO) {
7711 
7712     Register DstReg = MRI.createVirtualRegister(RI.getBoolRC());
7713     Register BackEdgeReg = MRI.createVirtualRegister(RI.getBoolRC());
7714     MachineInstrBuilder HeaderPHIBuilder =
7715         BuildMI(*(MF), Branch->getDebugLoc(), get(TargetOpcode::PHI), DstReg);
7716     for (MachineBasicBlock *PMBB : LoopEntry->predecessors()) {
7717       if (PMBB == LoopEnd) {
7718         HeaderPHIBuilder.addReg(BackEdgeReg);
7719       } else {
7720         Register ZeroReg = MRI.createVirtualRegister(RI.getBoolRC());
7721         materializeImmediate(*PMBB, PMBB->getFirstTerminator(), DebugLoc(),
7722                              ZeroReg, 0);
7723         HeaderPHIBuilder.addReg(ZeroReg);
7724       }
7725       HeaderPHIBuilder.addMBB(PMBB);
7726     }
7727     MachineInstr *HeaderPhi = HeaderPHIBuilder;
7728     MachineInstr *SIIFBREAK = BuildMI(*(MF), Branch->getDebugLoc(),
7729                                       get(AMDGPU::SI_IF_BREAK), BackEdgeReg)
7730                                   .addReg(DstReg)
7731                                   .add(Branch->getOperand(0));
7732     MachineInstr *SILOOP =
7733         BuildMI(*(MF), Branch->getDebugLoc(), get(AMDGPU::SI_LOOP))
7734             .addReg(BackEdgeReg)
7735             .addMBB(LoopEntry);
7736 
7737     LoopEntry->insert(LoopEntry->begin(), HeaderPhi);
7738     LoopEnd->erase(TI);
7739     LoopEnd->insert(LoopEnd->end(), SIIFBREAK);
7740     LoopEnd->insert(LoopEnd->end(), SILOOP);
7741   }
7742 }
7743 
7744 ArrayRef<std::pair<int, const char *>>
7745 SIInstrInfo::getSerializableTargetIndices() const {
7746   static const std::pair<int, const char *> TargetIndices[] = {
7747       {AMDGPU::TI_CONSTDATA_START, "amdgpu-constdata-start"},
7748       {AMDGPU::TI_SCRATCH_RSRC_DWORD0, "amdgpu-scratch-rsrc-dword0"},
7749       {AMDGPU::TI_SCRATCH_RSRC_DWORD1, "amdgpu-scratch-rsrc-dword1"},
7750       {AMDGPU::TI_SCRATCH_RSRC_DWORD2, "amdgpu-scratch-rsrc-dword2"},
7751       {AMDGPU::TI_SCRATCH_RSRC_DWORD3, "amdgpu-scratch-rsrc-dword3"}};
7752   return ArrayRef(TargetIndices);
7753 }
7754 
7755 /// This is used by the post-RA scheduler (SchedulePostRAList.cpp).  The
7756 /// post-RA version of misched uses CreateTargetMIHazardRecognizer.
7757 ScheduleHazardRecognizer *
7758 SIInstrInfo::CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II,
7759                                             const ScheduleDAG *DAG) const {
7760   return new GCNHazardRecognizer(DAG->MF);
7761 }
7762 
7763 /// This is the hazard recognizer used at -O0 by the PostRAHazardRecognizer
7764 /// pass.
7765 ScheduleHazardRecognizer *
7766 SIInstrInfo::CreateTargetPostRAHazardRecognizer(const MachineFunction &MF) const {
7767   return new GCNHazardRecognizer(MF);
7768 }
7769 
7770 // Called during:
7771 // - pre-RA scheduling and post-RA scheduling
7772 ScheduleHazardRecognizer *
7773 SIInstrInfo::CreateTargetMIHazardRecognizer(const InstrItineraryData *II,
7774                                             const ScheduleDAGMI *DAG) const {
7775   // Borrowed from Arm Target
7776   // We would like to restrict this hazard recognizer to only
7777   // post-RA scheduling; we can tell that we're post-RA because we don't
7778   // track VRegLiveness.
7779   if (!DAG->hasVRegLiveness())
7780     return new GCNHazardRecognizer(DAG->MF);
7781   return TargetInstrInfo::CreateTargetMIHazardRecognizer(II, DAG);
7782 }
7783 
7784 std::pair<unsigned, unsigned>
7785 SIInstrInfo::decomposeMachineOperandsTargetFlags(unsigned TF) const {
7786   return std::pair(TF & MO_MASK, TF & ~MO_MASK);
7787 }
7788 
7789 ArrayRef<std::pair<unsigned, const char *>>
7790 SIInstrInfo::getSerializableDirectMachineOperandTargetFlags() const {
7791   static const std::pair<unsigned, const char *> TargetFlags[] = {
7792     { MO_GOTPCREL, "amdgpu-gotprel" },
7793     { MO_GOTPCREL32_LO, "amdgpu-gotprel32-lo" },
7794     { MO_GOTPCREL32_HI, "amdgpu-gotprel32-hi" },
7795     { MO_REL32_LO, "amdgpu-rel32-lo" },
7796     { MO_REL32_HI, "amdgpu-rel32-hi" },
7797     { MO_ABS32_LO, "amdgpu-abs32-lo" },
7798     { MO_ABS32_HI, "amdgpu-abs32-hi" },
7799   };
7800 
7801   return ArrayRef(TargetFlags);
7802 }
7803 
7804 ArrayRef<std::pair<MachineMemOperand::Flags, const char *>>
7805 SIInstrInfo::getSerializableMachineMemOperandTargetFlags() const {
7806   static const std::pair<MachineMemOperand::Flags, const char *> TargetFlags[] =
7807       {
7808           {MONoClobber, "amdgpu-noclobber"},
7809       };
7810 
7811   return ArrayRef(TargetFlags);
7812 }
7813 
7814 bool SIInstrInfo::isBasicBlockPrologue(const MachineInstr &MI) const {
7815   return !MI.isTerminator() && MI.getOpcode() != AMDGPU::COPY &&
7816          MI.modifiesRegister(AMDGPU::EXEC, &RI);
7817 }
7818 
7819 MachineInstrBuilder
7820 SIInstrInfo::getAddNoCarry(MachineBasicBlock &MBB,
7821                            MachineBasicBlock::iterator I,
7822                            const DebugLoc &DL,
7823                            Register DestReg) const {
7824   if (ST.hasAddNoCarry())
7825     return BuildMI(MBB, I, DL, get(AMDGPU::V_ADD_U32_e64), DestReg);
7826 
7827   MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
7828   Register UnusedCarry = MRI.createVirtualRegister(RI.getBoolRC());
7829   MRI.setRegAllocationHint(UnusedCarry, 0, RI.getVCC());
7830 
7831   return BuildMI(MBB, I, DL, get(AMDGPU::V_ADD_CO_U32_e64), DestReg)
7832            .addReg(UnusedCarry, RegState::Define | RegState::Dead);
7833 }
7834 
7835 MachineInstrBuilder SIInstrInfo::getAddNoCarry(MachineBasicBlock &MBB,
7836                                                MachineBasicBlock::iterator I,
7837                                                const DebugLoc &DL,
7838                                                Register DestReg,
7839                                                RegScavenger &RS) const {
7840   if (ST.hasAddNoCarry())
7841     return BuildMI(MBB, I, DL, get(AMDGPU::V_ADD_U32_e32), DestReg);
7842 
7843   // If available, prefer to use vcc.
7844   Register UnusedCarry = !RS.isRegUsed(AMDGPU::VCC)
7845                              ? Register(RI.getVCC())
7846                              : RS.scavengeRegister(RI.getBoolRC(), I, 0, false);
7847 
7848   // TODO: Users need to deal with this.
7849   if (!UnusedCarry.isValid())
7850     return MachineInstrBuilder();
7851 
7852   return BuildMI(MBB, I, DL, get(AMDGPU::V_ADD_CO_U32_e64), DestReg)
7853            .addReg(UnusedCarry, RegState::Define | RegState::Dead);
7854 }
7855 
7856 bool SIInstrInfo::isKillTerminator(unsigned Opcode) {
7857   switch (Opcode) {
7858   case AMDGPU::SI_KILL_F32_COND_IMM_TERMINATOR:
7859   case AMDGPU::SI_KILL_I1_TERMINATOR:
7860     return true;
7861   default:
7862     return false;
7863   }
7864 }
7865 
7866 const MCInstrDesc &SIInstrInfo::getKillTerminatorFromPseudo(unsigned Opcode) const {
7867   switch (Opcode) {
7868   case AMDGPU::SI_KILL_F32_COND_IMM_PSEUDO:
7869     return get(AMDGPU::SI_KILL_F32_COND_IMM_TERMINATOR);
7870   case AMDGPU::SI_KILL_I1_PSEUDO:
7871     return get(AMDGPU::SI_KILL_I1_TERMINATOR);
7872   default:
7873     llvm_unreachable("invalid opcode, expected SI_KILL_*_PSEUDO");
7874   }
7875 }
7876 
7877 void SIInstrInfo::fixImplicitOperands(MachineInstr &MI) const {
7878   if (!ST.isWave32())
7879     return;
7880 
7881   for (auto &Op : MI.implicit_operands()) {
7882     if (Op.isReg() && Op.getReg() == AMDGPU::VCC)
7883       Op.setReg(AMDGPU::VCC_LO);
7884   }
7885 }
7886 
7887 bool SIInstrInfo::isBufferSMRD(const MachineInstr &MI) const {
7888   if (!isSMRD(MI))
7889     return false;
7890 
7891   // Check that it is using a buffer resource.
7892   int Idx = AMDGPU::getNamedOperandIdx(MI.getOpcode(), AMDGPU::OpName::sbase);
7893   if (Idx == -1) // e.g. s_memtime
7894     return false;
7895 
7896   const auto RCID = MI.getDesc().operands()[Idx].RegClass;
7897   return RI.getRegClass(RCID)->hasSubClassEq(&AMDGPU::SGPR_128RegClass);
7898 }
7899 
7900 // Depending on the used address space and instructions, some immediate offsets
7901 // are allowed and some are not.
7902 // In general, flat instruction offsets can only be non-negative, global and
7903 // scratch instruction offsets can also be negative.
7904 //
7905 // There are several bugs related to these offsets:
7906 // On gfx10.1, flat instructions that go into the global address space cannot
7907 // use an offset.
7908 //
7909 // For scratch instructions, the address can be either an SGPR or a VGPR.
7910 // The following offsets can be used, depending on the architecture (x means
7911 // cannot be used):
7912 // +----------------------------+------+------+
7913 // | Address-Mode               | SGPR | VGPR |
7914 // +----------------------------+------+------+
7915 // | gfx9                       |      |      |
7916 // | negative, 4-aligned offset | x    | ok   |
7917 // | negative, unaligned offset | x    | ok   |
7918 // +----------------------------+------+------+
7919 // | gfx10                      |      |      |
7920 // | negative, 4-aligned offset | ok   | ok   |
7921 // | negative, unaligned offset | ok   | x    |
7922 // +----------------------------+------+------+
7923 // | gfx10.3                    |      |      |
7924 // | negative, 4-aligned offset | ok   | ok   |
7925 // | negative, unaligned offset | ok   | ok   |
7926 // +----------------------------+------+------+
7927 //
7928 // This function ignores the addressing mode, so if an offset cannot be used in
7929 // one addressing mode, it is considered illegal.
7930 bool SIInstrInfo::isLegalFLATOffset(int64_t Offset, unsigned AddrSpace,
7931                                     uint64_t FlatVariant) const {
7932   // TODO: Should 0 be special cased?
7933   if (!ST.hasFlatInstOffsets())
7934     return false;
7935 
7936   if (ST.hasFlatSegmentOffsetBug() && FlatVariant == SIInstrFlags::FLAT &&
7937       (AddrSpace == AMDGPUAS::FLAT_ADDRESS ||
7938        AddrSpace == AMDGPUAS::GLOBAL_ADDRESS))
7939     return false;
7940 
7941   bool AllowNegative = FlatVariant != SIInstrFlags::FLAT;
7942   if (ST.hasNegativeScratchOffsetBug() &&
7943       FlatVariant == SIInstrFlags::FlatScratch)
7944     AllowNegative = false;
7945   if (ST.hasNegativeUnalignedScratchOffsetBug() &&
7946       FlatVariant == SIInstrFlags::FlatScratch && Offset < 0 &&
7947       (Offset % 4) != 0) {
7948     return false;
7949   }
7950 
7951   unsigned N = AMDGPU::getNumFlatOffsetBits(ST);
7952   return isIntN(N, Offset) && (AllowNegative || Offset >= 0);
7953 }
7954 
7955 // See comment on SIInstrInfo::isLegalFLATOffset for what is legal and what not.
7956 std::pair<int64_t, int64_t>
7957 SIInstrInfo::splitFlatOffset(int64_t COffsetVal, unsigned AddrSpace,
7958                              uint64_t FlatVariant) const {
7959   int64_t RemainderOffset = COffsetVal;
7960   int64_t ImmField = 0;
7961   bool AllowNegative = FlatVariant != SIInstrFlags::FLAT;
7962   if (ST.hasNegativeScratchOffsetBug() &&
7963       FlatVariant == SIInstrFlags::FlatScratch)
7964     AllowNegative = false;
7965 
7966   const unsigned NumBits = AMDGPU::getNumFlatOffsetBits(ST) - 1;
7967   if (AllowNegative) {
7968     // Use signed division by a power of two to truncate towards 0.
7969     int64_t D = 1LL << NumBits;
7970     RemainderOffset = (COffsetVal / D) * D;
7971     ImmField = COffsetVal - RemainderOffset;
7972 
7973     if (ST.hasNegativeUnalignedScratchOffsetBug() &&
7974         FlatVariant == SIInstrFlags::FlatScratch && ImmField < 0 &&
7975         (ImmField % 4) != 0) {
7976       // Make ImmField a multiple of 4
7977       RemainderOffset += ImmField % 4;
7978       ImmField -= ImmField % 4;
7979     }
7980   } else if (COffsetVal >= 0) {
7981     ImmField = COffsetVal & maskTrailingOnes<uint64_t>(NumBits);
7982     RemainderOffset = COffsetVal - ImmField;
7983   }
7984 
7985   assert(isLegalFLATOffset(ImmField, AddrSpace, FlatVariant));
7986   assert(RemainderOffset + ImmField == COffsetVal);
7987   return {ImmField, RemainderOffset};
7988 }
7989 
7990 // This must be kept in sync with the SIEncodingFamily class in SIInstrInfo.td
7991 // and the columns of the getMCOpcodeGen table.
7992 enum SIEncodingFamily {
7993   SI = 0,
7994   VI = 1,
7995   SDWA = 2,
7996   SDWA9 = 3,
7997   GFX80 = 4,
7998   GFX9 = 5,
7999   GFX10 = 6,
8000   SDWA10 = 7,
8001   GFX90A = 8,
8002   GFX940 = 9,
8003   GFX11 = 10,
8004 };
8005 
8006 static SIEncodingFamily subtargetEncodingFamily(const GCNSubtarget &ST) {
8007   switch (ST.getGeneration()) {
8008   default:
8009     break;
8010   case AMDGPUSubtarget::SOUTHERN_ISLANDS:
8011   case AMDGPUSubtarget::SEA_ISLANDS:
8012     return SIEncodingFamily::SI;
8013   case AMDGPUSubtarget::VOLCANIC_ISLANDS:
8014   case AMDGPUSubtarget::GFX9:
8015     return SIEncodingFamily::VI;
8016   case AMDGPUSubtarget::GFX10:
8017     return SIEncodingFamily::GFX10;
8018   case AMDGPUSubtarget::GFX11:
8019     return SIEncodingFamily::GFX11;
8020   }
8021   llvm_unreachable("Unknown subtarget generation!");
8022 }
8023 
8024 bool SIInstrInfo::isAsmOnlyOpcode(int MCOp) const {
8025   switch(MCOp) {
8026   // These opcodes use indirect register addressing so
8027   // they need special handling by codegen (currently missing).
8028   // Therefore it is too risky to allow these opcodes
8029   // to be selected by dpp combiner or sdwa peepholer.
8030   case AMDGPU::V_MOVRELS_B32_dpp_gfx10:
8031   case AMDGPU::V_MOVRELS_B32_sdwa_gfx10:
8032   case AMDGPU::V_MOVRELD_B32_dpp_gfx10:
8033   case AMDGPU::V_MOVRELD_B32_sdwa_gfx10:
8034   case AMDGPU::V_MOVRELSD_B32_dpp_gfx10:
8035   case AMDGPU::V_MOVRELSD_B32_sdwa_gfx10:
8036   case AMDGPU::V_MOVRELSD_2_B32_dpp_gfx10:
8037   case AMDGPU::V_MOVRELSD_2_B32_sdwa_gfx10:
8038     return true;
8039   default:
8040     return false;
8041   }
8042 }
8043 
8044 int SIInstrInfo::pseudoToMCOpcode(int Opcode) const {
8045   SIEncodingFamily Gen = subtargetEncodingFamily(ST);
8046 
8047   if ((get(Opcode).TSFlags & SIInstrFlags::renamedInGFX9) != 0 &&
8048     ST.getGeneration() == AMDGPUSubtarget::GFX9)
8049     Gen = SIEncodingFamily::GFX9;
8050 
8051   // Adjust the encoding family to GFX80 for D16 buffer instructions when the
8052   // subtarget has UnpackedD16VMem feature.
8053   // TODO: remove this when we discard GFX80 encoding.
8054   if (ST.hasUnpackedD16VMem() && (get(Opcode).TSFlags & SIInstrFlags::D16Buf))
8055     Gen = SIEncodingFamily::GFX80;
8056 
8057   if (get(Opcode).TSFlags & SIInstrFlags::SDWA) {
8058     switch (ST.getGeneration()) {
8059     default:
8060       Gen = SIEncodingFamily::SDWA;
8061       break;
8062     case AMDGPUSubtarget::GFX9:
8063       Gen = SIEncodingFamily::SDWA9;
8064       break;
8065     case AMDGPUSubtarget::GFX10:
8066       Gen = SIEncodingFamily::SDWA10;
8067       break;
8068     }
8069   }
8070 
8071   if (isMAI(Opcode)) {
8072     int MFMAOp = AMDGPU::getMFMAEarlyClobberOp(Opcode);
8073     if (MFMAOp != -1)
8074       Opcode = MFMAOp;
8075   }
8076 
8077   int MCOp = AMDGPU::getMCOpcode(Opcode, Gen);
8078 
8079   // -1 means that Opcode is already a native instruction.
8080   if (MCOp == -1)
8081     return Opcode;
8082 
8083   if (ST.hasGFX90AInsts()) {
8084     uint16_t NMCOp = (uint16_t)-1;
8085     if (ST.hasGFX940Insts())
8086       NMCOp = AMDGPU::getMCOpcode(Opcode, SIEncodingFamily::GFX940);
8087     if (NMCOp == (uint16_t)-1)
8088       NMCOp = AMDGPU::getMCOpcode(Opcode, SIEncodingFamily::GFX90A);
8089     if (NMCOp == (uint16_t)-1)
8090       NMCOp = AMDGPU::getMCOpcode(Opcode, SIEncodingFamily::GFX9);
8091     if (NMCOp != (uint16_t)-1)
8092       MCOp = NMCOp;
8093   }
8094 
8095   // (uint16_t)-1 means that Opcode is a pseudo instruction that has
8096   // no encoding in the given subtarget generation.
8097   if (MCOp == (uint16_t)-1)
8098     return -1;
8099 
8100   if (isAsmOnlyOpcode(MCOp))
8101     return -1;
8102 
8103   return MCOp;
8104 }
8105 
8106 static
8107 TargetInstrInfo::RegSubRegPair getRegOrUndef(const MachineOperand &RegOpnd) {
8108   assert(RegOpnd.isReg());
8109   return RegOpnd.isUndef() ? TargetInstrInfo::RegSubRegPair() :
8110                              getRegSubRegPair(RegOpnd);
8111 }
8112 
8113 TargetInstrInfo::RegSubRegPair
8114 llvm::getRegSequenceSubReg(MachineInstr &MI, unsigned SubReg) {
8115   assert(MI.isRegSequence());
8116   for (unsigned I = 0, E = (MI.getNumOperands() - 1)/ 2; I < E; ++I)
8117     if (MI.getOperand(1 + 2 * I + 1).getImm() == SubReg) {
8118       auto &RegOp = MI.getOperand(1 + 2 * I);
8119       return getRegOrUndef(RegOp);
8120     }
8121   return TargetInstrInfo::RegSubRegPair();
8122 }
8123 
8124 // Try to find the definition of reg:subreg in subreg-manipulation pseudos
8125 // Following a subreg of reg:subreg isn't supported
8126 static bool followSubRegDef(MachineInstr &MI,
8127                             TargetInstrInfo::RegSubRegPair &RSR) {
8128   if (!RSR.SubReg)
8129     return false;
8130   switch (MI.getOpcode()) {
8131   default: break;
8132   case AMDGPU::REG_SEQUENCE:
8133     RSR = getRegSequenceSubReg(MI, RSR.SubReg);
8134     return true;
8135   // EXTRACT_SUBREG ins't supported as this would follow a subreg of subreg
8136   case AMDGPU::INSERT_SUBREG:
8137     if (RSR.SubReg == (unsigned)MI.getOperand(3).getImm())
8138       // inserted the subreg we're looking for
8139       RSR = getRegOrUndef(MI.getOperand(2));
8140     else { // the subreg in the rest of the reg
8141       auto R1 = getRegOrUndef(MI.getOperand(1));
8142       if (R1.SubReg) // subreg of subreg isn't supported
8143         return false;
8144       RSR.Reg = R1.Reg;
8145     }
8146     return true;
8147   }
8148   return false;
8149 }
8150 
8151 MachineInstr *llvm::getVRegSubRegDef(const TargetInstrInfo::RegSubRegPair &P,
8152                                      MachineRegisterInfo &MRI) {
8153   assert(MRI.isSSA());
8154   if (!P.Reg.isVirtual())
8155     return nullptr;
8156 
8157   auto RSR = P;
8158   auto *DefInst = MRI.getVRegDef(RSR.Reg);
8159   while (auto *MI = DefInst) {
8160     DefInst = nullptr;
8161     switch (MI->getOpcode()) {
8162     case AMDGPU::COPY:
8163     case AMDGPU::V_MOV_B32_e32: {
8164       auto &Op1 = MI->getOperand(1);
8165       if (Op1.isReg() && Op1.getReg().isVirtual()) {
8166         if (Op1.isUndef())
8167           return nullptr;
8168         RSR = getRegSubRegPair(Op1);
8169         DefInst = MRI.getVRegDef(RSR.Reg);
8170       }
8171       break;
8172     }
8173     default:
8174       if (followSubRegDef(*MI, RSR)) {
8175         if (!RSR.Reg)
8176           return nullptr;
8177         DefInst = MRI.getVRegDef(RSR.Reg);
8178       }
8179     }
8180     if (!DefInst)
8181       return MI;
8182   }
8183   return nullptr;
8184 }
8185 
8186 bool llvm::execMayBeModifiedBeforeUse(const MachineRegisterInfo &MRI,
8187                                       Register VReg,
8188                                       const MachineInstr &DefMI,
8189                                       const MachineInstr &UseMI) {
8190   assert(MRI.isSSA() && "Must be run on SSA");
8191 
8192   auto *TRI = MRI.getTargetRegisterInfo();
8193   auto *DefBB = DefMI.getParent();
8194 
8195   // Don't bother searching between blocks, although it is possible this block
8196   // doesn't modify exec.
8197   if (UseMI.getParent() != DefBB)
8198     return true;
8199 
8200   const int MaxInstScan = 20;
8201   int NumInst = 0;
8202 
8203   // Stop scan at the use.
8204   auto E = UseMI.getIterator();
8205   for (auto I = std::next(DefMI.getIterator()); I != E; ++I) {
8206     if (I->isDebugInstr())
8207       continue;
8208 
8209     if (++NumInst > MaxInstScan)
8210       return true;
8211 
8212     if (I->modifiesRegister(AMDGPU::EXEC, TRI))
8213       return true;
8214   }
8215 
8216   return false;
8217 }
8218 
8219 bool llvm::execMayBeModifiedBeforeAnyUse(const MachineRegisterInfo &MRI,
8220                                          Register VReg,
8221                                          const MachineInstr &DefMI) {
8222   assert(MRI.isSSA() && "Must be run on SSA");
8223 
8224   auto *TRI = MRI.getTargetRegisterInfo();
8225   auto *DefBB = DefMI.getParent();
8226 
8227   const int MaxUseScan = 10;
8228   int NumUse = 0;
8229 
8230   for (auto &Use : MRI.use_nodbg_operands(VReg)) {
8231     auto &UseInst = *Use.getParent();
8232     // Don't bother searching between blocks, although it is possible this block
8233     // doesn't modify exec.
8234     if (UseInst.getParent() != DefBB || UseInst.isPHI())
8235       return true;
8236 
8237     if (++NumUse > MaxUseScan)
8238       return true;
8239   }
8240 
8241   if (NumUse == 0)
8242     return false;
8243 
8244   const int MaxInstScan = 20;
8245   int NumInst = 0;
8246 
8247   // Stop scan when we have seen all the uses.
8248   for (auto I = std::next(DefMI.getIterator()); ; ++I) {
8249     assert(I != DefBB->end());
8250 
8251     if (I->isDebugInstr())
8252       continue;
8253 
8254     if (++NumInst > MaxInstScan)
8255       return true;
8256 
8257     for (const MachineOperand &Op : I->operands()) {
8258       // We don't check reg masks here as they're used only on calls:
8259       // 1. EXEC is only considered const within one BB
8260       // 2. Call should be a terminator instruction if present in a BB
8261 
8262       if (!Op.isReg())
8263         continue;
8264 
8265       Register Reg = Op.getReg();
8266       if (Op.isUse()) {
8267         if (Reg == VReg && --NumUse == 0)
8268           return false;
8269       } else if (TRI->regsOverlap(Reg, AMDGPU::EXEC))
8270         return true;
8271     }
8272   }
8273 }
8274 
8275 MachineInstr *SIInstrInfo::createPHIDestinationCopy(
8276     MachineBasicBlock &MBB, MachineBasicBlock::iterator LastPHIIt,
8277     const DebugLoc &DL, Register Src, Register Dst) const {
8278   auto Cur = MBB.begin();
8279   if (Cur != MBB.end())
8280     do {
8281       if (!Cur->isPHI() && Cur->readsRegister(Dst))
8282         return BuildMI(MBB, Cur, DL, get(TargetOpcode::COPY), Dst).addReg(Src);
8283       ++Cur;
8284     } while (Cur != MBB.end() && Cur != LastPHIIt);
8285 
8286   return TargetInstrInfo::createPHIDestinationCopy(MBB, LastPHIIt, DL, Src,
8287                                                    Dst);
8288 }
8289 
8290 MachineInstr *SIInstrInfo::createPHISourceCopy(
8291     MachineBasicBlock &MBB, MachineBasicBlock::iterator InsPt,
8292     const DebugLoc &DL, Register Src, unsigned SrcSubReg, Register Dst) const {
8293   if (InsPt != MBB.end() &&
8294       (InsPt->getOpcode() == AMDGPU::SI_IF ||
8295        InsPt->getOpcode() == AMDGPU::SI_ELSE ||
8296        InsPt->getOpcode() == AMDGPU::SI_IF_BREAK) &&
8297       InsPt->definesRegister(Src)) {
8298     InsPt++;
8299     return BuildMI(MBB, InsPt, DL,
8300                    get(ST.isWave32() ? AMDGPU::S_MOV_B32_term
8301                                      : AMDGPU::S_MOV_B64_term),
8302                    Dst)
8303         .addReg(Src, 0, SrcSubReg)
8304         .addReg(AMDGPU::EXEC, RegState::Implicit);
8305   }
8306   return TargetInstrInfo::createPHISourceCopy(MBB, InsPt, DL, Src, SrcSubReg,
8307                                               Dst);
8308 }
8309 
8310 bool llvm::SIInstrInfo::isWave32() const { return ST.isWave32(); }
8311 
8312 MachineInstr *SIInstrInfo::foldMemoryOperandImpl(
8313     MachineFunction &MF, MachineInstr &MI, ArrayRef<unsigned> Ops,
8314     MachineBasicBlock::iterator InsertPt, int FrameIndex, LiveIntervals *LIS,
8315     VirtRegMap *VRM) const {
8316   // This is a bit of a hack (copied from AArch64). Consider this instruction:
8317   //
8318   //   %0:sreg_32 = COPY $m0
8319   //
8320   // We explicitly chose SReg_32 for the virtual register so such a copy might
8321   // be eliminated by RegisterCoalescer. However, that may not be possible, and
8322   // %0 may even spill. We can't spill $m0 normally (it would require copying to
8323   // a numbered SGPR anyway), and since it is in the SReg_32 register class,
8324   // TargetInstrInfo::foldMemoryOperand() is going to try.
8325   // A similar issue also exists with spilling and reloading $exec registers.
8326   //
8327   // To prevent that, constrain the %0 register class here.
8328   if (MI.isFullCopy()) {
8329     Register DstReg = MI.getOperand(0).getReg();
8330     Register SrcReg = MI.getOperand(1).getReg();
8331     if ((DstReg.isVirtual() || SrcReg.isVirtual()) &&
8332         (DstReg.isVirtual() != SrcReg.isVirtual())) {
8333       MachineRegisterInfo &MRI = MF.getRegInfo();
8334       Register VirtReg = DstReg.isVirtual() ? DstReg : SrcReg;
8335       const TargetRegisterClass *RC = MRI.getRegClass(VirtReg);
8336       if (RC->hasSuperClassEq(&AMDGPU::SReg_32RegClass)) {
8337         MRI.constrainRegClass(VirtReg, &AMDGPU::SReg_32_XM0_XEXECRegClass);
8338         return nullptr;
8339       } else if (RC->hasSuperClassEq(&AMDGPU::SReg_64RegClass)) {
8340         MRI.constrainRegClass(VirtReg, &AMDGPU::SReg_64_XEXECRegClass);
8341         return nullptr;
8342       }
8343     }
8344   }
8345 
8346   return nullptr;
8347 }
8348 
8349 unsigned SIInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
8350                                       const MachineInstr &MI,
8351                                       unsigned *PredCost) const {
8352   if (MI.isBundle()) {
8353     MachineBasicBlock::const_instr_iterator I(MI.getIterator());
8354     MachineBasicBlock::const_instr_iterator E(MI.getParent()->instr_end());
8355     unsigned Lat = 0, Count = 0;
8356     for (++I; I != E && I->isBundledWithPred(); ++I) {
8357       ++Count;
8358       Lat = std::max(Lat, SchedModel.computeInstrLatency(&*I));
8359     }
8360     return Lat + Count - 1;
8361   }
8362 
8363   return SchedModel.computeInstrLatency(&MI);
8364 }
8365 
8366 InstructionUniformity
8367 SIInstrInfo::getGenericInstructionUniformity(const MachineInstr &MI) const {
8368   unsigned opcode = MI.getOpcode();
8369   if (opcode == AMDGPU::G_INTRINSIC ||
8370       opcode == AMDGPU::G_INTRINSIC_W_SIDE_EFFECTS) {
8371     return AMDGPU::isIntrinsicSourceOfDivergence(MI.getIntrinsicID())
8372                ? InstructionUniformity::NeverUniform
8373                : InstructionUniformity::AlwaysUniform;
8374   }
8375 
8376   // Loads from the private and flat address spaces are divergent, because
8377   // threads can execute the load instruction with the same inputs and get
8378   // different results.
8379   //
8380   // All other loads are not divergent, because if threads issue loads with the
8381   // same arguments, they will always get the same result.
8382   if (opcode == AMDGPU::G_LOAD) {
8383     if (MI.memoperands_empty())
8384       return InstructionUniformity::NeverUniform; // conservative assumption
8385 
8386     if (llvm::any_of(MI.memoperands(), [](const MachineMemOperand *mmo) {
8387           return mmo->getAddrSpace() == AMDGPUAS::PRIVATE_ADDRESS ||
8388                  mmo->getAddrSpace() == AMDGPUAS::FLAT_ADDRESS;
8389         })) {
8390       // At least one MMO in a non-global address space.
8391       return InstructionUniformity::NeverUniform;
8392     }
8393     return InstructionUniformity::Default;
8394   }
8395 
8396   if (SIInstrInfo::isGenericAtomicRMWOpcode(opcode) ||
8397       opcode == AMDGPU::G_ATOMIC_CMPXCHG ||
8398       opcode == AMDGPU::G_ATOMIC_CMPXCHG_WITH_SUCCESS) {
8399     return InstructionUniformity::NeverUniform;
8400   }
8401   return InstructionUniformity::Default;
8402 }
8403 
8404 InstructionUniformity
8405 SIInstrInfo::getInstructionUniformity(const MachineInstr &MI) const {
8406   // Atomics are divergent because they are executed sequentially: when an
8407   // atomic operation refers to the same address in each thread, then each
8408   // thread after the first sees the value written by the previous thread as
8409   // original value.
8410 
8411   if (isAtomic(MI))
8412     return InstructionUniformity::NeverUniform;
8413 
8414   // Loads from the private and flat address spaces are divergent, because
8415   // threads can execute the load instruction with the same inputs and get
8416   // different results.
8417   if (isFLAT(MI) && MI.mayLoad()) {
8418     if (MI.memoperands_empty())
8419       return InstructionUniformity::NeverUniform; // conservative assumption
8420 
8421     if (llvm::any_of(MI.memoperands(), [](const MachineMemOperand *mmo) {
8422           return mmo->getAddrSpace() == AMDGPUAS::PRIVATE_ADDRESS ||
8423                  mmo->getAddrSpace() == AMDGPUAS::FLAT_ADDRESS;
8424         })) {
8425       // At least one MMO in a non-global address space.
8426       return InstructionUniformity::NeverUniform;
8427     }
8428 
8429     return InstructionUniformity::Default;
8430   }
8431 
8432   unsigned opcode = MI.getOpcode();
8433   if (opcode == AMDGPU::COPY) {
8434     const MachineOperand &srcOp = MI.getOperand(1);
8435     if (srcOp.isReg() && srcOp.getReg().isPhysical()) {
8436       const TargetRegisterClass *regClass = RI.getPhysRegBaseClass(srcOp.getReg());
8437       return RI.isSGPRClass(regClass) ? InstructionUniformity::AlwaysUniform
8438                                       : InstructionUniformity::NeverUniform;
8439     }
8440     return InstructionUniformity::Default;
8441   }
8442   if (opcode == AMDGPU::INLINEASM || opcode == AMDGPU::INLINEASM_BR) {
8443     const MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
8444     for (auto &op : MI.operands()) {
8445       if (!op.isReg() || !op.isDef())
8446         continue;
8447       auto *RC = MRI.getRegClass(op.getReg());
8448       if (!RC || RI.isDivergentRegClass(RC))
8449         return InstructionUniformity::NeverUniform;
8450     }
8451     return InstructionUniformity::AlwaysUniform;
8452   }
8453   if (opcode == AMDGPU::V_READLANE_B32 || opcode == AMDGPU::V_READFIRSTLANE_B32)
8454     return InstructionUniformity::AlwaysUniform;
8455 
8456   if (opcode == AMDGPU::V_WRITELANE_B32)
8457     return InstructionUniformity::NeverUniform;
8458 
8459   // GMIR handling
8460   if (SIInstrInfo::isGenericOpcode(opcode))
8461     return SIInstrInfo::getGenericInstructionUniformity(MI);
8462 
8463   // Handling $vpgr reads
8464   for (auto srcOp : MI.operands()) {
8465     if (srcOp.isReg() && srcOp.getReg().isPhysical()) {
8466       const TargetRegisterClass *regClass = RI.getPhysRegBaseClass(srcOp.getReg());
8467       if (RI.isVGPRClass(regClass))
8468         return InstructionUniformity::NeverUniform;
8469     }
8470   }
8471 
8472   // TODO: Uniformity check condtions above can be rearranged for more
8473   // redability
8474 
8475   // TODO: amdgcn.{ballot, [if]cmp} should be AlwaysUniform, but they are
8476   //       currently turned into no-op COPYs by SelectionDAG ISel and are
8477   //       therefore no longer recognizable.
8478 
8479   return InstructionUniformity::Default;
8480 }
8481 
8482 unsigned SIInstrInfo::getDSShaderTypeValue(const MachineFunction &MF) {
8483   switch (MF.getFunction().getCallingConv()) {
8484   case CallingConv::AMDGPU_PS:
8485     return 1;
8486   case CallingConv::AMDGPU_VS:
8487     return 2;
8488   case CallingConv::AMDGPU_GS:
8489     return 3;
8490   case CallingConv::AMDGPU_HS:
8491   case CallingConv::AMDGPU_LS:
8492   case CallingConv::AMDGPU_ES:
8493     report_fatal_error("ds_ordered_count unsupported for this calling conv");
8494   case CallingConv::AMDGPU_CS:
8495   case CallingConv::AMDGPU_KERNEL:
8496   case CallingConv::C:
8497   case CallingConv::Fast:
8498   default:
8499     // Assume other calling conventions are various compute callable functions
8500     return 0;
8501   }
8502 }
8503 
8504 bool SIInstrInfo::analyzeCompare(const MachineInstr &MI, Register &SrcReg,
8505                                  Register &SrcReg2, int64_t &CmpMask,
8506                                  int64_t &CmpValue) const {
8507   if (!MI.getOperand(0).isReg() || MI.getOperand(0).getSubReg())
8508     return false;
8509 
8510   switch (MI.getOpcode()) {
8511   default:
8512     break;
8513   case AMDGPU::S_CMP_EQ_U32:
8514   case AMDGPU::S_CMP_EQ_I32:
8515   case AMDGPU::S_CMP_LG_U32:
8516   case AMDGPU::S_CMP_LG_I32:
8517   case AMDGPU::S_CMP_LT_U32:
8518   case AMDGPU::S_CMP_LT_I32:
8519   case AMDGPU::S_CMP_GT_U32:
8520   case AMDGPU::S_CMP_GT_I32:
8521   case AMDGPU::S_CMP_LE_U32:
8522   case AMDGPU::S_CMP_LE_I32:
8523   case AMDGPU::S_CMP_GE_U32:
8524   case AMDGPU::S_CMP_GE_I32:
8525   case AMDGPU::S_CMP_EQ_U64:
8526   case AMDGPU::S_CMP_LG_U64:
8527     SrcReg = MI.getOperand(0).getReg();
8528     if (MI.getOperand(1).isReg()) {
8529       if (MI.getOperand(1).getSubReg())
8530         return false;
8531       SrcReg2 = MI.getOperand(1).getReg();
8532       CmpValue = 0;
8533     } else if (MI.getOperand(1).isImm()) {
8534       SrcReg2 = Register();
8535       CmpValue = MI.getOperand(1).getImm();
8536     } else {
8537       return false;
8538     }
8539     CmpMask = ~0;
8540     return true;
8541   case AMDGPU::S_CMPK_EQ_U32:
8542   case AMDGPU::S_CMPK_EQ_I32:
8543   case AMDGPU::S_CMPK_LG_U32:
8544   case AMDGPU::S_CMPK_LG_I32:
8545   case AMDGPU::S_CMPK_LT_U32:
8546   case AMDGPU::S_CMPK_LT_I32:
8547   case AMDGPU::S_CMPK_GT_U32:
8548   case AMDGPU::S_CMPK_GT_I32:
8549   case AMDGPU::S_CMPK_LE_U32:
8550   case AMDGPU::S_CMPK_LE_I32:
8551   case AMDGPU::S_CMPK_GE_U32:
8552   case AMDGPU::S_CMPK_GE_I32:
8553     SrcReg = MI.getOperand(0).getReg();
8554     SrcReg2 = Register();
8555     CmpValue = MI.getOperand(1).getImm();
8556     CmpMask = ~0;
8557     return true;
8558   }
8559 
8560   return false;
8561 }
8562 
8563 bool SIInstrInfo::optimizeCompareInstr(MachineInstr &CmpInstr, Register SrcReg,
8564                                        Register SrcReg2, int64_t CmpMask,
8565                                        int64_t CmpValue,
8566                                        const MachineRegisterInfo *MRI) const {
8567   if (!SrcReg || SrcReg.isPhysical())
8568     return false;
8569 
8570   if (SrcReg2 && !getFoldableImm(SrcReg2, *MRI, CmpValue))
8571     return false;
8572 
8573   const auto optimizeCmpAnd = [&CmpInstr, SrcReg, CmpValue, MRI,
8574                                this](int64_t ExpectedValue, unsigned SrcSize,
8575                                      bool IsReversible, bool IsSigned) -> bool {
8576     // s_cmp_eq_u32 (s_and_b32 $src, 1 << n), 1 << n => s_and_b32 $src, 1 << n
8577     // s_cmp_eq_i32 (s_and_b32 $src, 1 << n), 1 << n => s_and_b32 $src, 1 << n
8578     // s_cmp_ge_u32 (s_and_b32 $src, 1 << n), 1 << n => s_and_b32 $src, 1 << n
8579     // s_cmp_ge_i32 (s_and_b32 $src, 1 << n), 1 << n => s_and_b32 $src, 1 << n
8580     // s_cmp_eq_u64 (s_and_b64 $src, 1 << n), 1 << n => s_and_b64 $src, 1 << n
8581     // s_cmp_lg_u32 (s_and_b32 $src, 1 << n), 0 => s_and_b32 $src, 1 << n
8582     // s_cmp_lg_i32 (s_and_b32 $src, 1 << n), 0 => s_and_b32 $src, 1 << n
8583     // s_cmp_gt_u32 (s_and_b32 $src, 1 << n), 0 => s_and_b32 $src, 1 << n
8584     // s_cmp_gt_i32 (s_and_b32 $src, 1 << n), 0 => s_and_b32 $src, 1 << n
8585     // s_cmp_lg_u64 (s_and_b64 $src, 1 << n), 0 => s_and_b64 $src, 1 << n
8586     //
8587     // Signed ge/gt are not used for the sign bit.
8588     //
8589     // If result of the AND is unused except in the compare:
8590     // s_and_b(32|64) $src, 1 << n => s_bitcmp1_b(32|64) $src, n
8591     //
8592     // s_cmp_eq_u32 (s_and_b32 $src, 1 << n), 0 => s_bitcmp0_b32 $src, n
8593     // s_cmp_eq_i32 (s_and_b32 $src, 1 << n), 0 => s_bitcmp0_b32 $src, n
8594     // s_cmp_eq_u64 (s_and_b64 $src, 1 << n), 0 => s_bitcmp0_b64 $src, n
8595     // s_cmp_lg_u32 (s_and_b32 $src, 1 << n), 1 << n => s_bitcmp0_b32 $src, n
8596     // s_cmp_lg_i32 (s_and_b32 $src, 1 << n), 1 << n => s_bitcmp0_b32 $src, n
8597     // s_cmp_lg_u64 (s_and_b64 $src, 1 << n), 1 << n => s_bitcmp0_b64 $src, n
8598 
8599     MachineInstr *Def = MRI->getUniqueVRegDef(SrcReg);
8600     if (!Def || Def->getParent() != CmpInstr.getParent())
8601       return false;
8602 
8603     if (Def->getOpcode() != AMDGPU::S_AND_B32 &&
8604         Def->getOpcode() != AMDGPU::S_AND_B64)
8605       return false;
8606 
8607     int64_t Mask;
8608     const auto isMask = [&Mask, SrcSize](const MachineOperand *MO) -> bool {
8609       if (MO->isImm())
8610         Mask = MO->getImm();
8611       else if (!getFoldableImm(MO, Mask))
8612         return false;
8613       Mask &= maxUIntN(SrcSize);
8614       return isPowerOf2_64(Mask);
8615     };
8616 
8617     MachineOperand *SrcOp = &Def->getOperand(1);
8618     if (isMask(SrcOp))
8619       SrcOp = &Def->getOperand(2);
8620     else if (isMask(&Def->getOperand(2)))
8621       SrcOp = &Def->getOperand(1);
8622     else
8623       return false;
8624 
8625     unsigned BitNo = countTrailingZeros((uint64_t)Mask);
8626     if (IsSigned && BitNo == SrcSize - 1)
8627       return false;
8628 
8629     ExpectedValue <<= BitNo;
8630 
8631     bool IsReversedCC = false;
8632     if (CmpValue != ExpectedValue) {
8633       if (!IsReversible)
8634         return false;
8635       IsReversedCC = CmpValue == (ExpectedValue ^ Mask);
8636       if (!IsReversedCC)
8637         return false;
8638     }
8639 
8640     Register DefReg = Def->getOperand(0).getReg();
8641     if (IsReversedCC && !MRI->hasOneNonDBGUse(DefReg))
8642       return false;
8643 
8644     for (auto I = std::next(Def->getIterator()), E = CmpInstr.getIterator();
8645          I != E; ++I) {
8646       if (I->modifiesRegister(AMDGPU::SCC, &RI) ||
8647           I->killsRegister(AMDGPU::SCC, &RI))
8648         return false;
8649     }
8650 
8651     MachineOperand *SccDef = Def->findRegisterDefOperand(AMDGPU::SCC);
8652     SccDef->setIsDead(false);
8653     CmpInstr.eraseFromParent();
8654 
8655     if (!MRI->use_nodbg_empty(DefReg)) {
8656       assert(!IsReversedCC);
8657       return true;
8658     }
8659 
8660     // Replace AND with unused result with a S_BITCMP.
8661     MachineBasicBlock *MBB = Def->getParent();
8662 
8663     unsigned NewOpc = (SrcSize == 32) ? IsReversedCC ? AMDGPU::S_BITCMP0_B32
8664                                                      : AMDGPU::S_BITCMP1_B32
8665                                       : IsReversedCC ? AMDGPU::S_BITCMP0_B64
8666                                                      : AMDGPU::S_BITCMP1_B64;
8667 
8668     BuildMI(*MBB, Def, Def->getDebugLoc(), get(NewOpc))
8669       .add(*SrcOp)
8670       .addImm(BitNo);
8671     Def->eraseFromParent();
8672 
8673     return true;
8674   };
8675 
8676   switch (CmpInstr.getOpcode()) {
8677   default:
8678     break;
8679   case AMDGPU::S_CMP_EQ_U32:
8680   case AMDGPU::S_CMP_EQ_I32:
8681   case AMDGPU::S_CMPK_EQ_U32:
8682   case AMDGPU::S_CMPK_EQ_I32:
8683     return optimizeCmpAnd(1, 32, true, false);
8684   case AMDGPU::S_CMP_GE_U32:
8685   case AMDGPU::S_CMPK_GE_U32:
8686     return optimizeCmpAnd(1, 32, false, false);
8687   case AMDGPU::S_CMP_GE_I32:
8688   case AMDGPU::S_CMPK_GE_I32:
8689     return optimizeCmpAnd(1, 32, false, true);
8690   case AMDGPU::S_CMP_EQ_U64:
8691     return optimizeCmpAnd(1, 64, true, false);
8692   case AMDGPU::S_CMP_LG_U32:
8693   case AMDGPU::S_CMP_LG_I32:
8694   case AMDGPU::S_CMPK_LG_U32:
8695   case AMDGPU::S_CMPK_LG_I32:
8696     return optimizeCmpAnd(0, 32, true, false);
8697   case AMDGPU::S_CMP_GT_U32:
8698   case AMDGPU::S_CMPK_GT_U32:
8699     return optimizeCmpAnd(0, 32, false, false);
8700   case AMDGPU::S_CMP_GT_I32:
8701   case AMDGPU::S_CMPK_GT_I32:
8702     return optimizeCmpAnd(0, 32, false, true);
8703   case AMDGPU::S_CMP_LG_U64:
8704     return optimizeCmpAnd(0, 64, true, false);
8705   }
8706 
8707   return false;
8708 }
8709 
8710 void SIInstrInfo::enforceOperandRCAlignment(MachineInstr &MI,
8711                                             unsigned OpName) const {
8712   if (!ST.needsAlignedVGPRs())
8713     return;
8714 
8715   int OpNo = AMDGPU::getNamedOperandIdx(MI.getOpcode(), OpName);
8716   if (OpNo < 0)
8717     return;
8718   MachineOperand &Op = MI.getOperand(OpNo);
8719   if (getOpSize(MI, OpNo) > 4)
8720     return;
8721 
8722   // Add implicit aligned super-reg to force alignment on the data operand.
8723   const DebugLoc &DL = MI.getDebugLoc();
8724   MachineBasicBlock *BB = MI.getParent();
8725   MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
8726   Register DataReg = Op.getReg();
8727   bool IsAGPR = RI.isAGPR(MRI, DataReg);
8728   Register Undef = MRI.createVirtualRegister(
8729       IsAGPR ? &AMDGPU::AGPR_32RegClass : &AMDGPU::VGPR_32RegClass);
8730   BuildMI(*BB, MI, DL, get(AMDGPU::IMPLICIT_DEF), Undef);
8731   Register NewVR =
8732       MRI.createVirtualRegister(IsAGPR ? &AMDGPU::AReg_64_Align2RegClass
8733                                        : &AMDGPU::VReg_64_Align2RegClass);
8734   BuildMI(*BB, MI, DL, get(AMDGPU::REG_SEQUENCE), NewVR)
8735       .addReg(DataReg, 0, Op.getSubReg())
8736       .addImm(AMDGPU::sub0)
8737       .addReg(Undef)
8738       .addImm(AMDGPU::sub1);
8739   Op.setReg(NewVR);
8740   Op.setSubReg(AMDGPU::sub0);
8741   MI.addOperand(MachineOperand::CreateReg(NewVR, false, true));
8742 }
8743