xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AMDGPU/AMDGPUAtomicOptimizer.cpp (revision cfd6422a5217410fbd66f7a7a8a64d9d85e61229)
1 //===-- AMDGPUAtomicOptimizer.cpp -----------------------------------------===//
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 /// This pass optimizes atomic operations by using a single lane of a wavefront
11 /// to perform the atomic operation, thus reducing contention on that memory
12 /// location.
13 //
14 //===----------------------------------------------------------------------===//
15 
16 #include "AMDGPU.h"
17 #include "AMDGPUSubtarget.h"
18 #include "SIDefines.h"
19 #include "llvm/Analysis/LegacyDivergenceAnalysis.h"
20 #include "llvm/CodeGen/TargetPassConfig.h"
21 #include "llvm/IR/IRBuilder.h"
22 #include "llvm/IR/InstVisitor.h"
23 #include "llvm/InitializePasses.h"
24 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
25 
26 #define DEBUG_TYPE "amdgpu-atomic-optimizer"
27 
28 using namespace llvm;
29 using namespace llvm::AMDGPU;
30 
31 namespace {
32 
33 struct ReplacementInfo {
34   Instruction *I;
35   AtomicRMWInst::BinOp Op;
36   unsigned ValIdx;
37   bool ValDivergent;
38 };
39 
40 class AMDGPUAtomicOptimizer : public FunctionPass,
41                               public InstVisitor<AMDGPUAtomicOptimizer> {
42 private:
43   SmallVector<ReplacementInfo, 8> ToReplace;
44   const LegacyDivergenceAnalysis *DA;
45   const DataLayout *DL;
46   DominatorTree *DT;
47   const GCNSubtarget *ST;
48   bool IsPixelShader;
49 
50   Value *buildScan(IRBuilder<> &B, AtomicRMWInst::BinOp Op, Value *V,
51                    Value *const Identity) const;
52   Value *buildShiftRight(IRBuilder<> &B, Value *V, Value *const Identity) const;
53   void optimizeAtomic(Instruction &I, AtomicRMWInst::BinOp Op, unsigned ValIdx,
54                       bool ValDivergent) const;
55 
56 public:
57   static char ID;
58 
59   AMDGPUAtomicOptimizer() : FunctionPass(ID) {}
60 
61   bool runOnFunction(Function &F) override;
62 
63   void getAnalysisUsage(AnalysisUsage &AU) const override {
64     AU.addPreserved<DominatorTreeWrapperPass>();
65     AU.addRequired<LegacyDivergenceAnalysis>();
66     AU.addRequired<TargetPassConfig>();
67   }
68 
69   void visitAtomicRMWInst(AtomicRMWInst &I);
70   void visitIntrinsicInst(IntrinsicInst &I);
71 };
72 
73 } // namespace
74 
75 char AMDGPUAtomicOptimizer::ID = 0;
76 
77 char &llvm::AMDGPUAtomicOptimizerID = AMDGPUAtomicOptimizer::ID;
78 
79 bool AMDGPUAtomicOptimizer::runOnFunction(Function &F) {
80   if (skipFunction(F)) {
81     return false;
82   }
83 
84   DA = &getAnalysis<LegacyDivergenceAnalysis>();
85   DL = &F.getParent()->getDataLayout();
86   DominatorTreeWrapperPass *const DTW =
87       getAnalysisIfAvailable<DominatorTreeWrapperPass>();
88   DT = DTW ? &DTW->getDomTree() : nullptr;
89   const TargetPassConfig &TPC = getAnalysis<TargetPassConfig>();
90   const TargetMachine &TM = TPC.getTM<TargetMachine>();
91   ST = &TM.getSubtarget<GCNSubtarget>(F);
92   IsPixelShader = F.getCallingConv() == CallingConv::AMDGPU_PS;
93 
94   visit(F);
95 
96   const bool Changed = !ToReplace.empty();
97 
98   for (ReplacementInfo &Info : ToReplace) {
99     optimizeAtomic(*Info.I, Info.Op, Info.ValIdx, Info.ValDivergent);
100   }
101 
102   ToReplace.clear();
103 
104   return Changed;
105 }
106 
107 void AMDGPUAtomicOptimizer::visitAtomicRMWInst(AtomicRMWInst &I) {
108   // Early exit for unhandled address space atomic instructions.
109   switch (I.getPointerAddressSpace()) {
110   default:
111     return;
112   case AMDGPUAS::GLOBAL_ADDRESS:
113   case AMDGPUAS::LOCAL_ADDRESS:
114     break;
115   }
116 
117   AtomicRMWInst::BinOp Op = I.getOperation();
118 
119   switch (Op) {
120   default:
121     return;
122   case AtomicRMWInst::Add:
123   case AtomicRMWInst::Sub:
124   case AtomicRMWInst::And:
125   case AtomicRMWInst::Or:
126   case AtomicRMWInst::Xor:
127   case AtomicRMWInst::Max:
128   case AtomicRMWInst::Min:
129   case AtomicRMWInst::UMax:
130   case AtomicRMWInst::UMin:
131     break;
132   }
133 
134   const unsigned PtrIdx = 0;
135   const unsigned ValIdx = 1;
136 
137   // If the pointer operand is divergent, then each lane is doing an atomic
138   // operation on a different address, and we cannot optimize that.
139   if (DA->isDivergentUse(&I.getOperandUse(PtrIdx))) {
140     return;
141   }
142 
143   const bool ValDivergent = DA->isDivergentUse(&I.getOperandUse(ValIdx));
144 
145   // If the value operand is divergent, each lane is contributing a different
146   // value to the atomic calculation. We can only optimize divergent values if
147   // we have DPP available on our subtarget, and the atomic operation is 32
148   // bits.
149   if (ValDivergent &&
150       (!ST->hasDPP() || DL->getTypeSizeInBits(I.getType()) != 32)) {
151     return;
152   }
153 
154   // If we get here, we can optimize the atomic using a single wavefront-wide
155   // atomic operation to do the calculation for the entire wavefront, so
156   // remember the instruction so we can come back to it.
157   const ReplacementInfo Info = {&I, Op, ValIdx, ValDivergent};
158 
159   ToReplace.push_back(Info);
160 }
161 
162 void AMDGPUAtomicOptimizer::visitIntrinsicInst(IntrinsicInst &I) {
163   AtomicRMWInst::BinOp Op;
164 
165   switch (I.getIntrinsicID()) {
166   default:
167     return;
168   case Intrinsic::amdgcn_buffer_atomic_add:
169   case Intrinsic::amdgcn_struct_buffer_atomic_add:
170   case Intrinsic::amdgcn_raw_buffer_atomic_add:
171     Op = AtomicRMWInst::Add;
172     break;
173   case Intrinsic::amdgcn_buffer_atomic_sub:
174   case Intrinsic::amdgcn_struct_buffer_atomic_sub:
175   case Intrinsic::amdgcn_raw_buffer_atomic_sub:
176     Op = AtomicRMWInst::Sub;
177     break;
178   case Intrinsic::amdgcn_buffer_atomic_and:
179   case Intrinsic::amdgcn_struct_buffer_atomic_and:
180   case Intrinsic::amdgcn_raw_buffer_atomic_and:
181     Op = AtomicRMWInst::And;
182     break;
183   case Intrinsic::amdgcn_buffer_atomic_or:
184   case Intrinsic::amdgcn_struct_buffer_atomic_or:
185   case Intrinsic::amdgcn_raw_buffer_atomic_or:
186     Op = AtomicRMWInst::Or;
187     break;
188   case Intrinsic::amdgcn_buffer_atomic_xor:
189   case Intrinsic::amdgcn_struct_buffer_atomic_xor:
190   case Intrinsic::amdgcn_raw_buffer_atomic_xor:
191     Op = AtomicRMWInst::Xor;
192     break;
193   case Intrinsic::amdgcn_buffer_atomic_smin:
194   case Intrinsic::amdgcn_struct_buffer_atomic_smin:
195   case Intrinsic::amdgcn_raw_buffer_atomic_smin:
196     Op = AtomicRMWInst::Min;
197     break;
198   case Intrinsic::amdgcn_buffer_atomic_umin:
199   case Intrinsic::amdgcn_struct_buffer_atomic_umin:
200   case Intrinsic::amdgcn_raw_buffer_atomic_umin:
201     Op = AtomicRMWInst::UMin;
202     break;
203   case Intrinsic::amdgcn_buffer_atomic_smax:
204   case Intrinsic::amdgcn_struct_buffer_atomic_smax:
205   case Intrinsic::amdgcn_raw_buffer_atomic_smax:
206     Op = AtomicRMWInst::Max;
207     break;
208   case Intrinsic::amdgcn_buffer_atomic_umax:
209   case Intrinsic::amdgcn_struct_buffer_atomic_umax:
210   case Intrinsic::amdgcn_raw_buffer_atomic_umax:
211     Op = AtomicRMWInst::UMax;
212     break;
213   }
214 
215   const unsigned ValIdx = 0;
216 
217   const bool ValDivergent = DA->isDivergentUse(&I.getOperandUse(ValIdx));
218 
219   // If the value operand is divergent, each lane is contributing a different
220   // value to the atomic calculation. We can only optimize divergent values if
221   // we have DPP available on our subtarget, and the atomic operation is 32
222   // bits.
223   if (ValDivergent &&
224       (!ST->hasDPP() || DL->getTypeSizeInBits(I.getType()) != 32)) {
225     return;
226   }
227 
228   // If any of the other arguments to the intrinsic are divergent, we can't
229   // optimize the operation.
230   for (unsigned Idx = 1; Idx < I.getNumOperands(); Idx++) {
231     if (DA->isDivergentUse(&I.getOperandUse(Idx))) {
232       return;
233     }
234   }
235 
236   // If we get here, we can optimize the atomic using a single wavefront-wide
237   // atomic operation to do the calculation for the entire wavefront, so
238   // remember the instruction so we can come back to it.
239   const ReplacementInfo Info = {&I, Op, ValIdx, ValDivergent};
240 
241   ToReplace.push_back(Info);
242 }
243 
244 // Use the builder to create the non-atomic counterpart of the specified
245 // atomicrmw binary op.
246 static Value *buildNonAtomicBinOp(IRBuilder<> &B, AtomicRMWInst::BinOp Op,
247                                   Value *LHS, Value *RHS) {
248   CmpInst::Predicate Pred;
249 
250   switch (Op) {
251   default:
252     llvm_unreachable("Unhandled atomic op");
253   case AtomicRMWInst::Add:
254     return B.CreateBinOp(Instruction::Add, LHS, RHS);
255   case AtomicRMWInst::Sub:
256     return B.CreateBinOp(Instruction::Sub, LHS, RHS);
257   case AtomicRMWInst::And:
258     return B.CreateBinOp(Instruction::And, LHS, RHS);
259   case AtomicRMWInst::Or:
260     return B.CreateBinOp(Instruction::Or, LHS, RHS);
261   case AtomicRMWInst::Xor:
262     return B.CreateBinOp(Instruction::Xor, LHS, RHS);
263 
264   case AtomicRMWInst::Max:
265     Pred = CmpInst::ICMP_SGT;
266     break;
267   case AtomicRMWInst::Min:
268     Pred = CmpInst::ICMP_SLT;
269     break;
270   case AtomicRMWInst::UMax:
271     Pred = CmpInst::ICMP_UGT;
272     break;
273   case AtomicRMWInst::UMin:
274     Pred = CmpInst::ICMP_ULT;
275     break;
276   }
277   Value *Cond = B.CreateICmp(Pred, LHS, RHS);
278   return B.CreateSelect(Cond, LHS, RHS);
279 }
280 
281 // Use the builder to create an inclusive scan of V across the wavefront, with
282 // all lanes active.
283 Value *AMDGPUAtomicOptimizer::buildScan(IRBuilder<> &B, AtomicRMWInst::BinOp Op,
284                                         Value *V, Value *const Identity) const {
285   Type *const Ty = V->getType();
286   Module *M = B.GetInsertBlock()->getModule();
287   Function *UpdateDPP =
288       Intrinsic::getDeclaration(M, Intrinsic::amdgcn_update_dpp, Ty);
289   Function *PermLaneX16 =
290       Intrinsic::getDeclaration(M, Intrinsic::amdgcn_permlanex16, {});
291   Function *ReadLane =
292       Intrinsic::getDeclaration(M, Intrinsic::amdgcn_readlane, {});
293 
294   for (unsigned Idx = 0; Idx < 4; Idx++) {
295     V = buildNonAtomicBinOp(
296         B, Op, V,
297         B.CreateCall(UpdateDPP,
298                      {Identity, V, B.getInt32(DPP::ROW_SHR0 | 1 << Idx),
299                       B.getInt32(0xf), B.getInt32(0xf), B.getFalse()}));
300   }
301   if (ST->hasDPPBroadcasts()) {
302     // GFX9 has DPP row broadcast operations.
303     V = buildNonAtomicBinOp(
304         B, Op, V,
305         B.CreateCall(UpdateDPP,
306                      {Identity, V, B.getInt32(DPP::BCAST15), B.getInt32(0xa),
307                       B.getInt32(0xf), B.getFalse()}));
308     V = buildNonAtomicBinOp(
309         B, Op, V,
310         B.CreateCall(UpdateDPP,
311                      {Identity, V, B.getInt32(DPP::BCAST31), B.getInt32(0xc),
312                       B.getInt32(0xf), B.getFalse()}));
313   } else {
314     // On GFX10 all DPP operations are confined to a single row. To get cross-
315     // row operations we have to use permlane or readlane.
316 
317     // Combine lane 15 into lanes 16..31 (and, for wave 64, lane 47 into lanes
318     // 48..63).
319     Value *const PermX =
320         B.CreateCall(PermLaneX16, {V, V, B.getInt32(-1), B.getInt32(-1),
321                                    B.getFalse(), B.getFalse()});
322     V = buildNonAtomicBinOp(
323         B, Op, V,
324         B.CreateCall(UpdateDPP,
325                      {Identity, PermX, B.getInt32(DPP::QUAD_PERM_ID),
326                       B.getInt32(0xa), B.getInt32(0xf), B.getFalse()}));
327     if (!ST->isWave32()) {
328       // Combine lane 31 into lanes 32..63.
329       Value *const Lane31 = B.CreateCall(ReadLane, {V, B.getInt32(31)});
330       V = buildNonAtomicBinOp(
331           B, Op, V,
332           B.CreateCall(UpdateDPP,
333                        {Identity, Lane31, B.getInt32(DPP::QUAD_PERM_ID),
334                         B.getInt32(0xc), B.getInt32(0xf), B.getFalse()}));
335     }
336   }
337   return V;
338 }
339 
340 // Use the builder to create a shift right of V across the wavefront, with all
341 // lanes active, to turn an inclusive scan into an exclusive scan.
342 Value *AMDGPUAtomicOptimizer::buildShiftRight(IRBuilder<> &B, Value *V,
343                                               Value *const Identity) const {
344   Type *const Ty = V->getType();
345   Module *M = B.GetInsertBlock()->getModule();
346   Function *UpdateDPP =
347       Intrinsic::getDeclaration(M, Intrinsic::amdgcn_update_dpp, Ty);
348   Function *ReadLane =
349       Intrinsic::getDeclaration(M, Intrinsic::amdgcn_readlane, {});
350   Function *WriteLane =
351       Intrinsic::getDeclaration(M, Intrinsic::amdgcn_writelane, {});
352 
353   if (ST->hasDPPWavefrontShifts()) {
354     // GFX9 has DPP wavefront shift operations.
355     V = B.CreateCall(UpdateDPP,
356                      {Identity, V, B.getInt32(DPP::WAVE_SHR1), B.getInt32(0xf),
357                       B.getInt32(0xf), B.getFalse()});
358   } else {
359     // On GFX10 all DPP operations are confined to a single row. To get cross-
360     // row operations we have to use permlane or readlane.
361     Value *Old = V;
362     V = B.CreateCall(UpdateDPP,
363                      {Identity, V, B.getInt32(DPP::ROW_SHR0 + 1),
364                       B.getInt32(0xf), B.getInt32(0xf), B.getFalse()});
365 
366     // Copy the old lane 15 to the new lane 16.
367     V = B.CreateCall(WriteLane, {B.CreateCall(ReadLane, {Old, B.getInt32(15)}),
368                                  B.getInt32(16), V});
369 
370     if (!ST->isWave32()) {
371       // Copy the old lane 31 to the new lane 32.
372       V = B.CreateCall(
373           WriteLane,
374           {B.CreateCall(ReadLane, {Old, B.getInt32(31)}), B.getInt32(32), V});
375 
376       // Copy the old lane 47 to the new lane 48.
377       V = B.CreateCall(
378           WriteLane,
379           {B.CreateCall(ReadLane, {Old, B.getInt32(47)}), B.getInt32(48), V});
380     }
381   }
382 
383   return V;
384 }
385 
386 static APInt getIdentityValueForAtomicOp(AtomicRMWInst::BinOp Op,
387                                          unsigned BitWidth) {
388   switch (Op) {
389   default:
390     llvm_unreachable("Unhandled atomic op");
391   case AtomicRMWInst::Add:
392   case AtomicRMWInst::Sub:
393   case AtomicRMWInst::Or:
394   case AtomicRMWInst::Xor:
395   case AtomicRMWInst::UMax:
396     return APInt::getMinValue(BitWidth);
397   case AtomicRMWInst::And:
398   case AtomicRMWInst::UMin:
399     return APInt::getMaxValue(BitWidth);
400   case AtomicRMWInst::Max:
401     return APInt::getSignedMinValue(BitWidth);
402   case AtomicRMWInst::Min:
403     return APInt::getSignedMaxValue(BitWidth);
404   }
405 }
406 
407 void AMDGPUAtomicOptimizer::optimizeAtomic(Instruction &I,
408                                            AtomicRMWInst::BinOp Op,
409                                            unsigned ValIdx,
410                                            bool ValDivergent) const {
411   // Start building just before the instruction.
412   IRBuilder<> B(&I);
413 
414   // If we are in a pixel shader, because of how we have to mask out helper
415   // lane invocations, we need to record the entry and exit BB's.
416   BasicBlock *PixelEntryBB = nullptr;
417   BasicBlock *PixelExitBB = nullptr;
418 
419   // If we're optimizing an atomic within a pixel shader, we need to wrap the
420   // entire atomic operation in a helper-lane check. We do not want any helper
421   // lanes that are around only for the purposes of derivatives to take part
422   // in any cross-lane communication, and we use a branch on whether the lane is
423   // live to do this.
424   if (IsPixelShader) {
425     // Record I's original position as the entry block.
426     PixelEntryBB = I.getParent();
427 
428     Value *const Cond = B.CreateIntrinsic(Intrinsic::amdgcn_ps_live, {}, {});
429     Instruction *const NonHelperTerminator =
430         SplitBlockAndInsertIfThen(Cond, &I, false, nullptr, DT, nullptr);
431 
432     // Record I's new position as the exit block.
433     PixelExitBB = I.getParent();
434 
435     I.moveBefore(NonHelperTerminator);
436     B.SetInsertPoint(&I);
437   }
438 
439   Type *const Ty = I.getType();
440   const unsigned TyBitWidth = DL->getTypeSizeInBits(Ty);
441   auto *const VecTy = FixedVectorType::get(B.getInt32Ty(), 2);
442 
443   // This is the value in the atomic operation we need to combine in order to
444   // reduce the number of atomic operations.
445   Value *const V = I.getOperand(ValIdx);
446 
447   // We need to know how many lanes are active within the wavefront, and we do
448   // this by doing a ballot of active lanes.
449   Type *const WaveTy = B.getIntNTy(ST->getWavefrontSize());
450   CallInst *const Ballot =
451       B.CreateIntrinsic(Intrinsic::amdgcn_ballot, WaveTy, B.getTrue());
452 
453   // We need to know how many lanes are active within the wavefront that are
454   // below us. If we counted each lane linearly starting from 0, a lane is
455   // below us only if its associated index was less than ours. We do this by
456   // using the mbcnt intrinsic.
457   Value *Mbcnt;
458   if (ST->isWave32()) {
459     Mbcnt = B.CreateIntrinsic(Intrinsic::amdgcn_mbcnt_lo, {},
460                               {Ballot, B.getInt32(0)});
461   } else {
462     Value *const BitCast = B.CreateBitCast(Ballot, VecTy);
463     Value *const ExtractLo = B.CreateExtractElement(BitCast, B.getInt32(0));
464     Value *const ExtractHi = B.CreateExtractElement(BitCast, B.getInt32(1));
465     Mbcnt = B.CreateIntrinsic(Intrinsic::amdgcn_mbcnt_lo, {},
466                               {ExtractLo, B.getInt32(0)});
467     Mbcnt =
468         B.CreateIntrinsic(Intrinsic::amdgcn_mbcnt_hi, {}, {ExtractHi, Mbcnt});
469   }
470   Mbcnt = B.CreateIntCast(Mbcnt, Ty, false);
471 
472   Value *const Identity = B.getInt(getIdentityValueForAtomicOp(Op, TyBitWidth));
473 
474   Value *ExclScan = nullptr;
475   Value *NewV = nullptr;
476 
477   // If we have a divergent value in each lane, we need to combine the value
478   // using DPP.
479   if (ValDivergent) {
480     // First we need to set all inactive invocations to the identity value, so
481     // that they can correctly contribute to the final result.
482     NewV = B.CreateIntrinsic(Intrinsic::amdgcn_set_inactive, Ty, {V, Identity});
483 
484     const AtomicRMWInst::BinOp ScanOp =
485         Op == AtomicRMWInst::Sub ? AtomicRMWInst::Add : Op;
486     NewV = buildScan(B, ScanOp, NewV, Identity);
487     ExclScan = buildShiftRight(B, NewV, Identity);
488 
489     // Read the value from the last lane, which has accumlated the values of
490     // each active lane in the wavefront. This will be our new value which we
491     // will provide to the atomic operation.
492     Value *const LastLaneIdx = B.getInt32(ST->getWavefrontSize() - 1);
493     if (TyBitWidth == 64) {
494       Value *const ExtractLo = B.CreateTrunc(NewV, B.getInt32Ty());
495       Value *const ExtractHi =
496           B.CreateTrunc(B.CreateLShr(NewV, 32), B.getInt32Ty());
497       CallInst *const ReadLaneLo = B.CreateIntrinsic(
498           Intrinsic::amdgcn_readlane, {}, {ExtractLo, LastLaneIdx});
499       CallInst *const ReadLaneHi = B.CreateIntrinsic(
500           Intrinsic::amdgcn_readlane, {}, {ExtractHi, LastLaneIdx});
501       Value *const PartialInsert = B.CreateInsertElement(
502           UndefValue::get(VecTy), ReadLaneLo, B.getInt32(0));
503       Value *const Insert =
504           B.CreateInsertElement(PartialInsert, ReadLaneHi, B.getInt32(1));
505       NewV = B.CreateBitCast(Insert, Ty);
506     } else if (TyBitWidth == 32) {
507       NewV = B.CreateIntrinsic(Intrinsic::amdgcn_readlane, {},
508                                {NewV, LastLaneIdx});
509     } else {
510       llvm_unreachable("Unhandled atomic bit width");
511     }
512 
513     // Finally mark the readlanes in the WWM section.
514     NewV = B.CreateIntrinsic(Intrinsic::amdgcn_wwm, Ty, NewV);
515   } else {
516     switch (Op) {
517     default:
518       llvm_unreachable("Unhandled atomic op");
519 
520     case AtomicRMWInst::Add:
521     case AtomicRMWInst::Sub: {
522       // The new value we will be contributing to the atomic operation is the
523       // old value times the number of active lanes.
524       Value *const Ctpop = B.CreateIntCast(
525           B.CreateUnaryIntrinsic(Intrinsic::ctpop, Ballot), Ty, false);
526       NewV = B.CreateMul(V, Ctpop);
527       break;
528     }
529 
530     case AtomicRMWInst::And:
531     case AtomicRMWInst::Or:
532     case AtomicRMWInst::Max:
533     case AtomicRMWInst::Min:
534     case AtomicRMWInst::UMax:
535     case AtomicRMWInst::UMin:
536       // These operations with a uniform value are idempotent: doing the atomic
537       // operation multiple times has the same effect as doing it once.
538       NewV = V;
539       break;
540 
541     case AtomicRMWInst::Xor:
542       // The new value we will be contributing to the atomic operation is the
543       // old value times the parity of the number of active lanes.
544       Value *const Ctpop = B.CreateIntCast(
545           B.CreateUnaryIntrinsic(Intrinsic::ctpop, Ballot), Ty, false);
546       NewV = B.CreateMul(V, B.CreateAnd(Ctpop, 1));
547       break;
548     }
549   }
550 
551   // We only want a single lane to enter our new control flow, and we do this
552   // by checking if there are any active lanes below us. Only one lane will
553   // have 0 active lanes below us, so that will be the only one to progress.
554   Value *const Cond = B.CreateICmpEQ(Mbcnt, B.getIntN(TyBitWidth, 0));
555 
556   // Store I's original basic block before we split the block.
557   BasicBlock *const EntryBB = I.getParent();
558 
559   // We need to introduce some new control flow to force a single lane to be
560   // active. We do this by splitting I's basic block at I, and introducing the
561   // new block such that:
562   // entry --> single_lane -\
563   //       \------------------> exit
564   Instruction *const SingleLaneTerminator =
565       SplitBlockAndInsertIfThen(Cond, &I, false, nullptr, DT, nullptr);
566 
567   // Move the IR builder into single_lane next.
568   B.SetInsertPoint(SingleLaneTerminator);
569 
570   // Clone the original atomic operation into single lane, replacing the
571   // original value with our newly created one.
572   Instruction *const NewI = I.clone();
573   B.Insert(NewI);
574   NewI->setOperand(ValIdx, NewV);
575 
576   // Move the IR builder into exit next, and start inserting just before the
577   // original instruction.
578   B.SetInsertPoint(&I);
579 
580   const bool NeedResult = !I.use_empty();
581   if (NeedResult) {
582     // Create a PHI node to get our new atomic result into the exit block.
583     PHINode *const PHI = B.CreatePHI(Ty, 2);
584     PHI->addIncoming(UndefValue::get(Ty), EntryBB);
585     PHI->addIncoming(NewI, SingleLaneTerminator->getParent());
586 
587     // We need to broadcast the value who was the lowest active lane (the first
588     // lane) to all other lanes in the wavefront. We use an intrinsic for this,
589     // but have to handle 64-bit broadcasts with two calls to this intrinsic.
590     Value *BroadcastI = nullptr;
591 
592     if (TyBitWidth == 64) {
593       Value *const ExtractLo = B.CreateTrunc(PHI, B.getInt32Ty());
594       Value *const ExtractHi =
595           B.CreateTrunc(B.CreateLShr(PHI, 32), B.getInt32Ty());
596       CallInst *const ReadFirstLaneLo =
597           B.CreateIntrinsic(Intrinsic::amdgcn_readfirstlane, {}, ExtractLo);
598       CallInst *const ReadFirstLaneHi =
599           B.CreateIntrinsic(Intrinsic::amdgcn_readfirstlane, {}, ExtractHi);
600       Value *const PartialInsert = B.CreateInsertElement(
601           UndefValue::get(VecTy), ReadFirstLaneLo, B.getInt32(0));
602       Value *const Insert =
603           B.CreateInsertElement(PartialInsert, ReadFirstLaneHi, B.getInt32(1));
604       BroadcastI = B.CreateBitCast(Insert, Ty);
605     } else if (TyBitWidth == 32) {
606 
607       BroadcastI = B.CreateIntrinsic(Intrinsic::amdgcn_readfirstlane, {}, PHI);
608     } else {
609       llvm_unreachable("Unhandled atomic bit width");
610     }
611 
612     // Now that we have the result of our single atomic operation, we need to
613     // get our individual lane's slice into the result. We use the lane offset
614     // we previously calculated combined with the atomic result value we got
615     // from the first lane, to get our lane's index into the atomic result.
616     Value *LaneOffset = nullptr;
617     if (ValDivergent) {
618       LaneOffset = B.CreateIntrinsic(Intrinsic::amdgcn_wwm, Ty, ExclScan);
619     } else {
620       switch (Op) {
621       default:
622         llvm_unreachable("Unhandled atomic op");
623       case AtomicRMWInst::Add:
624       case AtomicRMWInst::Sub:
625         LaneOffset = B.CreateMul(V, Mbcnt);
626         break;
627       case AtomicRMWInst::And:
628       case AtomicRMWInst::Or:
629       case AtomicRMWInst::Max:
630       case AtomicRMWInst::Min:
631       case AtomicRMWInst::UMax:
632       case AtomicRMWInst::UMin:
633         LaneOffset = B.CreateSelect(Cond, Identity, V);
634         break;
635       case AtomicRMWInst::Xor:
636         LaneOffset = B.CreateMul(V, B.CreateAnd(Mbcnt, 1));
637         break;
638       }
639     }
640     Value *const Result = buildNonAtomicBinOp(B, Op, BroadcastI, LaneOffset);
641 
642     if (IsPixelShader) {
643       // Need a final PHI to reconverge to above the helper lane branch mask.
644       B.SetInsertPoint(PixelExitBB->getFirstNonPHI());
645 
646       PHINode *const PHI = B.CreatePHI(Ty, 2);
647       PHI->addIncoming(UndefValue::get(Ty), PixelEntryBB);
648       PHI->addIncoming(Result, I.getParent());
649       I.replaceAllUsesWith(PHI);
650     } else {
651       // Replace the original atomic instruction with the new one.
652       I.replaceAllUsesWith(Result);
653     }
654   }
655 
656   // And delete the original.
657   I.eraseFromParent();
658 }
659 
660 INITIALIZE_PASS_BEGIN(AMDGPUAtomicOptimizer, DEBUG_TYPE,
661                       "AMDGPU atomic optimizations", false, false)
662 INITIALIZE_PASS_DEPENDENCY(LegacyDivergenceAnalysis)
663 INITIALIZE_PASS_DEPENDENCY(TargetPassConfig)
664 INITIALIZE_PASS_END(AMDGPUAtomicOptimizer, DEBUG_TYPE,
665                     "AMDGPU atomic optimizations", false, false)
666 
667 FunctionPass *llvm::createAMDGPUAtomicOptimizerPass() {
668   return new AMDGPUAtomicOptimizer();
669 }
670