1 //===-- AMDGPULowerKernelArguments.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 This pass replaces accesses to kernel arguments with loads from
10 /// offsets from the kernarg base pointer.
11 //
12 //===----------------------------------------------------------------------===//
13
14 #include "AMDGPU.h"
15 #include "GCNSubtarget.h"
16 #include "llvm/CodeGen/TargetPassConfig.h"
17 #include "llvm/IR/IRBuilder.h"
18 #include "llvm/IR/IntrinsicsAMDGPU.h"
19 #include "llvm/IR/MDBuilder.h"
20 #include "llvm/Target/TargetMachine.h"
21
22 #define DEBUG_TYPE "amdgpu-lower-kernel-arguments"
23
24 using namespace llvm;
25
26 namespace {
27
28 class PreloadKernelArgInfo {
29 private:
30 Function &F;
31 const GCNSubtarget &ST;
32 unsigned NumFreeUserSGPRs;
33
34 public:
35 SmallVector<llvm::Metadata *, 8> KernelArgMetadata;
36
PreloadKernelArgInfo(Function & F,const GCNSubtarget & ST)37 PreloadKernelArgInfo(Function &F, const GCNSubtarget &ST) : F(F), ST(ST) {
38 setInitialFreeUserSGPRsCount();
39 }
40
41 // Returns the maximum number of user SGPRs that we have available to preload
42 // arguments.
setInitialFreeUserSGPRsCount()43 void setInitialFreeUserSGPRsCount() {
44 const unsigned MaxUserSGPRs = ST.getMaxNumUserSGPRs();
45 GCNUserSGPRUsageInfo UserSGPRInfo(F, ST);
46
47 NumFreeUserSGPRs = MaxUserSGPRs - UserSGPRInfo.getNumUsedUserSGPRs();
48 }
49
tryAllocPreloadSGPRs(unsigned AllocSize,uint64_t ArgOffset,uint64_t LastExplicitArgOffset)50 bool tryAllocPreloadSGPRs(unsigned AllocSize, uint64_t ArgOffset,
51 uint64_t LastExplicitArgOffset) {
52 // Check if this argument may be loaded into the same register as the
53 // previous argument.
54 if (!isAligned(Align(4), ArgOffset) && AllocSize < 4)
55 return true;
56
57 // Pad SGPRs for kernarg alignment.
58 unsigned Padding = ArgOffset - LastExplicitArgOffset;
59 unsigned PaddingSGPRs = alignTo(Padding, 4) / 4;
60 unsigned NumPreloadSGPRs = alignTo(AllocSize, 4) / 4;
61 if (NumPreloadSGPRs + PaddingSGPRs > NumFreeUserSGPRs)
62 return false;
63
64 NumFreeUserSGPRs -= (NumPreloadSGPRs + PaddingSGPRs);
65 return true;
66 }
67 };
68
69 class AMDGPULowerKernelArguments : public FunctionPass {
70 public:
71 static char ID;
72
AMDGPULowerKernelArguments()73 AMDGPULowerKernelArguments() : FunctionPass(ID) {}
74
75 bool runOnFunction(Function &F) override;
76
getAnalysisUsage(AnalysisUsage & AU) const77 void getAnalysisUsage(AnalysisUsage &AU) const override {
78 AU.addRequired<TargetPassConfig>();
79 AU.setPreservesAll();
80 }
81 };
82
83 } // end anonymous namespace
84
85 // skip allocas
getInsertPt(BasicBlock & BB)86 static BasicBlock::iterator getInsertPt(BasicBlock &BB) {
87 BasicBlock::iterator InsPt = BB.getFirstInsertionPt();
88 for (BasicBlock::iterator E = BB.end(); InsPt != E; ++InsPt) {
89 AllocaInst *AI = dyn_cast<AllocaInst>(&*InsPt);
90
91 // If this is a dynamic alloca, the value may depend on the loaded kernargs,
92 // so loads will need to be inserted before it.
93 if (!AI || !AI->isStaticAlloca())
94 break;
95 }
96
97 return InsPt;
98 }
99
lowerKernelArguments(Function & F,const TargetMachine & TM)100 static bool lowerKernelArguments(Function &F, const TargetMachine &TM) {
101 CallingConv::ID CC = F.getCallingConv();
102 if (CC != CallingConv::AMDGPU_KERNEL || F.arg_empty())
103 return false;
104
105 const GCNSubtarget &ST = TM.getSubtarget<GCNSubtarget>(F);
106 LLVMContext &Ctx = F.getParent()->getContext();
107 const DataLayout &DL = F.getDataLayout();
108 BasicBlock &EntryBlock = *F.begin();
109 IRBuilder<> Builder(&EntryBlock, getInsertPt(EntryBlock));
110
111 const Align KernArgBaseAlign(16); // FIXME: Increase if necessary
112 const uint64_t BaseOffset = ST.getExplicitKernelArgOffset();
113
114 Align MaxAlign;
115 // FIXME: Alignment is broken with explicit arg offset.;
116 const uint64_t TotalKernArgSize = ST.getKernArgSegmentSize(F, MaxAlign);
117 if (TotalKernArgSize == 0)
118 return false;
119
120 CallInst *KernArgSegment =
121 Builder.CreateIntrinsic(Intrinsic::amdgcn_kernarg_segment_ptr, {}, {},
122 nullptr, F.getName() + ".kernarg.segment");
123 KernArgSegment->addRetAttr(Attribute::NonNull);
124 KernArgSegment->addRetAttr(
125 Attribute::getWithDereferenceableBytes(Ctx, TotalKernArgSize));
126
127 uint64_t ExplicitArgOffset = 0;
128 // Preloaded kernel arguments must be sequential.
129 bool InPreloadSequence = true;
130 PreloadKernelArgInfo PreloadInfo(F, ST);
131
132 for (Argument &Arg : F.args()) {
133 const bool IsByRef = Arg.hasByRefAttr();
134 Type *ArgTy = IsByRef ? Arg.getParamByRefType() : Arg.getType();
135 MaybeAlign ParamAlign = IsByRef ? Arg.getParamAlign() : std::nullopt;
136 Align ABITypeAlign = DL.getValueOrABITypeAlignment(ParamAlign, ArgTy);
137
138 uint64_t Size = DL.getTypeSizeInBits(ArgTy);
139 uint64_t AllocSize = DL.getTypeAllocSize(ArgTy);
140
141 uint64_t EltOffset = alignTo(ExplicitArgOffset, ABITypeAlign) + BaseOffset;
142 uint64_t LastExplicitArgOffset = ExplicitArgOffset;
143 ExplicitArgOffset = alignTo(ExplicitArgOffset, ABITypeAlign) + AllocSize;
144
145 // Try to preload this argument into user SGPRs.
146 if (Arg.hasInRegAttr() && InPreloadSequence && ST.hasKernargPreload() &&
147 !Arg.getType()->isAggregateType())
148 if (PreloadInfo.tryAllocPreloadSGPRs(AllocSize, EltOffset,
149 LastExplicitArgOffset))
150 continue;
151
152 InPreloadSequence = false;
153
154 if (Arg.use_empty())
155 continue;
156
157 // If this is byval, the loads are already explicit in the function. We just
158 // need to rewrite the pointer values.
159 if (IsByRef) {
160 Value *ArgOffsetPtr = Builder.CreateConstInBoundsGEP1_64(
161 Builder.getInt8Ty(), KernArgSegment, EltOffset,
162 Arg.getName() + ".byval.kernarg.offset");
163
164 Value *CastOffsetPtr =
165 Builder.CreateAddrSpaceCast(ArgOffsetPtr, Arg.getType());
166 Arg.replaceAllUsesWith(CastOffsetPtr);
167 continue;
168 }
169
170 if (PointerType *PT = dyn_cast<PointerType>(ArgTy)) {
171 // FIXME: Hack. We rely on AssertZext to be able to fold DS addressing
172 // modes on SI to know the high bits are 0 so pointer adds don't wrap. We
173 // can't represent this with range metadata because it's only allowed for
174 // integer types.
175 if ((PT->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
176 PT->getAddressSpace() == AMDGPUAS::REGION_ADDRESS) &&
177 !ST.hasUsableDSOffset())
178 continue;
179
180 // FIXME: We can replace this with equivalent alias.scope/noalias
181 // metadata, but this appears to be a lot of work.
182 if (Arg.hasNoAliasAttr())
183 continue;
184 }
185
186 auto *VT = dyn_cast<FixedVectorType>(ArgTy);
187 bool IsV3 = VT && VT->getNumElements() == 3;
188 bool DoShiftOpt = Size < 32 && !ArgTy->isAggregateType();
189
190 VectorType *V4Ty = nullptr;
191
192 int64_t AlignDownOffset = alignDown(EltOffset, 4);
193 int64_t OffsetDiff = EltOffset - AlignDownOffset;
194 Align AdjustedAlign = commonAlignment(
195 KernArgBaseAlign, DoShiftOpt ? AlignDownOffset : EltOffset);
196
197 Value *ArgPtr;
198 Type *AdjustedArgTy;
199 if (DoShiftOpt) { // FIXME: Handle aggregate types
200 // Since we don't have sub-dword scalar loads, avoid doing an extload by
201 // loading earlier than the argument address, and extracting the relevant
202 // bits.
203 // TODO: Update this for GFX12 which does have scalar sub-dword loads.
204 //
205 // Additionally widen any sub-dword load to i32 even if suitably aligned,
206 // so that CSE between different argument loads works easily.
207 ArgPtr = Builder.CreateConstInBoundsGEP1_64(
208 Builder.getInt8Ty(), KernArgSegment, AlignDownOffset,
209 Arg.getName() + ".kernarg.offset.align.down");
210 AdjustedArgTy = Builder.getInt32Ty();
211 } else {
212 ArgPtr = Builder.CreateConstInBoundsGEP1_64(
213 Builder.getInt8Ty(), KernArgSegment, EltOffset,
214 Arg.getName() + ".kernarg.offset");
215 AdjustedArgTy = ArgTy;
216 }
217
218 if (IsV3 && Size >= 32) {
219 V4Ty = FixedVectorType::get(VT->getElementType(), 4);
220 // Use the hack that clang uses to avoid SelectionDAG ruining v3 loads
221 AdjustedArgTy = V4Ty;
222 }
223
224 LoadInst *Load =
225 Builder.CreateAlignedLoad(AdjustedArgTy, ArgPtr, AdjustedAlign);
226 Load->setMetadata(LLVMContext::MD_invariant_load, MDNode::get(Ctx, {}));
227
228 MDBuilder MDB(Ctx);
229
230 if (isa<PointerType>(ArgTy)) {
231 if (Arg.hasNonNullAttr())
232 Load->setMetadata(LLVMContext::MD_nonnull, MDNode::get(Ctx, {}));
233
234 uint64_t DerefBytes = Arg.getDereferenceableBytes();
235 if (DerefBytes != 0) {
236 Load->setMetadata(
237 LLVMContext::MD_dereferenceable,
238 MDNode::get(Ctx,
239 MDB.createConstant(
240 ConstantInt::get(Builder.getInt64Ty(), DerefBytes))));
241 }
242
243 uint64_t DerefOrNullBytes = Arg.getDereferenceableOrNullBytes();
244 if (DerefOrNullBytes != 0) {
245 Load->setMetadata(
246 LLVMContext::MD_dereferenceable_or_null,
247 MDNode::get(Ctx,
248 MDB.createConstant(ConstantInt::get(Builder.getInt64Ty(),
249 DerefOrNullBytes))));
250 }
251
252 if (MaybeAlign ParamAlign = Arg.getParamAlign()) {
253 Load->setMetadata(
254 LLVMContext::MD_align,
255 MDNode::get(Ctx, MDB.createConstant(ConstantInt::get(
256 Builder.getInt64Ty(), ParamAlign->value()))));
257 }
258 }
259
260 // TODO: Convert noalias arg to !noalias
261
262 if (DoShiftOpt) {
263 Value *ExtractBits = OffsetDiff == 0 ?
264 Load : Builder.CreateLShr(Load, OffsetDiff * 8);
265
266 IntegerType *ArgIntTy = Builder.getIntNTy(Size);
267 Value *Trunc = Builder.CreateTrunc(ExtractBits, ArgIntTy);
268 Value *NewVal = Builder.CreateBitCast(Trunc, ArgTy,
269 Arg.getName() + ".load");
270 Arg.replaceAllUsesWith(NewVal);
271 } else if (IsV3) {
272 Value *Shuf = Builder.CreateShuffleVector(Load, ArrayRef<int>{0, 1, 2},
273 Arg.getName() + ".load");
274 Arg.replaceAllUsesWith(Shuf);
275 } else {
276 Load->setName(Arg.getName() + ".load");
277 Arg.replaceAllUsesWith(Load);
278 }
279 }
280
281 KernArgSegment->addRetAttr(
282 Attribute::getWithAlignment(Ctx, std::max(KernArgBaseAlign, MaxAlign)));
283
284 return true;
285 }
286
runOnFunction(Function & F)287 bool AMDGPULowerKernelArguments::runOnFunction(Function &F) {
288 auto &TPC = getAnalysis<TargetPassConfig>();
289 const TargetMachine &TM = TPC.getTM<TargetMachine>();
290 return lowerKernelArguments(F, TM);
291 }
292
293 INITIALIZE_PASS_BEGIN(AMDGPULowerKernelArguments, DEBUG_TYPE,
294 "AMDGPU Lower Kernel Arguments", false, false)
295 INITIALIZE_PASS_END(AMDGPULowerKernelArguments, DEBUG_TYPE, "AMDGPU Lower Kernel Arguments",
296 false, false)
297
298 char AMDGPULowerKernelArguments::ID = 0;
299
createAMDGPULowerKernelArgumentsPass()300 FunctionPass *llvm::createAMDGPULowerKernelArgumentsPass() {
301 return new AMDGPULowerKernelArguments();
302 }
303
304 PreservedAnalyses
run(Function & F,FunctionAnalysisManager & AM)305 AMDGPULowerKernelArgumentsPass::run(Function &F, FunctionAnalysisManager &AM) {
306 bool Changed = lowerKernelArguments(F, TM);
307 if (Changed) {
308 // TODO: Preserves a lot more.
309 PreservedAnalyses PA;
310 PA.preserveSet<CFGAnalyses>();
311 return PA;
312 }
313
314 return PreservedAnalyses::all();
315 }
316