xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AMDGPU/AMDGPULowerKernelArguments.cpp (revision 770cf0a5f02dc8983a89c6568d741fbc25baa999)
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/Analysis/ValueTracking.h"
17 #include "llvm/CodeGen/TargetPassConfig.h"
18 #include "llvm/IR/Attributes.h"
19 #include "llvm/IR/IRBuilder.h"
20 #include "llvm/IR/IntrinsicsAMDGPU.h"
21 #include "llvm/IR/MDBuilder.h"
22 #include "llvm/Target/TargetMachine.h"
23 
24 #define DEBUG_TYPE "amdgpu-lower-kernel-arguments"
25 
26 using namespace llvm;
27 
28 namespace {
29 
30 class AMDGPULowerKernelArguments : public FunctionPass {
31 public:
32   static char ID;
33 
34   AMDGPULowerKernelArguments() : FunctionPass(ID) {}
35 
36   bool runOnFunction(Function &F) override;
37 
38   void getAnalysisUsage(AnalysisUsage &AU) const override {
39     AU.addRequired<TargetPassConfig>();
40     AU.setPreservesAll();
41  }
42 };
43 
44 } // end anonymous namespace
45 
46 // skip allocas
47 static BasicBlock::iterator getInsertPt(BasicBlock &BB) {
48   BasicBlock::iterator InsPt = BB.getFirstInsertionPt();
49   for (BasicBlock::iterator E = BB.end(); InsPt != E; ++InsPt) {
50     AllocaInst *AI = dyn_cast<AllocaInst>(&*InsPt);
51 
52     // If this is a dynamic alloca, the value may depend on the loaded kernargs,
53     // so loads will need to be inserted before it.
54     if (!AI || !AI->isStaticAlloca())
55       break;
56   }
57 
58   return InsPt;
59 }
60 
61 static bool lowerKernelArguments(Function &F, const TargetMachine &TM) {
62   CallingConv::ID CC = F.getCallingConv();
63   if (CC != CallingConv::AMDGPU_KERNEL || F.arg_empty())
64     return false;
65 
66   const GCNSubtarget &ST = TM.getSubtarget<GCNSubtarget>(F);
67   LLVMContext &Ctx = F.getParent()->getContext();
68   const DataLayout &DL = F.getDataLayout();
69   BasicBlock &EntryBlock = *F.begin();
70   IRBuilder<> Builder(&EntryBlock, getInsertPt(EntryBlock));
71 
72   const Align KernArgBaseAlign(16); // FIXME: Increase if necessary
73   const uint64_t BaseOffset = ST.getExplicitKernelArgOffset();
74 
75   Align MaxAlign;
76   // FIXME: Alignment is broken with explicit arg offset.;
77   const uint64_t TotalKernArgSize = ST.getKernArgSegmentSize(F, MaxAlign);
78   if (TotalKernArgSize == 0)
79     return false;
80 
81   CallInst *KernArgSegment =
82       Builder.CreateIntrinsic(Intrinsic::amdgcn_kernarg_segment_ptr, {},
83                               nullptr, F.getName() + ".kernarg.segment");
84   KernArgSegment->addRetAttr(Attribute::NonNull);
85   KernArgSegment->addRetAttr(
86       Attribute::getWithDereferenceableBytes(Ctx, TotalKernArgSize));
87 
88   uint64_t ExplicitArgOffset = 0;
89   for (Argument &Arg : F.args()) {
90     const bool IsByRef = Arg.hasByRefAttr();
91     Type *ArgTy = IsByRef ? Arg.getParamByRefType() : Arg.getType();
92     MaybeAlign ParamAlign = IsByRef ? Arg.getParamAlign() : std::nullopt;
93     Align ABITypeAlign = DL.getValueOrABITypeAlignment(ParamAlign, ArgTy);
94 
95     uint64_t Size = DL.getTypeSizeInBits(ArgTy);
96     uint64_t AllocSize = DL.getTypeAllocSize(ArgTy);
97 
98     uint64_t EltOffset = alignTo(ExplicitArgOffset, ABITypeAlign) + BaseOffset;
99     ExplicitArgOffset = alignTo(ExplicitArgOffset, ABITypeAlign) + AllocSize;
100 
101     // Skip inreg arguments which should be preloaded.
102     if (Arg.use_empty() || Arg.hasInRegAttr())
103       continue;
104 
105     // If this is byval, the loads are already explicit in the function. We just
106     // need to rewrite the pointer values.
107     if (IsByRef) {
108       Value *ArgOffsetPtr = Builder.CreateConstInBoundsGEP1_64(
109           Builder.getInt8Ty(), KernArgSegment, EltOffset,
110           Arg.getName() + ".byval.kernarg.offset");
111 
112       Value *CastOffsetPtr =
113           Builder.CreateAddrSpaceCast(ArgOffsetPtr, Arg.getType());
114       Arg.replaceAllUsesWith(CastOffsetPtr);
115       continue;
116     }
117 
118     if (PointerType *PT = dyn_cast<PointerType>(ArgTy)) {
119       // FIXME: Hack. We rely on AssertZext to be able to fold DS addressing
120       // modes on SI to know the high bits are 0 so pointer adds don't wrap. We
121       // can't represent this with range metadata because it's only allowed for
122       // integer types.
123       if ((PT->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
124            PT->getAddressSpace() == AMDGPUAS::REGION_ADDRESS) &&
125           !ST.hasUsableDSOffset())
126         continue;
127 
128       // FIXME: We can replace this with equivalent alias.scope/noalias
129       // metadata, but this appears to be a lot of work.
130       if (Arg.hasNoAliasAttr())
131         continue;
132     }
133 
134     auto *VT = dyn_cast<FixedVectorType>(ArgTy);
135     bool IsV3 = VT && VT->getNumElements() == 3;
136     bool DoShiftOpt = Size < 32 && !ArgTy->isAggregateType();
137 
138     VectorType *V4Ty = nullptr;
139 
140     int64_t AlignDownOffset = alignDown(EltOffset, 4);
141     int64_t OffsetDiff = EltOffset - AlignDownOffset;
142     Align AdjustedAlign = commonAlignment(
143         KernArgBaseAlign, DoShiftOpt ? AlignDownOffset : EltOffset);
144 
145     Value *ArgPtr;
146     Type *AdjustedArgTy;
147     if (DoShiftOpt) { // FIXME: Handle aggregate types
148       // Since we don't have sub-dword scalar loads, avoid doing an extload by
149       // loading earlier than the argument address, and extracting the relevant
150       // bits.
151       // TODO: Update this for GFX12 which does have scalar sub-dword loads.
152       //
153       // Additionally widen any sub-dword load to i32 even if suitably aligned,
154       // so that CSE between different argument loads works easily.
155       ArgPtr = Builder.CreateConstInBoundsGEP1_64(
156           Builder.getInt8Ty(), KernArgSegment, AlignDownOffset,
157           Arg.getName() + ".kernarg.offset.align.down");
158       AdjustedArgTy = Builder.getInt32Ty();
159     } else {
160       ArgPtr = Builder.CreateConstInBoundsGEP1_64(
161           Builder.getInt8Ty(), KernArgSegment, EltOffset,
162           Arg.getName() + ".kernarg.offset");
163       AdjustedArgTy = ArgTy;
164     }
165 
166     if (IsV3 && Size >= 32) {
167       V4Ty = FixedVectorType::get(VT->getElementType(), 4);
168       // Use the hack that clang uses to avoid SelectionDAG ruining v3 loads
169       AdjustedArgTy = V4Ty;
170     }
171 
172     LoadInst *Load =
173         Builder.CreateAlignedLoad(AdjustedArgTy, ArgPtr, AdjustedAlign);
174     Load->setMetadata(LLVMContext::MD_invariant_load, MDNode::get(Ctx, {}));
175 
176     MDBuilder MDB(Ctx);
177 
178     if (Arg.hasAttribute(Attribute::NoUndef))
179       Load->setMetadata(LLVMContext::MD_noundef, MDNode::get(Ctx, {}));
180 
181     if (Arg.hasAttribute(Attribute::Range)) {
182       const ConstantRange &Range =
183           Arg.getAttribute(Attribute::Range).getValueAsConstantRange();
184       Load->setMetadata(LLVMContext::MD_range,
185                         MDB.createRange(Range.getLower(), Range.getUpper()));
186     }
187 
188     if (isa<PointerType>(ArgTy)) {
189       if (Arg.hasNonNullAttr())
190         Load->setMetadata(LLVMContext::MD_nonnull, MDNode::get(Ctx, {}));
191 
192       uint64_t DerefBytes = Arg.getDereferenceableBytes();
193       if (DerefBytes != 0) {
194         Load->setMetadata(
195           LLVMContext::MD_dereferenceable,
196           MDNode::get(Ctx,
197                       MDB.createConstant(
198                         ConstantInt::get(Builder.getInt64Ty(), DerefBytes))));
199       }
200 
201       uint64_t DerefOrNullBytes = Arg.getDereferenceableOrNullBytes();
202       if (DerefOrNullBytes != 0) {
203         Load->setMetadata(
204           LLVMContext::MD_dereferenceable_or_null,
205           MDNode::get(Ctx,
206                       MDB.createConstant(ConstantInt::get(Builder.getInt64Ty(),
207                                                           DerefOrNullBytes))));
208       }
209 
210       if (MaybeAlign ParamAlign = Arg.getParamAlign()) {
211         Load->setMetadata(
212             LLVMContext::MD_align,
213             MDNode::get(Ctx, MDB.createConstant(ConstantInt::get(
214                                  Builder.getInt64Ty(), ParamAlign->value()))));
215       }
216     }
217 
218     // TODO: Convert noalias arg to !noalias
219 
220     if (DoShiftOpt) {
221       Value *ExtractBits = OffsetDiff == 0 ?
222         Load : Builder.CreateLShr(Load, OffsetDiff * 8);
223 
224       IntegerType *ArgIntTy = Builder.getIntNTy(Size);
225       Value *Trunc = Builder.CreateTrunc(ExtractBits, ArgIntTy);
226       Value *NewVal = Builder.CreateBitCast(Trunc, ArgTy,
227                                             Arg.getName() + ".load");
228       Arg.replaceAllUsesWith(NewVal);
229     } else if (IsV3) {
230       Value *Shuf = Builder.CreateShuffleVector(Load, ArrayRef<int>{0, 1, 2},
231                                                 Arg.getName() + ".load");
232       Arg.replaceAllUsesWith(Shuf);
233     } else {
234       Load->setName(Arg.getName() + ".load");
235       Arg.replaceAllUsesWith(Load);
236     }
237   }
238 
239   KernArgSegment->addRetAttr(
240       Attribute::getWithAlignment(Ctx, std::max(KernArgBaseAlign, MaxAlign)));
241 
242   return true;
243 }
244 
245 bool AMDGPULowerKernelArguments::runOnFunction(Function &F) {
246   auto &TPC = getAnalysis<TargetPassConfig>();
247   const TargetMachine &TM = TPC.getTM<TargetMachine>();
248   return lowerKernelArguments(F, TM);
249 }
250 
251 INITIALIZE_PASS_BEGIN(AMDGPULowerKernelArguments, DEBUG_TYPE,
252                       "AMDGPU Lower Kernel Arguments", false, false)
253 INITIALIZE_PASS_END(AMDGPULowerKernelArguments, DEBUG_TYPE, "AMDGPU Lower Kernel Arguments",
254                     false, false)
255 
256 char AMDGPULowerKernelArguments::ID = 0;
257 
258 FunctionPass *llvm::createAMDGPULowerKernelArgumentsPass() {
259   return new AMDGPULowerKernelArguments();
260 }
261 
262 PreservedAnalyses
263 AMDGPULowerKernelArgumentsPass::run(Function &F, FunctionAnalysisManager &AM) {
264   bool Changed = lowerKernelArguments(F, TM);
265   if (Changed) {
266     // TODO: Preserves a lot more.
267     PreservedAnalyses PA;
268     PA.preserveSet<CFGAnalyses>();
269     return PA;
270   }
271 
272   return PreservedAnalyses::all();
273 }
274