xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AMDGPU/AMDGPULateCodeGenPrepare.cpp (revision 06c3fb2749bda94cb5201f81ffdb8fa6c3161b2e)
1 //===-- AMDGPUCodeGenPrepare.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 does misc. AMDGPU optimizations on IR *just* before instruction
11 /// selection.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #include "AMDGPU.h"
16 #include "llvm/Analysis/AssumptionCache.h"
17 #include "llvm/Analysis/UniformityAnalysis.h"
18 #include "llvm/Analysis/ValueTracking.h"
19 #include "llvm/IR/IRBuilder.h"
20 #include "llvm/IR/InstVisitor.h"
21 #include "llvm/InitializePasses.h"
22 #include "llvm/Support/CommandLine.h"
23 #include "llvm/Support/KnownBits.h"
24 #include "llvm/Transforms/Utils/Local.h"
25 
26 #define DEBUG_TYPE "amdgpu-late-codegenprepare"
27 
28 using namespace llvm;
29 
30 // Scalar load widening needs running after load-store-vectorizer as that pass
31 // doesn't handle overlapping cases. In addition, this pass enhances the
32 // widening to handle cases where scalar sub-dword loads are naturally aligned
33 // only but not dword aligned.
34 static cl::opt<bool>
35     WidenLoads("amdgpu-late-codegenprepare-widen-constant-loads",
36                cl::desc("Widen sub-dword constant address space loads in "
37                         "AMDGPULateCodeGenPrepare"),
38                cl::ReallyHidden, cl::init(true));
39 
40 namespace {
41 
42 class AMDGPULateCodeGenPrepare
43     : public FunctionPass,
44       public InstVisitor<AMDGPULateCodeGenPrepare, bool> {
45   Module *Mod = nullptr;
46   const DataLayout *DL = nullptr;
47 
48   AssumptionCache *AC = nullptr;
49   UniformityInfo *UA = nullptr;
50 
51 public:
52   static char ID;
53 
54   AMDGPULateCodeGenPrepare() : FunctionPass(ID) {}
55 
56   StringRef getPassName() const override {
57     return "AMDGPU IR late optimizations";
58   }
59 
60   void getAnalysisUsage(AnalysisUsage &AU) const override {
61     AU.addRequired<AssumptionCacheTracker>();
62     AU.addRequired<UniformityInfoWrapperPass>();
63     AU.setPreservesAll();
64   }
65 
66   bool doInitialization(Module &M) override;
67   bool runOnFunction(Function &F) override;
68 
69   bool visitInstruction(Instruction &) { return false; }
70 
71   // Check if the specified value is at least DWORD aligned.
72   bool isDWORDAligned(const Value *V) const {
73     KnownBits Known = computeKnownBits(V, *DL, 0, AC);
74     return Known.countMinTrailingZeros() >= 2;
75   }
76 
77   bool canWidenScalarExtLoad(LoadInst &LI) const;
78   bool visitLoadInst(LoadInst &LI);
79 };
80 
81 } // end anonymous namespace
82 
83 bool AMDGPULateCodeGenPrepare::doInitialization(Module &M) {
84   Mod = &M;
85   DL = &Mod->getDataLayout();
86   return false;
87 }
88 
89 bool AMDGPULateCodeGenPrepare::runOnFunction(Function &F) {
90   if (skipFunction(F))
91     return false;
92 
93   AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
94   UA = &getAnalysis<UniformityInfoWrapperPass>().getUniformityInfo();
95 
96   bool Changed = false;
97   for (auto &BB : F)
98     for (Instruction &I : llvm::make_early_inc_range(BB))
99       Changed |= visit(I);
100 
101   return Changed;
102 }
103 
104 bool AMDGPULateCodeGenPrepare::canWidenScalarExtLoad(LoadInst &LI) const {
105   unsigned AS = LI.getPointerAddressSpace();
106   // Skip non-constant address space.
107   if (AS != AMDGPUAS::CONSTANT_ADDRESS &&
108       AS != AMDGPUAS::CONSTANT_ADDRESS_32BIT)
109     return false;
110   // Skip non-simple loads.
111   if (!LI.isSimple())
112     return false;
113   auto *Ty = LI.getType();
114   // Skip aggregate types.
115   if (Ty->isAggregateType())
116     return false;
117   unsigned TySize = DL->getTypeStoreSize(Ty);
118   // Only handle sub-DWORD loads.
119   if (TySize >= 4)
120     return false;
121   // That load must be at least naturally aligned.
122   if (LI.getAlign() < DL->getABITypeAlign(Ty))
123     return false;
124   // It should be uniform, i.e. a scalar load.
125   return UA->isUniform(&LI);
126 }
127 
128 bool AMDGPULateCodeGenPrepare::visitLoadInst(LoadInst &LI) {
129   if (!WidenLoads)
130     return false;
131 
132   // Skip if that load is already aligned on DWORD at least as it's handled in
133   // SDAG.
134   if (LI.getAlign() >= 4)
135     return false;
136 
137   if (!canWidenScalarExtLoad(LI))
138     return false;
139 
140   int64_t Offset = 0;
141   auto *Base =
142       GetPointerBaseWithConstantOffset(LI.getPointerOperand(), Offset, *DL);
143   // If that base is not DWORD aligned, it's not safe to perform the following
144   // transforms.
145   if (!isDWORDAligned(Base))
146     return false;
147 
148   int64_t Adjust = Offset & 0x3;
149   if (Adjust == 0) {
150     // With a zero adjust, the original alignment could be promoted with a
151     // better one.
152     LI.setAlignment(Align(4));
153     return true;
154   }
155 
156   IRBuilder<> IRB(&LI);
157   IRB.SetCurrentDebugLocation(LI.getDebugLoc());
158 
159   unsigned LdBits = DL->getTypeStoreSizeInBits(LI.getType());
160   auto IntNTy = Type::getIntNTy(LI.getContext(), LdBits);
161 
162   auto *NewPtr = IRB.CreateConstGEP1_64(
163       IRB.getInt8Ty(),
164       IRB.CreateAddrSpaceCast(Base, LI.getPointerOperand()->getType()),
165       Offset - Adjust);
166 
167   LoadInst *NewLd = IRB.CreateAlignedLoad(IRB.getInt32Ty(), NewPtr, Align(4));
168   NewLd->copyMetadata(LI);
169   NewLd->setMetadata(LLVMContext::MD_range, nullptr);
170 
171   unsigned ShAmt = Adjust * 8;
172   auto *NewVal = IRB.CreateBitCast(
173       IRB.CreateTrunc(IRB.CreateLShr(NewLd, ShAmt), IntNTy), LI.getType());
174   LI.replaceAllUsesWith(NewVal);
175   RecursivelyDeleteTriviallyDeadInstructions(&LI);
176 
177   return true;
178 }
179 
180 INITIALIZE_PASS_BEGIN(AMDGPULateCodeGenPrepare, DEBUG_TYPE,
181                       "AMDGPU IR late optimizations", false, false)
182 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
183 INITIALIZE_PASS_DEPENDENCY(UniformityInfoWrapperPass)
184 INITIALIZE_PASS_END(AMDGPULateCodeGenPrepare, DEBUG_TYPE,
185                     "AMDGPU IR late optimizations", false, false)
186 
187 char AMDGPULateCodeGenPrepare::ID = 0;
188 
189 FunctionPass *llvm::createAMDGPULateCodeGenPreparePass() {
190   return new AMDGPULateCodeGenPrepare();
191 }
192