| 1 | //===----- SVEIntrinsicOpts - SVE ACLE Intrinsics Opts --------------------===// |
| 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 | // Performs general IR level optimizations on SVE intrinsics. |
| 10 | // |
| 11 | // This pass performs the following optimizations: |
| 12 | // |
| 13 | // - removes unnecessary ptrue intrinsics (llvm.aarch64.sve.ptrue), e.g: |
| 14 | // %1 = @llvm.aarch64.sve.ptrue.nxv4i1(i32 31) |
| 15 | // %2 = @llvm.aarch64.sve.ptrue.nxv8i1(i32 31) |
| 16 | // ; (%1 can be replaced with a reinterpret of %2) |
| 17 | // |
| 18 | // - optimizes ptest intrinsics where the operands are being needlessly |
| 19 | // converted to and from svbool_t. |
| 20 | // |
| 21 | //===----------------------------------------------------------------------===// |
| 22 | |
| 23 | #include "AArch64.h" |
| 24 | #include "Utils/AArch64BaseInfo.h" |
| 25 | #include "llvm/ADT/PostOrderIterator.h" |
| 26 | #include "llvm/ADT/SetVector.h" |
| 27 | #include "llvm/IR/Constants.h" |
| 28 | #include "llvm/IR/Dominators.h" |
| 29 | #include "llvm/IR/IRBuilder.h" |
| 30 | #include "llvm/IR/Instructions.h" |
| 31 | #include "llvm/IR/IntrinsicInst.h" |
| 32 | #include "llvm/IR/IntrinsicsAArch64.h" |
| 33 | #include "llvm/IR/LLVMContext.h" |
| 34 | #include "llvm/IR/Module.h" |
| 35 | #include "llvm/IR/PatternMatch.h" |
| 36 | #include "llvm/InitializePasses.h" |
| 37 | #include <optional> |
| 38 | |
| 39 | using namespace llvm; |
| 40 | using namespace llvm::PatternMatch; |
| 41 | |
| 42 | #define DEBUG_TYPE "aarch64-sve-intrinsic-opts" |
| 43 | |
| 44 | namespace { |
| 45 | struct SVEIntrinsicOpts : public ModulePass { |
| 46 | static char ID; // Pass identification, replacement for typeid |
| 47 | SVEIntrinsicOpts() : ModulePass(ID) {} |
| 48 | |
| 49 | bool runOnModule(Module &M) override; |
| 50 | void getAnalysisUsage(AnalysisUsage &AU) const override; |
| 51 | |
| 52 | private: |
| 53 | bool coalescePTrueIntrinsicCalls(BasicBlock &BB, |
| 54 | SmallSetVector<IntrinsicInst *, 4> &PTrues); |
| 55 | bool optimizePTrueIntrinsicCalls(SmallSetVector<Function *, 4> &Functions); |
| 56 | }; |
| 57 | } // end anonymous namespace |
| 58 | |
| 59 | void SVEIntrinsicOpts::getAnalysisUsage(AnalysisUsage &AU) const { |
| 60 | AU.addRequired<DominatorTreeWrapperPass>(); |
| 61 | AU.setPreservesCFG(); |
| 62 | } |
| 63 | |
| 64 | char SVEIntrinsicOpts::ID = 0; |
| 65 | static const char *name = "SVE intrinsics optimizations" ; |
| 66 | INITIALIZE_PASS_BEGIN(SVEIntrinsicOpts, DEBUG_TYPE, name, false, false) |
| 67 | INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass); |
| 68 | INITIALIZE_PASS_END(SVEIntrinsicOpts, DEBUG_TYPE, name, false, false) |
| 69 | |
| 70 | ModulePass *llvm::createSVEIntrinsicOptsPass() { |
| 71 | return new SVEIntrinsicOpts(); |
| 72 | } |
| 73 | |
| 74 | /// Checks if a ptrue intrinsic call is promoted. The act of promoting a |
| 75 | /// ptrue will introduce zeroing. For example: |
| 76 | /// |
| 77 | /// %1 = <vscale x 4 x i1> call @llvm.aarch64.sve.ptrue.nxv4i1(i32 31) |
| 78 | /// %2 = <vscale x 16 x i1> call @llvm.aarch64.sve.convert.to.svbool.nxv4i1(<vscale x 4 x i1> %1) |
| 79 | /// %3 = <vscale x 8 x i1> call @llvm.aarch64.sve.convert.from.svbool.nxv8i1(<vscale x 16 x i1> %2) |
| 80 | /// |
| 81 | /// %1 is promoted, because it is converted: |
| 82 | /// |
| 83 | /// <vscale x 4 x i1> => <vscale x 16 x i1> => <vscale x 8 x i1> |
| 84 | /// |
| 85 | /// via a sequence of the SVE reinterpret intrinsics convert.{to,from}.svbool. |
| 86 | static bool isPTruePromoted(IntrinsicInst *PTrue) { |
| 87 | // Find all users of this intrinsic that are calls to convert-to-svbool |
| 88 | // reinterpret intrinsics. |
| 89 | SmallVector<IntrinsicInst *, 4> ConvertToUses; |
| 90 | for (User *User : PTrue->users()) { |
| 91 | if (match(V: User, P: m_Intrinsic<Intrinsic::aarch64_sve_convert_to_svbool>())) { |
| 92 | ConvertToUses.push_back(Elt: cast<IntrinsicInst>(Val: User)); |
| 93 | } |
| 94 | } |
| 95 | |
| 96 | // If no such calls were found, this is ptrue is not promoted. |
| 97 | if (ConvertToUses.empty()) |
| 98 | return false; |
| 99 | |
| 100 | // Otherwise, try to find users of the convert-to-svbool intrinsics that are |
| 101 | // calls to the convert-from-svbool intrinsic, and would result in some lanes |
| 102 | // being zeroed. |
| 103 | const auto *PTrueVTy = cast<ScalableVectorType>(Val: PTrue->getType()); |
| 104 | for (IntrinsicInst *ConvertToUse : ConvertToUses) { |
| 105 | for (User *User : ConvertToUse->users()) { |
| 106 | auto *IntrUser = dyn_cast<IntrinsicInst>(Val: User); |
| 107 | if (IntrUser && IntrUser->getIntrinsicID() == |
| 108 | Intrinsic::aarch64_sve_convert_from_svbool) { |
| 109 | const auto *IntrUserVTy = cast<ScalableVectorType>(Val: IntrUser->getType()); |
| 110 | |
| 111 | // Would some lanes become zeroed by the conversion? |
| 112 | if (IntrUserVTy->getElementCount().getKnownMinValue() > |
| 113 | PTrueVTy->getElementCount().getKnownMinValue()) |
| 114 | // This is a promoted ptrue. |
| 115 | return true; |
| 116 | } |
| 117 | } |
| 118 | } |
| 119 | |
| 120 | // If no matching calls were found, this is not a promoted ptrue. |
| 121 | return false; |
| 122 | } |
| 123 | |
| 124 | /// Attempts to coalesce ptrues in a basic block. |
| 125 | bool SVEIntrinsicOpts::coalescePTrueIntrinsicCalls( |
| 126 | BasicBlock &BB, SmallSetVector<IntrinsicInst *, 4> &PTrues) { |
| 127 | if (PTrues.size() <= 1) |
| 128 | return false; |
| 129 | |
| 130 | // Find the ptrue with the most lanes. |
| 131 | auto *MostEncompassingPTrue = |
| 132 | *llvm::max_element(Range&: PTrues, C: [](auto *PTrue1, auto *PTrue2) { |
| 133 | auto *PTrue1VTy = cast<ScalableVectorType>(PTrue1->getType()); |
| 134 | auto *PTrue2VTy = cast<ScalableVectorType>(PTrue2->getType()); |
| 135 | return PTrue1VTy->getElementCount().getKnownMinValue() < |
| 136 | PTrue2VTy->getElementCount().getKnownMinValue(); |
| 137 | }); |
| 138 | |
| 139 | // Remove the most encompassing ptrue, as well as any promoted ptrues, leaving |
| 140 | // behind only the ptrues to be coalesced. |
| 141 | PTrues.remove(X: MostEncompassingPTrue); |
| 142 | PTrues.remove_if(P: isPTruePromoted); |
| 143 | |
| 144 | // Hoist MostEncompassingPTrue to the start of the basic block. It is always |
| 145 | // safe to do this, since ptrue intrinsic calls are guaranteed to have no |
| 146 | // predecessors. |
| 147 | MostEncompassingPTrue->moveBefore(BB, I: BB.getFirstInsertionPt()); |
| 148 | |
| 149 | LLVMContext &Ctx = BB.getContext(); |
| 150 | IRBuilder<> Builder(Ctx); |
| 151 | Builder.SetInsertPoint(TheBB: &BB, IP: ++MostEncompassingPTrue->getIterator()); |
| 152 | |
| 153 | auto *MostEncompassingPTrueVTy = |
| 154 | cast<VectorType>(Val: MostEncompassingPTrue->getType()); |
| 155 | auto *ConvertToSVBool = Builder.CreateIntrinsicWithoutFolding( |
| 156 | ID: Intrinsic::aarch64_sve_convert_to_svbool, OverloadTypes: {MostEncompassingPTrueVTy}, |
| 157 | Args: {MostEncompassingPTrue}); |
| 158 | |
| 159 | bool ConvertFromCreated = false; |
| 160 | for (auto *PTrue : PTrues) { |
| 161 | auto *PTrueVTy = cast<VectorType>(Val: PTrue->getType()); |
| 162 | |
| 163 | // Only create the converts if the types are not already the same, otherwise |
| 164 | // just use the most encompassing ptrue. |
| 165 | if (MostEncompassingPTrueVTy != PTrueVTy) { |
| 166 | ConvertFromCreated = true; |
| 167 | |
| 168 | Builder.SetInsertPoint(TheBB: &BB, IP: ++ConvertToSVBool->getIterator()); |
| 169 | auto *ConvertFromSVBool = |
| 170 | Builder.CreateIntrinsic(ID: Intrinsic::aarch64_sve_convert_from_svbool, |
| 171 | OverloadTypes: {PTrueVTy}, Args: {ConvertToSVBool}); |
| 172 | PTrue->replaceAllUsesWith(V: ConvertFromSVBool); |
| 173 | } else |
| 174 | PTrue->replaceAllUsesWith(V: MostEncompassingPTrue); |
| 175 | |
| 176 | PTrue->eraseFromParent(); |
| 177 | } |
| 178 | |
| 179 | // We never used the ConvertTo so remove it |
| 180 | if (!ConvertFromCreated) |
| 181 | ConvertToSVBool->eraseFromParent(); |
| 182 | |
| 183 | return true; |
| 184 | } |
| 185 | |
| 186 | /// The goal of this function is to remove redundant calls to the SVE ptrue |
| 187 | /// intrinsic in each basic block within the given functions. |
| 188 | /// |
| 189 | /// SVE ptrues have two representations in LLVM IR: |
| 190 | /// - a logical representation -- an arbitrary-width scalable vector of i1s, |
| 191 | /// i.e. <vscale x N x i1>. |
| 192 | /// - a physical representation (svbool, <vscale x 16 x i1>) -- a 16-element |
| 193 | /// scalable vector of i1s, i.e. <vscale x 16 x i1>. |
| 194 | /// |
| 195 | /// The SVE ptrue intrinsic is used to create a logical representation of an SVE |
| 196 | /// predicate. Suppose that we have two SVE ptrue intrinsic calls: P1 and P2. If |
| 197 | /// P1 creates a logical SVE predicate that is at least as wide as the logical |
| 198 | /// SVE predicate created by P2, then all of the bits that are true in the |
| 199 | /// physical representation of P2 are necessarily also true in the physical |
| 200 | /// representation of P1. P1 'encompasses' P2, therefore, the intrinsic call to |
| 201 | /// P2 is redundant and can be replaced by an SVE reinterpret of P1 via |
| 202 | /// convert.{to,from}.svbool. |
| 203 | /// |
| 204 | /// Currently, this pass only coalesces calls to SVE ptrue intrinsics |
| 205 | /// if they match the following conditions: |
| 206 | /// |
| 207 | /// - the call to the intrinsic uses either the SV_ALL or SV_POW2 patterns. |
| 208 | /// SV_ALL indicates that all bits of the predicate vector are to be set to |
| 209 | /// true. SV_POW2 indicates that all bits of the predicate vector up to the |
| 210 | /// largest power-of-two are to be set to true. |
| 211 | /// - the result of the call to the intrinsic is not promoted to a wider |
| 212 | /// predicate. In this case, keeping the extra ptrue leads to better codegen |
| 213 | /// -- coalescing here would create an irreducible chain of SVE reinterprets |
| 214 | /// via convert.{to,from}.svbool. |
| 215 | /// |
| 216 | /// EXAMPLE: |
| 217 | /// |
| 218 | /// %1 = <vscale x 8 x i1> ptrue(i32 SV_ALL) |
| 219 | /// ; Logical: <1, 1, 1, 1, 1, 1, 1, 1> |
| 220 | /// ; Physical: <1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0> |
| 221 | /// ... |
| 222 | /// |
| 223 | /// %2 = <vscale x 4 x i1> ptrue(i32 SV_ALL) |
| 224 | /// ; Logical: <1, 1, 1, 1> |
| 225 | /// ; Physical: <1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0> |
| 226 | /// ... |
| 227 | /// |
| 228 | /// Here, %2 can be replaced by an SVE reinterpret of %1, giving, for instance: |
| 229 | /// |
| 230 | /// %1 = <vscale x 8 x i1> ptrue(i32 i31) |
| 231 | /// %2 = <vscale x 16 x i1> convert.to.svbool(<vscale x 8 x i1> %1) |
| 232 | /// %3 = <vscale x 4 x i1> convert.from.svbool(<vscale x 16 x i1> %2) |
| 233 | /// |
| 234 | bool SVEIntrinsicOpts::optimizePTrueIntrinsicCalls( |
| 235 | SmallSetVector<Function *, 4> &Functions) { |
| 236 | bool Changed = false; |
| 237 | |
| 238 | for (auto *F : Functions) { |
| 239 | for (auto &BB : *F) { |
| 240 | SmallSetVector<IntrinsicInst *, 4> SVAllPTrues; |
| 241 | SmallSetVector<IntrinsicInst *, 4> SVPow2PTrues; |
| 242 | |
| 243 | // For each basic block, collect the used ptrues and try to coalesce them. |
| 244 | for (Instruction &I : BB) { |
| 245 | if (I.use_empty()) |
| 246 | continue; |
| 247 | |
| 248 | auto *IntrI = dyn_cast<IntrinsicInst>(Val: &I); |
| 249 | if (!IntrI || IntrI->getIntrinsicID() != Intrinsic::aarch64_sve_ptrue) |
| 250 | continue; |
| 251 | |
| 252 | const auto PTruePattern = |
| 253 | cast<ConstantInt>(Val: IntrI->getOperand(i_nocapture: 0))->getZExtValue(); |
| 254 | |
| 255 | if (PTruePattern == AArch64SVEPredPattern::all) |
| 256 | SVAllPTrues.insert(X: IntrI); |
| 257 | if (PTruePattern == AArch64SVEPredPattern::pow2) |
| 258 | SVPow2PTrues.insert(X: IntrI); |
| 259 | } |
| 260 | |
| 261 | Changed |= coalescePTrueIntrinsicCalls(BB, PTrues&: SVAllPTrues); |
| 262 | Changed |= coalescePTrueIntrinsicCalls(BB, PTrues&: SVPow2PTrues); |
| 263 | } |
| 264 | } |
| 265 | |
| 266 | return Changed; |
| 267 | } |
| 268 | |
| 269 | bool SVEIntrinsicOpts::runOnModule(Module &M) { |
| 270 | bool Changed = false; |
| 271 | SmallSetVector<Function *, 4> Functions; |
| 272 | |
| 273 | // Check for SVE intrinsic declarations first so that we only iterate over |
| 274 | // relevant functions. Where an appropriate declaration is found, store the |
| 275 | // function(s) where it is used so we can target these only. |
| 276 | for (auto &F : M.getFunctionList()) { |
| 277 | if (!F.isDeclaration()) |
| 278 | continue; |
| 279 | |
| 280 | switch (F.getIntrinsicID()) { |
| 281 | case Intrinsic::aarch64_sve_ptrue: |
| 282 | for (User *U : F.users()) |
| 283 | Functions.insert(X: cast<Instruction>(Val: U)->getFunction()); |
| 284 | break; |
| 285 | default: |
| 286 | break; |
| 287 | } |
| 288 | } |
| 289 | |
| 290 | if (!Functions.empty()) |
| 291 | Changed |= optimizePTrueIntrinsicCalls(Functions); |
| 292 | |
| 293 | return Changed; |
| 294 | } |
| 295 | |