[RISCV] Introduce pass to promote double constants to a global array #160536
+365
−16
There are no files selected for viewing
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,192 @@ | ||
| //==- RISCVPromoteConstant.cpp - Promote constant fp to global for RISC-V --==// | ||
| // | ||
| // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. | ||
| // See https://llvm.org/LICENSE.txt for license information. | ||
| // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception | ||
| // | ||
| //===----------------------------------------------------------------------===// | ||
|
|
||
| #include "RISCV.h" | ||
| #include "RISCVSubtarget.h" | ||
| #include "llvm/ADT/DenseMap.h" | ||
| #include "llvm/ADT/SmallVector.h" | ||
| #include "llvm/ADT/Statistic.h" | ||
| #include "llvm/CodeGen/TargetLowering.h" | ||
| #include "llvm/CodeGen/TargetPassConfig.h" | ||
| #include "llvm/IR/BasicBlock.h" | ||
| #include "llvm/IR/Constant.h" | ||
| #include "llvm/IR/Constants.h" | ||
| #include "llvm/IR/Function.h" | ||
| #include "llvm/IR/GlobalValue.h" | ||
| #include "llvm/IR/GlobalVariable.h" | ||
| #include "llvm/IR/IRBuilder.h" | ||
| #include "llvm/IR/InstIterator.h" | ||
| #include "llvm/IR/Instruction.h" | ||
| #include "llvm/IR/Instructions.h" | ||
| #include "llvm/IR/Module.h" | ||
| #include "llvm/IR/Type.h" | ||
| #include "llvm/InitializePasses.h" | ||
| #include "llvm/Pass.h" | ||
| #include "llvm/Support/Casting.h" | ||
| #include "llvm/Support/Debug.h" | ||
|
|
||
| using namespace llvm; | ||
|
|
||
| #define DEBUG_TYPE "riscv-promote-const" | ||
| #define RISCV_PROMOTE_CONSTANT_NAME "RISC-V Promote Constants" | ||
|
|
||
| STATISTIC(NumPromoted, "Number of constant literals promoted to globals"); | ||
| STATISTIC(NumPromotedUses, "Number of uses of promoted literal constants"); | ||
|
|
||
| namespace { | ||
|
|
||
| class RISCVPromoteConstant : public ModulePass { | ||
|
There was a problem hiding this comment. Also add new pass manager version There was a problem hiding this comment. Based on my understanding form There was a problem hiding this comment. On this particular suggestion, I would actively prefer this be done in a follow on change. We do not have new pass manager versions of most RISCV codegen passes, and we should do these together. |
||
| public: | ||
| static char ID; | ||
| RISCVPromoteConstant() : ModulePass(ID) {} | ||
|
|
||
| StringRef getPassName() const override { return RISCV_PROMOTE_CONSTANT_NAME; } | ||
|
|
||
| void getAnalysisUsage(AnalysisUsage &AU) const override { | ||
| AU.addRequired<TargetPassConfig>(); | ||
| AU.setPreservesCFG(); | ||
| } | ||
|
|
||
| /// Iterate over the functions and promote the double fp constants that | ||
| /// would otherwise go into the constant pool to a constant array. | ||
| bool runOnModule(Module &M) override { | ||
| if (skipModule(M)) | ||
| return false; | ||
| // TargetMachine and Subtarget are needed to query isFPImmlegal. | ||
| const TargetPassConfig &TPC = getAnalysis<TargetPassConfig>(); | ||
| const TargetMachine &TM = TPC.getTM<TargetMachine>(); | ||
| bool Changed = false; | ||
| for (Function &F : M) { | ||
| const RISCVSubtarget &ST = TM.getSubtarget<RISCVSubtarget>(F); | ||
| const RISCVTargetLowering *TLI = ST.getTargetLowering(); | ||
| Changed |= runOnFunction(F, TLI); | ||
| } | ||
| return Changed; | ||
| } | ||
|
|
||
| private: | ||
| bool runOnFunction(Function &F, const RISCVTargetLowering *TLI); | ||
| }; | ||
| } // end anonymous namespace | ||
|
|
||
| char RISCVPromoteConstant::ID = 0; | ||
|
|
||
| INITIALIZE_PASS(RISCVPromoteConstant, DEBUG_TYPE, RISCV_PROMOTE_CONSTANT_NAME, | ||
| false, false) | ||
|
|
||
| ModulePass *llvm::createRISCVPromoteConstantPass() { | ||
| return new RISCVPromoteConstant(); | ||
| } | ||
|
|
||
| bool RISCVPromoteConstant::runOnFunction(Function &F, | ||
| const RISCVTargetLowering *TLI) { | ||
| if (F.hasOptNone() || F.hasOptSize()) | ||
| return false; | ||
|
|
||
| // Bail out and make no transformation if the target doesn't support | ||
|
There was a problem hiding this comment. Need to skip optnone functions / skipFunction There was a problem hiding this comment. skipFunction isn't available in a ModulePass, but I've added a skip if OptNone (and you can see if call check skipModule in runOnModule). |
||
| // doubles, or if we're not targeting RV64 as we currently see some | ||
| // regressions for those targets. | ||
| if (!TLI->isTypeLegal(MVT::f64) || !TLI->isTypeLegal(MVT::i64)) | ||
| return false; | ||
|
|
||
| // Collect all unique double constants and their uses in the function. Use | ||
| // MapVector to preserve insertion order. | ||
| MapVector<ConstantFP *, SmallVector<Use *, 8>> ConstUsesMap; | ||
|
|
||
| for (Instruction &I : instructions(F)) { | ||
| for (Use &U : I.operands()) { | ||
| auto *C = dyn_cast<ConstantFP>(U.get()); | ||
| if (!C || !C->getType()->isDoubleTy()) | ||
| continue; | ||
| if (TLI->isFPImmLegal(C->getValueAPF(), MVT::f64, | ||
| /*ForCodeSize=*/false)) | ||
| continue; | ||
| ConstUsesMap[C].push_back(&U); | ||
| } | ||
| } | ||
|
|
||
| int PromotableConstants = ConstUsesMap.size(); | ||
| LLVM_DEBUG(dbgs() << "Found " << PromotableConstants | ||
| << " promotable constants in " << F.getName() << "\n"); | ||
| // Bail out if no promotable constants found, or if only one is found. | ||
| if (PromotableConstants < 2) { | ||
| LLVM_DEBUG(dbgs() << "Performing no promotions as insufficient promotable " | ||
| "constants found\n"); | ||
| return false; | ||
| } | ||
|
|
||
| NumPromoted += PromotableConstants; | ||
|
|
||
| // Create a global array containing the promoted constants. | ||
| Module *M = F.getParent(); | ||
| Type *DoubleTy = Type::getDoubleTy(M->getContext()); | ||
|
|
||
| SmallVector<Constant *, 16> ConstantVector; | ||
| for (auto const &Pair : ConstUsesMap) | ||
| ConstantVector.push_back(Pair.first); | ||
|
|
||
| ArrayType *ArrayTy = ArrayType::get(DoubleTy, ConstantVector.size()); | ||
| Constant *GlobalArrayInitializer = | ||
| ConstantArray::get(ArrayTy, ConstantVector); | ||
|
|
||
| auto *GlobalArray = new GlobalVariable( | ||
| *M, ArrayTy, | ||
| /*isConstant=*/true, GlobalValue::InternalLinkage, GlobalArrayInitializer, | ||
| ".promoted_doubles." + F.getName()); | ||
|
|
||
| // A cache to hold the loaded value for a given constant within a basic block. | ||
| DenseMap<std::pair<ConstantFP *, BasicBlock *>, Value *> LocalLoads; | ||
|
|
||
| // Replace all uses with the loaded value. | ||
| unsigned Idx = 0; | ||
| for (auto const &Pair : ConstUsesMap) { | ||
| ConstantFP *Const = Pair.first; | ||
| const SmallVector<Use *, 8> &Uses = Pair.second; | ||
|
|
||
| for (Use *U : Uses) { | ||
| Instruction *UserInst = cast<Instruction>(U->getUser()); | ||
| BasicBlock *InsertionBB; | ||
| BasicBlock::iterator InsertionPt; | ||
|
|
||
| // If the user is a PHI node, we must insert the load in the | ||
| // corresponding predecessor basic block. Otherwise, it's inserted into | ||
| // the same block as the use. | ||
| if (auto *PN = dyn_cast<PHINode>(UserInst)) | ||
| InsertionBB = PN->getIncomingBlock(*U); | ||
| else | ||
| InsertionBB = UserInst->getParent(); | ||
|
|
||
| // It is always safe to insert in the first insertion point in the BB, | ||
| // so do that and let other passes reorder. | ||
| InsertionPt = InsertionBB->getFirstInsertionPt(); | ||
|
|
||
| auto CacheKey = std::make_pair(Const, InsertionBB); | ||
asb marked this conversation as resolved.
Show resolved
Hide resolved
|
||
| Value *LoadedVal = nullptr; | ||
|
|
||
| // Re-use a load if it exists in the insertion block. | ||
| if (LocalLoads.count(CacheKey)) { | ||
| LoadedVal = LocalLoads.at(CacheKey); | ||
| } else { | ||
| // Otherwise, create a new GEP and Load at the correct insertion point. | ||
| IRBuilder<> Builder(InsertionBB, InsertionPt); | ||
| Value *ElementPtr = Builder.CreateConstInBoundsGEP2_64( | ||
| GlobalArray->getValueType(), GlobalArray, 0, Idx, "double.addr"); | ||
| LoadedVal = Builder.CreateLoad(DoubleTy, ElementPtr, "double.val"); | ||
|
There was a problem hiding this comment. The loads will be created with appropraite alignment - but of course the lack of an IR->IR test means that wasn't obvious! I've added tests now, which confirms this. |
||
|
|
||
| // Cache the newly created load for this block. | ||
| LocalLoads[CacheKey] = LoadedVal; | ||
| } | ||
|
|
||
| U->set(LoadedVal); | ||
| ++NumPromotedUses; | ||
| } | ||
| ++Idx; | ||
| } | ||
|
|
||
| return true; | ||
| } | ||
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Oops, something went wrong.
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Uh oh!
There was an error while loading. Please reload this page.