LLVM 19.0.0git
RISCVCodeGenPrepare.cpp
Go to the documentation of this file.
1//===----- RISCVCodeGenPrepare.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// This is a RISC-V specific version of CodeGenPrepare.
10// It munges the code in the input function to better prepare it for
11// SelectionDAG-based code generation. This works around limitations in it's
12// basic-block-at-a-time approach.
13//
14//===----------------------------------------------------------------------===//
15
16#include "RISCV.h"
17#include "RISCVTargetMachine.h"
18#include "llvm/ADT/Statistic.h"
21#include "llvm/IR/IRBuilder.h"
22#include "llvm/IR/InstVisitor.h"
23#include "llvm/IR/Intrinsics.h"
26#include "llvm/Pass.h"
27
28using namespace llvm;
29
30#define DEBUG_TYPE "riscv-codegenprepare"
31#define PASS_NAME "RISC-V CodeGenPrepare"
32
33namespace {
34
35class RISCVCodeGenPrepare : public FunctionPass,
36 public InstVisitor<RISCVCodeGenPrepare, bool> {
37 const DataLayout *DL;
38 const RISCVSubtarget *ST;
39
40public:
41 static char ID;
42
43 RISCVCodeGenPrepare() : FunctionPass(ID) {}
44
45 bool runOnFunction(Function &F) override;
46
47 StringRef getPassName() const override { return PASS_NAME; }
48
49 void getAnalysisUsage(AnalysisUsage &AU) const override {
50 AU.setPreservesCFG();
52 }
53
54 bool visitInstruction(Instruction &I) { return false; }
55 bool visitAnd(BinaryOperator &BO);
57};
58
59} // end anonymous namespace
60
61// Try to optimize (i64 (and (zext/sext (i32 X), C1))) if C1 has bit 31 set,
62// but bits 63:32 are zero. If we know that bit 31 of X is 0, we can fill
63// the upper 32 bits with ones.
64bool RISCVCodeGenPrepare::visitAnd(BinaryOperator &BO) {
65 if (!ST->is64Bit())
66 return false;
67
68 if (!BO.getType()->isIntegerTy(64))
69 return false;
70
71 using namespace PatternMatch;
72
73 // Left hand side should be a zext nneg.
74 Value *LHSSrc;
75 if (!match(BO.getOperand(0), m_NNegZExt(m_Value(LHSSrc))))
76 return false;
77
78 if (!LHSSrc->getType()->isIntegerTy(32))
79 return false;
80
81 // Right hand side should be a constant.
82 Value *RHS = BO.getOperand(1);
83
84 auto *CI = dyn_cast<ConstantInt>(RHS);
85 if (!CI)
86 return false;
87 uint64_t C = CI->getZExtValue();
88
89 // Look for constants that fit in 32 bits but not simm12, and can be made
90 // into simm12 by sign extending bit 31. This will allow use of ANDI.
91 // TODO: Is worth making simm32?
92 if (!isUInt<32>(C) || isInt<12>(C) || !isInt<12>(SignExtend64<32>(C)))
93 return false;
94
95 // Sign extend the constant and replace the And operand.
96 C = SignExtend64<32>(C);
97 BO.setOperand(1, ConstantInt::get(RHS->getType(), C));
98
99 return true;
100}
101
102// LLVM vector reduction intrinsics return a scalar result, but on RISC-V vector
103// reduction instructions write the result in the first element of a vector
104// register. So when a reduction in a loop uses a scalar phi, we end up with
105// unnecessary scalar moves:
106//
107// loop:
108// vfmv.s.f v10, fa0
109// vfredosum.vs v8, v8, v10
110// vfmv.f.s fa0, v8
111//
112// This mainly affects ordered fadd reductions, since other types of reduction
113// typically use element-wise vectorisation in the loop body. This tries to
114// vectorize any scalar phis that feed into a fadd reduction:
115//
116// loop:
117// %phi = phi <float> [ ..., %entry ], [ %acc, %loop ]
118// %acc = call float @llvm.vector.reduce.fadd.nxv4f32(float %phi, <vscale x 2 x float> %vec)
119//
120// ->
121//
122// loop:
123// %phi = phi <vscale x 2 x float> [ ..., %entry ], [ %acc.vec, %loop ]
124// %phi.scalar = extractelement <vscale x 2 x float> %phi, i64 0
125// %acc = call float @llvm.vector.reduce.fadd.nxv4f32(float %x, <vscale x 2 x float> %vec)
126// %acc.vec = insertelement <vscale x 2 x float> poison, float %acc.next, i64 0
127//
128// Which eliminates the scalar -> vector -> scalar crossing during instruction
129// selection.
130bool RISCVCodeGenPrepare::visitIntrinsicInst(IntrinsicInst &I) {
131 if (I.getIntrinsicID() != Intrinsic::vector_reduce_fadd)
132 return false;
133
134 auto *PHI = dyn_cast<PHINode>(I.getOperand(0));
135 if (!PHI || !PHI->hasOneUse() ||
136 !llvm::is_contained(PHI->incoming_values(), &I))
137 return false;
138
139 Type *VecTy = I.getOperand(1)->getType();
140 IRBuilder<> Builder(PHI);
141 auto *VecPHI = Builder.CreatePHI(VecTy, PHI->getNumIncomingValues());
142
143 for (auto *BB : PHI->blocks()) {
144 Builder.SetInsertPoint(BB->getTerminator());
145 Value *InsertElt = Builder.CreateInsertElement(
146 VecTy, PHI->getIncomingValueForBlock(BB), (uint64_t)0);
147 VecPHI->addIncoming(InsertElt, BB);
148 }
149
150 Builder.SetInsertPoint(&I);
151 I.setOperand(0, Builder.CreateExtractElement(VecPHI, (uint64_t)0));
152
153 PHI->eraseFromParent();
154
155 return true;
156}
157
158bool RISCVCodeGenPrepare::runOnFunction(Function &F) {
159 if (skipFunction(F))
160 return false;
161
162 auto &TPC = getAnalysis<TargetPassConfig>();
163 auto &TM = TPC.getTM<RISCVTargetMachine>();
164 ST = &TM.getSubtarget<RISCVSubtarget>(F);
165
166 DL = &F.getParent()->getDataLayout();
167
168 bool MadeChange = false;
169 for (auto &BB : F)
171 MadeChange |= visit(I);
172
173 return MadeChange;
174}
175
176INITIALIZE_PASS_BEGIN(RISCVCodeGenPrepare, DEBUG_TYPE, PASS_NAME, false, false)
178INITIALIZE_PASS_END(RISCVCodeGenPrepare, DEBUG_TYPE, PASS_NAME, false, false)
179
180char RISCVCodeGenPrepare::ID = 0;
181
183 return new RISCVCodeGenPrepare();
184}
MachineBasicBlock MachineBasicBlock::iterator DebugLoc DL
Rewrite undef for PHI
#define F(x, y, z)
Definition: MD5.cpp:55
#define I(x, y, z)
Definition: MD5.cpp:58
const char LLVMTargetMachineRef TM
#define INITIALIZE_PASS_DEPENDENCY(depName)
Definition: PassSupport.h:55
#define INITIALIZE_PASS_END(passName, arg, name, cfg, analysis)
Definition: PassSupport.h:59
#define INITIALIZE_PASS_BEGIN(passName, arg, name, cfg, analysis)
Definition: PassSupport.h:52
#define PASS_NAME
#define DEBUG_TYPE
This file defines the 'Statistic' class, which is designed to be an easy way to expose various metric...
Target-Independent Code Generator Pass Configuration Options pass.
#define PASS_NAME
Value * RHS
Represent the analysis usage information of a pass.
AnalysisUsage & addRequired()
void setPreservesCFG()
This function should be called by the pass, iff they do not:
Definition: Pass.cpp:269
A parsed version of the target data layout string in and methods for querying it.
Definition: DataLayout.h:110
FunctionPass class - This class is used to implement most global optimizations.
Definition: Pass.h:311
virtual bool runOnFunction(Function &F)=0
runOnFunction - Virtual method overriden by subclasses to do the per-function processing of the pass.
This provides a uniform API for creating instructions and inserting them into a basic block: either a...
Definition: IRBuilder.h:2649
Base class for instruction visitors.
Definition: InstVisitor.h:78
RetTy visitIntrinsicInst(IntrinsicInst &I)
Definition: InstVisitor.h:219
void visitInstruction(Instruction &I)
Definition: InstVisitor.h:280
A wrapper class for inspecting calls to intrinsic functions.
Definition: IntrinsicInst.h:47
virtual void getAnalysisUsage(AnalysisUsage &) const
getAnalysisUsage - This function should be overriden by passes that need analysis information to do t...
Definition: Pass.cpp:98
virtual StringRef getPassName() const
getPassName - Return a nice clean name for a pass.
Definition: Pass.cpp:81
StringRef - Represent a constant reference to a string, i.e.
Definition: StringRef.h:50
Target-Independent Code Generator Pass Configuration Options.
The instances of the Type class are immutable: once they are created, they are never changed.
Definition: Type.h:45
bool isIntegerTy() const
True if this is an instance of IntegerType.
Definition: Type.h:228
void setOperand(unsigned i, Value *Val)
Definition: User.h:174
Value * getOperand(unsigned i) const
Definition: User.h:169
LLVM Value Representation.
Definition: Value.h:74
Type * getType() const
All values are typed, get the type of this value.
Definition: Value.h:255
unsigned ID
LLVM IR allows to use arbitrary numbers as calling convention identifiers.
Definition: CallingConv.h:24
@ C
The default llvm calling convention, compatible with C.
Definition: CallingConv.h:34
bool match(Val *V, const Pattern &P)
Definition: PatternMatch.h:49
NNegZExt_match< OpTy > m_NNegZExt(const OpTy &Op)
class_match< Value > m_Value()
Match an arbitrary value and ignore it.
Definition: PatternMatch.h:76
This is an optimization pass for GlobalISel generic memory operations.
Definition: AddressRanges.h:18
iterator_range< early_inc_iterator_impl< detail::IterOfRange< RangeT > > > make_early_inc_range(RangeT &&Range)
Make a range that does early increment to allow mutation of the underlying range without disrupting i...
Definition: STLExtras.h:665
FunctionPass * createRISCVCodeGenPreparePass()
bool is_contained(R &&Range, const E &Element)
Returns true if Element is found in Range.
Definition: STLExtras.h:1888