File: PrepareLoadsStoresUtils.cpp

package info (click to toggle)
intel-graphics-compiler 1.0.12504.6-1%2Bdeb12u1
  • links: PTS, VCS
  • area: main
  • in suites: bookworm
  • size: 83,912 kB
  • sloc: cpp: 910,147; lisp: 202,655; ansic: 15,197; python: 4,025; yacc: 2,241; lex: 1,570; pascal: 244; sh: 104; makefile: 25
file content (201 lines) | stat: -rw-r--r-- 6,999 bytes parent folder | download
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
/*========================== begin_copyright_notice ============================

Copyright (C) 2021 Intel Corporation

SPDX-License-Identifier: MIT

============================= end_copyright_notice ===========================*/

//===----------------------------------------------------------------------===//
///
/// These functions convert load and store instructions that have 64-bit
/// elements (whether scalar or vector) to vectors of dwords.
///
/// This is used by PrepareLoadsStoresPass to run prior to MemOpt to aid
/// vectorization of values with different types.
///
/// For example:
///
/// %x = load i64, i64 addrspace(1)* %"&x"
/// ===>
/// %7 = bitcast i64 addrspace(1)* %"&x" to <2 x i32> addrspace(1)*
/// 8 = load <2 x i32>, <2 x i32> addrspace(1)* %7
/// %9 = bitcast <2 x i32> %8 to i64
///
//===----------------------------------------------------------------------===//

#include "Compiler/IGCPassSupport.h"
#include "Compiler/CodeGenPublicEnums.h"
#include "Compiler/CodeGenPublic.h"
#include "AdaptorCommon/RayTracing/RTBuilder.h"
#include "PrepareLoadsStoresUtils.h"

#include "common/LLVMWarningsPush.hpp"
#include "llvmWrapper/Support/Alignment.h"
#include "llvmWrapper/IR/DerivedTypes.h"
#include <llvm/IR/IRBuilder.h>
#include <llvm/IR/InstIterator.h>
#include <llvm/IR/NoFolder.h>
#include "common/LLVMWarningsPop.hpp"

using namespace llvm;
using namespace IGC;
using IGCLLVM::getAlign;

template <typename T>
static Value* getIntToPtr(IRBuilder<T> &IRB, Value* NewVal, Type* Ty)
{
    // We must emit scalar inttoptr this late in compilation otherwise it won't
    // be handled correctly (by e.g., Emu64Ops).

    if (NewVal->getType()->isIntegerTy())
        return IRB.CreateIntToPtr(NewVal, Ty);

    // otherwise, 'NewVal' and 'Ty' are both vectors and will have the same
    // length.
    IGCLLVM::FixedVectorType* newVecType = dyn_cast<IGCLLVM::FixedVectorType>(NewVal->getType());
    IGCLLVM::FixedVectorType* vecType = dyn_cast<IGCLLVM::FixedVectorType>(Ty);
    uint64_t NumElts = newVecType->getNumElements();
    IGC_ASSERT(NumElts == vecType->getNumElements());
    Value* ResultVec = UndefValue::get(vecType);
    auto* ResEltTy = vecType->getElementType();
    for (uint32_t i = 0; i < NumElts; i++)
    {
        auto* Elt = IRB.CreateExtractElement(NewVal, i);
        auto* Cast = IRB.CreateIntToPtr(Elt, ResEltTy);
        ResultVec = IRB.CreateInsertElement(ResultVec, Cast, i);
    }

    return ResultVec;
}

template <typename T>
static Value* getPtrToInt(IRBuilder<T> &IRB, Value* NewVal, Type* Ty)
{
    // We must emit scalar ptrtoint this late in compilation otherwise it won't
    // be handled correctly (by e.g., Emu64Ops).

    if (NewVal->getType()->isPointerTy())
        return IRB.CreatePtrToInt(NewVal, Ty);

    // otherwise, 'NewVal' and 'Ty' are both vectors and will have the same
    // length.
    IGCLLVM::FixedVectorType* newVecType = dyn_cast<IGCLLVM::FixedVectorType>(NewVal->getType());
    IGCLLVM::FixedVectorType* vecType = dyn_cast<IGCLLVM::FixedVectorType>(Ty);
    uint64_t NumElts = newVecType->getNumElements();
    IGC_ASSERT(NumElts == vecType->getNumElements());
    Value* ResultVec = UndefValue::get(Ty);
    auto* ResEltTy = vecType->getElementType();
    for (uint32_t i = 0; i < NumElts; i++)
    {
        auto* Elt = IRB.CreateExtractElement(NewVal, i);
        auto* Cast = IRB.CreatePtrToInt(Elt, ResEltTy);
        ResultVec = IRB.CreateInsertElement(ResultVec, Cast, i);
    }

    return ResultVec;
}

namespace IGC {

    template <typename T>
    std::pair<Value*, LoadInst*>
    expand64BitLoad(IRBuilder<T>& IRB, const DataLayout &DL, LoadInst* LI)
    {
        auto* Ty = LI->getType();
        if (Ty->isAggregateType())
            return {};

        auto* EltTy = Ty->getScalarType();
        auto* OldPtrTy = LI->getPointerOperandType();
        uint64_t NumEltBytes = DL.getTypeStoreSize(EltTy);

        if (NumEltBytes != 8)
            return {};

        uint32_t NumElts = isa<IGCLLVM::FixedVectorType>(Ty) ?
            (uint32_t)cast<IGCLLVM::FixedVectorType>(Ty)->getNumElements() : 1;

        uint32_t NewNumElts = NumElts * 2;

        auto* NewTy = IGCLLVM::FixedVectorType::get(IRB.getInt32Ty(), NewNumElts);
        auto* PtrTy = PointerType::get(
            NewTy, OldPtrTy->getPointerAddressSpace());

        auto* NewPtr = IRB.CreateBitCast(LI->getPointerOperand(), PtrTy);

        LoadInst* NewLI =
            IRB.CreateAlignedLoad(NewTy, NewPtr, IGCLLVM::getCorrectAlign(LI->getAlignment()));
        NewLI->copyMetadata(*LI);
        Value* NewVal = NewLI;

        if (Ty->isPtrOrPtrVectorTy())
        {
            Type* NewTy = (NumElts == 1) ?
                (Type*)IRB.getInt64Ty() :
                (Type*)IGCLLVM::FixedVectorType::get(IRB.getInt64Ty(), NumElts);
            NewVal = IRB.CreateBitCast(NewVal, NewTy);
            NewVal = getIntToPtr(IRB, NewVal, Ty);
        }
        else
        {
            NewVal = IRB.CreateBitCast(NewVal, Ty);
        }

        return { NewVal, NewLI };
    }

    template <typename T>
    StoreInst* expand64BitStore(IRBuilder<T>& IRB, const DataLayout &DL, StoreInst* SI)
    {
        auto* Ty = SI->getValueOperand()->getType();
        if (Ty->isAggregateType())
            return nullptr;

        auto* EltTy = Ty->getScalarType();
        auto* OldPtrTy = SI->getPointerOperandType();
        uint64_t NumEltBytes = DL.getTypeStoreSize(EltTy);

        if (NumEltBytes != 8)
            return nullptr;

        uint32_t NumElts = isa<IGCLLVM::FixedVectorType>(Ty) ?
            (uint32_t)cast<IGCLLVM::FixedVectorType>(Ty)->getNumElements() : 1;

        uint32_t NewNumElts = NumElts * 2;

        auto* NewTy = IGCLLVM::FixedVectorType::get(IRB.getInt32Ty(), NewNumElts);

        Value* NewVal = SI->getValueOperand();
        if (Ty->isPtrOrPtrVectorTy())
        {
            Type* NewTy = (NumElts == 1) ?
                (Type*)IRB.getInt64Ty() :
                (Type*)IGCLLVM::FixedVectorType::get(IRB.getInt64Ty(), NumElts);
            NewVal = getPtrToInt(IRB, NewVal, NewTy);
        }
        NewVal = IRB.CreateBitCast(NewVal, NewTy);
        auto* NewPtr = IRB.CreateBitCast(
            SI->getPointerOperand(),
            NewVal->getType()->getPointerTo(
                OldPtrTy->getPointerAddressSpace()));
        auto* NewST =
            IRB.CreateAlignedStore(NewVal, NewPtr, IGCLLVM::getCorrectAlign(SI->getAlignment()));
        NewST->copyMetadata(*SI);

        return NewST;
    }

    template std::pair<Value*, LoadInst*>
        expand64BitLoad(IRBuilder<>& IRB, const DataLayout& DL, LoadInst* LI);

    template StoreInst* expand64BitStore(
        IRBuilder<>& IRB, const DataLayout& DL, StoreInst* SI);

    template std::pair<Value*, LoadInst*>
        expand64BitLoad(IRBuilder<NoFolder>& IRB, const DataLayout& DL, LoadInst* LI);

    template StoreInst* expand64BitStore(
        IRBuilder<NoFolder>& IRB, const DataLayout& DL, StoreInst* SI);
}