File: SetFastMathFlags.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 (139 lines) | stat: -rw-r--r-- 4,459 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
/*========================== begin_copyright_notice ============================

Copyright (C) 2017-2021 Intel Corporation

SPDX-License-Identifier: MIT

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

#include "Compiler/Optimizer/OpenCLPasses/SetFastMathFlags/SetFastMathFlags.hpp"
#include "Compiler/MetaDataApi/IGCMetaDataHelper.h"
#include "Compiler/IGCPassSupport.h"

#include "common/LLVMWarningsPush.hpp"
#include <llvm/IR/InstIterator.h>
#include <llvm/IR/IntrinsicInst.h>
#include "common/LLVMWarningsPop.hpp"

using namespace llvm;
using namespace IGC;
using namespace IGC::IGCMD;

// Register pass to igc-opt
#define PASS_FLAG "igc-set-fast-math-flags"
#define PASS_DESCRIPTION "Set llvm fast math flags according to compiler options"
#define PASS_CFG_ONLY false
#define PASS_ANALYSIS false
IGC_INITIALIZE_PASS_BEGIN(SetFastMathFlags, PASS_FLAG, PASS_DESCRIPTION, PASS_CFG_ONLY, PASS_ANALYSIS)
IGC_INITIALIZE_PASS_DEPENDENCY(MetaDataUtilsWrapper)
IGC_INITIALIZE_PASS_END(SetFastMathFlags, PASS_FLAG, PASS_DESCRIPTION, PASS_CFG_ONLY, PASS_ANALYSIS)

char SetFastMathFlags::ID = 0;

SetFastMathFlags::SetFastMathFlags() : ModulePass(ID)
{
    m_Mask.setFast();
    initializeSetFastMathFlagsPass(*PassRegistry::getPassRegistry());
}

SetFastMathFlags::SetFastMathFlags(FastMathFlags Mask) : ModulePass(ID)
{
    m_Mask = Mask;
    initializeSetFastMathFlagsPass(*PassRegistry::getPassRegistry());
}

SetFastMathFlags::SetFastMathFlags(FastMathFlags Mask, bool skipUnsafeFpMathAttr) : ModulePass(ID)
{
    m_Mask = Mask;
    m_skipUnsafeFpMathAttr = skipUnsafeFpMathAttr;
    initializeSetFastMathFlagsPass(*PassRegistry::getPassRegistry());
}

bool SetFastMathFlags::runOnModule(Module& M)
{
    auto hasFnAttributeSet = [](Function& F, StringRef Attr) -> bool
    {
        return F.hasFnAttribute(Attr) && F.getFnAttribute(Attr).getValueAsString() == "true";
    };

    const ModuleMetaData& modMD = *(getAnalysis<MetaDataUtilsWrapper>().getModuleMetaData());
    bool changed = false;
    for (Function& F : M) {
        FastMathFlags fmfs;
        // We determine if we have to use the unsafe-fp-math attr or we can omit it
        bool unsafeFpMathAttrCheck = m_skipUnsafeFpMathAttr || hasFnAttributeSet(F, "unsafe-fp-math");
        // Fast relaxed math implies all other flags.
        if (modMD.compOpt.FastRelaxedMath && unsafeFpMathAttrCheck) {
            fmfs.setFast();
            fmfs &= m_Mask;
            changed |= setFlags(F, fmfs);
            continue;
        }
        // Unsafe math implies no signed zeros.
        if ((modMD.compOpt.NoSignedZeros || modMD.compOpt.UnsafeMathOptimizations) && (hasFnAttributeSet(F, "no-signed-zeros-fp-math"))) {
            fmfs.setNoSignedZeros();
        }
        // Finite math implies no infs and nans.
        if (modMD.compOpt.FiniteMathOnly) {
            if (hasFnAttributeSet(F, "no-infs-fp-math")) {
                fmfs.setNoInfs();
            }
            if (hasFnAttributeSet(F, "no-nans-fp-math")) {
                fmfs.setNoNaNs();
            }
        }
        fmfs &= m_Mask;
        changed |= setFlags(F, fmfs);
    }
    return changed;
}

static bool setMathFlags(IntrinsicInst* II) {
    switch (II->getIntrinsicID()) {
    case Intrinsic::pow:
    case Intrinsic::exp2:
    case Intrinsic::log:
    case Intrinsic::sqrt:
        II->setFast(true);
        return true;
    default:
        break;
    }
    return false;
}

bool SetFastMathFlags::setFlags(Function& F, FastMathFlags fmfs)
{
    if (!fmfs.any())
    {
        return false;
    }
    bool changed = false;
    for (inst_iterator i = inst_begin(&F), e = inst_end(&F); i != e; ++i)
    {
        unsigned int op = i->getOpcode();
        switch (op) {
        case Instruction::FAdd:
        case Instruction::FSub:
        case Instruction::FMul:
        case Instruction::FDiv:
        case Instruction::FRem:
            i->setFastMathFlags(fmfs); // this actually does an OR between flags.
            changed = true;
            break;
        case Instruction::Call:
            if (auto II = dyn_cast<IntrinsicInst>(&*i)) {
                if (fmfs.isFast())
                    changed |= setMathFlags(II);
            }
            else if (isa<FPMathOperator>(&*i)) {
                i->setFastMathFlags(fmfs);
                changed = true;
            }
            break;
        default:
            break;
        }
    }
    return changed;
}