aboutsummaryrefslogtreecommitdiff
path: root/llvm/lib/Transforms/Vectorize/VPlanUtils.cpp
blob: 81bd21bb904c0a688ed6b3822bdcf105dffa741a (plain)
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
//===- VPlanUtils.cpp - VPlan-related utilities ---------------------------===//
//
// 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 "VPlanUtils.h"
#include "VPlanCFG.h"
#include "VPlanPatternMatch.h"
#include "llvm/ADT/TypeSwitch.h"
#include "llvm/Analysis/ScalarEvolutionExpressions.h"

using namespace llvm;

bool vputils::onlyFirstLaneUsed(const VPValue *Def) {
  return all_of(Def->users(),
                [Def](const VPUser *U) { return U->onlyFirstLaneUsed(Def); });
}

bool vputils::onlyFirstPartUsed(const VPValue *Def) {
  return all_of(Def->users(),
                [Def](const VPUser *U) { return U->onlyFirstPartUsed(Def); });
}

VPValue *vputils::getOrCreateVPValueForSCEVExpr(VPlan &Plan, const SCEV *Expr,
                                                ScalarEvolution &SE) {
  if (auto *Expanded = Plan.getSCEVExpansion(Expr))
    return Expanded;
  VPValue *Expanded = nullptr;
  if (auto *E = dyn_cast<SCEVConstant>(Expr))
    Expanded = Plan.getOrAddLiveIn(E->getValue());
  else {
    auto *U = dyn_cast<SCEVUnknown>(Expr);
    // Skip SCEV expansion if Expr is a SCEVUnknown wrapping a non-instruction
    // value. Otherwise the value may be defined in a loop and using it directly
    // will break LCSSA form. The SCEV expansion takes care of preserving LCSSA
    // form.
    if (U && !isa<Instruction>(U->getValue())) {
      Expanded = Plan.getOrAddLiveIn(U->getValue());
    } else {
      Expanded = new VPExpandSCEVRecipe(Expr, SE);
      Plan.getEntry()->appendRecipe(Expanded->getDefiningRecipe());
    }
  }
  Plan.addSCEVExpansion(Expr, Expanded);
  return Expanded;
}

bool vputils::isHeaderMask(const VPValue *V, VPlan &Plan) {
  if (isa<VPActiveLaneMaskPHIRecipe>(V))
    return true;

  auto IsWideCanonicalIV = [](VPValue *A) {
    return isa<VPWidenCanonicalIVRecipe>(A) ||
           (isa<VPWidenIntOrFpInductionRecipe>(A) &&
            cast<VPWidenIntOrFpInductionRecipe>(A)->isCanonical());
  };

  VPValue *A, *B;
  using namespace VPlanPatternMatch;

  if (match(V, m_ActiveLaneMask(m_VPValue(A), m_VPValue(B))))
    return B == Plan.getTripCount() &&
           (match(A, m_ScalarIVSteps(m_Specific(Plan.getCanonicalIV()),
                                     m_SpecificInt(1),
                                     m_Specific(&Plan.getVF()))) ||
            IsWideCanonicalIV(A));

  return match(V, m_Binary<Instruction::ICmp>(m_VPValue(A), m_VPValue(B))) &&
         IsWideCanonicalIV(A) && B == Plan.getOrCreateBackedgeTakenCount();
}

const SCEV *vputils::getSCEVExprForVPValue(VPValue *V, ScalarEvolution &SE) {
  if (V->isLiveIn())
    return SE.getSCEV(V->getLiveInIRValue());

  // TODO: Support constructing SCEVs for more recipes as needed.
  return TypeSwitch<const VPRecipeBase *, const SCEV *>(V->getDefiningRecipe())
      .Case<VPExpandSCEVRecipe>(
          [](const VPExpandSCEVRecipe *R) { return R->getSCEV(); })
      .Default([&SE](const VPRecipeBase *) { return SE.getCouldNotCompute(); });
}

bool vputils::isUniformAcrossVFsAndUFs(VPValue *V) {
  using namespace VPlanPatternMatch;
  // Live-ins are uniform.
  if (V->isLiveIn())
    return true;

  VPRecipeBase *R = V->getDefiningRecipe();
  if (R && V->isDefinedOutsideLoopRegions()) {
    if (match(V->getDefiningRecipe(),
              m_VPInstruction<VPInstruction::CanonicalIVIncrementForPart>(
                  m_VPValue())))
      return false;
    return all_of(R->operands(), isUniformAcrossVFsAndUFs);
  }

  auto *CanonicalIV = R->getParent()->getPlan()->getCanonicalIV();
  // Canonical IV chain is uniform.
  if (V == CanonicalIV || V == CanonicalIV->getBackedgeValue())
    return true;

  return TypeSwitch<const VPRecipeBase *, bool>(R)
      .Case<VPDerivedIVRecipe>([](const auto *R) { return true; })
      .Case<VPReplicateRecipe>([](const auto *R) {
        // Loads and stores that are uniform across VF lanes are handled by
        // VPReplicateRecipe.IsUniform. They are also uniform across UF parts if
        // all their operands are invariant.
        // TODO: Further relax the restrictions.
        return R->isSingleScalar() &&
               (isa<LoadInst, StoreInst>(R->getUnderlyingValue())) &&
               all_of(R->operands(), isUniformAcrossVFsAndUFs);
      })
      .Case<VPInstruction>([](const auto *VPI) {
        return VPI->isScalarCast() &&
               isUniformAcrossVFsAndUFs(VPI->getOperand(0));
      })
      .Case<VPWidenCastRecipe>([](const auto *R) {
        // A cast is uniform according to its operand.
        return isUniformAcrossVFsAndUFs(R->getOperand(0));
      })
      .Default([](const VPRecipeBase *) { // A value is considered non-uniform
                                          // unless proven otherwise.
        return false;
      });
}

VPBasicBlock *vputils::getFirstLoopHeader(VPlan &Plan, VPDominatorTree &VPDT) {
  auto DepthFirst = vp_depth_first_shallow(Plan.getEntry());
  auto I = find_if(DepthFirst, [&VPDT](VPBlockBase *VPB) {
    return VPBlockUtils::isHeader(VPB, VPDT);
  });
  return I == DepthFirst.end() ? nullptr : cast<VPBasicBlock>(*I);
}