[Struct Opt] A structure array accesses its member variables, the base address of the structure array is recalculated each time

[LiqinWeng]

Test case from spec2006 sjen, pls see this link： https://godbolt.org/z/hbnKz33r6

I debugged it and found that the main reason was that the alias analysis (TBAA) between the numb_move variable and MustNotAliasInsts (storeInst in the figure below) resulted in MayAlias, making it impossible to optimize.

I haven't thought of a better solution yet. Does anyone have any ideas?：）

[llvmbot]

@llvm/issue-subscribers-backend-risc-v

Author: LiqinWeng (LiqinWeng)

Test case from spec2006 sjen: **test case**： https://godbolt.org/z/hbnKz33r6

[LiqinWeng]

@nikic @nikita-kud @fhahn

[LiqinWeng]

@topperc

[topperc]

@nikic and I looked at this in person at the LLVM developer meeting.

We initially thought maybe this code from BasicAA.cpp should handle it.

  // If an inbounds GEP would have to start from an out of bounds address
  // for the two to alias, then we can assume noalias.
  // TODO: Remove !isScalable() once BasicAA fully support scalable location
  // size
  if (*DecompGEP1.InBounds && DecompGEP1.VarIndices.empty() &&
      V2Size.hasValue() && !V2Size.isScalable() &&
      DecompGEP1.Offset.sge(V2Size.getValue()) &&
      isBaseOfObject(DecompGEP2.Base))
    return AliasResult::NoAlias;

  if (isa<GEPOperator>(V2)) {
    // Symmetric case to above.
    if (*DecompGEP2.InBounds && DecompGEP1.VarIndices.empty() &&
        V1Size.hasValue() && !V1Size.isScalable() &&
        DecompGEP1.Offset.sle(-V1Size.getValue()) &&
        isBaseOfObject(DecompGEP1.Base))
      return AliasResult::NoAlias;
  }

But we determined that DecompGEP1.VarIndices.empty() was false so the code doesn't trigger.

I don't think we came up with anything after that.