[llvm] Revert "[InstCombine/RISCV] Constant-fold bitcast(vmv.v.x) (#182630)" (PR #194292)

[Ramkumar Ramachandra via llvm-commits]

================
@@ -3654,54 +3653,3 @@ bool RISCVTTIImpl::shouldCopyAttributeWhenOutliningFrom(
 
   return BaseT::shouldCopyAttributeWhenOutliningFrom(Caller, Attr);
 }
-
-std::optional<Instruction *>
-RISCVTTIImpl::instCombineIntrinsic(InstCombiner &IC, IntrinsicInst &II) const {
-  // If all operands of a vmv.v.x are constant, fold a bitcast(vmv.v.x) to scale
-  // the vmv.v.x, enabling removal of the bitcast. The transform helps avoid
-  // creating redundant masks.
-  const DataLayout &DL = IC.getDataLayout();
-  if (II.user_empty())
-    return {};
-  auto *TargetVecTy = dyn_cast<ScalableVectorType>(II.user_back()->getType());
-  if (!TargetVecTy)
-    return {};
-  const APInt *Scalar;
-  uint64_t VL;
-  if (!match(&II, m_Intrinsic<Intrinsic::riscv_vmv_v_x>(
-                      m_Poison(), m_APInt(Scalar), m_ConstantInt(VL))) ||
-      !all_of(II.users(), [TargetVecTy](User *U) {
-        return U->getType() == TargetVecTy && match(U, m_BitCast(m_Value()));
-      }))
-    return {};
-  auto *SourceVecTy = cast<ScalableVectorType>(II.getType());
-  unsigned TargetEltBW = DL.getTypeSizeInBits(TargetVecTy->getElementType());
-  unsigned SourceEltBW = DL.getTypeSizeInBits(SourceVecTy->getElementType());
----------------
artagnon wrote:

If I understand correctly, the infinite loop can be fixed by simply checking that TargetEltBW == SourceEltBW, as sizeof(float) == sizeof(i32), which causes the issue? Identity casts are always stripped anyway?

https://github.com/llvm/llvm-project/pull/194292