[llvm] [VectorCombine][AMDGPU] Shrink demanded vector loads (PR #202501)

[Matt Arsenault via llvm-commits]

================
@@ -2157,6 +2160,231 @@ bool VectorCombine::scalarizeLoadExtract(LoadInst *LI, VectorType *VecTy,
   return true;
 }
 
+/// Try to shrink vector loads feeding extractelement instructions when some
+/// of the loaded lanes are not demanded.
+bool VectorCombine::shrinkLoadForExtracts(Instruction &I) {
+  if (!DB)
+    return false;
+
+  auto *OldLoad = dyn_cast<LoadInst>(&I);
+  if (!OldLoad || !OldLoad->isSimple())
+    return false;
+
+  auto *OldLoadTy = dyn_cast<FixedVectorType>(OldLoad->getType());
+  if (!OldLoadTy)
+    return false;
+
+  Type *OldEltTy = OldLoadTy->getElementType();
+  bool IsIntegerElement = OldEltTy->isIntegerTy();
+  bool IsFloatingPointElement = OldEltTy->isFloatingPointTy();
+  if (!IsIntegerElement && !IsFloatingPointElement)
+    return false;
+
+  constexpr unsigned ByteWidth = 8;
+  TypeSize OldEltSizeInBits = DL->getTypeSizeInBits(OldEltTy);
+  if (OldEltSizeInBits.isScalable())
+    return false;
+  unsigned OldEltSize = OldEltSizeInBits.getFixedValue();
+  if (OldEltSize < ByteWidth || OldEltSize % ByteWidth != 0)
+    return false;
+
+  unsigned NumElts = OldLoadTy->getNumElements();
+  SmallVector<ExtractElementInst *, 8> Extracts;
+  SmallVector<APInt, 8> ExtractDemandedBits;
+
+  for (User *U : OldLoad->users()) {
+    auto *Ext = dyn_cast<ExtractElementInst>(U);
+    if (!Ext || Ext->use_empty())
+      return false;
+
+    auto *Idx = dyn_cast<ConstantInt>(Ext->getIndexOperand());
+    if (!Idx || Idx->getValue().uge(NumElts))
+      return false;
----------------
arsenm wrote:

I don't see negative tests for the variable or OOB index

https://github.com/llvm/llvm-project/pull/202501