[llvm] [SLP]Improve minbitwidth analysis. (PR #78976)

[Alexey Bataev via llvm-commits]

================
@@ -13206,36 +13208,58 @@ unsigned BoUpSLP::getVectorElementSize(Value *V) {
 // smaller type with a truncation. We collect the values that will be demoted
 // in ToDemote and additional roots that require investigating in Roots.
 bool BoUpSLP::collectValuesToDemote(
-    Value *V, SmallVectorImpl<Value *> &ToDemote,
+    Value *V, bool IsProfitableToDemoteRoot, unsigned &BitWidth,
+    SmallVectorImpl<Value *> &ToDemote,
     DenseMap<Instruction *, SmallVector<unsigned>> &DemotedConsts,
-    SmallVectorImpl<Value *> &Roots, DenseSet<Value *> &Visited) const {
+    DenseSet<Value *> &Visited, unsigned &MaxDepthLevel,
+    bool &IsProfitableToDemote) const {
   // We can always demote constants.
-  if (isa<Constant>(V))
+  if (isa<Constant>(V)) {
+    MaxDepthLevel = 1;
     return true;
+  }
 
   // If the value is not a vectorized instruction in the expression and not used
   // by the insertelement instruction and not used in multiple vector nodes, it
   // cannot be demoted.
+  // TODO: improve handling of gathered values and others.
   auto *I = dyn_cast<Instruction>(V);
   if (!I || !getTreeEntry(I) || MultiNodeScalars.contains(I) ||
       !Visited.insert(I).second || all_of(I->users(), [&](User *U) {
         return isa<InsertElementInst>(U) && !getTreeEntry(U);
       }))
     return false;
 
+  auto IsPotentiallyTruncated = [&](Value *V, unsigned &BitWidth) -> bool {
+    if (MultiNodeScalars.contains(V))
+      return false;
+    uint32_t OrigBitWidth = DL->getTypeSizeInBits(V->getType());
+    APInt Mask = APInt::getBitsSetFrom(OrigBitWidth, BitWidth);
+    if (MaskedValueIsZero(V, Mask, SimplifyQuery(*DL)))
+      return true;
+    auto NumSignBits = ComputeNumSignBits(V, *DL, 0, AC, nullptr, DT);
+    unsigned BitWidth1 = OrigBitWidth - NumSignBits;
+    KnownBits Known = computeKnownBits(V, *DL);
----------------
alexey-bataev wrote:

> You should be able to replace MaskedValueIsZero with something like:
> 
> ```
> KnownBits Known = computeKnownBits(V, *DL);
> if (Known.countMinLeadingZeros() >= (OrigBitWidth - NumSignBits))
>   return true;
> ```

Not sure this is better. Prefer to use standard functions rather than reinvent them here. Such kind of "manual inlining" is very to hard to maintain.

https://github.com/llvm/llvm-project/pull/78976