[InstCombine] Fold Minimum over Trailing/Leading Bits Counts (#90402)

The new transformation folds `umin(cttz(x), c)` to `cttz(x | (1 << c))`
and `umin(ctlz(x), c)` to `ctlz(x | ((1 << (bitwidth - 1)) >> c))`. The
transformation is only implemented for constant `c` to not increase the
number of instructions.
    
The idea of the transformation is to set the c-th lowest (for `cttz`) or
highest (for `ctlz`) bit in the operand. In this way, the `cttz` or
`ctlz` instruction always returns at most `c`.
    
Alive2 proofs: https://alive2.llvm.org/ce/z/y8Hdb8

Fixes #90000

Author: mskamp

llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp

@@ -1456,6 +1456,43 @@ static Value *simplifyReductionOperand(Value *Arg, bool CanReorderLanes) {
   return UsedIndices.all() ? V : nullptr;
 }
 
+/// Fold an unsigned minimum of trailing or leading zero bits counts:
+///   umin(cttz(CtOp, ZeroUndef), ConstOp) --> cttz(CtOp | (1 << ConstOp))
+///   umin(ctlz(CtOp, ZeroUndef), ConstOp) --> ctlz(CtOp | (SignedMin
+///                                              >> ConstOp))
+template <Intrinsic::ID IntrID>
+static Value *
+foldMinimumOverTrailingOrLeadingZeroCount(Value *I0, Value *I1,
+                                          const DataLayout &DL,
+                                          InstCombiner::BuilderTy &Builder) {
+  static_assert(IntrID == Intrinsic::cttz || IntrID == Intrinsic::ctlz,
+                "This helper only supports cttz and ctlz intrinsics");
+
+  Value *CtOp;
+  Value *ZeroUndef;
+  if (!match(I0,
+             m_OneUse(m_Intrinsic<IntrID>(m_Value(CtOp), m_Value(ZeroUndef)))))
+    return nullptr;
+
+  unsigned BitWidth = I1->getType()->getScalarSizeInBits();
+  auto LessBitWidth = [BitWidth](auto &C) { return C.ult(BitWidth); };
+  if (!match(I1, m_CheckedInt(LessBitWidth)))
+    // We have a constant >= BitWidth (which can be handled by CVP)
+    // or a non-splat vector with elements < and >= BitWidth
+    return nullptr;
+
+  Type *Ty = I1->getType();
+  Constant *NewConst = ConstantFoldBinaryOpOperands(
+      IntrID == Intrinsic::cttz ? Instruction::Shl : Instruction::LShr,
+      IntrID == Intrinsic::cttz
+          ? ConstantInt::get(Ty, 1)
+          : ConstantInt::get(Ty, APInt::getSignedMinValue(BitWidth)),
+      cast<Constant>(I1), DL);
+  return Builder.CreateBinaryIntrinsic(
+      IntrID, Builder.CreateOr(CtOp, NewConst),
+      ConstantInt::getTrue(ZeroUndef->getType()));
+}
+
 /// CallInst simplification. This mostly only handles folding of intrinsic
 /// instructions. For normal calls, it allows visitCallBase to do the heavy
 /// lifting.
@@ -1661,6 +1698,16 @@ Instruction *InstCombinerImpl::visitCallInst(CallInst &CI) {
       Value *Cmp = Builder.CreateICmpNE(I0, Zero);
       return CastInst::Create(Instruction::ZExt, Cmp, II->getType());
     }
+    // umin(cttz(x), const) --> cttz(x | (1 << const))
+    if (Value *FoldedCttz =
+            foldMinimumOverTrailingOrLeadingZeroCount<Intrinsic::cttz>(
+                I0, I1, DL, Builder))
+      return replaceInstUsesWith(*II, FoldedCttz);
+    // umin(ctlz(x), const) --> ctlz(x | (SignedMin >> const))
+    if (Value *FoldedCtlz =
+            foldMinimumOverTrailingOrLeadingZeroCount<Intrinsic::ctlz>(
+                I0, I1, DL, Builder))
+      return replaceInstUsesWith(*II, FoldedCtlz);
     [[fallthrough]];
   }
   case Intrinsic::umax: {