[llvm] [AggressiveInstCombine] Fold i64 x i64 -> i128 multiply-by-parts (PR #156879)

[David Green via llvm-commits]

================
@@ -1457,6 +1457,268 @@ static bool foldLibCalls(Instruction &I, TargetTransformInfo &TTI,
   return false;
 }
 
+/// Match low part of 128-bit multiplication.
+///
+/// Use counts are checked to prevent total instruction count increase as per
+/// contributors guide:
+/// https://llvm.org/docs/InstCombineContributorGuide.html#multi-use-handling
+static bool foldMul128Low(Instruction &I) {
+  auto *Ty = I.getType();
+  if (!Ty->isIntegerTy(64))
+    return false;
+
+  // (low_accum << 32) | lo(lo(y) * lo(x))
+  Value *LowAccum = nullptr, *YLowXLow = nullptr;
+  if (!match(&I, m_c_DisjointOr(
+                     m_OneUse(m_Shl(m_Value(LowAccum), m_SpecificInt(32))),
+                     m_OneUse(
+                         m_And(m_Value(YLowXLow), m_SpecificInt(0xffffffff))))))
+    return false;
+
+  // lo(cross_sum) + hi(lo(y) * lo(x))
+  Value *CrossSum = nullptr;
+  if (!match(
+          LowAccum,
+          m_c_Add(m_OneUse(m_And(m_Value(CrossSum), m_SpecificInt(0xffffffff))),
----------------
davemgreen wrote:

I think this And might be optional, as it only clears top bits that we will ignore in the shift.

https://github.com/llvm/llvm-project/pull/156879