[llvm] [CGP]: Optimize mul.overflow. (PR #148343)

[David Green via llvm-commits]

================
@@ -6389,6 +6395,183 @@ bool CodeGenPrepare::optimizeGatherScatterInst(Instruction *MemoryInst,
   return true;
 }
 
+// This is a helper for CodeGenPrepare::optimizeMulWithOverflow.
+// Check the pattern we are interested in where there are maximum 2 uses
+// of the intrinsic which are the extract instructions.
+static bool matchOverflowPattern(Instruction *&I, ExtractValueInst *&MulExtract,
+                                 ExtractValueInst *&OverflowExtract) {
+  if (I->getNumUses() > 2)
+    return false;
+
+  for (User *U : I->users()) {
+    auto *Extract = dyn_cast<ExtractValueInst>(U);
+    if (!Extract || Extract->getNumIndices() != 1)
+      return false;
+
+    unsigned Index = Extract->getIndices()[0];
+    if (Index == 0)
+      MulExtract = Extract;
+    else if (Index == 1)
+      OverflowExtract = Extract;
+    else
+      return false;
+  }
+  return true;
+}
+
+// Rewrite the mul_with_overflow intrinsic by checking if both of the
+// operands' value ranges are within the legal type. If so, we can optimize the
+// multiplication algorithm. This code is supposed to be written during the step
+// of type legalization, but given that we need to reconstruct the IR which is
+// not doable there, we do it here.
+// The IR after the optimization will look like:
+// entry:
+//   if signed:
+//     ( (lhs_lo>>BW-1) ^ lhs_hi) || ( (rhs_lo>>BW-1) ^ rhs_hi) ? overflow,
+//     overflow_no
+//   else:
+//     (lhs_hi != 0) || (rhs_hi != 0) ? overflow, overflow_no
+// overflow_no:
+// overflow:
+// overflow.res:
+// \returns true if optimization was applied
+// TODO: This optimization can be further improved but it will get more complex,
+// so we leave it for future work.
----------------
davemgreen wrote:

Perhaps "TODO: This optimization can be further improved to optimize branching on overflow." It helps explain what would change.

https://github.com/llvm/llvm-project/pull/148343