[Mlir-commits] [mlir] [mlir][SCF] Avoid generating unnecessary div/rem operations during coalescing (PR #91562)

[Quinn Dawkins]

================
@@ -562,19 +575,43 @@ static Value getProductOfIntsOrIndexes(RewriterBase &rewriter, Location loc,
 static std::pair<SmallVector<Value>, SmallPtrSet<Operation *, 2>>
 delinearizeInductionVariable(RewriterBase &rewriter, Location loc,
                              Value linearizedIv, ArrayRef<Value> ubs) {
-  Value previous = linearizedIv;
   SmallVector<Value> delinearizedIvs(ubs.size());
   SmallPtrSet<Operation *, 2> preservedUsers;
-  for (unsigned i = 0, e = ubs.size(); i < e; ++i) {
-    unsigned idx = ubs.size() - i - 1;
-    if (i != 0) {
+
+  llvm::BitVector isUbOne(ubs.size());
+  for (auto [index, ub] : llvm::enumerate(ubs)) {
+    auto ubCst = getConstantIntValue(ub);
+    if (ubCst && ubCst.value() == 1)
+      isUbOne.set(index);
+  }
----------------
qedawkins wrote:

Are the upper bounds already normalized to unit step and zero lower bound by this point?

https://github.com/llvm/llvm-project/pull/91562