[llvm] [CGP] [CodeGenPrepare] Folding `urem` with loop invariant value plus offset (PR #104724)

[via llvm-commits]

================
@@ -2066,6 +2094,29 @@ static bool foldURemOfLoopIncrement(Instruction *Rem, const DataLayout *DL,
 
   Loop *L = LI->getLoopFor(LoopIncrPN->getParent());
   Value *Start = LoopIncrPN->getIncomingValueForBlock(L->getLoopPreheader());
+  // If we have add/sub create initial value for remainder.
+  // The logic here is:
+  // (urem (add/sub nuw Start, IncrLoopInvariant), RemAmtLoopInvariant
+  //
+  // Only proceed if the expression simplifies (otherwise we can't fully
+  // optimize out the urem).
+  if (AddOrSub) {
+    assert(AddOrSubOffset && AddOrSubInst &&
+           "We found an add/sub but missing values");
+    // Without dom-condition/assumption cache we aren't likely to get much out
+    // of a context instruction.
+    const SimplifyQuery Q(*DL);
+    bool NSW = cast<OverflowingBinaryOperator>(AddOrSubInst)->hasNoSignedWrap();
----------------
goldsteinn wrote:

Fixed + Tests.

https://github.com/llvm/llvm-project/pull/104724