[llvm] [CodeGenPrepare] Folding `urem` with loop invariant value as remainder (PR #96625)

[Nikita Popov via llvm-commits]

================
@@ -1974,6 +1975,133 @@ static bool foldFCmpToFPClassTest(CmpInst *Cmp, const TargetLowering &TLI,
   return true;
 }
 
+static bool isRemOfLoopIncrementWithLoopInvariant(Instruction *Rem,
+                                                  const LoopInfo *LI,
+                                                  Value *&RemAmtOut,
+                                                  PHINode *&LoopIncrPNOut) {
+  Value *Incr, *RemAmt;
+  // NB: If RemAmt is a power of 2 it *should* have been transformed by now.
+  if (!match(Rem, m_URem(m_Value(Incr), m_Value(RemAmt))))
+    return false;
+
+  // Only trivially analyzable loops.
+  Loop *L = LI->getLoopFor(Rem->getParent());
+  if (!L || !L->getLoopPreheader() || !L->getLoopLatch())
+    return false;
+
+  // Find out loop increment PHI.
+  auto *PN = dyn_cast<PHINode>(Incr);
+
+  if (!PN)
+    return false;
+
+  // This isn't strictly necessary, what we really need is one increment and any
+  // amount of initial values all being the same.
+  if (PN->getNumIncomingValues() != 2)
+    return false;
+
+  // Only works if the remainder amount is a loop invaraint
+  if (!L->isLoopInvariant(RemAmt))
+    return false;
+
+  // Is the PHI a loop increment?
+  auto LoopIncrInfo = getIVIncrement(PN, LI);
+  if (!LoopIncrInfo)
+    return false;
+
+  // getIVIncrement finds the loop at PN->getParent(). This might be a different
+  // loop from the loop with Rem->getParent().
+  if (L->getHeader() != PN->getParent())
----------------
nikic wrote:

I think technically the right thing to do would be to have `L` be the loop of the phi node, not the rem, and only check that L also contains rem (which is not guaranteed as we don't require LCSSA in CGP). The loop of the phi node is really what you care about, the urem itself can be in a nested loop without issue (I think?)

But it's fine to ignore that case for now.

https://github.com/llvm/llvm-project/pull/96625