[llvm-branch-commits] [llvm] [LoopInterchange] Identify unsafe instructions for interchange (PR #201402)

[Nikita Popov via llvm-branch-commits]

================
@@ -1472,25 +1472,28 @@ bool LoopInterchangeLegality::canInterchangeLoops(unsigned InnerLoopId,
   }
   // Check if outer and inner loop contain legal instructions only.
   for (auto *BB : OuterLoop->blocks())
-    for (Instruction &I : *BB)
-      if (CallInst *CI = dyn_cast<CallInst>(&I)) {
-        // Functions which don't access memory and don't diverge do not prevent
-        // interchanging.
-        if ((CI->doesNotAccessMemory() && !CI->mayHaveSideEffects()) ||
-            isa<PseudoProbeInst>(CI))
-          continue;
-        LLVM_DEBUG(
-            dbgs() << "Loops with call instructions cannot be interchanged "
-                   << "safely.");
-        ORE->emit([&]() {
-          return OptimizationRemarkMissed(DEBUG_TYPE, "CallInst",
-                                          CI->getDebugLoc(),
-                                          CI->getParent())
-                 << "Cannot interchange loops due to call instruction.";
-        });
+    for (Instruction &I : *BB) {
+      // Loads and stores are checked separately, so we can skip them here.
+      if (isa<LoadInst, StoreInst, PseudoProbeInst>(&I))
+        continue;
 
-        return false;
-      }
+      // We cannot ignore potential memory reads, e.g., loads inside the called
+      // function.
+      if (!I.mayHaveSideEffects() && !I.mayReadFromMemory() && !I.isAtomic())
----------------
nikic wrote:

The isAtomic() check here is redundant, it's implied by the others.

https://github.com/llvm/llvm-project/pull/201402