[Mlir-commits] [mlir] [mlir][inliner] optimize self-recursive function detection (PR #88452)

Mehdi Amini llvmlistbot at llvm.org
Sat May 11 20:12:21 PDT 2024


================
@@ -705,24 +715,30 @@ Inliner::Impl::inlineCallsInSCC(InlinerInterfaceImpl &inlinerIface,
   return success(inlinedAnyCalls);
 }
 
+static bool isSelfRecursiveFunction(CallGraphNode *node) {
+  return llvm::find_if(*node, [&](const CallGraphNode::Edge &edge) -> bool {
+           return edge.getTarget() == node;
+         }) != node->end();
+}
+
 /// Returns true if the given call should be inlined.
-bool Inliner::Impl::shouldInline(ResolvedCall &resolvedCall) {
+bool Inliner::Impl::shouldInline(
+    ResolvedCall &resolvedCall,
+    const llvm::SmallPtrSetImpl<Region *> &recursiveCallRegions) {
   // Don't allow inlining terminator calls. We currently don't support this
   // case.
   if (resolvedCall.call->hasTrait<OpTrait::IsTerminator>())
     return false;
 
-  // Don't allow inlining if the target is a self-recursive function.
-  if (llvm::count_if(*resolvedCall.targetNode,
-                     [&](CallGraphNode::Edge const &edge) -> bool {
-                       return edge.getTarget() == resolvedCall.targetNode;
-                     }) > 0)
-    return false;
-
-  // Don't allow inlining if the target is an ancestor of the call. This
-  // prevents inlining recursively.
   Region *callableRegion = resolvedCall.targetNode->getCallableRegion();
-  if (callableRegion->isAncestor(resolvedCall.call->getParentRegion()))
+
+  // Don't allow inlining this following cases to prevent inlining recursively.
+  // 1. target has at least an edge back to itself in original call graph.
----------------
joker-eph wrote:

This seems overly conservative I would think, wouldn't it prevent to inline a case like:

```
void foo(int value) {
  if (value > 20) foo(value-1);
}
void bar() {
  foo(10);
}
```

Here foo is self-recursive, but inlining it into bar is still desirable.


https://github.com/llvm/llvm-project/pull/88452


More information about the Mlir-commits mailing list