[mlir][inliner] Return early if the inliningThreshold is 0U or -1U.

[naibaf7]

Computing the inlinling profitability can be costly due to walking the graph when counting the number of operations.

This PR addresses that by returning early if the threshold is set to never or always inline.

[llvmbot]

@llvm/pr-subscribers-mlir

@llvm/pr-subscribers-mlir-core

Author: Fabian Tschopp (naibaf7)

Changes

Computing the inlinling profitability can be costly due to walking the graph when counting the number of operations.

This PR addresses that by returning early if the threshold is set to never or always inline.

---

Full diff: https://github.com/llvm/llvm-project/pull/86287.diff

1 Files Affected:

• (modified) mlir/lib/Transforms/InlinerPass.cpp (+6-1)

diff --git a/mlir/lib/Transforms/InlinerPass.cpp b/mlir/lib/Transforms/InlinerPass.cpp
index 08d8dbf73a6a1d..8be7868b576721 100644
--- a/mlir/lib/Transforms/InlinerPass.cpp
+++ b/mlir/lib/Transforms/InlinerPass.cpp
@@ -93,12 +93,17 @@ InlinerPass::InlinerPass(std::function<void(OpPassManager &)> defaultPipeline,
 // Return true if the inlining ratio does not exceed the threshold.
 static bool isProfitableToInline(const Inliner::ResolvedCall &resolvedCall,
                                  unsigned inliningThreshold) {
+  if (inliningThreshold == 0U)
+    return false;
+  if (inliningThreshold == -1U)
+    return true;
+
   Region *callerRegion = resolvedCall.sourceNode->getCallableRegion();
   Region *calleeRegion = resolvedCall.targetNode->getCallableRegion();
 
   // We should not get external nodes here, but just return true
   // for now to preserve the original behavior of the inliner pass.
-  if (!calleeRegion || !calleeRegion)
+  if (!callerRegion || !calleeRegion)
     return true;
 
   auto countOps = [](Region *region) {

[weiweichen]

LGTM

[dukebw]

LGTM!

[github-actions]

✅ With the latest revision this PR passed the C/C++ code formatter.

[vzakhari]

Thank you for the fix!

[joker-eph]

And I thought I had checked for this on the original revision…. Apparently not, thanks for the fix!