[PartialInlining] Enable recursive partial inlining.
authorMark Lacey <mark.lacey@apple.com>
Wed, 19 Oct 2022 20:20:56 +0000 (13:20 -0700)
committerMark Lacey <mark.lacey@apple.com>
Tue, 6 Dec 2022 07:10:37 +0000 (23:10 -0800)
It seems unnecessarily limiting to disallow recursive partial
inlining, and there are clearly cases where it can benefit
code by avoiding a function call and potentially enabling
other transformations like dead argument elimination
in cases where an argument is only used prior to the early-out
test at the top of the function.

The pass already properly rewrites the recursive calls
within the body of the freshly cloned function, so the only
change here is removing the bail-out when recursion is
detected.

Reviewed By: efriedma

Differential Revision: https://reviews.llvm.org/D136383

llvm/lib/Transforms/IPO/PartialInlining.cpp
llvm/test/Transforms/PartialInlining/recursive_partial_inlining.ll [new file with mode: 0644]

index 97c49c0..1aa737b 100644 (file)
@@ -1489,16 +1489,6 @@ bool PartialInlinerImpl::run(Module &M) {
     if (CurrFunc->use_empty())
       continue;
 
-    bool Recursive = false;
-    for (User *U : CurrFunc->users())
-      if (Instruction *I = dyn_cast<Instruction>(U))
-        if (I->getParent()->getParent() == CurrFunc) {
-          Recursive = true;
-          break;
-        }
-    if (Recursive)
-      continue;
-
     std::pair<bool, Function *> Result = unswitchFunction(*CurrFunc);
     if (Result.second)
       Worklist.push_back(Result.second);
diff --git a/llvm/test/Transforms/PartialInlining/recursive_partial_inlining.ll b/llvm/test/Transforms/PartialInlining/recursive_partial_inlining.ll
new file mode 100644 (file)
index 0000000..fea1714
--- /dev/null
@@ -0,0 +1,31 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -partial-inliner -skip-partial-inlining-cost-analysis -S < %s | FileCheck %s
+define void @_Z26recursive_partial_inliningi(i32 noundef %i) local_unnamed_addr {
+; CHECK-LABEL: @_Z26recursive_partial_inliningi(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[CMP:%.*]] = icmp slt i32 [[I:%.*]], 0
+; CHECK-NEXT:    br i1 [[CMP]], label [[COMMON_RET2:%.*]], label [[IF_END:%.*]]
+; CHECK:       common.ret2:
+; CHECK-NEXT:    ret void
+; CHECK:       if.end:
+; CHECK-NEXT:    [[SUB:%.*]] = add nsw i32 [[I]], -1
+; CHECK-NEXT:    [[CMP_I:%.*]] = icmp slt i32 [[SUB]], 0
+; CHECK-NEXT:    br i1 [[CMP_I]], label [[_Z26RECURSIVE_PARTIAL_INLININGI_1_EXIT:%.*]], label [[CODEREPL_I:%.*]]
+; CHECK:       codeRepl.i:
+; CHECK-NEXT:    call void @_Z26recursive_partial_inliningi.1.if.end(i32 [[SUB]])
+; CHECK-NEXT:    br label [[_Z26RECURSIVE_PARTIAL_INLININGI_1_EXIT]]
+; CHECK:       _Z26recursive_partial_inliningi.1.exit:
+; CHECK-NEXT:    br label [[COMMON_RET2]]
+;
+entry:
+  %cmp = icmp slt i32 %i, 0
+  br i1 %cmp, label %common.ret2, label %if.end
+
+common.ret2:                                      ; preds = %entry, %if.end
+  ret void
+
+if.end:                                           ; preds = %entry
+  %sub = add nsw i32 %i, -1
+  tail call void @_Z26recursive_partial_inliningi(i32 noundef %sub)
+  br label %common.ret2
+}