[Transforms][CodeExtraction] bug fix regions with stackrestore (#118564)

Ensure code extraction for outlining to a function does not create a function with stacksave of caller to restore stack (e.g. tail call).

Author: AbhayKanhere

llvm/lib/Transforms/Utils/CodeExtractor.cpp

@@ -627,6 +627,24 @@ bool CodeExtractor::isEligible() const {
         return false;
     }
   }
+  // stacksave as input implies stackrestore in the outlined function.
+  // This can confuse prolog epilog insertion phase.
+  // stacksave's uses must not cross outlined function.
+  for (BasicBlock *BB : Blocks) {
+    for (Instruction &I : *BB) {
+      IntrinsicInst *II = dyn_cast<IntrinsicInst>(&I);
+      if (!II)
+        continue;
+      bool IsSave = II->getIntrinsicID() == Intrinsic::stacksave;
+      bool IsRestore = II->getIntrinsicID() == Intrinsic::stackrestore;
+      if (IsSave && any_of(II->users(), [&Blks = this->Blocks](User *U) {
+            return !definedInRegion(Blks, U);
+          }))
+        return false;
+      if (IsRestore && !definedInRegion(Blocks, II->getArgOperand(0)))
+        return false;
+    }
+  }
   return true;
 }
 

llvm/test/Transforms/HotColdSplit/outline-inner-region-stacktoocomplex.ll

@@ -0,0 +1,71 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+; RUN: opt -S --passes=hotcoldsplit --hotcoldsplit-max-params=1 < %s | FileCheck %s
+
+target datalayout = "E-m:a-p:32:32-i64:64-n32"
+
+define void @foo(i32 %cond) {
+; CHECK-LABEL: define void @foo(
+; CHECK-SAME: i32 [[COND:%.*]]) {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    [[COND_ADDR:%.*]] = alloca i32, align 4
+; CHECK-NEXT:    store i32 [[COND]], ptr [[COND_ADDR]], align 4
+; CHECK-NEXT:    [[TMP0:%.*]] = load i32, ptr [[COND_ADDR]], align 4
+; CHECK-NEXT:    [[STKS:%.*]] = tail call ptr @llvm.stacksave.p0()
+; CHECK-NEXT:    [[TOBOOL:%.*]] = icmp ne i32 [[TMP0]], 0
+; CHECK-NEXT:    br i1 [[TOBOOL]], label %[[IF_THEN:.*]], label %[[IF_END2:.*]]
+; CHECK:       [[IF_THEN]]:
+; CHECK-NEXT:    [[TMP1:%.*]] = load i32, ptr [[COND_ADDR]], align 4
+; CHECK-NEXT:    call void @sink(i32 [[TMP0]])
+; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt i32 [[TMP1]], 10
+; CHECK-NEXT:    br i1 [[CMP]], label %[[IF_THEN1:.*]], label %[[IF_ELSE:.*]]
+; CHECK:       [[IF_THEN1]]:
+; CHECK-NEXT:    call void @sideeffect(i32 2)
+; CHECK-NEXT:    br label %[[IF_END:.*]]
+; CHECK:       [[IF_ELSE]]:
+; CHECK-NEXT:    call void @sink(i32 0)
+; CHECK-NEXT:    call void @sideeffect(i32 0)
+; CHECK-NEXT:    call void @llvm.stackrestore.p0(ptr [[STKS]])
+; CHECK-NEXT:    br label %[[IF_END]]
+; CHECK:       [[IF_END]]:
+; CHECK-NEXT:    br label %[[IF_END2]]
+; CHECK:       [[IF_END2]]:
+; CHECK-NEXT:    call void @sideeffect(i32 1)
+; CHECK-NEXT:    ret void
+;
+entry:
+  %cond.addr = alloca i32
+  store i32 %cond, ptr %cond.addr
+  %0 = load i32, ptr %cond.addr
+  %stks =  tail call ptr @llvm.stacksave.p0()
+  %tobool = icmp ne i32 %0, 0
+  br i1 %tobool, label %if.then, label %if.end2
+
+if.then:                                          ; preds = %entry
+  %1 = load i32, ptr %cond.addr
+  call void @sink(i32 %0)
+  %cmp = icmp sgt i32 %1, 10
+  br i1 %cmp, label %if.then1, label %if.else
+
+if.then1:                                         ; preds = %if.then
+  call void @sideeffect(i32 2)
+  br label %if.end
+
+if.else:                                          ; preds = %if.then
+  call void @sink(i32 0)
+  call void @sideeffect(i32 0)
+  call void @llvm.stackrestore.p0(ptr %stks)
+  br label %if.end
+
+
+if.end:                                           ; preds = %if.else, %if.then1
+  br label %if.end2
+
+if.end2:                                          ; preds = %entry
+  call void @sideeffect(i32 1)
+  ret void
+}
+
+
+declare void @sideeffect(i32)
+
+declare void @sink(i32) cold