[Inline][Cloning] Drop incompatible attributes from NewFunc

Performing `instSimplify` while cloning is unsafe due to incomplete
remapping (as reported in #87534). Ideally, `instSimplify` ought to
reason on the updated newly-cloned function, after returns have been
rewritten and callee entry basic block / call-site have been fixed up.
This is in contrast to `CloneAndPruneIntoFromInst` behaviour, which
is inherently expected to clone basic blocks, with pruning on top of
– if any –, and not actually fixing up returns / CFG, which should be
up to the Inliner. We may solve this by letting `instSimplify` work on
the newly-cloned function, while maintaining old function attributes,
so as to avoid inconsistencies between the yet-to-be-solved return
type, and new function ret type attributes.

Author: antoniofrighetto

llvm/lib/Transforms/Utils/CloneFunction.cpp

@@ -18,6 +18,7 @@
 #include "llvm/Analysis/DomTreeUpdater.h"
 #include "llvm/Analysis/InstructionSimplify.h"
 #include "llvm/Analysis/LoopInfo.h"
+#include "llvm/IR/AttributeMask.h"
 #include "llvm/IR/CFG.h"
 #include "llvm/IR/Constants.h"
 #include "llvm/IR/DebugInfo.h"
@@ -819,6 +820,14 @@ void llvm::CloneAndPruneIntoFromInst(Function *NewFunc, const Function *OldFunc,
     }
   }
 
+  // Drop all incompatible return attributes that cannot be applied to NewFunc
+  // during cloning, so as to allow instruction simplification to reason on the
+  // old state of the function. The original attributes are restored later.
+  AttributeMask IncompatibleAttrs =
+      AttributeFuncs::typeIncompatible(OldFunc->getReturnType());
+  AttributeList Attrs = NewFunc->getAttributes();
+  NewFunc->removeRetAttrs(IncompatibleAttrs);
+
   // As phi-nodes have been now remapped, allow incremental simplification of
   // newly-cloned instructions.
   const DataLayout &DL = NewFunc->getParent()->getDataLayout();
@@ -849,6 +858,9 @@ void llvm::CloneAndPruneIntoFromInst(Function *NewFunc, const Function *OldFunc,
     }
   }
 
+  // Restore attributes.
+  NewFunc->setAttributes(Attrs);
+
   // Remap debug intrinsic operands now that all values have been mapped.
   // Doing this now (late) preserves use-before-defs in debug intrinsics. If
   // we didn't do this, ValueAsMetadata(use-before-def) operands would be

llvm/test/Transforms/Inline/inline-drop-attributes.ll

@@ -0,0 +1,32 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 4
+; RUN: opt < %s -passes=inline -S | FileCheck %s
+; RUN: opt < %s -passes='cgscc(inline)' -S | FileCheck %s
+
+define void @callee() {
+; CHECK-LABEL: define void @callee() {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[VAL_PTR:%.*]] = load ptr, ptr null, align 8
+; CHECK-NEXT:    [[CMP:%.*]] = icmp eq ptr [[VAL_PTR]], null
+; CHECK-NEXT:    [[VAL:%.*]] = load i64, ptr null, align 8
+; CHECK-NEXT:    [[SEL:%.*]] = select i1 [[CMP]], i64 undef, i64 [[VAL]]
+; CHECK-NEXT:    ret void
+;
+entry:
+  %val_ptr = load ptr, ptr null, align 8
+  %cmp = icmp eq ptr %val_ptr, null
+  %val = load i64, ptr null, align 8
+  %sel = select i1 %cmp, i64 undef, i64 %val
+  ret void
+}
+
+define noundef i1 @caller() {
+; CHECK-LABEL: define noundef i1 @caller() {
+; CHECK-NEXT:    [[VAL_PTR_I:%.*]] = load ptr, ptr null, align 8
+; CHECK-NEXT:    [[CMP_I:%.*]] = icmp eq ptr [[VAL_PTR_I]], null
+; CHECK-NEXT:    [[VAL_I:%.*]] = load i64, ptr null, align 8
+; CHECK-NEXT:    [[SEL_I:%.*]] = select i1 [[CMP_I]], i64 undef, i64 [[VAL_I]]
+; CHECK-NEXT:    ret i1 false
+;
+  call void @callee()
+  ret i1 false
+}