[CodeExtractor] Emit lifetime markers around reloads of outputs

[vsk]

Summary

CodeExtractor permits extracting a region of blocks from a function even
when values defined within the region are used outside of it.

This is typically done by creating an alloca in the original function
and reloading the alloca after a call to the extracted function.

Wrap the reload in lifetime start/end markers to promote stack coloring.

Depends on D55967.

Suggested by Sergei Kachkov!

[vsk]

• Tighten up the tests a bit.

[vsk]

+ @gyiu, as I believe he's worked on the partial inliner.

[vsk]

Gentle ping (this isn't urgent, but it seems nice to have).

[kachkov98]

lgtm

llvm/trunk/lib/Transforms/Utils/CodeExtractor.cpp

for (unsigned i = 0, e = Users.size(); i != e; ++i)
     if (Instruction *I = dyn_cast<Instruction>(Users[i]))
       if (I->isTerminator() && !Blocks.count(I->getParent()) &&
           I->getParent()->getParent() == oldFunction)
         I->replaceUsesOfWith(header, newHeader);
 
   return newFunction;
 }
 
+/// Scan the extraction region for lifetime markers which reference inputs.
+/// Erase these markers. Return the inputs which were referenced.
+///
+/// The extraction region is defined by a set of blocks (\p Blocks), and a set
+/// of allocas which will be moved from the caller function into the extracted
+/// function (\p SunkAllocas).
+static SetVector<Value *>
+eraseLifetimeMarkersOnInputs(const SetVector<BasicBlock *> &Blocks,
+                             const SetVector<Value *> &SunkAllocas) {
+  SetVector<Value *> InputObjectsWithLifetime;
+  for (BasicBlock *BB : Blocks) {
+    for (auto It = BB->begin(), End = BB->end(); It != End;) {
+      auto *II = dyn_cast<IntrinsicInst>(&*It);
+      ++It;
+      if (!II || !II->isLifetimeStartOrEnd())
+        continue;
+
+      // Get the memory operand of the lifetime marker. If the underlying
+      // object is a sunk alloca, or is otherwise defined in the extraction
+      // region, the lifetime marker must not be erased.
+      Value *Mem = II->getOperand(1)->stripInBoundsOffsets();
+      if (SunkAllocas.count(Mem) || definedInRegion(Blocks, Mem))
+        continue;
+
+      InputObjectsWithLifetime.insert(Mem);
+      II->eraseFromParent();
+    }
+  }
+  return InputObjectsWithLifetime;
+}
+
+/// Insert lifetime start/end markers surrounding the call to the new function
+/// for objects defined in the caller.
+static void insertLifetimeMarkersSurroundingCall(Module *M,
+                                                 ArrayRef<Value *> Objects,
+                                                 CallInst *TheCall) {
+  if (Objects.empty())
+    return;
+
+  LLVMContext &Ctx = M->getContext();
+  auto Int8PtrTy = Type::getInt8PtrTy(Ctx);
+  auto NegativeOne = ConstantInt::getSigned(Type::getInt64Ty(Ctx), -1);
+  auto StartFn = llvm::Intrinsic::getDeclaration(
+      M, llvm::Intrinsic::lifetime_start, Int8PtrTy);
+  auto EndFn = llvm::Intrinsic::getDeclaration(M, llvm::Intrinsic::lifetime_end,
+                                               Int8PtrTy);
+  Instruction *Term = TheCall->getParent()->getTerminator();
+  for (Value *Mem : Objects) {
+    assert((!isa<Instruction>(Mem) ||
+            cast<Instruction>(Mem)->getFunction() == TheCall->getFunction()) &&
+           "Input memory not defined in original function");
+    Value *MemAsI8Ptr = nullptr;
+    if (Mem->getType() == Int8PtrTy)
+      MemAsI8Ptr = Mem;
+    else
+      MemAsI8Ptr =
+          CastInst::CreatePointerCast(Mem, Int8PtrTy, "lt.cast", TheCall);
+
+    auto StartMarker = CallInst::Create(StartFn, {NegativeOne, MemAsI8Ptr});
+    StartMarker->insertBefore(TheCall);
+    auto EndMarker = CallInst::Create(EndFn, {NegativeOne, MemAsI8Ptr});
+    EndMarker->insertBefore(Term);
+  }
+}
+
 /// emitCallAndSwitchStatement - This method sets up the caller side by adding
 /// the call instruction, splitting any PHI nodes in the header block as
 /// necessary.
 CallInst *CodeExtractor::emitCallAndSwitchStatement(Function *newFunction,
                                                     BasicBlock *codeReplacer,
                                                     ValueSet &inputs,
                                                     ValueSet &outputs) {
   // Emit a call to the new function, passing in: *pointer to struct (if
default:
     // of the other successors.
     TheSwitch->setCondition(call);
     TheSwitch->setDefaultDest(TheSwitch->getSuccessor(NumExitBlocks));
     // Remove redundant case
     TheSwitch->removeCase(SwitchInst::CaseIt(TheSwitch, NumExitBlocks-1));
     break;
   }
 
+  // Insert lifetime markers around the reloads of any output values. The
+  // allocas output values are stored in are only in-use in the codeRepl block.
+  insertLifetimeMarkersSurroundingCall(M, ReloadOutputs, call);
+
   return call;
 }
 
 void CodeExtractor::moveCodeToFunction(Function *newFunction) {
   Function *oldFunc = (*Blocks.begin())->getParent();
   Function::BasicBlockListType &oldBlocks = oldFunc->getBasicBlockList();
   Function::BasicBlockListType &newBlocks = newFunction->getBasicBlockList();
 
for (unsigned I = 0, E = BranchDist.Weights.size(); I < E; ++I) {
     BranchProbability BP(Weight.Amount, BranchDist.Total);
     BPI->setEdgeProbability(CodeReplacer, Weight.TargetNode.Index, BP);
   }
   TI->setMetadata(
       LLVMContext::MD_prof,
       MDBuilder(TI->getContext()).createBranchWeights(BranchWeights));
 }
 
-/// Scan the extraction region for lifetime markers which reference inputs.
-/// Erase these markers. Return the inputs which were referenced.
-///
-/// The extraction region is defined by a set of blocks (\p Blocks), and a set
-/// of allocas which will be moved from the caller function into the extracted
-/// function (\p SunkAllocas).
-static SetVector<Value *>
-eraseLifetimeMarkersOnInputs(const SetVector<BasicBlock *> &Blocks,
-                             const SetVector<Value *> &SunkAllocas) {
-  SetVector<Value *> InputObjectsWithLifetime;
-  for (BasicBlock *BB : Blocks) {
-    for (auto It = BB->begin(), End = BB->end(); It != End;) {
-      auto *II = dyn_cast<IntrinsicInst>(&*It);
-      ++It;
-      if (!II || !II->isLifetimeStartOrEnd())
-        continue;
-
-      // Get the memory operand of the lifetime marker. If the underlying
-      // object is a sunk alloca, or is otherwise defined in the extraction
-      // region, the lifetime marker must not be erased.
-      Value *Mem = II->getOperand(1)->stripInBoundsOffsets();
-      if (SunkAllocas.count(Mem) || definedInRegion(Blocks, Mem))
-        continue;
-
-      InputObjectsWithLifetime.insert(Mem);
-      II->eraseFromParent();
-    }
-  }
-  return InputObjectsWithLifetime;
-}
-
-/// Insert lifetime start/end markers surrounding the call to the new function
-/// for objects defined in the caller.
-static void insertLifetimeMarkersSurroundingCall(
-    Module *M, const SetVector<Value *> &InputObjectsWithLifetime,
-    CallInst *TheCall) {
-  if (InputObjectsWithLifetime.empty())
-    return;
-
-  LLVMContext &Ctx = M->getContext();
-  auto Int8PtrTy = Type::getInt8PtrTy(Ctx);
-  auto NegativeOne = ConstantInt::getSigned(Type::getInt64Ty(Ctx), -1);
-  auto LifetimeStartFn = llvm::Intrinsic::getDeclaration(
-      M, llvm::Intrinsic::lifetime_start, Int8PtrTy);
-  auto LifetimeEndFn = llvm::Intrinsic::getDeclaration(
-      M, llvm::Intrinsic::lifetime_end, Int8PtrTy);
-  for (Value *Mem : InputObjectsWithLifetime) {
-    assert((!isa<Instruction>(Mem) ||
-            cast<Instruction>(Mem)->getFunction() == TheCall->getFunction()) &&
-           "Input memory not defined in original function");
-    Value *MemAsI8Ptr = nullptr;
-    if (Mem->getType() == Int8PtrTy)
-      MemAsI8Ptr = Mem;
-    else
-      MemAsI8Ptr =
-          CastInst::CreatePointerCast(Mem, Int8PtrTy, "lt.cast", TheCall);
-
-    auto StartMarker =
-        CallInst::Create(LifetimeStartFn, {NegativeOne, MemAsI8Ptr});
-    StartMarker->insertBefore(TheCall);
-    auto EndMarker = CallInst::Create(LifetimeEndFn, {NegativeOne, MemAsI8Ptr});
-    EndMarker->insertAfter(TheCall);
-  }
-}
-
 Function *CodeExtractor::extractCodeRegion() {
   if (!isEligible())
     return nullptr;
 
   // Assumption: this is a single-entry code region, and the header is the first
   // block in the region.
   BasicBlock *header = *Blocks.begin();
   Function *oldFunction = header->getParent();
Function *CodeExtractor::extractCodeRegion() {
   CallInst *TheCall =
       emitCallAndSwitchStatement(newFunction, codeReplacer, inputs, outputs);
 
   moveCodeToFunction(newFunction);
 
   // Replicate the effects of any lifetime start/end markers which referenced
   // input objects in the extraction region by placing markers around the call.
   insertLifetimeMarkersSurroundingCall(oldFunction->getParent(),
-                                       InputObjectsWithLifetime, TheCall);
+                                       InputObjectsWithLifetime.getArrayRef(),
+                                       TheCall);
 
   // Propagate personality info to the new function if there is one.
   if (oldFunction->hasPersonalityFn())
     newFunction->setPersonalityFn(oldFunction->getPersonalityFn());
 
   // Update the branch weights for the exit block.
   if (BFI && NumExitBlocks > 1)
     calculateNewCallTerminatorWeights(codeReplacer, ExitWeights, BPI);

llvm/trunk/test/Transforms/CodeExtractor/PartialInlineDebug.ll

 if.end:                                           ; preds = %if.then, %entry
   %v2 = phi i32 [ %v, %entry ], [ %mul, %if.then2 ]
   %add = add nsw i32 %v2, 200
   ret i32 %add
 }
 
 ; CHECK-LABEL: @caller
 ; CHECK: codeRepl.i:
-; CHECK-NEXT: call void @callee.2.if.then(i32 %v, i32* %mul.loc.i), !dbg ![[DBG2:[0-9]+]]
+; CHECK-NOT: br label
+; CHECK: call void @callee.2.if.then(i32 %v, i32* %mul.loc.i), !dbg ![[DBG2:[0-9]+]]
 define i32 @caller(i32 %v) !dbg !8 {
 entry:
   %call = call i32 @callee(i32 %v), !dbg !14
   ret i32 %call
 }
 
 
 ; CHECK-LABEL: @callee2

 if.end:
   %v2 = phi i32 [ %v, %entry ], [ %sub, %if.then2 ]
   %add = add nsw i32 %v2, 200
   ret i32 %add
 }
 
 ; CHECK-LABEL: @caller2
 ; CHECK: codeRepl.i:
-; CHECK-NEXT: call void @callee2.1.if.then(i32 %v, i32* %sub.loc.i), !dbg ![[DBG4:[0-9]+]]
+; CHECK-NOT: br label
+; CHECK: call void @callee2.1.if.then(i32 %v, i32* %sub.loc.i), !dbg ![[DBG4:[0-9]+]]
 define i32 @caller2(i32 %v) !dbg !21 {
 entry:
   %call = call i32 @callee2(i32 %v), !dbg !22
   ret i32 %call
 }
 
 ; CHECK-LABEL: define internal void @callee2.1.if.then
 ; CHECK: br label %if.then, !dbg ![[DBG5:[0-9]+]]

llvm/trunk/test/Transforms/CodeExtractor/PartialInlineInvokeProducesOutVal.ll

bb5:                                              ; preds = %bb4, %bb1, %bb
   %retval = phi i8* [ %call26, %cont ], [ undef, %bb]
   ret i8* %retval
 }
 
 ; CHECK-LABEL: @dummy_caller
 ; CHECK-LABEL: bb:
 ; CHECK-NEXT:  [[CALL26LOC:%.*]] = alloca i8*
 ; CHECK-LABEL: codeRepl.i:
+; CHECK-NEXT:   %lt.cast.i = bitcast i8** [[CALL26LOC]] to i8*
+; CHECK-NEXT:   call void @llvm.lifetime.start.p0i8(i64 -1, i8* %lt.cast.i)
 ; CHECK-NEXT:   call void @bar.1.bb1(i8** [[CALL26LOC]])
+; CHECK-NEXT:   %call26.reload.i = load i8*, i8** [[CALL26LOC]]
+; CHECK-NEXT:   call void @llvm.lifetime.end.p0i8(i64 -1, i8* %lt.cast.i)
 define i8* @dummy_caller(i32 %arg) {
 bb:
   %tmp = tail call i8* @bar(i32 %arg)
   ret i8* %tmp
 }
 
 ; CHECK-LABEL: define internal void @bar.1.bb1
 ; CHECK-LABEL: bb1:

llvm/trunk/test/Transforms/CodeExtractor/PartialInlineVarArgsDebug.ll

 if.end:                                           ; preds = %if.then, %entry
   %v2 = phi i32 [ %v, %entry ], [ %mul, %if.then ]
   %add = add nsw i32 %v2, 200, !dbg !22
   ret i32 %add, !dbg !23
 }
 
 ; CHECK-LABEL: @caller
 ; CHECK: codeRepl.i:
-; CHECK-NEXT: call void (i32, i32*, ...) @callee.1.if.then(i32 %v, i32* %mul.loc.i, i32 99), !dbg ![[DBG2:[0-9]+]]
+; CHECK-NOT: br label
+; CHECK: call void (i32, i32*, ...) @callee.1.if.then(i32 %v, i32* %mul.loc.i, i32 99), !dbg ![[DBG2:[0-9]+]]
 define i32 @caller(i32 %v) !dbg !8 {
 entry:
   %call = call i32 (i32, ...) @callee(i32 %v, i32 99), !dbg !14
   ret i32 %call, !dbg !15
 }
 
 ; CHECK-LABEL: define internal void @callee.1.if.then
 ; CHECK: br label %if.then, !dbg ![[DBG3:[0-9]+]]

llvm/trunk/test/Transforms/HotColdSplit/duplicate-phi-preds-crash.ll

 declare void @free(i8* %ptr)
 
 declare void @sink() cold
 
 ; CHECK-LABEL: define {{.*}}@realloc2(
 ; CHECK: call {{.*}}@sideeffect(
 ; CHECK: call {{.*}}@realloc(
 ; CHECK-LABEL: codeRepl:
-; CHECK-NEXT: call {{.*}}@realloc2.cold.1(i64 %size, i8* %ptr, i8** %retval.0.ce.loc)
+; CHECK: call {{.*}}@realloc2.cold.1(i64 %size, i8* %ptr, i8** %retval.0.ce.loc)
 ; CHECK-LABEL: cleanup:
 ; CHECK-NEXT: phi i8* [ null, %if.then ], [ %call, %if.end ], [ %retval.0.ce.reload, %codeRepl ]
 define i8* @realloc2(i8* %ptr, i64 %size) {
 entry:
   %0 = add i64 %size, -1
   %1 = icmp ugt i64 %0, 184549375
   br i1 %1, label %if.then, label %if.end
 

llvm/trunk/test/Transforms/HotColdSplit/lifetime-markers-on-inputs.ll

normalPath:
   call void @llvm.lifetime.end.p0i8(i64 1, i8* %local1_cast)
   call void @llvm.lifetime.start.p0i8(i64 1, i8* %local2_cast)
   call void @use(i8* %local2_cast)
   call void @llvm.lifetime.end.p0i8(i64 1, i8* %local2_cast)
   ret void
 
 ; CHECK-LABEL: codeRepl:
 ; CHECK: [[local1_cast:%.*]] = bitcast i256* %local1 to i8*
-; CHECK: call void @llvm.lifetime.start.p0i8(i64 -1, i8* [[local1_cast]])
-; CHECK: [[local2_cast:%.*]] = bitcast i256* %local2 to i8*
-; CHECK: call void @llvm.lifetime.start.p0i8(i64 -1, i8* [[local2_cast]])
-; CHECK: call i1 @foo.cold.1(i8* %local1_cast, i8* %local2_cast)
-; CHECK: call void @llvm.lifetime.end.p0i8(i64 -1, i8* [[local2_cast]])
-; CHECK: call void @llvm.lifetime.end.p0i8(i64 -1, i8* [[local1_cast]])
-; CHECK: br i1
+; CHECK-NEXT: call void @llvm.lifetime.start.p0i8(i64 -1, i8* [[local1_cast]])
+; CHECK-NEXT: [[local2_cast:%.*]] = bitcast i256* %local2 to i8*
+; CHECK-NEXT: call void @llvm.lifetime.start.p0i8(i64 -1, i8* [[local2_cast]])
+; CHECK-NEXT: call i1 @foo.cold.1(i8* %local1_cast, i8* %local2_cast)
+; CHECK-NEXT: call void @llvm.lifetime.end.p0i8(i64 -1, i8* [[local1_cast]])
+; CHECK-NEXT: call void @llvm.lifetime.end.p0i8(i64 -1, i8* [[local2_cast]])
+; CHECK-NEXT: br i1
 
 outlinedPath:
   ; These two uses of stack slots are overlapping. This should prevent
   ; merging of stack slots. CodeExtractor must replicate the effects of
   ; these markers in the caller to inhibit stack coloring.
   %gep1 = getelementptr inbounds i8, i8* %local1_cast, i64 1
   call void @llvm.lifetime.start.p0i8(i64 1, i8* %gep1)
   call void @llvm.lifetime.start.p0i8(i64 1, i8* %local2_cast)

llvm/trunk/test/Transforms/HotColdSplit/split-phis-in-exit-blocks.ll

 ; RUN: opt -hotcoldsplit-threshold=0 -hotcoldsplit -S < %s | FileCheck %s
 
 target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
 target triple = "x86_64-apple-macosx10.14.0"
 
 ; CHECK-LABEL: define {{.*}}@pluto(
 ; CHECK-NEXT: bb:
 ; CHECK-NEXT:  %tmp8.ce.loc = alloca i1
 ; CHECK-NEXT:  switch i8 undef, label %codeRepl [
 ; CHECK-NEXT:    i8 0, label %bb7
 ; CHECK-NEXT:    i8 1, label %bb7
 ; CHECK-NEXT:  ]
 ;
 ; CHECK:  codeRepl:
+; CHECK-NEXT:    bitcast
+; CHECK-NEXT:    lifetime.start
 ; CHECK-NEXT:    call void @pluto.cold.1(i1* %tmp8.ce.loc)
 ; CHECK-NEXT:    %tmp8.ce.reload = load i1, i1* %tmp8.ce.loc
+; CHECK-NEXT:    lifetime.end
 ; CHECK-NEXT:    br label %bb7
 ;
 ; CHECK:  bb7:
 ; CHECK:    %tmp8 = phi i1 [ true, %bb ], [ true, %bb ], [ %tmp8.ce.reload, %codeRepl ]
 ; CHECK:    ret void
 
 ; CHECK-LABEL: define {{.*}}@pluto.cold.1(
 ; CHECK: call {{.*}}@sideeffect(i32 1)