[llvm] InlineFunction: Split inlining into predicate and apply functions (PR #134213)
Matt Arsenault via llvm-commits
llvm-commits at lists.llvm.org
Thu Apr 3 09:54:15 PDT 2025
https://github.com/arsenm updated https://github.com/llvm/llvm-project/pull/134213
>From 61ae6efe18fa58381f52092b301e8834c2a63ef6 Mon Sep 17 00:00:00 2001
From: Matt Arsenault <Matthew.Arsenault at amd.com>
Date: Sat, 29 Mar 2025 11:18:19 +0700
Subject: [PATCH] InlineFunction: Split inlining into predicate and apply
functions
This is to support a new inline function reduction in llvm-reduce,
which should pre-filter callsites that are not eligible for inlining.
This code was mostly structured as a match and apply, with a few
exceptions. The ugliest piece is for propagating and verifying compatible
getGC and personalities. Also collection of EHPad and the convergence token
to use are now cached in InlineFunctionInfo.
I was initially confused by the split between the checks performed here
and isInlineViable, so better document how this system is supposed to work.
It turns out this split does make sense, in that isInlineViable checks
if it's possible based on the callee content and the ultimate inline
depended on the callsite context. I think more renames of these functions
would help, and isInlineViable should probably move out of InlineCost to be
with these transfoms.
---
llvm/include/llvm/Analysis/InlineCost.h | 6 +-
llvm/include/llvm/Transforms/Utils/Cloning.h | 28 +++++
llvm/lib/Transforms/Utils/InlineFunction.cpp | 123 ++++++++++++-------
3 files changed, 114 insertions(+), 43 deletions(-)
diff --git a/llvm/include/llvm/Analysis/InlineCost.h b/llvm/include/llvm/Analysis/InlineCost.h
index 90ee75773957a..ec59d54954e16 100644
--- a/llvm/include/llvm/Analysis/InlineCost.h
+++ b/llvm/include/llvm/Analysis/InlineCost.h
@@ -334,7 +334,11 @@ std::optional<InlineCostFeatures> getInliningCostFeatures(
ProfileSummaryInfo *PSI = nullptr,
OptimizationRemarkEmitter *ORE = nullptr);
-/// Minimal filter to detect invalid constructs for inlining.
+/// Check if it is mechanically possible to inline the function \p Callee, based
+/// on the contents of the function.
+///
+/// See also \p CanInlineCallSite as an additional precondition necessary to
+/// perform a valid inline in a particular use context.
InlineResult isInlineViable(Function &Callee);
// This pass is used to annotate instructions during the inline process for
diff --git a/llvm/include/llvm/Transforms/Utils/Cloning.h b/llvm/include/llvm/Transforms/Utils/Cloning.h
index ec1a1d5faa7e9..201e6ba2b491f 100644
--- a/llvm/include/llvm/Transforms/Utils/Cloning.h
+++ b/llvm/include/llvm/Transforms/Utils/Cloning.h
@@ -263,6 +263,9 @@ class InlineFunctionInfo {
/// `InlinedCalls` above is used.
SmallVector<CallBase *, 8> InlinedCallSites;
+ Value *ConvergenceControlToken = nullptr;
+ Instruction *CallSiteEHPad = nullptr;
+
/// Update profile for callee as well as cloned version. We need to do this
/// for regular inlining, but not for inlining from sample profile loader.
bool UpdateProfile;
@@ -271,9 +274,34 @@ class InlineFunctionInfo {
StaticAllocas.clear();
InlinedCalls.clear();
InlinedCallSites.clear();
+ ConvergenceControlToken = nullptr;
+ CallSiteEHPad = nullptr;
}
};
+/// Check if it is legal to perform inlining of the function called by \p CB
+/// into the caller at this particular use, and sets fields in \p IFI.
+///
+/// This does not consider whether it is possible for the function callee itself
+/// to be inlined; for that see isInlineViable.
+InlineResult CanInlineCallSite(const CallBase &CB, InlineFunctionInfo &IFI);
+
+/// This should generally not be used, use InlineFunction instead.
+///
+/// Perform mechanical inlining of \p CB into the caller.
+///
+/// This does not perform any legality or profitability checks for the
+/// inlining. This assumes that CanInlineCallSite was already called, populated
+/// \p IFI, and returned InlineResult::success.
+///
+/// Also assumes that isInlineViable returned InlineResult::success for the
+/// called function.
+void InlineFunctionImpl(CallBase &CB, InlineFunctionInfo &IFI,
+ bool MergeAttributes = false,
+ AAResults *CalleeAAR = nullptr,
+ bool InsertLifetime = true,
+ Function *ForwardVarArgsTo = nullptr);
+
/// This function inlines the called function into the basic
/// block of the caller. This returns false if it is not possible to inline
/// this call. The program is still in a well defined state if this occurs
diff --git a/llvm/lib/Transforms/Utils/InlineFunction.cpp b/llvm/lib/Transforms/Utils/InlineFunction.cpp
index 131fbe654c11c..7236cc0131eb9 100644
--- a/llvm/lib/Transforms/Utils/InlineFunction.cpp
+++ b/llvm/lib/Transforms/Utils/InlineFunction.cpp
@@ -2446,19 +2446,8 @@ llvm::InlineResult llvm::InlineFunction(CallBase &CB, InlineFunctionInfo &IFI,
return Ret;
}
-/// This function inlines the called function into the basic block of the
-/// caller. This returns false if it is not possible to inline this call.
-/// The program is still in a well defined state if this occurs though.
-///
-/// Note that this only does one level of inlining. For example, if the
-/// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now
-/// exists in the instruction stream. Similarly this will inline a recursive
-/// function by one level.
-llvm::InlineResult llvm::InlineFunction(CallBase &CB, InlineFunctionInfo &IFI,
- bool MergeAttributes,
- AAResults *CalleeAAR,
- bool InsertLifetime,
- Function *ForwardVarArgsTo) {
+llvm::InlineResult llvm::CanInlineCallSite(const CallBase &CB,
+ InlineFunctionInfo &IFI) {
assert(CB.getParent() && CB.getFunction() && "Instruction not in function!");
// FIXME: we don't inline callbr yet.
@@ -2475,7 +2464,6 @@ llvm::InlineResult llvm::InlineFunction(CallBase &CB, InlineFunctionInfo &IFI,
// The inliner does not know how to inline through calls with operand bundles
// in general ...
- Value *ConvergenceControlToken = nullptr;
if (CB.hasOperandBundles()) {
for (int i = 0, e = CB.getNumOperandBundles(); i != e; ++i) {
auto OBUse = CB.getOperandBundleAt(i);
@@ -2491,7 +2479,7 @@ llvm::InlineResult llvm::InlineFunction(CallBase &CB, InlineFunctionInfo &IFI,
if (Tag == LLVMContext::OB_kcfi)
continue;
if (Tag == LLVMContext::OB_convergencectrl) {
- ConvergenceControlToken = OBUse.Inputs[0].get();
+ IFI.ConvergenceControlToken = OBUse.Inputs[0].get();
continue;
}
@@ -2509,28 +2497,22 @@ llvm::InlineResult llvm::InlineFunction(CallBase &CB, InlineFunctionInfo &IFI,
// fully implements convergence control tokens, there is no mixing of
// controlled and uncontrolled convergent operations in the whole program.
if (CB.isConvergent()) {
- if (!ConvergenceControlToken &&
+ if (!IFI.ConvergenceControlToken &&
getConvergenceEntry(CalledFunc->getEntryBlock())) {
return InlineResult::failure(
"convergent call needs convergencectrl operand");
}
}
- // If the call to the callee cannot throw, set the 'nounwind' flag on any
- // calls that we inline.
- bool MarkNoUnwind = CB.doesNotThrow();
-
- BasicBlock *OrigBB = CB.getParent();
- Function *Caller = OrigBB->getParent();
+ const BasicBlock *OrigBB = CB.getParent();
+ const Function *Caller = OrigBB->getParent();
// GC poses two hazards to inlining, which only occur when the callee has GC:
// 1. If the caller has no GC, then the callee's GC must be propagated to the
// caller.
// 2. If the caller has a differing GC, it is invalid to inline.
if (CalledFunc->hasGC()) {
- if (!Caller->hasGC())
- Caller->setGC(CalledFunc->getGC());
- else if (CalledFunc->getGC() != Caller->getGC())
+ if (Caller->hasGC() && CalledFunc->getGC() != Caller->getGC())
return InlineResult::failure("incompatible GC");
}
@@ -2548,34 +2530,31 @@ llvm::InlineResult llvm::InlineFunction(CallBase &CB, InlineFunctionInfo &IFI,
? Caller->getPersonalityFn()->stripPointerCasts()
: nullptr;
if (CalledPersonality) {
- if (!CallerPersonality)
- Caller->setPersonalityFn(CalledPersonality);
// If the personality functions match, then we can perform the
// inlining. Otherwise, we can't inline.
// TODO: This isn't 100% true. Some personality functions are proper
// supersets of others and can be used in place of the other.
- else if (CalledPersonality != CallerPersonality)
+ if (CallerPersonality && CalledPersonality != CallerPersonality)
return InlineResult::failure("incompatible personality");
}
// We need to figure out which funclet the callsite was in so that we may
// properly nest the callee.
- Instruction *CallSiteEHPad = nullptr;
if (CallerPersonality) {
EHPersonality Personality = classifyEHPersonality(CallerPersonality);
if (isScopedEHPersonality(Personality)) {
std::optional<OperandBundleUse> ParentFunclet =
CB.getOperandBundle(LLVMContext::OB_funclet);
if (ParentFunclet)
- CallSiteEHPad = cast<FuncletPadInst>(ParentFunclet->Inputs.front());
+ IFI.CallSiteEHPad = cast<FuncletPadInst>(ParentFunclet->Inputs.front());
// OK, the inlining site is legal. What about the target function?
- if (CallSiteEHPad) {
+ if (IFI.CallSiteEHPad) {
if (Personality == EHPersonality::MSVC_CXX) {
// The MSVC personality cannot tolerate catches getting inlined into
// cleanup funclets.
- if (isa<CleanupPadInst>(CallSiteEHPad)) {
+ if (isa<CleanupPadInst>(IFI.CallSiteEHPad)) {
// Ok, the call site is within a cleanuppad. Let's check the callee
// for catchpads.
for (const BasicBlock &CalledBB : *CalledFunc) {
@@ -2595,13 +2574,33 @@ llvm::InlineResult llvm::InlineFunction(CallBase &CB, InlineFunctionInfo &IFI,
}
}
+ return InlineResult::success();
+}
+
+/// This function inlines the called function into the basic block of the
+/// caller. This returns false if it is not possible to inline this call.
+/// The program is still in a well defined state if this occurs though.
+///
+/// Note that this only does one level of inlining. For example, if the
+/// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now
+/// exists in the instruction stream. Similarly this will inline a recursive
+/// function by one level.
+void llvm::InlineFunctionImpl(CallBase &CB, InlineFunctionInfo &IFI,
+ bool MergeAttributes, AAResults *CalleeAAR,
+ bool InsertLifetime, Function *ForwardVarArgsTo) {
+ BasicBlock *OrigBB = CB.getParent();
+ Function *Caller = OrigBB->getParent();
+ Function *CalledFunc = CB.getCalledFunction();
+ assert(CalledFunc && !CalledFunc->isDeclaration() &&
+ "CanInlineCallSite should have verified direct call to definition");
+
// Determine if we are dealing with a call in an EHPad which does not unwind
// to caller.
bool EHPadForCallUnwindsLocally = false;
- if (CallSiteEHPad && isa<CallInst>(CB)) {
+ if (IFI.CallSiteEHPad && isa<CallInst>(CB)) {
UnwindDestMemoTy FuncletUnwindMap;
Value *CallSiteUnwindDestToken =
- getUnwindDestToken(CallSiteEHPad, FuncletUnwindMap);
+ getUnwindDestToken(IFI.CallSiteEHPad, FuncletUnwindMap);
EHPadForCallUnwindsLocally =
CallSiteUnwindDestToken &&
@@ -2618,6 +2617,30 @@ llvm::InlineResult llvm::InlineFunction(CallBase &CB, InlineFunctionInfo &IFI,
ClonedCodeInfo InlinedFunctionInfo;
Function::iterator FirstNewBlock;
+ // GC poses two hazards to inlining, which only occur when the callee has GC:
+ // 1. If the caller has no GC, then the callee's GC must be propagated to the
+ // caller.
+ // 2. If the caller has a differing GC, it is invalid to inline.
+ if (CalledFunc->hasGC()) {
+ if (!Caller->hasGC())
+ Caller->setGC(CalledFunc->getGC());
+ else {
+ assert(CalledFunc->getGC() == Caller->getGC() &&
+ "CanInlineCallSite should have verified compatible GCs");
+ }
+ }
+
+ if (CalledFunc->hasPersonalityFn()) {
+ Constant *CalledPersonality =
+ CalledFunc->getPersonalityFn()->stripPointerCasts();
+ if (!Caller->hasPersonalityFn()) {
+ Caller->setPersonalityFn(CalledPersonality);
+ } else
+ assert(Caller->getPersonalityFn()->stripPointerCasts() ==
+ CalledPersonality &&
+ "CanInlineCallSite should have verified compatible personality");
+ }
+
{ // Scope to destroy VMap after cloning.
ValueToValueMapTy VMap;
struct ByValInit {
@@ -2802,10 +2825,10 @@ llvm::InlineResult llvm::InlineFunction(CallBase &CB, InlineFunctionInfo &IFI,
IFI.GetAssumptionCache(*Caller).registerAssumption(II);
}
- if (ConvergenceControlToken) {
+ if (IFI.ConvergenceControlToken) {
IntrinsicInst *IntrinsicCall = getConvergenceEntry(*FirstNewBlock);
if (IntrinsicCall) {
- IntrinsicCall->replaceAllUsesWith(ConvergenceControlToken);
+ IntrinsicCall->replaceAllUsesWith(IFI.ConvergenceControlToken);
IntrinsicCall->eraseFromParent();
}
}
@@ -2852,6 +2875,10 @@ llvm::InlineResult llvm::InlineFunction(CallBase &CB, InlineFunctionInfo &IFI,
}
}
+ // If the call to the callee cannot throw, set the 'nounwind' flag on any
+ // calls that we inline.
+ bool MarkNoUnwind = CB.doesNotThrow();
+
SmallVector<Value*,4> VarArgsToForward;
SmallVector<AttributeSet, 4> VarArgsAttrs;
for (unsigned i = CalledFunc->getFunctionType()->getNumParams();
@@ -3038,12 +3065,12 @@ llvm::InlineResult llvm::InlineFunction(CallBase &CB, InlineFunctionInfo &IFI,
// Update the lexical scopes of the new funclets and callsites.
// Anything that had 'none' as its parent is now nested inside the callsite's
// EHPad.
- if (CallSiteEHPad) {
+ if (IFI.CallSiteEHPad) {
for (Function::iterator BB = FirstNewBlock->getIterator(),
E = Caller->end();
BB != E; ++BB) {
// Add bundle operands to inlined call sites.
- PropagateOperandBundles(BB, CallSiteEHPad);
+ PropagateOperandBundles(BB, IFI.CallSiteEHPad);
// It is problematic if the inlinee has a cleanupret which unwinds to
// caller and we inline it into a call site which doesn't unwind but into
@@ -3059,11 +3086,11 @@ llvm::InlineResult llvm::InlineFunction(CallBase &CB, InlineFunctionInfo &IFI,
if (auto *CatchSwitch = dyn_cast<CatchSwitchInst>(I)) {
if (isa<ConstantTokenNone>(CatchSwitch->getParentPad()))
- CatchSwitch->setParentPad(CallSiteEHPad);
+ CatchSwitch->setParentPad(IFI.CallSiteEHPad);
} else {
auto *FPI = cast<FuncletPadInst>(I);
if (isa<ConstantTokenNone>(FPI->getParentPad()))
- FPI->setParentPad(CallSiteEHPad);
+ FPI->setParentPad(IFI.CallSiteEHPad);
}
}
}
@@ -3219,7 +3246,7 @@ llvm::InlineResult llvm::InlineFunction(CallBase &CB, InlineFunctionInfo &IFI,
AttributeFuncs::mergeAttributesForInlining(*Caller, *CalledFunc);
// We are now done with the inlining.
- return InlineResult::success();
+ return;
}
// Otherwise, we have the normal case, of more than one block to inline or
@@ -3379,6 +3406,18 @@ llvm::InlineResult llvm::InlineFunction(CallBase &CB, InlineFunctionInfo &IFI,
if (MergeAttributes)
AttributeFuncs::mergeAttributesForInlining(*Caller, *CalledFunc);
+}
- return InlineResult::success();
+llvm::InlineResult llvm::InlineFunction(CallBase &CB, InlineFunctionInfo &IFI,
+ bool MergeAttributes,
+ AAResults *CalleeAAR,
+ bool InsertLifetime,
+ Function *ForwardVarArgsTo) {
+ llvm::InlineResult Result = CanInlineCallSite(CB, IFI);
+ if (Result.isSuccess()) {
+ InlineFunctionImpl(CB, IFI, MergeAttributes, CalleeAAR, InsertLifetime,
+ ForwardVarArgsTo);
+ }
+
+ return Result;
}
More information about the llvm-commits
mailing list