-
Notifications
You must be signed in to change notification settings - Fork 13.5k
[InlineCost]: Optimize inlining of recursive function. #139982
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Open
hassnaaHamdi
wants to merge
3
commits into
llvm:main
Choose a base branch
from
hassnaaHamdi:improve_inlining_decision
base: main
Could not load branches
Branch not found: {{ refName }}
Loading
Could not load tags
Nothing to show
Loading
Are you sure you want to change the base?
Some commits from the old base branch may be removed from the timeline,
and old review comments may become outdated.
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
- Consider inlining recursive function of depth 1 only when the caller is the function itself instead of inlining it for each callsite so that we avoid redundant work. - Use CondContext instead of DomTree for better compilation time.
@llvm/pr-subscribers-llvm-transforms @llvm/pr-subscribers-llvm-analysis Author: Hassnaa Hamdi (hassnaaHamdi) Changes
Full diff: https://github.com/llvm/llvm-project/pull/139982.diff 4 Files Affected:
diff --git a/llvm/include/llvm/Analysis/SimplifyQuery.h b/llvm/include/llvm/Analysis/SimplifyQuery.h
index e8f43c8c2e91f..063ca4eaa9db0 100644
--- a/llvm/include/llvm/Analysis/SimplifyQuery.h
+++ b/llvm/include/llvm/Analysis/SimplifyQuery.h
@@ -62,6 +62,8 @@ struct InstrInfoQuery {
struct CondContext {
Value *Cond;
bool Invert = false;
+ // Condition is true if CxtI is in the true successor of Cond.
+ bool CondIsTrue = false;
SmallPtrSet<Value *, 4> AffectedValues;
CondContext(Value *Cond) : Cond(Cond) {}
diff --git a/llvm/lib/Analysis/InlineCost.cpp b/llvm/lib/Analysis/InlineCost.cpp
index 8ddfa1e4eb6f7..12f780fea1620 100644
--- a/llvm/lib/Analysis/InlineCost.cpp
+++ b/llvm/lib/Analysis/InlineCost.cpp
@@ -1688,66 +1688,52 @@ bool CallAnalyzer::simplifyCmpInstForRecCall(CmpInst &Cmp) {
if (!isa<Argument>(Cmp.getOperand(0)) || !isa<Constant>(Cmp.getOperand(1)))
return false;
auto *CmpOp = Cmp.getOperand(0);
- Function *F = Cmp.getFunction();
- // Iterate over the users of the function to check if it's a recursive
- // function:
- for (auto *U : F->users()) {
- CallInst *Call = dyn_cast<CallInst>(U);
- if (!Call || Call->getFunction() != F || Call->getCalledFunction() != F)
- continue;
- auto *CallBB = Call->getParent();
- auto *Predecessor = CallBB->getSinglePredecessor();
- // Only handle the case when the callsite has a single predecessor:
- if (!Predecessor)
- continue;
+ // Make sure that the callsite is recursive:
+ if (CandidateCall.getCaller() != &F)
+ return false;
+ CallInst *CallInstr = dyn_cast<CallInst>(&CandidateCall);
+ // Only handle the case when the callsite has a single predecessor:
+ auto *CallBB = CallInstr->getParent();
+ auto *Predecessor = CallBB->getSinglePredecessor();
+ if (!Predecessor)
+ return false;
+ // Check if the callsite is guarded by the same Cmp instruction:
+ auto *Br = dyn_cast<BranchInst>(Predecessor->getTerminator());
+ if (!Br || Br->isUnconditional() || Br->getCondition() != &Cmp)
+ return false;
- auto *Br = dyn_cast<BranchInst>(Predecessor->getTerminator());
- if (!Br || Br->isUnconditional())
- continue;
- // Check if the Br condition is the same Cmp instr we are investigating:
- if (Br->getCondition() != &Cmp)
- continue;
- // Check if there are any arg of the recursive callsite is affecting the cmp
- // instr:
- bool ArgFound = false;
- Value *FuncArg = nullptr, *CallArg = nullptr;
- for (unsigned ArgNum = 0;
- ArgNum < F->arg_size() && ArgNum < Call->arg_size(); ArgNum++) {
- FuncArg = F->getArg(ArgNum);
- CallArg = Call->getArgOperand(ArgNum);
- if (FuncArg == CmpOp && CallArg != CmpOp) {
- ArgFound = true;
- break;
- }
- }
- if (!ArgFound)
- continue;
- // Now we have a recursive call that is guarded by a cmp instruction.
- // Check if this cmp can be simplified:
- SimplifyQuery SQ(DL, dyn_cast<Instruction>(CallArg));
- DomConditionCache DC;
- DC.registerBranch(Br);
- SQ.DC = &DC;
- if (DT.root_size() == 0) {
- // Dominator tree was never constructed for any function yet.
- DT.recalculate(*F);
- } else if (DT.getRoot()->getParent() != F) {
- // Dominator tree was constructed for a different function, recalculate
- // it for the current function.
- DT.recalculate(*F);
+ // Check if there is any arg of the recursive callsite is affecting the cmp
+ // instr:
+ bool ArgFound = false;
+ Value *FuncArg = nullptr, *CallArg = nullptr;
+ for (unsigned ArgNum = 0;
+ ArgNum < F.arg_size() && ArgNum < CallInstr->arg_size(); ArgNum++) {
+ FuncArg = F.getArg(ArgNum);
+ CallArg = CallInstr->getArgOperand(ArgNum);
+ if (FuncArg == CmpOp && CallArg != CmpOp) {
+ ArgFound = true;
+ break;
}
- SQ.DT = &DT;
- Value *SimplifiedInstruction = llvm::simplifyInstructionWithOperands(
- cast<CmpInst>(&Cmp), {CallArg, Cmp.getOperand(1)}, SQ);
- if (auto *ConstVal = dyn_cast_or_null<ConstantInt>(SimplifiedInstruction)) {
- bool IsTrueSuccessor = CallBB == Br->getSuccessor(0);
- // Make sure that the BB of the recursive call is NOT the next successor
- // of the icmp. In other words, make sure that the recursion depth is 1.
- if ((ConstVal->isOne() && !IsTrueSuccessor) ||
- (ConstVal->isZero() && IsTrueSuccessor)) {
- SimplifiedValues[&Cmp] = ConstVal;
- return true;
- }
+ }
+ if (!ArgFound)
+ return false;
+
+ // Now we have a recursive call that is guarded by a cmp instruction.
+ // Check if this cmp can be simplified:
+ SimplifyQuery SQ(DL, dyn_cast<Instruction>(CallArg));
+ CondContext CC(cast<Value>(&Cmp));
+ CC.CondIsTrue = CallBB == Br->getSuccessor(0);
+ SQ.CC = &CC;
+ CC.AffectedValues.insert(FuncArg);
+ Value *SimplifiedInstruction = llvm::simplifyInstructionWithOperands(
+ cast<CmpInst>(&Cmp), {CallArg, Cmp.getOperand(1)}, SQ);
+ if (auto *ConstVal = dyn_cast_or_null<ConstantInt>(SimplifiedInstruction)) {
+ // Make sure that the BB of the recursive call is NOT the true successor
+ // of the icmp. In other words, make sure that the recursion depth is 1.
+ if ((ConstVal->isOne() && !CC.CondIsTrue) ||
+ (ConstVal->isZero() && CC.CondIsTrue)) {
+ SimplifiedValues[&Cmp] = ConstVal;
+ return true;
}
}
return false;
diff --git a/llvm/lib/Analysis/ValueTracking.cpp b/llvm/lib/Analysis/ValueTracking.cpp
index 3d403531cea2f..e7d937a0893ab 100644
--- a/llvm/lib/Analysis/ValueTracking.cpp
+++ b/llvm/lib/Analysis/ValueTracking.cpp
@@ -5014,6 +5014,10 @@ static KnownFPClass computeKnownFPClassFromContext(const Value *V,
const SimplifyQuery &Q) {
KnownFPClass KnownFromContext;
+ if (Q.CC && Q.CC->AffectedValues.contains(V))
+ computeKnownFPClassFromCond(V, Q.CC->Cond, 0, Q.CC->CondIsTrue, Q.CxtI,
+ KnownFromContext);
+
if (!Q.CxtI)
return KnownFromContext;
diff --git a/llvm/test/Transforms/Inline/inline-recursive-fn2.ll b/llvm/test/Transforms/Inline/inline-recursive-fn2.ll
new file mode 100644
index 0000000000000..0323a6ee3a75a
--- /dev/null
+++ b/llvm/test/Transforms/Inline/inline-recursive-fn2.ll
@@ -0,0 +1,45 @@
+; RUN: opt -passes='cgscc(inline),instcombine,cgscc(inline)' -S -debug-only=inline -disable-output < %s 2>&1 | FileCheck %s
+
+; CHECK: Inlining calls in: test
+; CHECK: Function size: 2
+; CHECK: NOT Inlining (cost=never): recursive, Call: %call = tail call float @inline_rec_true_successor(float %x, float %scale)
+
+; CHECK: Inlining calls in: inline_rec_true_successor
+; CHECK: Function size: 10
+; CHECK: Inlining (cost=-35, threshold=337), Call: %call = tail call float @inline_rec_true_successor(float %fneg, float %scale)
+; CHECK: Size after inlining: 17
+; CHECK: NOT Inlining (cost=never): noinline function attribute, Call: %call_test = tail call float @test(float %fneg, float %common.ret18.op.i)
+; CHECK: NOT Inlining (cost=never): noinline function attribute, Call: %call_test.i = tail call float @test(float %x, float %call.i)
+; CHECK: Skipping inlining due to history: inline_rec_true_successor -> inline_rec_true_successor
+; CHECK: Updated inlining SCC: (test, inline_rec_true_successor)
+
+; CHECK: Inlining calls in: test
+; CHECK: Function size: 2
+; CHECK: Inlining (cost=25, threshold=225), Call: %call = tail call float @inline_rec_true_successor(float %x, float %scale)
+; CHECK: Size after inlining: 10
+
+define float @test(float %x, float %scale) noinline {
+entry:
+ %call = tail call float @inline_rec_true_successor(float %x, float %scale)
+ ret float %call
+}
+
+define float @inline_rec_true_successor(float %x, float %scale) {
+entry:
+ %cmp = fcmp olt float %x, 0.000000e+00
+ br i1 %cmp, label %if.then, label %if.end
+
+common.ret18: ; preds = %if.then, %if.end
+ %common.ret18.op = phi float [ %call_test, %if.then ], [ %mul, %if.end ]
+ ret float %common.ret18.op
+
+if.then: ; preds = %entry
+ %fneg = fneg float %x
+ %call = tail call float @inline_rec_true_successor(float %fneg, float %scale)
+ %call_test = tail call float @test(float %fneg, float %call)
+ br label %common.ret18
+
+if.end: ; preds = %entry
+ %mul = fmul float %x, %scale
+ br label %common.ret18
+}
|
66378f5
to
180244d
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
the caller is the function itself instead of inlining it
for each callsite so that we avoid redundant work.