Skip to content
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

[SCEV] Simplify SCEVExpr for PHI to SCEV for operand if operands are identical #115945

Merged
merged 7 commits into from
Dec 6, 2024
Merged
Show file tree
Hide file tree
Changes from 1 commit
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 4 additions & 0 deletions llvm/include/llvm/Analysis/ScalarEvolution.h
Original file line number Diff line number Diff line change
Expand Up @@ -1761,6 +1761,10 @@ class ScalarEvolution {
/// V.
const SCEV *getOperandsToCreate(Value *V, SmallVectorImpl<Value *> &Ops);

/// Returns SCEV for the first operand of a phi if all phi operands have
/// identical opcodes and operands
akshayrdeodhar marked this conversation as resolved.
Show resolved Hide resolved
const SCEV *createNodeForPHIWithIdenticalOperands(PHINode *PN);

/// Provide the special handling we need to analyze PHI SCEVs.
const SCEV *createNodeForPHI(PHINode *PN);

Expand Down
34 changes: 34 additions & 0 deletions llvm/lib/Analysis/ScalarEvolution.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -6019,6 +6019,37 @@ const SCEV *ScalarEvolution::createNodeFromSelectLikePHI(PHINode *PN) {
return nullptr;
}

// Returns SCEV for the first operand of a phi if all phi operands have
akshayrdeodhar marked this conversation as resolved.
Show resolved Hide resolved
// identical opcodes and operands
// eg.
// a: %add = %a + %b
// br %c
// b: %add1 = %a + %b
// br %c
// c: %phi = phi [%add, a], [%add1, b]
// scev(%phi) => scev(%add)
const SCEV *
ScalarEvolution::createNodeForPHIWithIdenticalOperands(PHINode *PN) {
BinaryOperator *CommonInst = nullptr;
for (Value *Incoming : PN->incoming_values()) {
BinaryOperator *IncomingInst = dyn_cast<BinaryOperator>(Incoming);
akshayrdeodhar marked this conversation as resolved.
Show resolved Hide resolved
if (CommonInst) {
if (!(IncomingInst && CommonInst->isIdenticalTo(IncomingInst))) {
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Should we check that all incoming instructions belong to the same loop?

Copy link
Contributor Author

@akshayrdeodhar akshayrdeodhar Nov 15, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

If the operands of the phi are identical, then the simplification is correct regardless of whether instructions belong to the same loop- I think there's no advantage to special-casing this.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Ping?

// Not identical, give up
CommonInst = nullptr;
break;
akshayrdeodhar marked this conversation as resolved.
Show resolved Hide resolved
}
} else if (IncomingInst) {
// Remember binary operator
CommonInst = IncomingInst;
} else {
// Not a binary operator, give up
akshayrdeodhar marked this conversation as resolved.
Show resolved Hide resolved
return nullptr;
}
}
return CommonInst ? getSCEV(CommonInst) : nullptr;
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This looks dangerous to me. Even if the instructions are the same, I think that SCEV would be allowed to use context-sensitive reasoning when constructing the SCEV node. It would be safer to call getSCEV for each incoming value in a second loop and make sure they're all the same.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Using a single loop with calls to getSCEV.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I expect that you do need both loops -- not for correctness, but as a profitability heuristic. Computing SCEVs is expensive, and always recursing through phis would likely add significant cost.

I tried to confirm this but the stage2 build crashes (https://llvm-compile-time-tracker.com/show_error.php?commit=a4e3a0e648d7a3664ca6269e846abf2e6ddcdb08). I didn't investigate, but it might be that the recursion causes a stack overflow.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Restored the original loop, and added an all_of check for SCEV exprs of the incoming values being identical.

}

const SCEV *ScalarEvolution::createNodeForPHI(PHINode *PN) {
if (const SCEV *S = createAddRecFromPHI(PN))
return S;
Expand All @@ -6030,6 +6061,9 @@ const SCEV *ScalarEvolution::createNodeForPHI(PHINode *PN) {
/*UseInstrInfo=*/true, /*CanUseUndef=*/false}))
return getSCEV(V);

if (const SCEV *S = createNodeForPHIWithIdenticalOperands(PN))
return S;

if (const SCEV *S = createNodeFromSelectLikePHI(PN))
return S;

Expand Down
46 changes: 46 additions & 0 deletions llvm/test/Analysis/ScalarEvolution/trip-count16.ll
akshayrdeodhar marked this conversation as resolved.
Show resolved Hide resolved
Original file line number Diff line number Diff line change
@@ -0,0 +1,46 @@
; NOTE: Assertions have been autogenerated by utils/update_analyze_test_checks.py UTC_ARGS: --version 5
; RUN: opt < %s -disable-output "-passes=print<scalar-evolution>" 2>&1 | FileCheck %s
define void @test(ptr %x, ptr %y) {
; CHECK-LABEL: 'test'
; CHECK-NEXT: Classifying expressions for: @test
; CHECK-NEXT: %v1.0 = phi i32 [ 0, %entry ], [ %k.0, %if.end ]
; CHECK-NEXT: --> {0,+,1}<nuw><nsw><%for.cond> U: [0,7) S: [0,7) Exits: 6 LoopDispositions: { %for.cond: Computable }
; CHECK-NEXT: %add = add nsw i32 %v1.0, 1
; CHECK-NEXT: --> {1,+,1}<nuw><nsw><%for.cond> U: [1,8) S: [1,8) Exits: 7 LoopDispositions: { %for.cond: Computable }
; CHECK-NEXT: %add6 = add nsw i32 %v1.0, 1
; CHECK-NEXT: --> {1,+,1}<nuw><nsw><%for.cond> U: [1,8) S: [1,8) Exits: 7 LoopDispositions: { %for.cond: Computable }
; CHECK-NEXT: %k.0 = phi i32 [ %add, %if.then ], [ %add6, %if.else ]
; CHECK-NEXT: --> {1,+,1}<nuw><nsw><%for.cond> U: [1,8) S: [1,8) Exits: 7 LoopDispositions: { %for.cond: Computable }
; CHECK-NEXT: Determining loop execution counts for: @test
; CHECK-NEXT: Loop %for.cond: backedge-taken count is i32 6
; CHECK-NEXT: Loop %for.cond: constant max backedge-taken count is i32 6
; CHECK-NEXT: Loop %for.cond: symbolic max backedge-taken count is i32 6
; CHECK-NEXT: Loop %for.cond: Trip multiple is 7
;
entry:
br label %for.cond

for.cond: ; preds = %6, %0
%v1.0 = phi i32 [ 0, %entry ], [ %k.0, %if.end ]
%cmp = icmp slt i32 %v1.0, 6
br i1 %cmp, label %for.body, label %exit

for.body: ; preds = %1
%cmp3 = icmp slt i32 %v1.0, 2
br i1 %cmp3, label %if.then, label %if.else

if.then: ; preds = %2
%add = add nsw i32 %v1.0, 1
br label %if.end

if.else: ; preds = %2
%add6 = add nsw i32 %v1.0, 1
br label %if.end

if.end: ; preds = %4, %3
%k.0 = phi i32 [ %add, %if.then ], [ %add6, %if.else ]
br label %for.cond

exit: ; preds = %5
ret void
}
Loading