Skip to content

[LP] Assign weights when peeling last iteration. #166858

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.

Sign up for GitHub

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
wants to merge 1 commit into
base: main
Choose a base branch
from
+121 −16
Open

[LP] Assign weights when peeling last iteration. #166858

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
23 changes: 22 additions & 1 deletion llvm/lib/Transforms/Utils/LoopPeel.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -54,6 +54,7 @@ using namespace llvm::SCEVPatternMatch;
STATISTIC(NumPeeled, "Number of loops peeled");
STATISTIC(NumPeeledEnd, "Number of loops peeled from end");

namespace llvm {
static cl::opt<unsigned> UnrollPeelCount(
"unroll-peel-count", cl::Hidden,
cl::desc("Set the unroll peeling count, for testing purposes"));
Expand Down Expand Up @@ -87,6 +88,9 @@ static cl::opt<bool> EnablePeelingForIV(

static const char *PeeledCountMetaData = "llvm.loop.peeled.count";

extern cl::opt<bool> ProfcheckDisableMetadataFixes;
} // namespace llvm

// Check whether we are capable of peeling this loop.
bool llvm::canPeel(const Loop *L) {
// Make sure the loop is in simplified form
Expand Down Expand Up @@ -1190,7 +1194,24 @@ bool llvm::peelLoop(Loop *L, unsigned PeelCount, bool PeelLast, LoopInfo *LI,
IRBuilder<> B(PreHeaderBR);
Value *Cond =
B.CreateICmpNE(BTCValue, ConstantInt::get(BTCValue->getType(), 0));
B.CreateCondBr(Cond, NewPreHeader, InsertTop);
auto *BI = B.CreateCondBr(Cond, NewPreHeader, InsertTop);
SmallVector<uint32_t> Weights;
auto *OrigLatchBr = Latch->getTerminator();
auto HasBranchWeights = !ProfcheckDisableMetadataFixes &&
extractBranchWeights(*OrigLatchBr, Weights);
if (HasBranchWeights) {
// The probability that the new guard skips the loop to execute just one
// iteration is the original loop's probability of exiting at the latch
// after any iteration. That should maintain the original loop body
// frequency. Upon arriving at the loop, due to the guard, the
// probability of reaching iteration i of the new loop is the
// probability of reaching iteration i+1 of the original loop. The
// probability of reaching the peeled iteration is 1, which is the
// probability of reaching iteration 0 of the original loop.
if (L->getExitBlock() == OrigLatchBr->getSuccessor(0))
std::swap(Weights[0], Weights[1]);
setBranchWeights(*BI, Weights, /*IsExpected=*/false);
}
PreHeaderBR->eraseFromParent();

// PreHeader now dominates InsertTop.
Expand Down
66 changes: 66 additions & 0 deletions llvm/test/Transforms/LoopUnroll/peel-last-iteration-bfi.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,66 @@
; Disable this test in profcheck because the first run would cause profcheck to fail.
; REQUIRES: !profcheck
; RUN: opt -p "print<block-freq>,loop-unroll,print<block-freq>" -scev-cheap-expansion-budget=3 -S %s -profcheck-disable-metadata-fixes 2>&1 | FileCheck %s --check-prefixes=COMMON,BAD
Copy link
Collaborator

@jdenny-ornl jdenny-ornl Nov 14, 2025

Choose a reason for hiding this comment

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

Sorry, remind me again: The BAD version goes away at some point in the future, right?

Copy link
Member Author

@mtrofin mtrofin Nov 15, 2025

Choose a reason for hiding this comment

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

yes, and could remove it now, too, but seemed like easy to show the difference. I'll just remove it at that later date.

; RUN: opt -p "print<block-freq>,loop-unroll,print<block-freq>" -scev-cheap-expansion-budget=3 -S %s 2>&1 | FileCheck %s --check-prefixes=COMMON,GOOD

define i32 @test_expansion_cost_2(i32 %start, i32 %end) !prof !0 {
entry:
%sub = add i32 %end, -1
br label %loop.header

loop.header:
%iv = phi i32 [ %start, %entry ], [ %iv.next, %loop.latch ]
%c = icmp eq i32 %iv, %sub
br i1 %c, label %then, label %loop.latch, !prof !1

then:
br label %loop.latch

loop.latch:
%iv.next = add nsw i32 %iv, 1
%ec = icmp eq i32 %iv.next, %end
br i1 %ec, label %exit, label %loop.header, !prof !2

exit:
ret i32 0
}

!0 = !{!"function_entry_count", i32 10}
!1 = !{!"branch_weights", i32 2, i32 3}
!2 = !{!"branch_weights", i32 1, i32 50}

; COMMON: block-frequency-info: test_expansion_cost_2
; COMMON-NEXT: entry: float = 1.0
; COMMON-NEXT: loop.header: float = 51.0
; COMMON-NEXT: then: float = 20.4
; COMMON-NEXT: loop.latch: float = 51.0
; COMMON-NEXT: exit: float = 1.0

; COMMON: block-frequency-info: test_expansion_cost_2
; GOOD-NEXT: entry: float = 1.0
; GOOD-NEXT: entry.split: float = 0.98039
; GOOD-NEXT: loop.header: float = 50.0
; GOOD-NEXT: then: float = 20.0
; GOOD-NEXT: loop.latch: float = 50.0
; GOOD-NEXT: exit.peel.begin.loopexit: float = 0.98039
; GOOD-NEXT: exit.peel.begin: float = 1.0
; GOOD-NEXT: loop.header.peel: float = 1.0
; GOOD-NEXT: then.peel: float = 0.4
; GOOD-NEXT: loop.latch.peel: float = 1.0
; GOOD-NEXT: exit.peel.next: float = 1.0
; GOOD-NEXT: loop.header.peel.next: float = 1.0
; GOOD-NEXT: exit: float = 1.0

; BAD-NEXT: entry: float = 1.0
; BAD-NEXT: entry.split: float = 0.625
; BAD-NEXT: loop.header: float = 31.875
; BAD-NEXT: then: float = 12.75
; BAD-NEXT: loop.latch: float = 31.875
; BAD-NEXT: exit.peel.begin.loopexit: float = 0.625
; BAD-NEXT: exit.peel.begin: float = 1.0
; BAD-NEXT: loop.header.peel: float = 1.0
; BAD-NEXT: then.peel: float = 0.4
; BAD-NEXT: loop.latch.peel: float = 1.0
; BAD-NEXT: exit.peel.next: float = 1.0
; BAD-NEXT: loop.header.peel.next: float = 1.0
; BAD-NEXT: exit: float = 1.0
47 changes: 32 additions & 15 deletions llvm/test/Transforms/LoopUnroll/peel-last-iteration-expansion-cost.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,46 +1,46 @@
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-globals all --version 5
; RUN: opt -p loop-unroll -scev-cheap-expansion-budget=2 -S %s | FileCheck --check-prefix=BUDGET2 %s
; RUN: opt -p loop-unroll -scev-cheap-expansion-budget=3 -S %s | FileCheck --check-prefix=BUDGET3 %s

define i32 @test_expansion_cost_2(i32 %start, i32 %end) {
define i32 @test_expansion_cost_2(i32 %start, i32 %end) !prof !0 {
; BUDGET2-LABEL: define i32 @test_expansion_cost_2(
; BUDGET2-SAME: i32 [[START:%.*]], i32 [[END:%.*]]) {
; BUDGET2-SAME: i32 [[START:%.*]], i32 [[END:%.*]]) !prof [[PROF0:![0-9]+]] {
; BUDGET2-NEXT: [[ENTRY:.*]]:
; BUDGET2-NEXT: [[SUB:%.*]] = add i32 [[END]], -1
; BUDGET2-NEXT: br label %[[LOOP_HEADER:.*]]
; BUDGET2: [[LOOP_HEADER]]:
; BUDGET2-NEXT: [[IV:%.*]] = phi i32 [ [[START]], %[[ENTRY]] ], [ [[IV_NEXT:%.*]], %[[LOOP_LATCH:.*]] ]
; BUDGET2-NEXT: [[C:%.*]] = icmp eq i32 [[IV]], [[SUB]]
; BUDGET2-NEXT: br i1 [[C]], label %[[THEN:.*]], label %[[LOOP_LATCH]]
; BUDGET2-NEXT: br i1 [[C]], label %[[THEN:.*]], label %[[LOOP_LATCH]], !prof [[PROF1:![0-9]+]]
; BUDGET2: [[THEN]]:
; BUDGET2-NEXT: br label %[[LOOP_LATCH]]
; BUDGET2: [[LOOP_LATCH]]:
; BUDGET2-NEXT: [[IV_NEXT]] = add nsw i32 [[IV]], 1
; BUDGET2-NEXT: [[EC:%.*]] = icmp eq i32 [[IV_NEXT]], [[END]]
; BUDGET2-NEXT: br i1 [[EC]], label %[[EXIT:.*]], label %[[LOOP_HEADER]]
; BUDGET2-NEXT: br i1 [[EC]], label %[[EXIT:.*]], label %[[LOOP_HEADER]], !prof [[PROF2:![0-9]+]], !llvm.loop [[LOOP3:![0-9]+]]
; BUDGET2: [[EXIT]]:
; BUDGET2-NEXT: ret i32 0
;
; BUDGET3-LABEL: define i32 @test_expansion_cost_2(
; BUDGET3-SAME: i32 [[START:%.*]], i32 [[END:%.*]]) {
; BUDGET3-SAME: i32 [[START:%.*]], i32 [[END:%.*]]) !prof [[PROF0:![0-9]+]] {
; BUDGET3-NEXT: [[ENTRY:.*]]:
; BUDGET3-NEXT: [[SUB:%.*]] = add i32 [[END]], -1
; BUDGET3-NEXT: [[TMP0:%.*]] = sub i32 [[SUB]], [[START]]
; BUDGET3-NEXT: [[TMP1:%.*]] = icmp ne i32 [[TMP0]], 0
; BUDGET3-NEXT: br i1 [[TMP1]], label %[[ENTRY_SPLIT:.*]], label %[[EXIT_PEEL_BEGIN:.*]]
; BUDGET3-NEXT: br i1 [[TMP1]], label %[[ENTRY_SPLIT:.*]], label %[[EXIT_PEEL_BEGIN:.*]], !prof [[PROF1:![0-9]+]]
; BUDGET3: [[ENTRY_SPLIT]]:
; BUDGET3-NEXT: br label %[[LOOP_HEADER:.*]]
; BUDGET3: [[LOOP_HEADER]]:
; BUDGET3-NEXT: [[IV:%.*]] = phi i32 [ [[START]], %[[ENTRY_SPLIT]] ], [ [[IV_NEXT:%.*]], %[[LOOP_LATCH:.*]] ]
; BUDGET3-NEXT: [[C:%.*]] = icmp eq i32 [[IV]], [[SUB]]
; BUDGET3-NEXT: br i1 [[C]], label %[[THEN:.*]], label %[[LOOP_LATCH]]
; BUDGET3-NEXT: br i1 [[C]], label %[[THEN:.*]], label %[[LOOP_LATCH]], !prof [[PROF2:![0-9]+]]
; BUDGET3: [[THEN]]:
; BUDGET3-NEXT: br label %[[LOOP_LATCH]]
; BUDGET3: [[LOOP_LATCH]]:
; BUDGET3-NEXT: [[IV_NEXT]] = add nsw i32 [[IV]], 1
; BUDGET3-NEXT: [[TMP2:%.*]] = sub i32 [[END]], 1
; BUDGET3-NEXT: [[EC:%.*]] = icmp eq i32 [[IV_NEXT]], [[TMP2]]
; BUDGET3-NEXT: br i1 [[EC]], label %[[EXIT_PEEL_BEGIN_LOOPEXIT:.*]], label %[[LOOP_HEADER]], !llvm.loop [[LOOP0:![0-9]+]]
; BUDGET3-NEXT: br i1 [[EC]], label %[[EXIT_PEEL_BEGIN_LOOPEXIT:.*]], label %[[LOOP_HEADER]], !prof [[PROF3:![0-9]+]], !llvm.loop [[LOOP4:![0-9]+]]
; BUDGET3: [[EXIT_PEEL_BEGIN_LOOPEXIT]]:
; BUDGET3-NEXT: [[DOTPH:%.*]] = phi i32 [ [[IV_NEXT]], %[[LOOP_LATCH]] ]
; BUDGET3-NEXT: br label %[[EXIT_PEEL_BEGIN]]
Expand All @@ -49,13 +49,13 @@ define i32 @test_expansion_cost_2(i32 %start, i32 %end) {
; BUDGET3-NEXT: br label %[[LOOP_HEADER_PEEL:.*]]
; BUDGET3: [[LOOP_HEADER_PEEL]]:
; BUDGET3-NEXT: [[C_PEEL:%.*]] = icmp eq i32 [[TMP3]], [[SUB]]
; BUDGET3-NEXT: br i1 [[C_PEEL]], label %[[THEN_PEEL:.*]], label %[[LOOP_LATCH_PEEL:.*]]
; BUDGET3-NEXT: br i1 [[C_PEEL]], label %[[THEN_PEEL:.*]], label %[[LOOP_LATCH_PEEL:.*]], !prof [[PROF2]]
; BUDGET3: [[THEN_PEEL]]:
; BUDGET3-NEXT: br label %[[LOOP_LATCH_PEEL]]
; BUDGET3: [[LOOP_LATCH_PEEL]]:
; BUDGET3-NEXT: [[IV_NEXT_PEEL:%.*]] = add nsw i32 [[TMP3]], 1
; BUDGET3-NEXT: [[EC_PEEL:%.*]] = icmp eq i32 [[IV_NEXT_PEEL]], [[END]]
; BUDGET3-NEXT: br i1 [[EC_PEEL]], label %[[EXIT_PEEL_NEXT:.*]], label %[[EXIT_PEEL_NEXT]]
; BUDGET3-NEXT: br i1 [[EC_PEEL]], label %[[EXIT_PEEL_NEXT:.*]], label %[[EXIT_PEEL_NEXT]], !prof [[PROF3]]
Copy link
Collaborator

@jdenny-ornl jdenny-ornl Nov 14, 2025

Choose a reason for hiding this comment

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

Given that both targets are the same, this becomes an unconditional exit, right? I just don't want that !prof influencing BFI somehow. Your BFI test suggests all is fine.

Copy link
Member Author

@mtrofin mtrofin Nov 15, 2025

Choose a reason for hiding this comment

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

Yup, and I think this is this way here because of some pass manager constraints.

; BUDGET3: [[EXIT_PEEL_NEXT]]:
; BUDGET3-NEXT: br label %[[LOOP_HEADER_PEEL_NEXT:.*]]
; BUDGET3: [[LOOP_HEADER_PEEL_NEXT]]:
Expand All @@ -70,20 +70,37 @@ entry:
loop.header:
%iv = phi i32 [ %start, %entry ], [ %iv.next, %loop.latch ]
%c = icmp eq i32 %iv, %sub
br i1 %c, label %then, label %loop.latch
br i1 %c, label %then, label %loop.latch, !prof !1

then:
br label %loop.latch

loop.latch:
%iv.next = add nsw i32 %iv, 1
%ec = icmp eq i32 %iv.next, %end
br i1 %ec, label %exit, label %loop.header
br i1 %ec, label %exit, label %loop.header, !prof !2, !llvm.loop !3

exit:
ret i32 0
}

!0 = !{!"function_entry_count", i32 10}
!1 = !{!"branch_weights", i32 2, i32 3}
!2 = !{!"branch_weights", i32 1, i32 10}
!3 = distinct !{!3, !4}
!4 = !{!"llvm.loop.estimated_trip_count", i32 11}
Copy link
Member Author

@mtrofin mtrofin Nov 15, 2025

Choose a reason for hiding this comment

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

@jdenny-ornl we could also set a value that's not derivable from the current profile, like 15. wdyt?

;.
; BUDGET2: [[PROF0]] = !{!"function_entry_count", i32 10}
; BUDGET2: [[PROF1]] = !{!"branch_weights", i32 2, i32 3}
; BUDGET2: [[PROF2]] = !{!"branch_weights", i32 1, i32 10}
; BUDGET2: [[LOOP3]] = distinct !{[[LOOP3]], [[META4:![0-9]+]]}
; BUDGET2: [[META4]] = !{!"llvm.loop.estimated_trip_count", i32 11}
;.
; BUDGET3: [[LOOP0]] = distinct !{[[LOOP0]], [[META1:![0-9]+]]}
; BUDGET3: [[META1]] = !{!"llvm.loop.peeled.count", i32 1}
; BUDGET3: [[PROF0]] = !{!"function_entry_count", i32 10}
; BUDGET3: [[PROF1]] = !{!"branch_weights", i32 10, i32 1}
; BUDGET3: [[PROF2]] = !{!"branch_weights", i32 2, i32 3}
; BUDGET3: [[PROF3]] = !{!"branch_weights", i32 1, i32 10}
; BUDGET3: [[LOOP4]] = distinct !{[[LOOP4]], [[META5:![0-9]+]], [[META6:![0-9]+]]}
; BUDGET3: [[META5]] = !{!"llvm.loop.peeled.count", i32 1}
; BUDGET3: [[META6]] = !{!"llvm.loop.estimated_trip_count", i32 10}
;.
1 change: 1 addition & 0 deletions llvm/test/lit.cfg.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -49,6 +49,7 @@
config.excludes = ["Inputs", "CMakeLists.txt", "README.txt", "LICENSE.txt"]

if config.enable_profcheck:
config.available_features.add("profcheck")
# Exclude llvm-reduce tests for profcheck because we substitute the FileCheck
# binary with a no-op command for profcheck, but llvm-reduce tests have RUN
# commands of the form llvm-reduce --test FileCheck, which explode if we
Expand Down
Loading