[llvm] [DFAJT][profcheck] Propagate `select` -> `br` profile metadata (PR #162213)
Mircea Trofin via llvm-commits
llvm-commits at lists.llvm.org
Tue Oct 7 09:27:59 PDT 2025
https://github.com/mtrofin updated https://github.com/llvm/llvm-project/pull/162213
>From a4ddaf20aa928f7c5f71f4bcb45caee1f2e47330 Mon Sep 17 00:00:00 2001
From: Mircea Trofin <mtrofin at google.com>
Date: Mon, 6 Oct 2025 18:46:54 -0700
Subject: [PATCH] [DFAJT][profcheck] Propagate `select` -> `br` profile
metadata
---
llvm/lib/Transforms/Scalar/DFAJumpThreading.cpp | 15 +++++++++++++--
.../dfa-jump-threading-analysis.ll | 13 ++++++++++---
.../dfa-jump-threading-transform.ll | 15 +++++++++++----
llvm/utils/profcheck-xfail.txt | 5 -----
4 files changed, 34 insertions(+), 14 deletions(-)
diff --git a/llvm/lib/Transforms/Scalar/DFAJumpThreading.cpp b/llvm/lib/Transforms/Scalar/DFAJumpThreading.cpp
index 04ffec998c56d..dbcc3c90a6307 100644
--- a/llvm/lib/Transforms/Scalar/DFAJumpThreading.cpp
+++ b/llvm/lib/Transforms/Scalar/DFAJumpThreading.cpp
@@ -120,6 +120,9 @@ static cl::opt<unsigned>
CostThreshold("dfa-cost-threshold",
cl::desc("Maximum cost accepted for the transformation"),
cl::Hidden, cl::init(50));
+
+extern cl::opt<bool> ProfcheckDisableMetadataFixes;
+
} // namespace llvm
static cl::opt<double> MaxClonedRate(
@@ -262,7 +265,11 @@ void unfold(DomTreeUpdater *DTU, LoopInfo *LI, SelectInstToUnfold SIToUnfold,
// Insert the real conditional branch based on the original condition.
StartBlockTerm->eraseFromParent();
- BranchInst::Create(EndBlock, NewBlock, SI->getCondition(), StartBlock);
+ auto *BI =
+ BranchInst::Create(EndBlock, NewBlock, SI->getCondition(), StartBlock);
+ if (!ProfcheckDisableMetadataFixes)
+ BI->setMetadata(LLVMContext::MD_prof,
+ SI->getMetadata(LLVMContext::MD_prof));
DTU->applyUpdates({{DominatorTree::Insert, StartBlock, EndBlock},
{DominatorTree::Insert, StartBlock, NewBlock}});
} else {
@@ -297,7 +304,11 @@ void unfold(DomTreeUpdater *DTU, LoopInfo *LI, SelectInstToUnfold SIToUnfold,
// (Use)
BranchInst::Create(EndBlock, NewBlockF);
// Insert the real conditional branch based on the original condition.
- BranchInst::Create(EndBlock, NewBlockF, SI->getCondition(), NewBlockT);
+ auto *BI =
+ BranchInst::Create(EndBlock, NewBlockF, SI->getCondition(), NewBlockT);
+ if (!ProfcheckDisableMetadataFixes)
+ BI->setMetadata(LLVMContext::MD_prof,
+ SI->getMetadata(LLVMContext::MD_prof));
DTU->applyUpdates({{DominatorTree::Insert, NewBlockT, NewBlockF},
{DominatorTree::Insert, NewBlockT, EndBlock},
{DominatorTree::Insert, NewBlockF, EndBlock}});
diff --git a/llvm/test/Transforms/DFAJumpThreading/dfa-jump-threading-analysis.ll b/llvm/test/Transforms/DFAJumpThreading/dfa-jump-threading-analysis.ll
index e7b7dffa516c6..4173c32af37f1 100644
--- a/llvm/test/Transforms/DFAJumpThreading/dfa-jump-threading-analysis.ll
+++ b/llvm/test/Transforms/DFAJumpThreading/dfa-jump-threading-analysis.ll
@@ -1,11 +1,12 @@
; REQUIRES: asserts
; RUN: opt -S -passes=dfa-jump-threading -debug-only=dfa-jump-threading -disable-output %s 2>&1 | FileCheck %s
+; RUN: opt -S -passes=dfa-jump-threading -print-prof-data %s -o - | FileCheck %s --check-prefix=PROFILE
; This test checks that the analysis identifies all threadable paths in a
; simple CFG. A threadable path includes a list of basic blocks, the exit
; state, and the block that determines the next state.
; < path of BBs that form a cycle > [ state, determinator ]
-define i32 @test1(i32 %num) {
+define i32 @test1(i32 %num) !prof !0{
; CHECK: < case2 for.inc for.body > [ 1, for.inc ]
; CHECK-NEXT: < for.inc for.body > [ 1, for.inc ]
; CHECK-NEXT: < case1 for.inc for.body > [ 2, for.inc ]
@@ -25,8 +26,11 @@ case1:
br label %for.inc
case2:
+ ; PROFILE-LABEL: @test1
+ ; PROFILE-LABEL: case2:
+ ; PROFILE: br i1 %cmp, label %for.inc.jt1, label %sel.si.unfold.false.jt2, !prof !1 ; !1 = !{!"branch_weights", i32 3, i32 5}
%cmp = icmp eq i32 %count, 50
- %sel = select i1 %cmp, i32 1, i32 2
+ %sel = select i1 %cmp, i32 1, i32 2, !prof !1
br label %for.inc
for.inc:
@@ -182,7 +186,7 @@ bb66: ; preds = %bb59
}
; Value %init is not predictable but it's okay since it is the value initial to the switch.
-define i32 @initial.value.positive1(i32 %init) {
+define i32 @initial.value.positive1(i32 %init) !prof !0 {
; CHECK: < loop.1.backedge loop.1 loop.2 loop.3 > [ 1, loop.1 ]
; CHECK-NEXT: < case4 loop.1.backedge state.1.be2.si.unfold.false loop.1 loop.2 loop.3 > [ 2, loop.1.backedge ]
; CHECK-NEXT: < case2 loop.1.backedge state.1.be2.si.unfold.false loop.1 loop.2 loop.3 > [ 4, loop.1.backedge ]
@@ -241,3 +245,6 @@ infloop.i:
exit:
ret i32 0
}
+
+!0 = !{!"function_entry_count", i32 10}
+!1 = !{!"branch_weights", i32 3, i32 5}
diff --git a/llvm/test/Transforms/DFAJumpThreading/dfa-jump-threading-transform.ll b/llvm/test/Transforms/DFAJumpThreading/dfa-jump-threading-transform.ll
index ad0568486396f..092c854890462 100644
--- a/llvm/test/Transforms/DFAJumpThreading/dfa-jump-threading-transform.ll
+++ b/llvm/test/Transforms/DFAJumpThreading/dfa-jump-threading-transform.ll
@@ -1,4 +1,4 @@
-; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-globals
; RUN: opt -S -passes=dfa-jump-threading %s | FileCheck %s
; These tests check that the DFA jump threading transformation is applied
@@ -301,7 +301,7 @@ end:
ret void
}
-define void @pr106083_invalidBBarg_fold(i1 %cmp1, i1 %cmp2, i1 %not, ptr %d) {
+define void @pr106083_invalidBBarg_fold(i1 %cmp1, i1 %cmp2, i1 %not, ptr %d) !prof !0 {
; CHECK-LABEL: @pr106083_invalidBBarg_fold(
; CHECK-NEXT: bb:
; CHECK-NEXT: br label [[BB1:%.*]]
@@ -310,7 +310,7 @@ define void @pr106083_invalidBBarg_fold(i1 %cmp1, i1 %cmp2, i1 %not, ptr %d) {
; CHECK-NEXT: br i1 [[NOT:%.*]], label [[BB7_JT0]], label [[BB2:%.*]]
; CHECK: BB2:
; CHECK-NEXT: store i16 0, ptr [[D:%.*]], align 2
-; CHECK-NEXT: br i1 [[CMP2:%.*]], label [[BB7:%.*]], label [[SPEC_SELECT_SI_UNFOLD_FALSE_JT0:%.*]]
+; CHECK-NEXT: br i1 [[CMP2:%.*]], label [[BB7:%.*]], label [[SPEC_SELECT_SI_UNFOLD_FALSE_JT0:%.*]], !prof [[PROF1:![0-9]+]]
; CHECK: spec.select.si.unfold.false:
; CHECK-NEXT: br label [[BB9]]
; CHECK: spec.select.si.unfold.false.jt0:
@@ -357,7 +357,7 @@ BB1: ; preds = %BB1.backedge, %BB7,
BB2: ; preds = %BB1
store i16 0, ptr %d, align 2
- %spec.select = select i1 %cmp2, i32 %sel, i32 0
+ %spec.select = select i1 %cmp2, i32 %sel, i32 0, !prof !1
br label %BB7
BB7: ; preds = %BB2, %BB1
@@ -444,3 +444,10 @@ select.unfold: ; preds = %bb1, %.loopexit6
bb2: ; preds = %select.unfold
unreachable
}
+
+!0 = !{!"function_entry_count", i32 10}
+!1 = !{!"branch_weights", i32 3, i32 5}
+;.
+; CHECK: [[META0:![0-9]+]] = !{!"function_entry_count", i32 10}
+; CHECK: [[PROF1]] = !{!"branch_weights", i32 3, i32 5}
+;.
diff --git a/llvm/utils/profcheck-xfail.txt b/llvm/utils/profcheck-xfail.txt
index bbc8f59509997..74ed1724fd382 100644
--- a/llvm/utils/profcheck-xfail.txt
+++ b/llvm/utils/profcheck-xfail.txt
@@ -711,11 +711,6 @@ Transforms/CorrelatedValuePropagation/urem.ll
Transforms/CrossDSOCFI/basic.ll
Transforms/CrossDSOCFI/cfi_functions.ll
Transforms/CrossDSOCFI/thumb.ll
-Transforms/DFAJumpThreading/dfa-jump-threading-analysis.ll
-Transforms/DFAJumpThreading/dfa-jump-threading-transform.ll
-Transforms/DFAJumpThreading/dfa-unfold-select.ll
-Transforms/DFAJumpThreading/max-path-length.ll
-Transforms/DFAJumpThreading/negative.ll
Transforms/ExpandFp/AMDGPU/frem-inf.ll
Transforms/ExpandFp/AMDGPU/frem.ll
Transforms/ExpandLargeDivRem/X86/sdiv129.ll
More information about the llvm-commits
mailing list