[llvm] JumpThreading: Propagate branch weights in tryToUnfoldSelectInCurrBB (PR #66116)

[via llvm-commits]

llvmbot wrote:

@llvm/pr-subscribers-llvm-transforms

<details>
<summary>Changes</summary>

Propagate "branch_weights" metadata whe turning a select into a conditional branch in tryToUnfoldSelectInCurrBB
--
Full diff: https://github.com/llvm/llvm-project/pull/66116.diff

2 Files Affected:

- (modified) llvm/lib/Transforms/Scalar/JumpThreading.cpp (+3-1) 
- (added) llvm/test/Transforms/JumpThreading/thread-prob-7.ll (+56) 


<pre>
diff --git a/llvm/lib/Transforms/Scalar/JumpThreading.cpp b/llvm/lib/Transforms/Scalar/JumpThreading.cpp
index 54402256e2cb644..6c1a1db279250f9 100644
--- a/llvm/lib/Transforms/Scalar/JumpThreading.cpp
+++ b/llvm/lib/Transforms/Scalar/JumpThreading.cpp
@@ -2926,7 +2926,9 @@ bool JumpThreadingPass::tryToUnfoldSelectInCurrBB(BasicBlock *BB) {
     Value *Cond = SI->getCondition();
     if (!isGuaranteedNotToBeUndefOrPoison(Cond, nullptr, SI))
       Cond = new FreezeInst(Cond, "cond.fr", SI);
-    Instruction *Term = SplitBlockAndInsertIfThen(Cond, SI, false);
+    MDNode *BranchWeights = getBranchWeightMDNode(*SI);
+    Instruction *Term =
+        SplitBlockAndInsertIfThen(Cond, SI, false, BranchWeights);
     BasicBlock *SplitBB = SI->getParent();
     BasicBlock *NewBB = Term->getParent();
     PHINode *NewPN = PHINode::Create(SI->getType(), 2, "", SI);
diff --git a/llvm/test/Transforms/JumpThreading/thread-prob-7.ll b/llvm/test/Transforms/JumpThreading/thread-prob-7.ll
new file mode 100644
index 000000000000000..f11bfd026688192
--- /dev/null
+++ b/llvm/test/Transforms/JumpThreading/thread-prob-7.ll
@@ -0,0 +1,56 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-globals
+; RUN: opt -S -passes="jump-threading" < %s 2>&1 | FileCheck %s
+
+declare void @foobar()
+
+define i32 @func0(i32 %a0, i32 %a1) !prof !0 {
+; CHECK-LABEL: @func0(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[CMP0:%.*]] = icmp ne i32 [[A0:%.*]], 0
+; CHECK-NEXT:    [[CX:%.*]] = icmp eq i32 [[A0]], 42
+; CHECK-NEXT:    br i1 [[CMP0]], label [[BB_JOIN:%.*]], label [[TEST2:%.*]], !prof [[PROF1:![0-9]+]]
+; CHECK:       test2:
+; CHECK-NEXT:    [[CMP1:%.*]] = icmp ne i32 [[A1:%.*]], 0
+; CHECK-NEXT:    br i1 [[CMP1]], label [[BB_JOIN_THREAD:%.*]], label [[TEST2_FALSE:%.*]], !prof [[PROF2:![0-9]+]]
+; CHECK:       test2_false:
+; CHECK-NEXT:    call void @foobar()
+; CHECK-NEXT:    br label [[TMP0:%.*]]
+; CHECK:       bb_join:
+; CHECK-NEXT:    [[C:%.*]] = phi i1 [ [[CX]], [[ENTRY:%.*]] ]
+; CHECK-NEXT:    [[COND_FR:%.*]] = freeze i1 [[C]]
+; CHECK-NEXT:    br i1 [[COND_FR]], label [[BB_JOIN_THREAD]], label [[TMP0]], !prof [[PROF3:![0-9]+]]
+; CHECK:       bb_join.thread:
+; CHECK-NEXT:    br label [[TMP0]]
+; CHECK:       0:
+; CHECK-NEXT:    [[TMP1:%.*]] = phi i32 [ 42, [[BB_JOIN_THREAD]] ], [ 7, [[BB_JOIN]] ], [ 7, [[TEST2_FALSE]] ]
+; CHECK-NEXT:    ret i32 [[TMP1]]
+;
+entry:
+  %cmp0 = icmp ne i32 %a0, 0
+  %cx = icmp eq i32 %a0, 42
+  br i1 %cmp0, label %bb_join, label %test2, !prof !1
+
+test2:
+  %cmp1 = icmp ne i32 %a1, 0
+  br i1 %cmp1, label %bb_join, label %test2_false, !prof !2
+
+test2_false:
+  call void @foobar()
+  br label %bb_join
+
+bb_join:
+  %c = phi i1 [%cx, %entry], [true, %test2], [%cx, %test2_false]
+  %val = select i1 %c, i32 42, i32 7, !prof !3
+  ret i32 %val
+}
+
+!0 = !{!"function_entry_count", i64 1000}
+!1 = !{!"branch_weights", i32 400, i32 600}
+!2 = !{!"branch_weights", i32 300, i32 300}
+!3 = !{!"branch_weights", i32 500, i32 500}
+;.
+; CHECK: [[META0:![0-9]+]] = !{!"function_entry_count", i64 1000}
+; CHECK: [[PROF1]] = !{!"branch_weights", i32 400, i32 600}
+; CHECK: [[PROF2]] = !{!"branch_weights", i32 300, i32 300}
+; CHECK: [[PROF3]] = !{!"branch_weights", i32 678152731, i32 1469330917}
+;.
</pre>

</details>

https://github.com/llvm/llvm-project/pull/66116