[Analysis] Ignore convergence tokens in dead branches in `CodeMetrics` by inbelic · Pull Request #193590 · llvm/llvm-project

inbelic · 2026-04-22T20:52:00Z

In extendsConvergenceOutsideLoop, skip users whose parent block is terminated by unreachable and which are statically known to not branch to anymore, as these represent dead branches. This prevents convergence control tokens in dead code from incorrectly marking a loop as having extended convergence, which could block loop unrolling.

This was discovered when #188792 caused the following failure: https://github.com/llvm/llvm-project/actions/runs/24577221310/job/71865579618. When emitting convergence control tokens, the inner loop successfully unrolls and marks the old body branch as unreachable, it then tries to unroll the outer loop, however the convergence.loop token in the outer loop header sees that a convergent op (in this case dx.resource.getpointer) in the dead branch is using it and marks it as Convergence::Extended, preventing a loop unroll from occurring. In this case it should still be possible to unroll.

Assisted by: Claude Opus 4.6

llvmbot · 2026-04-22T21:41:41Z

@llvm/pr-subscribers-llvm-transforms

@llvm/pr-subscribers-llvm-analysis

Author: Finn Plummer (inbelic)

Changes

In extendsConvergenceOutsideLoop, skip users whose parent block is terminated by unreachable, as these represent dead branches. This prevents convergence control tokens in dead code from incorrectly marking a loop as having extended convergence, which could block loop unrolling.

Assisted by: Claude Opus 4.6

Full diff: https://github.com/llvm/llvm-project/pull/193590.diff

2 Files Affected:

(modified) llvm/lib/Analysis/CodeMetrics.cpp (+7-1)
(modified) llvm/test/Transforms/LoopUnroll/convergent.controlled.ll (+58)

diff --git a/llvm/lib/Analysis/CodeMetrics.cpp b/llvm/lib/Analysis/CodeMetrics.cpp
index ea67b526423bf..21f5ff7319ceb 100644
--- a/llvm/lib/Analysis/CodeMetrics.cpp
+++ b/llvm/lib/Analysis/CodeMetrics.cpp
@@ -16,6 +16,7 @@
 #include "llvm/Analysis/LoopInfo.h"
 #include "llvm/Analysis/TargetTransformInfo.h"
 #include "llvm/IR/Function.h"
+#include "llvm/IR/Instructions.h"
 #include "llvm/IR/IntrinsicInst.h"
 #include "llvm/Support/Debug.h"
 #include "llvm/Support/InstructionCost.h"
@@ -118,7 +119,12 @@ static bool extendsConvergenceOutsideLoop(const Instruction &I, const Loop *L) {
   if (!isa<ConvergenceControlInst>(I))
     return false;
   for (const auto *U : I.users()) {
-    if (!L->contains(cast<Instruction>(U)))
+    const auto *UserInst = cast<Instruction>(U);
+    // Ignore users in dead branches, identified by blocks terminated with
+    // unreachable.
+    if (isa<UnreachableInst>(UserInst->getParent()->getTerminator()))
+      continue;
+    if (!L->contains(UserInst))
       return true;
   }
   return false;
diff --git a/llvm/test/Transforms/LoopUnroll/convergent.controlled.ll b/llvm/test/Transforms/LoopUnroll/convergent.controlled.ll
index 5dc613e733f00..6836aea1ff61f 100644
--- a/llvm/test/Transforms/LoopUnroll/convergent.controlled.ll
+++ b/llvm/test/Transforms/LoopUnroll/convergent.controlled.ll
@@ -555,6 +555,64 @@ exit:
   ret i32 0
 }
 
+; A convergence token defined in the loop is used outside the loop, but only
+; in a dead branch (block terminated by unreachable). This should not prevent
+; unrolling.
+define i32 @extended_loop_dead_branch(i32 %n) {
+; CHECK-LABEL: @extended_loop_dead_branch(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = freeze i32 [[N:%.*]]
+; CHECK-NEXT:    [[TMP1:%.*]] = add i32 [[TMP0]], -1
+; CHECK-NEXT:    [[XTRAITER:%.*]] = and i32 [[TMP0]], 1
+; CHECK-NEXT:    [[TMP2:%.*]] = icmp ult i32 [[TMP1]], 1
+; CHECK-NEXT:    br i1 [[TMP2]], label [[L3_EPIL_PREHEADER:%.*]], label [[ENTRY_NEW:%.*]]
+; CHECK:       entry.new:
+; CHECK-NEXT:    [[UNROLL_ITER:%.*]] = sub i32 [[TMP0]], [[XTRAITER]]
+; CHECK-NEXT:    br label [[L3:%.*]], !llvm.loop [[LOOP4]]
+; CHECK:       l3:
+; CHECK-NEXT:    [[X_0:%.*]] = phi i32 [ 0, [[ENTRY_NEW]] ], [ [[INC_1:%.*]], [[L3]] ]
+; CHECK-NEXT:    [[NITER:%.*]] = phi i32 [ 0, [[ENTRY_NEW]] ], [ [[NITER_NEXT_1:%.*]], [[L3]] ]
+; CHECK-NEXT:    [[TOK_LOOP:%.*]] = call token @llvm.experimental.convergence.anchor()
+; CHECK-NEXT:    call void @f() [ "convergencectrl"(token [[TOK_LOOP]]) ]
+; CHECK-NEXT:    [[TOK_LOOP_1:%.*]] = call token @llvm.experimental.convergence.anchor()
+; CHECK-NEXT:    call void @f() [ "convergencectrl"(token [[TOK_LOOP_1]]) ]
+; CHECK-NEXT:    [[INC_1]] = add nsw i32 [[X_0]], 2
+; CHECK-NEXT:    [[NITER_NEXT_1]] = add i32 [[NITER]], 2
+; CHECK-NEXT:    [[NITER_NCMP_1:%.*]] = icmp eq i32 [[NITER_NEXT_1]], [[UNROLL_ITER]]
+; CHECK-NEXT:    br i1 [[NITER_NCMP_1]], label [[EXIT_UNR_LCSSA:%.*]], label [[L3]], !llvm.loop [[LOOP10:![0-9]+]]
+; CHECK:       exit.unr-lcssa:
+; CHECK-NEXT:    [[LCMP_MOD:%.*]] = icmp ne i32 [[XTRAITER]], 0
+; CHECK-NEXT:    br i1 [[LCMP_MOD]], label [[L3_EPIL_PREHEADER]], label [[EXIT:%.*]]
+; CHECK:       l3.epil.preheader:
+; CHECK-NEXT:    [[LCMP_MOD1:%.*]] = icmp ne i32 [[XTRAITER]], 0
+; CHECK-NEXT:    call void @llvm.assume(i1 [[LCMP_MOD1]])
+; CHECK-NEXT:    br label [[L3_EPIL:%.*]]
+; CHECK:       l3.epil:
+; CHECK-NEXT:    [[TOK_LOOP_EPIL:%.*]] = call token @llvm.experimental.convergence.anchor()
+; CHECK-NEXT:    call void @f() [ "convergencectrl"(token [[TOK_LOOP_EPIL]]) ]
+; CHECK-NEXT:    br label [[EXIT]]
+; CHECK:       exit:
+; CHECK-NEXT:    ret i32 0
+;
+entry:
+  br label %l3, !llvm.loop !1
+
+l3:
+  %x.0 = phi i32 [ 0, %entry ], [ %inc, %l3 ]
+  %tok.loop = call token @llvm.experimental.convergence.anchor()
+  call void @f() [ "convergencectrl"(token %tok.loop) ]
+  %inc = add nsw i32 %x.0, 1
+  %exitcond = icmp eq i32 %inc, %n
+  br i1 %exitcond, label %exit, label %l3, !llvm.loop !1
+
+exit:
+  ret i32 0
+
+dead:
+  call void @f() [ "convergencectrl"(token %tok.loop) ]
+  unreachable
+}
+
 declare token @llvm.experimental.convergence.anchor()
 declare token @llvm.experimental.convergence.loop()

In `extendsConvergenceOutsideLoop`, skip users whose parent block is terminated by unreachable, as these represent dead branches. This prevents convergence control tokens in dead code from incorrectly marking a loop as having extended convergence, which could block loop unrolling.

bob80905 · 2026-04-30T18:37:58Z

 }

+/// Check if a block was previously marked dead by setting the terminator to
+/// `unreachable` and the condiational branch to the block is statically


nit: conditional

also perhaps a more accurate description:
"... and no predecessors exist with statically evaluatable terminators that branch to this block"

bob80905 · 2026-04-30T18:52:12Z

+  for (const BasicBlock *Pred : predecessors(BB)) {
+    auto *CondBr = dyn_cast<CondBrInst>(Pred->getTerminator());
+    if (!CondBr)
+      return false;


Are there other types of control flow instructions, other than condBrInst, that we care about that could switch control flow to this block?

The other kind would be an UncondBrInst, which if that is a predessor then this block isn't dead and we should return false

edit: (or as Deric pointed out offline a switch instruction)

inbelic · 2026-05-05T16:48:33Z

+/// Check if a block was previously marked dead by setting the terminator to
+/// `unreachable` and their is a statically evaluated conditional branch to NOT
+/// branch to the block. This is done for instance within the unroll pass,
+/// between unrolling inner/outer loops.


Suggested change

/// Check if a block was previously marked dead by setting the terminator to

/// `unreachable` and their is a statically evaluated conditional branch to NOT

/// branch to the block. This is done for instance within the unroll pass,

/// between unrolling inner/outer loops.

// Check if a block was previously marked dead by setting the terminator to

// `unreachable` and there are no branches that may branch to the block. We

// can only ensure a predecessor doesn't branch to this block if it is statically

// known not to. This is done for instance within the unroll pass, between

// unrolling inner/outer loops.

(removed the extra comment line as static functions shouldn't generate a doc)

bob80905 · 2026-05-12T21:19:19Z

 }

+/// Check if a block was previously marked dead by setting the terminator to
+/// `unreachable` and their is a statically evaluated conditional branch to NOT


nit: This comment could be slightly improved in accuracy and grammar

…g DirectX" (#194452) The initial landing surfaced 3 somewhat orthogonal issues related to loop unrolling. These are addressed: [here](#193592), [here](#193593) and [here](#193590). These caused these [tests](https://github.com/llvm/llvm-project/actions/runs/24577221310/job/71865579618#step:8:87913) to fail in the offload test suite. We can verify that these are now passing as expected (fixing any of the 3 issues would resolve this and allow us to reland) Some additional tests were added since the revert that are now accounted for and updated in the reland fixes commit. This relands #188792

…en targeting DirectX" (#194452) The initial landing surfaced 3 somewhat orthogonal issues related to loop unrolling. These are addressed: [here](llvm/llvm-project#193592), [here](llvm/llvm-project#193593) and [here](llvm/llvm-project#193590). These caused these [tests](https://github.com/llvm/llvm-project/actions/runs/24577221310/job/71865579618#step:8:87913) to fail in the offload test suite. We can verify that these are now passing as expected (fixing any of the 3 issues would resolve this and allow us to reland) Some additional tests were added since the revert that are now accounted for and updated in the reland fixes commit. This relands llvm/llvm-project#188792

inbelic changed the title ~~[Analysis] Ignore convergence tokens in dead branches in CodeMetrics~~ [Analysis] Ignore convergence tokens in dead branches in CodeMetrics Apr 22, 2026

inbelic commented Apr 22, 2026

View reviewed changes

Comment thread llvm/lib/Analysis/CodeMetrics.cpp Outdated

inbelic marked this pull request as ready for review April 22, 2026 21:41

inbelic requested a review from Keenuts April 22, 2026 21:41

llvmbot added llvm:analysis Includes value tracking, cost tables and constant folding llvm:transforms labels Apr 22, 2026

Icohedron reviewed Apr 23, 2026

View reviewed changes

Comment thread llvm/test/Transforms/LoopUnroll/convergent.controlled.ll

inbelic added 2 commits April 27, 2026 18:31

review: make dead branch check more strict and reflect test case changes

edcb29d

inbelic force-pushed the inbelic/unroll-conv branch from ccfea43 to edcb29d Compare April 27, 2026 20:08

Icohedron approved these changes Apr 27, 2026

View reviewed changes

Icohedron reviewed Apr 27, 2026

View reviewed changes

Comment thread llvm/lib/Analysis/CodeMetrics.cpp Outdated

Comment thread llvm/test/Transforms/LoopUnroll/convergent.controlled.ll

Comment thread llvm/lib/Analysis/CodeMetrics.cpp

inbelic added 3 commits April 29, 2026 17:38

review: add negative test

0ba850c

review: sanity check for entry

fa35fc2

review: typo

4f16e00

inbelic mentioned this pull request Apr 29, 2026

Reland "[HLSL][DirectX] Emit convergence control tokens when targeting DirectX" #194452

Merged

Icohedron reviewed Apr 29, 2026

View reviewed changes

Comment thread llvm/lib/Analysis/CodeMetrics.cpp Outdated

review: clarify function

0570ce9

bob80905 reviewed Apr 30, 2026

View reviewed changes

review: comment feedback

739b5e1

inbelic commented May 5, 2026

View reviewed changes

bob80905 approved these changes May 12, 2026

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[Analysis] Ignore convergence tokens in dead branches in `CodeMetrics`#193590

[Analysis] Ignore convergence tokens in dead branches in `CodeMetrics`#193590
inbelic wants to merge 7 commits into
llvm:mainfrom
inbelic:inbelic/unroll-conv

inbelic commented Apr 22, 2026 •

edited

Loading

Uh oh!

Uh oh!

llvmbot commented Apr 22, 2026 •

edited

Loading

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

bob80905 Apr 30, 2026

Uh oh!

bob80905 Apr 30, 2026

Uh oh!

bob80905 Apr 30, 2026

Uh oh!

inbelic May 4, 2026 •

edited

Loading

Uh oh!

inbelic May 5, 2026

Uh oh!

bob80905 May 12, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Conversation

inbelic commented Apr 22, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

llvmbot commented Apr 22, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

bob80905 Apr 30, 2026

Choose a reason for hiding this comment

Uh oh!

bob80905 Apr 30, 2026

Choose a reason for hiding this comment

Uh oh!

bob80905 Apr 30, 2026

Choose a reason for hiding this comment

Uh oh!

inbelic May 4, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

inbelic May 5, 2026

Choose a reason for hiding this comment

Uh oh!

bob80905 May 12, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

inbelic commented Apr 22, 2026 •

edited

Loading

llvmbot commented Apr 22, 2026 •

edited

Loading

inbelic May 4, 2026 •

edited

Loading