From 5b82c9ba31a05ce22a244e8b4ab6a345c9e8aecc Mon Sep 17 00:00:00 2001
From: Justin Lebar <jlebar@google.com>
Date: Mon, 22 Feb 2016 17:50:52 +0000
Subject: [PATCH] Don't tail-duplicate blocks that contain convergent
 instructions.

Summary:
Convergent instrs shouldn't be made control-dependent on other values,
but this is basically the whole point of tail duplication.  So just bail
if we see a convergent instruction.

Reviewers: iteratee

Subscribers: jholewinski, jhen, hfinkel, tra, jingyue, llvm-commits

Differential Revision: http://reviews.llvm.org/D17320

llvm-svn: 261540
---
 llvm/lib/CodeGen/TailDuplication.cpp               |  5 +++
 .../CodeGen/NVPTX/TailDuplication-convergent.ll    | 45 ++++++++++++++++++++++
 2 files changed, 50 insertions(+)
 create mode 100644 llvm/test/CodeGen/NVPTX/TailDuplication-convergent.ll

diff --git a/llvm/lib/CodeGen/TailDuplication.cpp b/llvm/lib/CodeGen/TailDuplication.cpp
index c2c7a1a..da593ce 100644
--- a/llvm/lib/CodeGen/TailDuplication.cpp
+++ b/llvm/lib/CodeGen/TailDuplication.cpp
@@ -590,6 +590,11 @@ TailDuplicatePass::shouldTailDuplicate(const MachineFunction &MF,
     if (MI.isNotDuplicable())
       return false;
 
+    // Convergent instructions can be duplicated only if doing so doesn't add
+    // new control dependencies, which is what we're going to do here.
+    if (MI.isConvergent())
+      return false;
+
     // Do not duplicate 'return' instructions if this is a pre-regalloc run.
     // A return may expand into a lot more instructions (e.g. reload of callee
     // saved registers) after PEI.
diff --git a/llvm/test/CodeGen/NVPTX/TailDuplication-convergent.ll b/llvm/test/CodeGen/NVPTX/TailDuplication-convergent.ll
new file mode 100644
index 0000000..02b562d
--- /dev/null
+++ b/llvm/test/CodeGen/NVPTX/TailDuplication-convergent.ll
@@ -0,0 +1,45 @@
+; RUN: llc -O2 -tail-dup-size=100 -enable-tail-merge=0 < %s | FileCheck %s
+target triple = "nvptx64-nvidia-cuda"
+
+declare void @foo()
+declare void @llvm.cuda.syncthreads()
+
+; syncthreads shouldn't be duplicated.
+; CHECK: .func call_syncthreads
+; CHECK: bar.sync
+; CHECK-NOT: bar.sync
+define void @call_syncthreads(i32* %a, i32* %b, i1 %cond, i1 %cond2) nounwind {
+  br i1 %cond, label %L1, label %L2
+  br i1 %cond2, label %Ret, label %L1
+Ret:
+  ret void
+L1:
+  store i32 0, i32* %a
+  br label %L42
+L2:
+  store i32 1, i32* %a
+  br label %L42
+L42:
+  call void @llvm.cuda.syncthreads()
+  br label %Ret
+}
+
+; Check that call_syncthreads really does trigger tail duplication.
+; CHECK: .func call_foo
+; CHECK: call
+; CHECK: call
+define void @call_foo(i32* %a, i32* %b, i1 %cond, i1 %cond2) nounwind {
+  br i1 %cond, label %L1, label %L2
+  br i1 %cond2, label %Ret, label %L1
+Ret:
+  ret void
+L1:
+  store i32 0, i32* %a
+  br label %L42
+L2:
+  store i32 1, i32* %a
+  br label %L42
+L42:
+  call void @foo()
+  br label %Ret
+}
-- 
2.7.4