From 653ef32216953f38c20706cc6824a552487e8d08 Mon Sep 17 00:00:00 2001
From: Tom Stellard <thomas.stellard@amd.com>
Date: Mon, 8 Dec 2014 23:36:48 +0000
Subject: [PATCH] MISched: Fix moving stores across barriers

This fixes an issue with ScheduleDAGInstrs::buildSchedGraph
where stores without an underlying object would not be added
as a predecessor to the current BarrierChain.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@223717 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/CodeGen/ScheduleDAGInstrs.cpp  | 13 ++++-----
 test/CodeGen/R600/store-barrier.ll | 42 ++++++++++++++++++++++++++++++
 2 files changed, 49 insertions(+), 6 deletions(-)
 create mode 100644 test/CodeGen/R600/store-barrier.ll

diff --git a/lib/CodeGen/ScheduleDAGInstrs.cpp b/lib/CodeGen/ScheduleDAGInstrs.cpp
index d8d8422f0ed..dece643c0d2 100644
--- a/lib/CodeGen/ScheduleDAGInstrs.cpp
+++ b/lib/CodeGen/ScheduleDAGInstrs.cpp
@@ -920,6 +920,13 @@ void ScheduleDAGInstrs::buildSchedGraph(AliasAnalysis *AA,
       AliasMemDefs.clear();
       AliasMemUses.clear();
     } else if (MI->mayStore()) {
+      // Add dependence on barrier chain, if needed.
+      // There is no point to check aliasing on barrier event. Even if
+      // SU and barrier _could_ be reordered, they should not. In addition,
+      // we have lost all RejectMemNodes below barrier.
+      if (BarrierChain)
+        BarrierChain->addPred(SDep(SU, SDep::Barrier));
+
       UnderlyingObjectsVector Objs;
       getUnderlyingObjectsForInstr(MI, MFI, Objs);
 
@@ -989,12 +996,6 @@ void ScheduleDAGInstrs::buildSchedGraph(AliasAnalysis *AA,
         adjustChainDeps(AA, MFI, SU, &ExitSU, RejectMemNodes,
                         TrueMemOrderLatency);
       }
-      // Add dependence on barrier chain, if needed.
-      // There is no point to check aliasing on barrier event. Even if
-      // SU and barrier _could_ be reordered, they should not. In addition,
-      // we have lost all RejectMemNodes below barrier.
-      if (BarrierChain)
-        BarrierChain->addPred(SDep(SU, SDep::Barrier));
     } else if (MI->mayLoad()) {
       bool MayAlias = true;
       if (MI->isInvariantLoad(AA)) {
diff --git a/test/CodeGen/R600/store-barrier.ll b/test/CodeGen/R600/store-barrier.ll
new file mode 100644
index 00000000000..350b006ba5e
--- /dev/null
+++ b/test/CodeGen/R600/store-barrier.ll
@@ -0,0 +1,42 @@
+; RUN: llc -march=r600 -mcpu=SI -verify-machineinstrs -mattr=+load-store-opt -enable-misched < %s | FileCheck  --check-prefix=CHECK %s
+; RUN: llc -march=r600 -mcpu=bonaire -verify-machineinstrs -mattr=+load-store-opt -enable-misched < %s | FileCheck  --check-prefix=CHECK %s
+
+; This test is for a bug in the machine scheduler where stores without
+; an underlying object would be moved across the barrier.  In this
+; test, the <2 x i8> store will be split into two i8 stores, so they
+; won't have an underlying object.
+
+; CHECK-LABEL: {{^}}test:
+; CHECK: ds_write_b8
+; CHECK: ds_write_b8
+; CHECK: s_barrier
+; CHECK: s_endpgm
+; Function Attrs: nounwind
+define void @test(<2 x i8> addrspace(3)* nocapture %arg, <2 x i8> addrspace(1)* nocapture readonly %arg1, i32 addrspace(1)* nocapture readonly %arg2, <2 x i8> addrspace(1)* nocapture %arg3, i32 %arg4, i64 %tmp9) {
+bb:
+  %tmp10 = getelementptr inbounds i32 addrspace(1)* %arg2, i64 %tmp9
+  %tmp13 = load i32 addrspace(1)* %tmp10, align 2
+  %tmp14 = getelementptr inbounds <2 x i8> addrspace(3)* %arg, i32 %tmp13
+  %tmp15 = load <2 x i8> addrspace(3)* %tmp14, align 2
+  %tmp16 = add i32 %tmp13, 1
+  %tmp17 = getelementptr inbounds <2 x i8> addrspace(3)* %arg, i32 %tmp16
+  store <2 x i8> %tmp15, <2 x i8> addrspace(3)* %tmp17, align 2
+  tail call void @llvm.AMDGPU.barrier.local() #2
+  %tmp25 = load i32 addrspace(1)* %tmp10, align 4
+  %tmp26 = sext i32 %tmp25 to i64
+  %tmp27 = sext i32 %arg4 to i64
+  %tmp28 = getelementptr inbounds <2 x i8> addrspace(3)* %arg, i32 %tmp25, i32 %arg4
+  %tmp29 = load i8 addrspace(3)* %tmp28, align 1
+  %tmp30 = getelementptr inbounds <2 x i8> addrspace(1)* %arg3, i64 %tmp26, i64 %tmp27
+  store i8 %tmp29, i8 addrspace(1)* %tmp30, align 1
+  %tmp32 = getelementptr inbounds <2 x i8> addrspace(3)* %arg, i32 %tmp25, i32 0
+  %tmp33 = load i8 addrspace(3)* %tmp32, align 1
+  %tmp35 = getelementptr inbounds <2 x i8> addrspace(1)* %arg3, i64 %tmp26, i64 0
+  store i8 %tmp33, i8 addrspace(1)* %tmp35, align 1
+  ret void
+}
+
+; Function Attrs: noduplicate nounwind
+declare void @llvm.AMDGPU.barrier.local() #2
+
+attributes #2 = { noduplicate nounwind }
-- 
2.34.1