[PATCH] MISched: Fix moving stores across barriers

[Andrew Trick]

Thanks Tom. Looks like a “goto” bug. buildSchedGraph needs to be rewritten but I never found the occasion to do it.

Hal already pointed out the minor issues with your patch that I noticed, so LGTM.

-Andy

> On Dec 2, 2014, at 12:32 PM, Tom Stellard <thomas.stellard at amd.com> wrote:
> 
> This fixes an issue with ScheduleDAGInstrs::buildSchedGraph
> where stores without an underlying object would not be added
> as a predecessor to the current BarrierChain.
> ---
> lib/CodeGen/ScheduleDAGInstrs.cpp  |  9 +++++--
> test/CodeGen/R600/store-barrier.ll | 52 ++++++++++++++++++++++++++++++++++++++
> 2 files changed, 59 insertions(+), 2 deletions(-)
> create mode 100644 test/CodeGen/R600/store-barrier.ll
> 
> diff --git a/lib/CodeGen/ScheduleDAGInstrs.cpp b/lib/CodeGen/ScheduleDAGInstrs.cpp
> index d8d8422..ee8b5c2 100644
> --- a/lib/CodeGen/ScheduleDAGInstrs.cpp
> +++ b/lib/CodeGen/ScheduleDAGInstrs.cpp
> @@ -794,6 +794,7 @@ void ScheduleDAGInstrs::buildSchedGraph(AliasAnalysis *AA,
>   for (MachineBasicBlock::iterator MII = RegionEnd, MIE = RegionBegin;
>        MII != MIE; --MII) {
>     MachineInstr *MI = std::prev(MII);
> +
>     if (MI && DbgMI) {
>       DbgValues.push_back(std::make_pair(DbgMI, MI));
>       DbgMI = nullptr;
> @@ -920,6 +921,12 @@ 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);
> 
> @@ -993,8 +1000,6 @@ void ScheduleDAGInstrs::buildSchedGraph(AliasAnalysis *AA,
>       // 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 0000000..229cd8f
> --- /dev/null
> +++ b/test/CodeGen/R600/store-barrier.ll
> @@ -0,0 +1,52 @@
> +; 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
> +; 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) #0 {
> +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 #0 = { nounwind "less-precise-fpmad"="false" "no-frame-pointer-elim"="false" "no-infs-fp-math"="false" "no-nans-fp-math"="false" "no-realign-stack" "stack-protector-buffer-size"="8" "unsafe-fp-math"="false" "use-soft-float"="false" }
> +attributes #1 = { nounwind readnone }
> +attributes #2 = { noduplicate nounwind }
> +
> +!opencl.kernels = !{!0}
> +
> +!0 = metadata !{void (<2 x i8> addrspace(3)*, <2 x i8> addrspace(1)*, i32 addrspace(1)*, <2 x i8> addrspace(1)*, i32, i64)* @test}
> +!3 = metadata !{metadata !4, metadata !4, i64 0}
> +!4 = metadata !{metadata !"int", metadata !5, i64 0}
> +!5 = metadata !{metadata !"omnipotent char", metadata !6, i64 0}
> +!6 = metadata !{metadata !"Simple C/C++ TBAA"}
> +!7 = metadata !{metadata !5, metadata !5, i64 0}
> -- 
> 2.0.4
>