AMDGPU: mark atomic instructions as sources of divergence

Summary:
As explained by the comment, threads will typically see different values
returned by atomic instructions even if the arguments are equal.

Reviewers: arsenm, tstellarAMD

Subscribers: arsenm, llvm-commits

Differential Revision: http://reviews.llvm.org/D18156

llvm-svn: 263719

2 files changed, 22 insertions, 0 deletions

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp b/llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp
index 540ca5f399f..ff5558bf4db 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp
@@ -186,6 +186,13 @@ bool AMDGPUTTIImpl::isSourceOfDivergence(const Value *V) const {
   if (const LoadInst *Load = dyn_cast<LoadInst>(V))
     return Load->getPointerAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS;
 
+  // Atomics are divergent because they are executed sequentially: when an
+  // atomic operation refers to the same address in each thread, then each
+  // thread after the first sees the value written by the previous thread as
+  // original value.
+  if (isa<AtomicRMWInst>(V) || isa<AtomicCmpXchgInst>(V))
+    return true;
+
   if (const IntrinsicInst *Intrinsic = dyn_cast<IntrinsicInst>(V)) {
     const TargetMachine &TM = getTLI()->getTargetMachine();
     return isIntrinsicSourceOfDivergence(TM.getIntrinsicInfo(), Intrinsic);
diff --git a/llvm/test/Analysis/DivergenceAnalysis/AMDGPU/atomics.ll b/llvm/test/Analysis/DivergenceAnalysis/AMDGPU/atomics.ll
new file mode 100644
index 00000000000..60d0de6035b
--- /dev/null
+++ b/llvm/test/Analysis/DivergenceAnalysis/AMDGPU/atomics.ll
@@ -0,0 +1,15 @@
+; RUN: opt -mtriple=amdgcn-- -analyze -divergence %s | FileCheck %s
+
+; CHECK: DIVERGENT: %orig = atomicrmw xchg i32* %ptr, i32 %val seq_cst
+define i32 @test1(i32* %ptr, i32 %val) #0 {
+  %orig = atomicrmw xchg i32* %ptr, i32 %val seq_cst
+  ret i32 %orig
+}
+
+; CHECK: DIVERGENT: %orig = cmpxchg i32* %ptr, i32 %cmp, i32 %new seq_cst seq_cst
+define {i32, i1} @test2(i32* %ptr, i32 %cmp, i32 %new) {
+  %orig = cmpxchg i32* %ptr, i32 %cmp, i32 %new seq_cst seq_cst
+  ret {i32, i1} %orig
+}
+
+attributes #0 = { "ShaderType"="0" }