AMDGPU: mark atomic instructions as sources of divergence

[nhaehnle]

Summary

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

[• tstellarAMD]

lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp

189–195

Are these still divergent if the return value isn't used?

[arsenm]

If the value is never used it doesn't matter since the use is what propagates divergence

[• tstellarAMD]

LGTM.

lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp

bool AMDGPUTTIImpl::isSourceOfDivergence(const Value *V) const {
   // can execute the load instruction with the same inputs and get different
   // results.
   //
   // All other loads are not divergent, because if threads issue loads with the
   // same arguments, they will always get the same result.
   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);
   }
 
   // Assume all function calls are a source of divergence.
   if (isa<CallInst>(V) || isa<InvokeInst>(V))
     return true;
 
   return false;
 }

test/Analysis/DivergenceAnalysis/AMDGPU/atomics.ll

+; 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" }