================
@@ -19273,9 +19269,14 @@ Value *CodeGenFunction::EmitAMDGPUBuiltinExpr(unsigned
BuiltinID,
ProcessOrderScopeAMDGCN(EmitScalarExpr(E->getArg(2)),
EmitScalarExpr(E->getArg(3)), AO, SSID);
} else {
- // The ds_atomic_fadd_* builtins do not have syncscope/order arguments.
- SSID = llvm::SyncScope::System;
- AO = AtomicOrdering::SequentiallyConsistent;
+ // Most of the builtins do not have syncscope/order arguments. For DS
+ // atomics the scope doesn't really matter, as they implicitly operate at
+ // workgroup scope.
+ //
+ // The global/flat cases need to use agent scope to consistently produce
+ // the native instruction instead of a cmpxchg expansion.
+ SSID = getLLVMContext().getOrInsertSyncScopeID("agent");
----------------
Pierre-vh wrote:
What happens with system (the default) ? I'm not sure I like using `agent` just
to force the right expansion when there is no memory model motivation behind
it. Do we have a precedent for this kind of thing?
Could codegen be fixed so you can just use `system`?
https://github.com/llvm/llvm-project/pull/96872
_______________________________________________
llvm-branch-commits mailing list
llvm-branch-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-branch-commits
