================
@@ -537,7 +537,11 @@ AMDGPUTargetCodeGenInfo::getLLVMSyncScopeID(const
LangOptions &LangOpts,
break;
}
- if (Ordering != llvm::AtomicOrdering::SequentiallyConsistent) {
+ // OpenCL assumes by default that atomic scopes are per-address space for
+ // non-sequentially consistent operations.
+ if (Scope >= SyncScope::OpenCLWorkGroup &&
----------------
t-tye wrote:
I am not sure what you mean by "each implementation of this "atomic scope"
thing". It looks like there is a single function involved here called
AMDGPUTargetCodeGenInfo::getLLVMSyncScopeID . It takes an argument called
scope. Its job is to correctly generate the sync-scope that should be used for
the AMD GPU target for the given value of the scope argument. It is up to the
callers of this function to pass the correct value for "scope". Are all the
callers doing that correctly?
This is not about an OpenCL implementation, it is about the compiler
implementing the correct semantics for the source code that clang is compiling.
Clang is responsible for honoring the semantics of the source language and any
directives it may have. Non-OpenCL source languages may want certain code to
use the OpenCL semantics and so "invent" directives to do that, and hence may
cause this function to be called with one of the OpenCL enumerator values.
https://github.com/llvm/llvm-project/pull/120095
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
