[clang] [llvm] [HLSL] Add QuadAll Emulation support (PR #240645)

Helena Kotas via cfe-commits cfe-commits at lists.llvm.org
Fri Nov 13 00:18:59 PST 2026


================
@@ -1000,6 +1001,37 @@ void CGHLSLRuntime::addHLSLBufferLayoutType(const RecordType *StructType,
   LayoutTypes[StructType] = LayoutTy;
 }
 
+static void addSPIRVQuadAllExecutionModes(llvm::Module &M) {
+  Function *QuadAll =
+      Intrinsic::getDeclarationIfExists(&M, Intrinsic::spv_quad_all);
+  if (!QuadAll)
+    return;
+
+  SmallVector<Function *, 8> Worklist{QuadAll};
+  SmallPtrSet<Function *, 16> Visited;
+  IRBuilder<> B(M);
+  while (!Worklist.empty()) {
+    Function *F = Worklist.pop_back_val();
+    if (!Visited.insert(F).second)
+      continue;
+
+    if (F->hasFnAttribute("hlsl.numthreads")) {
+      StringRef NumThreads =
+          F->getFnAttribute("hlsl.numthreads").getValueAsString();
+      // DerivativeGroupLinearKHR is 5290; DerivativeGroupQuadsKHR is 5289.
+      unsigned Mode = NumThreads.ends_with(",1,1") ? 5290 : 5289;
----------------
hekota wrote:

The `numthreads` values need more validation - `DerivativeGroupQuadsKHR` requires both X and Y to be even. So a shader with `[numthreads(1,4,1)]` should be rejected.
https://docs.vulkan.org/spec/latest/appendices/spirvenv.html#VUID-StandaloneSpirv-DerivativeGroupQuadsKHR-10151

https://github.com/llvm/llvm-project/pull/240645


More information about the cfe-commits mailing list