[clang-tools-extra] [clangd][HLSL] Prevent RootSignature internal identifier leak on hover (PR #214955)
Maria Fernanda GuimarĂ£es via cfe-commits
cfe-commits at lists.llvm.org
Fri Aug 21 04:29:44 PDT 2026
================
@@ -5431,6 +5431,23 @@ TEST(Hover, HLSLRegisterAttributeRange) {
}
}
+TEST(Hover, HLSLRootSignature) {
+ Annotations T(
+ R"hlsl(
+ #define RS_CBV "CBV(b0)"
+ [^RootSignature(RS_CBV)]
+ void main() {}
+ )hlsl",
+ Annotations::Markers().setRangeBegin("{{").setRangeEnd("}}"));
+
+ TestTU TU = TestTU::withCode(T.code());
+ configureHLSL(TU);
+ auto AST = TU.build();
+ auto H = getHover(AST, T.point(), format::getLLVMStyle(), nullptr);
+ ASSERT_TRUE(H) << "Hover should have been returned for RootSignature!";
+ EXPECT_EQ(H->Name, "RootSignature");
----------------
mafeguimaraes wrote:
Good questions, let me clarify both.
On the first point: the patch doesn't clear `Definition`, it just never sets it. `HoverInfo` starts empty, so `Name` and `Documentation` are populated (as before) and `Definition` simply stays unset for this attribute. So nothing about `Name`/`Documentation` changes.
On filling `Definition` with something meaningful instead of leaving it empty: I looked into this before choosing this approach. `RootSignatureAttr::getSignatureDecl()` does give us an `HLSLRootSignatureDecl*` with the fully parsed `RootElement`s (DescriptorTable, RootConstants, RootDescriptor, StaticSampler, etc.), so the information is there. The only existing printer for `RootElement` is its `operator<<` (used by `dumpRootElements()` for `-ast-dump`), and I prototyped calling it from hover via a small `getPrintedRootSignature()` helper. It works, but the output includes every field of every element, including defaulted ones (`numDescriptors = 1, space = 0, ...`), which is fine for `-ast-dump` but too verbose for a hover tooltip.
A compact hover-appropriate formatter (something like `SRV(t0), UAV(u0), DescriptorTable(2 clauses)`) is doable, but it means visiting each `RootElement` variant and deciding which fields matter for a quick-glance summary, a separate, non-trivial piece of work. Given the timeline of GSoC, I'd rather ship this suppress-and-show-docs fix now and file a follow-up issue to build the proper reconstruction. Happy to open that tracking issue right away if that sounds reasonable, or to scope it into this PR if you'd prefer it done now.
https://github.com/llvm/llvm-project/pull/214955
More information about the cfe-commits
mailing list