================
@@ -89,37 +88,178 @@ bool RootSignatureParser::parseDescriptorTableClause() {
CurToken.TokKind == TokenKind::kw_UAV ||
CurToken.TokKind == TokenKind::kw_Sampler) &&
"Expects to only be invoked starting at given keyword");
+ TokenKind ParamKind = CurToken.TokKind; // retain for diagnostics
DescriptorTableClause Clause;
- switch (CurToken.TokKind) {
+ TokenKind ExpectedRegister;
+ switch (ParamKind) {
default:
llvm_unreachable("Switch for consumed token was not provided");
case TokenKind::kw_CBV:
Clause.Type = ClauseType::CBuffer;
+ ExpectedRegister = TokenKind::bReg;
break;
case TokenKind::kw_SRV:
Clause.Type = ClauseType::SRV;
+ ExpectedRegister = TokenKind::tReg;
break;
case TokenKind::kw_UAV:
Clause.Type = ClauseType::UAV;
+ ExpectedRegister = TokenKind::uReg;
break;
case TokenKind::kw_Sampler:
Clause.Type = ClauseType::Sampler;
+ ExpectedRegister = TokenKind::sReg;
break;
}
if (consumeExpectedToken(TokenKind::pu_l_paren, diag::err_expected_after,
- CurToken.TokKind))
+ ParamKind))
return true;
- if (consumeExpectedToken(TokenKind::pu_r_paren, diag::err_expected_after,
- CurToken.TokKind))
+ llvm::SmallDenseMap<TokenKind, ParamType> Params = {
----------------
llvm-beanz wrote:
I'm a bit torn. The pattern used in Clang for things that can be parsed in
arbitrary order (like attributes) might be a bit overkill. Clang fully
separates parsing from semantic analysis, so it parses all your attributes
separate from converting them into AST nodes and attaching them to the proper
parent.
The structure-of-array approach storing all tokens has some scalability
concerns to me, and I'm not sure the genericness-gives benefit. Similarly I'm
not sure the map approach as implemented is the right approach.
My gut feeling is that we probably shouldn't use the same implementation for
parsing descriptor table parameters and sampler parameters.
I think a parameters structure for 6 parameters with either bools (which could
be uint32_t bitfield members) or optionals to signal if they are seen makes
sense. I suspect if they're mandatory should be something we know from context
and don't need additional tracking of.
https://github.com/llvm/llvm-project/pull/133800
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
