[llvm] [RISCV] Consolidate some DecoderNamespaces for standard extensions. (PR #128954)

[Craig Topper via llvm-commits]

================
@@ -721,6 +711,11 @@ DecodeStatus RISCVDisassembler::getInstruction32(MCInst &MI, uint64_t &Size,
   TRY_TO_DECODE_FEATURE_ANY(XRivosFeatureGroup, DecoderTableXRivos32, "Rivos");
 
   TRY_TO_DECODE(true, DecoderTable32, "RISCV32");
+  TRY_TO_DECODE(true, DecoderTableRV32GPRPair32,
+                "RV32GPRPair (rv32 and GPR pairs)");
+  TRY_TO_DECODE(true, DecoderTableZfinx32, "Zfinx (Float in Integer)");
+  TRY_TO_DECODE(true, DecoderTableZdinxRV32GPRPair32,
+                "ZdinxRV32GPRPair (rv32 and Double in Integer)");
----------------
topperc wrote:

> Why do we decode all the standard stuff last, rather than before custom extensions? I'd have thought we want to decode the standard extensions with higher priority than the custom extensions, but maybe not?

A good question. As far as I know all the vendor extensions are mutually exclusive with the standard extensions, except for Xqcisim which appears to use hint encodings of standard instructions. So they could go last and it would be faster for the common case.

https://github.com/llvm/llvm-project/pull/128954