[TableGen] Report a better error when an InstAlias does not use a RegClass

[arichardson]

Also fix the missing space in the error message. I notice while changing
RISC-V's loads and stores to use RegClassByHwMode and got a non-descriptive
error from T.getRegisterClass(OpRC) when parsing the InstAliases.

[llvmbot]

@llvm/pr-subscribers-tablegen

Author: Alexander Richardson (arichardson)

Changes

Also fix the missing space in the error message. I notice while changing
RISC-V's loads and stores to use RegClassByHwMode and got a non-descriptive
error when tablegen was parsing the InstAliases.

---

Full diff: https://github.com/llvm/llvm-project/pull/168444.diff

1 Files Affected:

• (modified) llvm/utils/TableGen/Common/CodeGenInstAlias.cpp (+3-2)

diff --git a/llvm/utils/TableGen/Common/CodeGenInstAlias.cpp b/llvm/utils/TableGen/Common/CodeGenInstAlias.cpp
index 6c1a3e9977c28..2de27986fa04a 100644
--- a/llvm/utils/TableGen/Common/CodeGenInstAlias.cpp
+++ b/llvm/utils/TableGen/Common/CodeGenInstAlias.cpp
@@ -55,9 +55,10 @@ static Expected<ResultOperand> matchSimpleOperand(const Init *Arg,
       // Match 'RegClass:$name' or 'RegOp:$name'.
       if (const Record *ArgRC = T.getInitValueAsRegClassLike(Arg)) {
         if (ArgRC->isSubClassOf("RegisterClass")) {
-          if (!T.getRegisterClass(OpRC).hasSubClass(&T.getRegisterClass(ArgRC)))
+          if (!OpRC->isSubClassOf("RegisterClass") ||
+              !T.getRegisterClass(OpRC).hasSubClass(&T.getRegisterClass(ArgRC)))
             return createStringError(
-                "argument register class" + ArgRC->getName() +
+                "argument register class " + ArgRC->getName() +
                 " is not a subclass of operand register class " +
                 OpRC->getName());
           if (!ArgName)

[arichardson]

Considering the typo wasn't noticed there does not seem to be a test for this error case. If there's an existing test I can add this to easily I can do so, otherwise I'll have to try create a minimal testcase from scratch.

[s-barannikov]

There are no tests for this, as far as I remember.

[s-barannikov]

Maybe instead if (OpRC->isSubclassOf("RegClassByHwMode")) then report "HwMode support is not implemented"?

[arichardson]

Yeah that would give slightly better results but may be too restrictive. It does seem to work with HWModes as long as the RegClassByHwMode argument is the same, but it doesn't handle the subset case.

[github-actions]

🐧 Linux x64 Test Results

• 186255 tests passed
• 4852 tests skipped