[clang] [clang][RISCV] support BITINT with mixed-type (PR #156592)

[Craig Topper via cfe-commits]

================
@@ -680,14 +680,11 @@ ABIArgInfo RISCVABIInfo::classifyArgumentType(QualType Ty, bool IsFixed,
     if (const auto *ED = Ty->getAsEnumDecl())
       Ty = ED->getIntegerType();
 
-    // All integral types are promoted to XLen width
-    if (Size < XLen && Ty->isIntegralOrEnumerationType()) {
-      return extendType(Ty, CGT.ConvertType(Ty));
-    }
-
     if (const auto *EIT = Ty->getAs<BitIntType>()) {
-      if (EIT->getNumBits() < XLen)
-        return extendType(Ty, CGT.ConvertType(Ty));
+      // FIXME: Maybe we should treat 32 as a special case and wait for
+      // the SPEC to decide.
+      if (EIT->getNumBits() <= 2 * XLen)
+        return ABIArgInfo::getExtend(Ty, CGT.ConvertType(Ty));
       if (EIT->getNumBits() > 128 ||
----------------
topperc wrote:

Isn't the `EIT->getNumBits() > 64` check always true after your change?

And it looks like for RV32, any size between 65 bits and 128 bits will pass directly when `-fforce-enable-int128` is passed. Why is _BitInt behavior affected by -fforce-enable-int128?

https://github.com/llvm/llvm-project/pull/156592