[clang] [RISCV][Sema] Add feature check for target attribute to VSETVL intrinsics (PR #126064)

[Craig Topper via cfe-commits]

================
@@ -623,13 +623,37 @@ bool SemaRISCV::CheckBuiltinFunctionCall(const TargetInfo &TI,
   }
   }
 
+  auto checkVsetvl = [&](unsigned SEWOffset,
+                                        unsigned LMULOffset) -> bool {
+    const FunctionDecl *FD = SemaRef.getCurFunctionDecl();
+    llvm::StringMap<bool> FunctionFeatureMap;
+    Context.getFunctionFeatureMap(FunctionFeatureMap, FD);
+    llvm::APSInt SEWResult;
+    llvm::APSInt LMULResult;
+    if (SemaRef.BuiltinConstantArg(TheCall, SEWOffset, SEWResult) ||
+        SemaRef.BuiltinConstantArg(TheCall, LMULOffset, LMULResult))
+      return true;
+    int SEWValue = SEWResult.getSExtValue();
+    int LMULValue = LMULResult.getSExtValue();
+    if (((SEWValue == 0 && LMULValue == 5) ||  // e8mf8
+         (SEWValue == 1 && LMULValue == 6) ||  // e16mf4
+         (SEWValue == 2 && LMULValue == 7) ||  // e32mf2
+         (SEWValue == 3 && LMULValue == 0) ||  // e64m1
+         (SEWValue == 3 && LMULValue == 1) ||  // e64m2
+         (SEWValue == 3 && LMULValue == 2) ||  // e64m4
+         (SEWValue == 3 && LMULValue == 3)) && // e64m8
+        (!TI.hasFeature("zve64x") && !FunctionFeatureMap.lookup("zve64x")))
----------------
topperc wrote:

The requirements for supported vtypes is a function of ELEN not VLEN. A Zvl64b+Zve32x implementation may support `e8,mf8`, `e16,mf4`, `e32,mf2` but it not required by the spec so the code would not be portable.

https://github.com/llvm/llvm-project/pull/126064