[clang] [X86][Clang] VectorExprEvaluator::VisitCallExpr / InterpretBuiltin - Allow AVX/AVX512 IFMA madd52 intrinsics to be used in constexpr (PR #161056)

[Simon Pilgrim via cfe-commits]

================
@@ -3523,6 +3523,26 @@ bool InterpretBuiltin(InterpState &S, CodePtr OpPC, const CallExpr *Call,
           return F;
         });
 
+  case X86::BI__builtin_ia32_vpmadd52luq128:
+  case X86::BI__builtin_ia32_vpmadd52luq256:
+  case X86::BI__builtin_ia32_vpmadd52luq512:
+    return interp__builtin_elementwise_triop(
+        S, OpPC, Call, [](const APSInt &A, const APSInt &B, const APSInt &C) {
+          APSInt Result(A + (B.trunc(52) * C.trunc(52)).trunc(52).zext(64),
+                        false);
+          return APSInt(Result.trunc(52).zext(64), false);
----------------
RKSimon wrote:

This is incorrect - only the multiply occurs as i52 - the accumulate is full i64:
```
return APSInt(A + (B.trunc(52) * C.trunc(52)).trunc(52).zext(64), false);
```
Same for the others.

https://github.com/llvm/llvm-project/pull/161056