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

[Timm Baeder via cfe-commits]

================
@@ -3564,6 +3564,28 @@ 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 + C;
+          APSInt mask(APSInt::getAllOnes(52).zext(64), false);
+          APSInt masked_result = result & mask;
+          return APSInt(masked_result, true); // unsigned result
+        });
+  case X86::BI__builtin_ia32_vpmadd52huq128:
+  case X86::BI__builtin_ia32_vpmadd52huq256:
+  case X86::BI__builtin_ia32_vpmadd52huq512:
+    return interp__builtin_elementwise_triop(
+        S, OpPC, Call, [](const APSInt &A, const APSInt &B, const APSInt &C) {
+          APSInt result = A * B + C;
+          APSInt mask(APSInt::getAllOnes(52).zext(64), false);
+          APSInt shifted_result = result >> 52;
----------------
tbaederr wrote:

Some explanation for the magic constants here would be nice I guess.

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