[clang] [Headers][X86] VectorExprEvaluator::VisitCallExpr - allow SSE/AVX2/AVX512 pack intrinsics to be used in constexpr (PR #156003)

Shafik Yaghmour via cfe-commits cfe-commits at lists.llvm.org
Fri Aug 29 10:40:17 PDT 2025


================
@@ -11599,6 +11599,89 @@ static bool handleVectorElementCast(EvalInfo &Info, const FPOptions FPO,
   return false;
 }
 
+enum class PackKind {
+  SSWB,
+  USWB,
+  SSDW,
+  USDW
+}; // 16→8 signed/unsigned; 32→16 signed/unsigned
+
+static bool evalPackBuiltin(const CallExpr *E, EvalInfo &Info, APValue &Result,
+                            PackKind K) {
+  APValue L, R;
+  if (!EvaluateAsRValue(Info, E->getArg(0), L) ||
+      !EvaluateAsRValue(Info, E->getArg(1), R))
+    return false;
+
+  unsigned SrcBits = (K == PackKind::SSWB || K == PackKind::USWB) ? 16 : 32;
+  unsigned DstBits = SrcBits / 2;
+
+  unsigned NL = L.getVectorLength();
+  unsigned NR = R.getVectorLength();
+  if (NL == 0 || NR == 0 || NL != NR)
----------------
shafik wrote:

Do we have a test that covers these cases?

https://github.com/llvm/llvm-project/pull/156003


More information about the cfe-commits mailing list