[clang] [Headers][X86] Allow MMX/SSE/AVX MOVMSK intrinsics to be used in constexpr (PR #161914)
Shawn K via cfe-commits
cfe-commits at lists.llvm.org
Tue Oct 7 11:38:05 PDT 2025
================
@@ -2817,6 +2817,46 @@ static bool interp__builtin_ia32_pshuf(InterpState &S, CodePtr OpPC,
return true;
}
+static bool interp__builtin_ia32_movmsk_op(InterpState &S, CodePtr OpPC,
+ const CallExpr *Call) {
+ assert(Call->getNumArgs() == 1);
+
+ const Pointer &Source = S.Stk.pop<Pointer>();
+
+ unsigned SourceLen = Source.getNumElems();
+ const QualType ElemQT = getElemType(Source);
+ const OptPrimType ElemPT = S.getContext().classify(ElemQT);
+ unsigned LaneWidth = S.getASTContext().getTypeSize(ElemQT);
+
+ if (ElemQT->isIntegerType()) {
+ unsigned Byte = 8;
+ unsigned ResultLen = (LaneWidth * SourceLen) / Byte;
+ APInt Result(ResultLen, 0);
+ unsigned ResultIdx = 0;
+ for (unsigned I = 0; I != SourceLen; ++I) {
+ APInt Lane;
+ INT_TYPE_SWITCH_NO_BOOL(*ElemPT,
+ { Lane = Source.elem<T>(I).toAPSInt(); });
+ for (unsigned J = 0; J != LaneWidth; J += Byte) {
+ Result.setBitVal(ResultIdx++, Lane[J + 7]);
+ }
----------------
kimsh02 wrote:
Calling `Lane.isNegative()` gets the MSB for the lane, not for every byte of the lane. I applied `isNegative()` for the floating type case however. 👍
https://github.com/llvm/llvm-project/pull/161914
More information about the cfe-commits
mailing list