[API Proposal]: ARM64 SVE: LoadVectorNonFaulting APIs require a mask due to FFR #108234
Description
Background and motivation
LoadVectorNonFaulting() does not take a mask argument because the load is guaranteed not to fault and therefore:
LoadVectorNonFaulting(mask, addr) === ConditionalSelect(mask, LoadVectorNonFaulting(addr), zero)
However, this is not entirely correct.
If a lane that would have been false is on a faulting memory address then that lane in the FFR register is cleared and successive loads to that lane with FFR become unknown/unpredictable.
See the pseudo code at the bottom of LDNF1B
I recommend adding a mask to all non-faulting SVE APIs.
API Proposal
Using the same T syntax as other SVE proposals.
These are all changes of existing API methods - they all differ from the existing methods due to adding an additional mask argument.
namespace System.Runtime.Intrinsics.Arm;
public partial class Sve
{
public static unsafe Vector<short> LoadVectorByteNonFaultingZeroExtendToInt16(Vector<short> mask, byte* address); // LDNF1B
public static unsafe Vector<int> LoadVectorByteNonFaultingZeroExtendToInt32(Vector<int> mask, byte* address); // LDNF1B
public static unsafe Vector<long> LoadVectorByteNonFaultingZeroExtendToInt64(Vector<long> mask, byte* address); // LDNF1B
public static unsafe Vector<ushort> LoadVectorByteNonFaultingZeroExtendToUInt16(Vector<ushort> mask, byte* address); // LDNF1B
public static unsafe Vector<uint> LoadVectorByteNonFaultingZeroExtendToUInt32(Vector<uint> mask, byte* address); // LDNF1B
public static unsafe Vector<ulong> LoadVectorByteNonFaultingZeroExtendToUInt64(Vector<ulong> mask, byte* address); // LDNF1B
public static unsafe Vector<int> LoadVectorInt16NonFaultingSignExtendToInt32(Vector<int> mask, short* address); // LDNF1SH
public static unsafe Vector<long> LoadVectorInt16NonFaultingSignExtendToInt64(Vector<long> mask, short* address); // LDNF1SH
public static unsafe Vector<uint> LoadVectorInt16NonFaultingSignExtendToUInt32(Vector<uint> mask, short* address); // LDNF1SH
public static unsafe Vector<ulong> LoadVectorInt16NonFaultingSignExtendToUInt64(Vector<ulong> mask, short* address); // LDNF1SH
public static unsafe Vector<long> LoadVectorInt32NonFaultingSignExtendToInt64(Vector<long> mask, int* address); // LDNF1SW
public static unsafe Vector<ulong> LoadVectorInt32NonFaultingSignExtendToUInt64(Vector<ulong> mask, int* address); // LDNF1SW
/// T: float, double, sbyte, short, int, long, byte, ushort, uint, ulong
public static unsafe Vector<T> LoadVectorNonFaulting(Vector<T> mask, T* address); // LDNF1W or LDNF1D or LDNF1B or LDNF1H
public static unsafe Vector<short> LoadVectorSByteNonFaultingSignExtendToInt16(Vector<short> mask, sbyte* address); // LDNF1SB
public static unsafe Vector<int> LoadVectorSByteNonFaultingSignExtendToInt32(Vector<int> mask, sbyte* address); // LDNF1SB
public static unsafe Vector<long> LoadVectorSByteNonFaultingSignExtendToInt64(Vector<long> mask, sbyte* address); // LDNF1SB
public static unsafe Vector<ushort> LoadVectorSByteNonFaultingSignExtendToUInt16(Vector<ushort> mask, sbyte* address); // LDNF1SB
public static unsafe Vector<uint> LoadVectorSByteNonFaultingSignExtendToUInt32(Vector<uint> mask, sbyte* address); // LDNF1SB
public static unsafe Vector<ulong> LoadVectorSByteNonFaultingSignExtendToUInt64(Vector<ulong> mask, sbyte* address); // LDNF1SB
public static unsafe Vector<int> LoadVectorUInt16NonFaultingZeroExtendToInt32(Vector<int> mask, ushort* address); // LDNF1H
public static unsafe Vector<long> LoadVectorUInt16NonFaultingZeroExtendToInt64(Vector<long> mask, ushort* address); // LDNF1H
public static unsafe Vector<uint> LoadVectorUInt16NonFaultingZeroExtendToUInt32(Vector<uint> mask, ushort* address); // LDNF1H
public static unsafe Vector<ulong> LoadVectorUInt16NonFaultingZeroExtendToUInt64(Vector<ulong> mask, ushort* address); // LDNF1H
public static unsafe Vector<long> LoadVectorUInt32NonFaultingZeroExtendToInt64(Vector<long> mask, uint* address); // LDNF1W
public static unsafe Vector<ulong> LoadVectorUInt32NonFaultingZeroExtendToUInt64(Vector<ulong> mask, uint* address); // LDNF1W
}