Optimize VectorX<T>.ConditionalSelect for constant masks (#104092)

* Optimize ConditionalSelect for const mask

This adds a check in the JIT for constant masks (`GT_CNS_VEC`, everything else gets lowered to it) and enables optimization to `BlendVariable` (`(v)pblendvb` instruction).
This currently does not work for masks loaded from an array in a field/variable.
Also this optimization is not triggered for platforms supporting AVX512F(/VL?) since it gets optimized earlier to `vpternlogd` instruction.

* Cleanup code and separate it into functions

* fix build

* Misc fixes

* Final build fixes

* Address review comments

* Address review comments again

* address the rest of the comments

* Remove scalar assertion

Co-authored-by: Tanner Gooding <tagoo@outlook.com>

---------

Co-authored-by: Tanner Gooding <tagoo@outlook.com>

src/coreclr/jit/gentree.cpp

@@ -29707,6 +29707,92 @@ bool GenTree::IsInvariant() const
     return OperIsConst() || OperIs(GT_LCL_ADDR) || OperIs(GT_FTN_ADDR);
 }
 
+//-------------------------------------------------------------------
+// IsVectorPerElementMask: returns true if this node is a vector constant per-element mask
+//                         (every element has either all bits set or none of them).
+//
+// Arguments:
+//    simdBaseType - the base type of the constant being checked.
+//    simdSize     - the size of the SIMD type of the intrinsic.
+//
+// Returns:
+//     True if this node is a vector constant per-element mask.
+//
+bool GenTree::IsVectorPerElementMask(var_types simdBaseType, unsigned simdSize) const
+{
+#ifdef FEATURE_SIMD
+    if (IsCnsVec())
+    {
+        const GenTreeVecCon* vecCon = AsVecCon();
+
+        int elementCount = vecCon->ElementCount(simdSize, simdBaseType);
+
+        switch (simdBaseType)
+        {
+            case TYP_BYTE:
+            case TYP_UBYTE:
+                return ElementsAreAllBitsSetOrZero(&vecCon->gtSimdVal.u8[0], elementCount);
+            case TYP_SHORT:
+            case TYP_USHORT:
+                return ElementsAreAllBitsSetOrZero(&vecCon->gtSimdVal.u16[0], elementCount);
+            case TYP_INT:
+            case TYP_UINT:
+            case TYP_FLOAT:
+                return ElementsAreAllBitsSetOrZero(&vecCon->gtSimdVal.u32[0], elementCount);
+            case TYP_LONG:
+            case TYP_ULONG:
+            case TYP_DOUBLE:
+                return ElementsAreAllBitsSetOrZero(&vecCon->gtSimdVal.u64[0], elementCount);
+            default:
+                unreached();
+        }
+    }
+    else if (OperIsHWIntrinsic())
+    {
+        const GenTreeHWIntrinsic* intrinsic   = AsHWIntrinsic();
+        const NamedIntrinsic      intrinsicId = intrinsic->GetHWIntrinsicId();
+
+        if (HWIntrinsicInfo::ReturnsPerElementMask(intrinsicId))
+        {
+            // We directly return a per-element mask
+            return true;
+        }
+
+        bool       isScalar = false;
+        genTreeOps oper     = intrinsic->HWOperGet(&isScalar);
+
+        switch (oper)
+        {
+            case GT_AND:
+            case GT_AND_NOT:
+            case GT_OR:
+            case GT_XOR:
+            {
+                // We are a binary bitwise operation where both inputs are per-element masks
+                return intrinsic->Op(1)->IsVectorPerElementMask(simdBaseType, simdSize) &&
+                       intrinsic->Op(2)->IsVectorPerElementMask(simdBaseType, simdSize);
+            }
+
+            case GT_NOT:
+            {
+                // We are an unary bitwise operation where the input is a per-element mask
+                return intrinsic->Op(1)->IsVectorPerElementMask(simdBaseType, simdSize);
+            }
+
+            default:
+            {
+                assert(!GenTreeHWIntrinsic::OperIsBitwiseHWIntrinsic(oper));
+                break;
+            }
+        }
+
+        return false;
+    }
+#endif // FEATURE_SIMD
+
+    return false;
+}
+
 //------------------------------------------------------------------------
 // IsNeverNegative: returns true if the given tree is known to be never
 //                  negative, i. e. the upper bit will always be zero.

src/coreclr/jit/gentree.h

@@ -2318,6 +2318,7 @@ struct GenTree
     bool Precedes(GenTree* other);
 
     bool IsInvariant() const;
+    bool IsVectorPerElementMask(var_types simdBaseType, unsigned simdSize) const;
 
     bool IsNeverNegative(Compiler* comp) const;
     bool IsNeverNegativeOne(Compiler* comp) const;

src/coreclr/jit/lowerxarch.cpp

@@ -3034,12 +3034,12 @@ GenTree* Lowering::LowerHWIntrinsicCndSel(GenTreeHWIntrinsic* node)
     GenTree* op3 = node->Op(3);
 
     // If the condition vector comes from a hardware intrinsic that
-    // returns a per-element mask (marked with HW_Flag_ReturnsPerElementMask),
-    // we can optimize the entire conditional select to
-    // a single BlendVariable instruction (if supported by the architecture)
+    // returns a per-element mask, we can optimize the entire
+    // conditional select to a single BlendVariable instruction
+    // (if supported by the architecture)
 
     // First, determine if the condition is a per-element mask
-    if (op1->OperIsHWIntrinsic() && HWIntrinsicInfo::ReturnsPerElementMask(op1->AsHWIntrinsic()->GetHWIntrinsicId()))
+    if (op1->IsVectorPerElementMask(simdBaseType, simdSize))
     {
         // Next, determine if the target architecture supports BlendVariable
         NamedIntrinsic blendVariableId = NI_Illegal;

src/coreclr/jit/simd.h

@@ -15,6 +15,17 @@ static bool ElementsAreSame(T* array, size_t size)
     return true;
 }
 
+template <typename T>
+static bool ElementsAreAllBitsSetOrZero(T* array, size_t size)
+{
+    for (size_t i = 0; i < size; i++)
+    {
+        if (array[i] != static_cast<T>(0) && array[i] != static_cast<T>(~0))
+            return false;
+    }
+    return true;
+}
+
 struct simd8_t
 {
     union