ir: Add opcodes form unsigned saturated accumulations of signed values

src/backend/x64/emit_x64_vector.cpp

@@ -3445,6 +3445,73 @@ void EmitX64::EmitVectorUnsignedRecipSqrtEstimate(EmitContext& ctx, IR::Inst* in
     });
 }
 
+// Simple generic case for 8, 16, and 32-bit values. 64-bit values
+// will need to be special-cased as we can't simply use a larger integral size.
+template <typename T, typename U = std::make_unsigned_t<T>>
+bool EmitVectorUnsignedSaturatedAccumulateSigned(VectorArray<U>& result, const VectorArray<T>& lhs, const VectorArray<T>& rhs) {
+    static_assert(std::is_signed_v<T>, "T must be signed.");
+    static_assert(sizeof(T) < 64, "T must be less than 64 bits in size.");
+
+    bool qc_flag = false;
+
+    for (size_t i = 0; i < result.size(); i++) {
+        // We treat rhs' members as unsigned, so cast to unsigned before signed to inhibit sign-extension.
+        // We use the unsigned equivalent of T, as we want zero-extension to occur, rather than a plain move.
+        const s64 x = s64{lhs[i]};
+        const s64 y = static_cast<s64>(static_cast<std::make_unsigned_t<U>>(rhs[i]));
+        const s64 sum = x + y;
+
+        if (sum > std::numeric_limits<U>::max()) {
+            result[i] = std::numeric_limits<U>::max();
+            qc_flag = true;
+        } else if (sum < 0) {
+            result[i] = std::numeric_limits<U>::min();
+            qc_flag = true;
+        } else {
+            result[i] = static_cast<U>(sum);
+        }
+    }
+
+    return qc_flag;
+}
+
+void EmitX64::EmitVectorUnsignedSaturatedAccumulateSigned8(EmitContext& ctx, IR::Inst* inst) {
+    EmitTwoArgumentFallbackWithSaturation(code, ctx, inst, EmitVectorUnsignedSaturatedAccumulateSigned<s8>);
+}
+
+void EmitX64::EmitVectorUnsignedSaturatedAccumulateSigned16(EmitContext& ctx, IR::Inst* inst) {
+    EmitTwoArgumentFallbackWithSaturation(code, ctx, inst, EmitVectorUnsignedSaturatedAccumulateSigned<s16>);
+}
+
+void EmitX64::EmitVectorUnsignedSaturatedAccumulateSigned32(EmitContext& ctx, IR::Inst* inst) {
+    EmitTwoArgumentFallbackWithSaturation(code, ctx, inst, EmitVectorUnsignedSaturatedAccumulateSigned<s32>);
+}
+
+void EmitX64::EmitVectorUnsignedSaturatedAccumulateSigned64([[maybe_unused]] EmitContext& ctx, [[maybe_unused]] IR::Inst* inst) {
+    EmitTwoArgumentFallbackWithSaturation(code, ctx, inst, [](VectorArray<u64>& result, const VectorArray<u64>& lhs, const VectorArray<u64>& rhs) {
+        bool qc_flag = false;
+
+        for (size_t i = 0; i < result.size(); i++) {
+            const u64 x = lhs[i];
+            const u64 y = rhs[i];
+            const u64 res = x + y;
+
+            // Check sign bits to determine if an overflow occurred.
+            if ((~x & y & ~res) & 0x8000000000000000) {
+                result[i] = UINT64_MAX;
+                qc_flag = true;
+            } else if ((x & ~y & res) & 0x8000000000000000) {
+                result[i] = 0;
+                qc_flag = true;
+            } else {
+                result[i] = res;
+            }
+        }
+
+        return qc_flag;
+    });
+}
+
 void EmitX64::EmitVectorUnsignedSaturatedNarrow16(EmitContext& ctx, IR::Inst* inst) {
     EmitOneArgumentFallbackWithSaturation(code, ctx, inst, [](VectorArray<u8>& result, const VectorArray<u16>& a) {
         bool qc_flag = false;

src/frontend/ir/ir_emitter.cpp

@@ -1642,6 +1642,21 @@ U128 IREmitter::VectorUnsignedRecipSqrtEstimate(const U128& a) {
     return Inst<U128>(Opcode::VectorUnsignedRecipSqrtEstimate, a);
 }
 
+U128 IREmitter::VectorUnsignedSaturatedAccumulateSigned(size_t esize, const U128& a, const U128& b) {
+    switch (esize) {
+    case 8:
+        return Inst<U128>(Opcode::VectorUnsignedSaturatedAccumulateSigned8, a, b);
+    case 16:
+        return Inst<U128>(Opcode::VectorUnsignedSaturatedAccumulateSigned16, a, b);
+    case 32:
+        return Inst<U128>(Opcode::VectorUnsignedSaturatedAccumulateSigned32, a, b);
+    case 64:
+        return Inst<U128>(Opcode::VectorUnsignedSaturatedAccumulateSigned64, a, b);
+    }
+    UNREACHABLE();
+    return {};
+}
+
 U128 IREmitter::VectorUnsignedSaturatedNarrow(size_t esize, const U128& a) {
     switch (esize) {
     case 16:

src/frontend/ir/ir_emitter.h

@@ -276,6 +276,7 @@ public:
     U128 VectorUnsignedAbsoluteDifference(size_t esize, const U128& a, const U128& b);
     U128 VectorUnsignedRecipEstimate(const U128& a);
     U128 VectorUnsignedRecipSqrtEstimate(const U128& a);
+    U128 VectorUnsignedSaturatedAccumulateSigned(size_t esize, const U128& a, const U128& b);
     U128 VectorUnsignedSaturatedNarrow(size_t esize, const U128& a);
     U128 VectorZeroExtend(size_t original_esize, const U128& a);
     U128 VectorZeroUpper(const U128& a);

src/frontend/ir/microinstruction.cpp

@@ -367,6 +367,10 @@ bool Inst::WritesToFPSRCumulativeSaturationBit() const {
     case Opcode::VectorSignedSaturatedNeg16:
     case Opcode::VectorSignedSaturatedNeg32:
     case Opcode::VectorSignedSaturatedNeg64:
+    case Opcode::VectorUnsignedSaturatedAccumulateSigned8:
+    case Opcode::VectorUnsignedSaturatedAccumulateSigned16:
+    case Opcode::VectorUnsignedSaturatedAccumulateSigned32:
+    case Opcode::VectorUnsignedSaturatedAccumulateSigned64:
     case Opcode::VectorUnsignedSaturatedNarrow16:
     case Opcode::VectorUnsignedSaturatedNarrow32:
     case Opcode::VectorUnsignedSaturatedNarrow64:

src/frontend/ir/opcodes.inc

@@ -424,6 +424,10 @@ OPCODE(VectorUnsignedAbsoluteDifference16,                 U128,           U128,
 OPCODE(VectorUnsignedAbsoluteDifference32,                 U128,           U128,           U128                                            )
 OPCODE(VectorUnsignedRecipEstimate,                        U128,           U128                                                            )
 OPCODE(VectorUnsignedRecipSqrtEstimate,                    U128,           U128                                                            )
+OPCODE(VectorUnsignedSaturatedAccumulateSigned8,           U128,           U128,           U128                                            )
+OPCODE(VectorUnsignedSaturatedAccumulateSigned16,          U128,           U128,           U128                                            )
+OPCODE(VectorUnsignedSaturatedAccumulateSigned32,          U128,           U128,           U128                                            )
+OPCODE(VectorUnsignedSaturatedAccumulateSigned64,          U128,           U128,           U128                                            )
 OPCODE(VectorUnsignedSaturatedNarrow16,                    U128,           U128                                                            )
 OPCODE(VectorUnsignedSaturatedNarrow32,                    U128,           U128                                                            )
 OPCODE(VectorUnsignedSaturatedNarrow64,                    U128,           U128                                                            )