Merge pull request #473 from lioncash/sqshlu

A64: Implement SQSHLU

src/backend/x64/emit_x64_vector.cpp

@@ -3871,6 +3871,60 @@ void EmitX64::EmitVectorSignedSaturatedShiftLeft64(EmitContext& ctx, IR::Inst* i
     EmitTwoArgumentFallbackWithSaturation(code, ctx, inst, VectorSignedSaturatedShiftLeft<s64>);
 }
 
+template <typename T, typename U = std::make_unsigned_t<T>>
+static bool VectorSignedSaturatedShiftLeftUnsigned(VectorArray<T>& dst, const VectorArray<T>& data, const VectorArray<T>& shift_values) {
+    static_assert(std::is_signed_v<T>, "T must be signed.");
+
+    constexpr size_t bit_size_minus_one = Common::BitSize<T>() - 1;
+
+    bool qc_flag = false;
+    for (size_t i = 0; i < dst.size(); i++) {
+        const T element = data[i];
+        const T shift = std::clamp<T>(static_cast<T>(Common::SignExtend<8>(shift_values[i] & 0xFF)),
+                                      -static_cast<T>(bit_size_minus_one), std::numeric_limits<T>::max());
+
+        if (element == 0) {
+            dst[i] = 0;
+        } else if (element < 0) {
+            dst[i] = 0;
+            qc_flag = true;
+        } else if (shift < 0) {
+            dst[i] = static_cast<T>(element >> -shift);
+        } else if (static_cast<U>(shift) > bit_size_minus_one) {
+            dst[i] = static_cast<T>(std::numeric_limits<U>::max());
+            qc_flag = true;
+        } else {
+            const U shifted = static_cast<U>(element) << static_cast<U>(shift);
+            const U shifted_test = shifted >> static_cast<U>(shift);
+
+            if (shifted_test != static_cast<U>(element)) {
+                dst[i] = static_cast<T>(std::numeric_limits<U>::max());
+                qc_flag = true;
+            } else {
+                dst[i] = shifted;
+            }
+        }
+    }
+
+    return qc_flag;
+}
+
+void EmitX64::EmitVectorSignedSaturatedShiftLeftUnsigned8(EmitContext& ctx, IR::Inst* inst) {
+    EmitTwoArgumentFallbackWithSaturation(code, ctx, inst, VectorSignedSaturatedShiftLeftUnsigned<s8>);
+}
+
+void EmitX64::EmitVectorSignedSaturatedShiftLeftUnsigned16(EmitContext& ctx, IR::Inst* inst) {
+    EmitTwoArgumentFallbackWithSaturation(code, ctx, inst, VectorSignedSaturatedShiftLeftUnsigned<s16>);
+}
+
+void EmitX64::EmitVectorSignedSaturatedShiftLeftUnsigned32(EmitContext& ctx, IR::Inst* inst) {
+    EmitTwoArgumentFallbackWithSaturation(code, ctx, inst, VectorSignedSaturatedShiftLeftUnsigned<s32>);
+}
+
+void EmitX64::EmitVectorSignedSaturatedShiftLeftUnsigned64(EmitContext& ctx, IR::Inst* inst) {
+    EmitTwoArgumentFallbackWithSaturation(code, ctx, inst, VectorSignedSaturatedShiftLeftUnsigned<s64>);
+}
+
 void EmitX64::EmitVectorSub8(EmitContext& ctx, IR::Inst* inst) {
     EmitVectorOperation(code, ctx, inst, &Xbyak::CodeGenerator::psubb);
 }

src/frontend/A64/decoder/a64.inc

@@ -523,7 +523,7 @@ INST(URSHR_1,                "URSHR",                                     "01111
 INST(URSRA_1,                "URSRA",                                     "011111110IIIIiii001101nnnnnddddd")
 INST(SRI_1,                  "SRI",                                       "011111110IIIIiii010001nnnnnddddd")
 INST(SLI_1,                  "SLI",                                       "011111110IIIIiii010101nnnnnddddd")
-//INST(SQSHLU_1,               "SQSHLU",                                    "011111110IIIIiii011001nnnnnddddd")
+INST(SQSHLU_1,               "SQSHLU",                                    "011111110IIIIiii011001nnnnnddddd")
 INST(UQSHL_imm_1,            "UQSHL (immediate)",                         "011111110IIIIiii011101nnnnnddddd")
 INST(SQSHRUN_1,              "SQSHRUN, SQSHRUN2",                         "011111110IIIIiii100001nnnnnddddd")
 //INST(SQRSHRUN_1,             "SQRSHRUN, SQRSHRUN2",                       "011111110IIIIiii100011nnnnnddddd")
@@ -855,7 +855,7 @@ INST(URSHR_2,                "URSHR",                                     "0Q101
 INST(URSRA_2,                "URSRA",                                     "0Q1011110IIIIiii001101nnnnnddddd")
 INST(SRI_2,                  "SRI",                                       "0Q1011110IIIIiii010001nnnnnddddd")
 INST(SLI_2,                  "SLI",                                       "0Q1011110IIIIiii010101nnnnnddddd")
-//INST(SQSHLU_2,               "SQSHLU",                                    "0Q1011110IIIIiii011001nnnnnddddd")
+INST(SQSHLU_2,               "SQSHLU",                                    "0Q1011110IIIIiii011001nnnnnddddd")
 INST(UQSHL_imm_2,            "UQSHL (immediate)",                         "0Q1011110IIIIiii011101nnnnnddddd")
 INST(SQSHRUN_2,              "SQSHRUN, SQSHRUN2",                         "0Q1011110IIIIiii100001nnnnnddddd")
 INST(SQRSHRUN_2,             "SQRSHRUN, SQRSHRUN2",                       "0Q1011110IIIIiii100011nnnnnddddd")

src/frontend/A64/translate/impl/simd_scalar_shift_by_immediate.cpp

@@ -15,6 +15,12 @@ enum class Narrowing {
     SaturateToSigned,
 };
 
+enum class SaturatingShiftLeftType {
+    Signed,
+    Unsigned,
+    SignedWithUnsignedSaturation,
+};
+
 enum class ShiftExtraBehavior {
     None,
     Accumulate,
@@ -30,7 +36,7 @@ enum class FloatConversionDirection {
     FloatToFixed,
 };
 
-bool SaturatingShiftLeft(TranslatorVisitor& v, Imm<4> immh, Imm<3> immb, Vec Vn, Vec Vd, Signedness sign) {
+bool SaturatingShiftLeft(TranslatorVisitor& v, Imm<4> immh, Imm<3> immb, Vec Vn, Vec Vd, SaturatingShiftLeftType type) {
     if (immh == 0b0000) {
         return v.ReservedValue();
     }
@@ -40,11 +46,16 @@ bool SaturatingShiftLeft(TranslatorVisitor& v, Imm<4> immh, Imm<3> immb, Vec Vn,
 
     const IR::U128 operand = v.ir.ZeroExtendToQuad(v.V_scalar(esize, Vn));
     const IR::U128 shift = v.ir.ZeroExtendToQuad(v.I(esize, shift_amount));
-    const IR::U128 result = [&v, esize, operand, shift, sign] {
+    const IR::U128 result = [&v, esize, operand, shift, type] {
-        if (sign == Signedness::Signed) {
+        if (type == SaturatingShiftLeftType::Signed) {
             return v.ir.VectorSignedSaturatedShiftLeft(esize, operand, shift);
         }
+
+        if (type == SaturatingShiftLeftType::Unsigned) {
             return v.ir.VectorUnsignedSaturatedShiftLeft(esize, operand, shift);
+        }
+
+        return v.ir.VectorSignedSaturatedShiftLeftUnsigned(esize, operand, shift);
     }();
 
     v.ir.SetQ(Vd, result);
@@ -275,7 +286,11 @@ bool TranslatorVisitor::SRI_1(Imm<4> immh, Imm<3> immb, Vec Vn, Vec Vd) {
 }
 
 bool TranslatorVisitor::SQSHL_imm_1(Imm<4> immh, Imm<3> immb, Vec Vn, Vec Vd) {
-    return SaturatingShiftLeft(*this, immh, immb, Vn, Vd, Signedness::Signed);
+    return SaturatingShiftLeft(*this, immh, immb, Vn, Vd, SaturatingShiftLeftType::Signed);
+}
+
+bool TranslatorVisitor::SQSHLU_1(Imm<4> immh, Imm<3> immb, Vec Vn, Vec Vd) {
+    return SaturatingShiftLeft(*this, immh, immb, Vn, Vd, SaturatingShiftLeftType::SignedWithUnsignedSaturation);
 }
 
 bool TranslatorVisitor::SQSHRN_1(Imm<4> immh, Imm<3> immb, Vec Vn, Vec Vd) {
@@ -318,7 +333,7 @@ bool TranslatorVisitor::SHL_1(Imm<4> immh, Imm<3> immb, Vec Vn, Vec Vd) {
 }
 
 bool TranslatorVisitor::UQSHL_imm_1(Imm<4> immh, Imm<3> immb, Vec Vn, Vec Vd) {
-    return SaturatingShiftLeft(*this, immh, immb, Vn, Vd, Signedness::Unsigned);
+    return SaturatingShiftLeft(*this, immh, immb, Vn, Vd, SaturatingShiftLeftType::Unsigned);
 }
 
 bool TranslatorVisitor::UQSHRN_1(Imm<4> immh, Imm<3> immb, Vec Vn, Vec Vd) {

src/frontend/A64/translate/impl/simd_shift_by_immediate.cpp

@@ -31,6 +31,12 @@ enum class Narrowing {
     SaturateToSigned,
 };
 
+enum class SaturatingShiftLeftType {
+    Signed,
+    Unsigned,
+    SignedWithUnsignedSaturation,
+};
+
 enum class FloatConversionDirection {
     FixedToFloat,
     FloatToFixed,
@@ -160,7 +166,7 @@ bool ShiftLeftLong(TranslatorVisitor& v, bool Q, Imm<4> immh, Imm<3> immb, Vec V
     return true;
 }
 
-bool SaturatingShiftLeft(TranslatorVisitor& v, bool Q, Imm<4> immh, Imm<3> immb, Vec Vn, Vec Vd, Signedness sign) {
+bool SaturatingShiftLeft(TranslatorVisitor& v, bool Q, Imm<4> immh, Imm<3> immb, Vec Vn, Vec Vd, SaturatingShiftLeftType type) {
     if (!Q && immh.Bit<3>()) {
         return v.ReservedValue();
     }
@@ -172,11 +178,15 @@ bool SaturatingShiftLeft(TranslatorVisitor& v, bool Q, Imm<4> immh, Imm<3> immb,
     const IR::U128 operand = v.V(datasize, Vn);
     const IR::U128 shift_vec = v.ir.VectorBroadcast(esize, v.I(esize, shift));
     const IR::U128 result = [&] {
-        if (sign == Signedness::Signed) {
+        if (type == SaturatingShiftLeftType::Signed) {
             return v.ir.VectorSignedSaturatedShiftLeft(esize, operand, shift_vec);
         }
 
+        if (type == SaturatingShiftLeftType::Unsigned) {
             return v.ir.VectorUnsignedSaturatedShiftLeft(esize, operand, shift_vec);
+        }
+
+        return v.ir.VectorSignedSaturatedShiftLeftUnsigned(esize, operand, shift_vec);
     }();
 
     v.V(datasize, Vd, result);
@@ -267,7 +277,11 @@ bool TranslatorVisitor::RSHRN(bool Q, Imm<4> immh, Imm<3> immb, Vec Vn, Vec Vd)
 }
 
 bool TranslatorVisitor::SQSHL_imm_2(bool Q, Imm<4> immh, Imm<3> immb, Vec Vn, Vec Vd) {
-    return SaturatingShiftLeft(*this, Q, immh, immb, Vn, Vd, Signedness::Signed);
+    return SaturatingShiftLeft(*this, Q, immh, immb, Vn, Vd, SaturatingShiftLeftType::Signed);
+}
+
+bool TranslatorVisitor::SQSHLU_2(bool Q, Imm<4> immh, Imm<3> immb, Vec Vn, Vec Vd) {
+    return SaturatingShiftLeft(*this, Q, immh, immb, Vn, Vd, SaturatingShiftLeftType::SignedWithUnsignedSaturation);
 }
 
 bool TranslatorVisitor::SQSHRN_2(bool Q, Imm<4> immh, Imm<3> immb, Vec Vn, Vec Vd) {
@@ -287,7 +301,7 @@ bool TranslatorVisitor::SQRSHRUN_2(bool Q, Imm<4> immh, Imm<3> immb, Vec Vn, Vec
 }
 
 bool TranslatorVisitor::UQSHL_imm_2(bool Q, Imm<4> immh, Imm<3> immb, Vec Vn, Vec Vd) {
-    return SaturatingShiftLeft(*this, Q, immh, immb, Vn, Vd, Signedness::Unsigned);
+    return SaturatingShiftLeft(*this, Q, immh, immb, Vn, Vd, SaturatingShiftLeftType::Unsigned);
 }
 
 bool TranslatorVisitor::UQSHRN_2(bool Q, Imm<4> immh, Imm<3> immb, Vec Vn, Vec Vd) {

src/frontend/ir/ir_emitter.cpp

@@ -1661,6 +1661,21 @@ U128 IREmitter::VectorSignedSaturatedShiftLeft(size_t esize, const U128& a, cons
     return {};
 }
 
+U128 IREmitter::VectorSignedSaturatedShiftLeftUnsigned(size_t esize, const U128& a, const U128& b) {
+    switch (esize) {
+    case 8:
+        return Inst<U128>(Opcode::VectorSignedSaturatedShiftLeftUnsigned8, a, b);
+    case 16:
+        return Inst<U128>(Opcode::VectorSignedSaturatedShiftLeftUnsigned16, a, b);
+    case 32:
+        return Inst<U128>(Opcode::VectorSignedSaturatedShiftLeftUnsigned32, a, b);
+    case 64:
+        return Inst<U128>(Opcode::VectorSignedSaturatedShiftLeftUnsigned64, a, b);
+    }
+    UNREACHABLE();
+    return {};
+}
+
 U128 IREmitter::VectorSub(size_t esize, const U128& a, const U128& b) {
     switch (esize) {
     case 8:

src/frontend/ir/ir_emitter.h

@@ -279,6 +279,7 @@ public:
     U128 VectorSignedSaturatedNarrowToUnsigned(size_t original_esize, const U128& a);
     U128 VectorSignedSaturatedNeg(size_t esize, const U128& a);
     U128 VectorSignedSaturatedShiftLeft(size_t esize, const U128& a, const U128& b);
+    U128 VectorSignedSaturatedShiftLeftUnsigned(size_t esize, const U128& a, const U128& b);
     U128 VectorSub(size_t esize, const U128& a, const U128& b);
     Table VectorTable(std::vector<U128> values);
     U128 VectorTableLookup(const U128& defaults, const Table& table, const U128& indices);

src/frontend/ir/microinstruction.cpp

@@ -384,6 +384,10 @@ bool Inst::WritesToFPSRCumulativeSaturationBit() const {
     case Opcode::VectorSignedSaturatedShiftLeft16:
     case Opcode::VectorSignedSaturatedShiftLeft32:
     case Opcode::VectorSignedSaturatedShiftLeft64:
+    case Opcode::VectorSignedSaturatedShiftLeftUnsigned8:
+    case Opcode::VectorSignedSaturatedShiftLeftUnsigned16:
+    case Opcode::VectorSignedSaturatedShiftLeftUnsigned32:
+    case Opcode::VectorSignedSaturatedShiftLeftUnsigned64:
     case Opcode::VectorUnsignedSaturatedAccumulateSigned8:
     case Opcode::VectorUnsignedSaturatedAccumulateSigned16:
     case Opcode::VectorUnsignedSaturatedAccumulateSigned32:

src/frontend/ir/opcodes.inc

@@ -428,6 +428,10 @@ OPCODE(VectorSignedSaturatedShiftLeft8,                     U128,           U128
 OPCODE(VectorSignedSaturatedShiftLeft16,                    U128,           U128,           U128                                            )
 OPCODE(VectorSignedSaturatedShiftLeft32,                    U128,           U128,           U128                                            )
 OPCODE(VectorSignedSaturatedShiftLeft64,                    U128,           U128,           U128                                            )
+OPCODE(VectorSignedSaturatedShiftLeftUnsigned8,             U128,           U128,           U128                                            )
+OPCODE(VectorSignedSaturatedShiftLeftUnsigned16,            U128,           U128,           U128                                            )
+OPCODE(VectorSignedSaturatedShiftLeftUnsigned32,            U128,           U128,           U128                                            )
+OPCODE(VectorSignedSaturatedShiftLeftUnsigned64,            U128,           U128,           U128                                            )
 OPCODE(VectorSub8,                                          U128,           U128,           U128                                            )
 OPCODE(VectorSub16,                                         U128,           U128,           U128                                            )
 OPCODE(VectorSub32,                                         U128,           U128,           U128                                            )