diff --git a/src/backend_x64/emit_x64_vector.cpp b/src/backend_x64/emit_x64_vector.cpp
index 23780db7..0145931f 100644
--- a/src/backend_x64/emit_x64_vector.cpp
+++ b/src/backend_x64/emit_x64_vector.cpp
@@ -849,6 +849,92 @@ void EmitX64::EmitVectorHalvingAddU32(EmitContext& ctx, IR::Inst* inst) {
     EmitVectorHalvingAddUnsigned(32, ctx, inst, code);
 }
 
+static void EmitVectorHalvingSubSigned(size_t esize, EmitContext& ctx, IR::Inst* inst, BlockOfCode& code) {
+    auto args = ctx.reg_alloc.GetArgumentInfo(inst);
+
+    const Xbyak::Xmm a = ctx.reg_alloc.UseScratchXmm(args[0]);
+    const Xbyak::Xmm b = ctx.reg_alloc.UseScratchXmm(args[1]);
+
+    switch (esize) {
+    case 8: {
+        const Xbyak::Xmm tmp = ctx.reg_alloc.ScratchXmm();
+        code.movdqa(tmp, code.MConst(xword, 0x8080808080808080, 0x8080808080808080));
+        code.pxor(a, tmp);
+        code.pxor(b, tmp);
+        code.pavgb(b, a);
+        code.psubb(a, b);
+        break;
+    }
+    case 16: {
+        const Xbyak::Xmm tmp = ctx.reg_alloc.ScratchXmm();
+        code.movdqa(tmp, code.MConst(xword, 0x8000800080008000, 0x8000800080008000));
+        code.pxor(a, tmp);
+        code.pxor(b, tmp);
+        code.pavgw(b, a);
+        code.psubw(a, b);
+        break;
+    }
+    case 32:
+        code.pxor(a, b);
+        code.pand(b, a);
+        code.psrad(a, 1);
+        code.psubd(a, b);
+        break;
+    }
+
+    ctx.reg_alloc.DefineValue(inst, a);
+}
+
+void EmitX64::EmitVectorHalvingSubS8(EmitContext& ctx, IR::Inst* inst) {
+    EmitVectorHalvingSubSigned(8, ctx, inst, code);
+}
+
+void EmitX64::EmitVectorHalvingSubS16(EmitContext& ctx, IR::Inst* inst) {
+    EmitVectorHalvingSubSigned(16, ctx, inst, code);
+}
+
+void EmitX64::EmitVectorHalvingSubS32(EmitContext& ctx, IR::Inst* inst) {
+    EmitVectorHalvingSubSigned(32, ctx, inst, code);
+}
+
+static void EmitVectorHalvingSubUnsigned(size_t esize, EmitContext& ctx, IR::Inst* inst, BlockOfCode& code) {
+    auto args = ctx.reg_alloc.GetArgumentInfo(inst);
+
+    const Xbyak::Xmm a = ctx.reg_alloc.UseScratchXmm(args[0]);
+    const Xbyak::Xmm b = ctx.reg_alloc.UseScratchXmm(args[1]);
+
+    switch (esize) {
+    case 8:
+        code.pavgb(b, a);
+        code.psubb(a, b);
+        break;
+    case 16:
+        code.pavgw(b, a);
+        code.psubw(a, b);
+        break;
+    case 32:
+        code.pxor(a, b);
+        code.pand(b, a);
+        code.psrld(a, 1);
+        code.psubd(a, b);
+        break;
+    }
+
+    ctx.reg_alloc.DefineValue(inst, a);
+}
+
+void EmitX64::EmitVectorHalvingSubU8(EmitContext& ctx, IR::Inst* inst) {
+    EmitVectorHalvingSubUnsigned(8, ctx, inst, code);
+}
+
+void EmitX64::EmitVectorHalvingSubU16(EmitContext& ctx, IR::Inst* inst) {
+    EmitVectorHalvingSubUnsigned(16, ctx, inst, code);
+}
+
+void EmitX64::EmitVectorHalvingSubU32(EmitContext& ctx, IR::Inst* inst) {
+    EmitVectorHalvingSubUnsigned(32, ctx, inst, code);
+}
+
 static void EmitVectorInterleaveLower(BlockOfCode& code, EmitContext& ctx, IR::Inst* inst, int size) {
     auto args = ctx.reg_alloc.GetArgumentInfo(inst);
 
diff --git a/src/frontend/ir/ir_emitter.cpp b/src/frontend/ir/ir_emitter.cpp
index 69acfd15..87a32049 100644
--- a/src/frontend/ir/ir_emitter.cpp
+++ b/src/frontend/ir/ir_emitter.cpp
@@ -912,6 +912,32 @@ U128 IREmitter::VectorHalvingAddUnsigned(size_t esize, const U128& a, const U128
     return {};
 }
 
+U128 IREmitter::VectorHalvingSubSigned(size_t esize, const U128& a, const U128& b) {
+    switch (esize) {
+    case 8:
+        return Inst<U128>(Opcode::VectorHalvingSubS8, a, b);
+    case 16:
+        return Inst<U128>(Opcode::VectorHalvingSubS16, a, b);
+    case 32:
+        return Inst<U128>(Opcode::VectorHalvingSubS32, a, b);
+    }
+    UNREACHABLE();
+    return {};
+}
+
+U128 IREmitter::VectorHalvingSubUnsigned(size_t esize, const U128& a, const U128& b) {
+    switch (esize) {
+    case 8:
+        return Inst<U128>(Opcode::VectorHalvingSubU8, a, b);
+    case 16:
+        return Inst<U128>(Opcode::VectorHalvingSubU16, a, b);
+    case 32:
+        return Inst<U128>(Opcode::VectorHalvingSubU32, a, b);
+    }
+    UNREACHABLE();
+    return {};
+}
+
 U128 IREmitter::VectorInterleaveLower(size_t esize, const U128& a, const U128& b) {
     switch (esize) {
     case 8:
diff --git a/src/frontend/ir/ir_emitter.h b/src/frontend/ir/ir_emitter.h
index 8d2c86d5..a95c7010 100644
--- a/src/frontend/ir/ir_emitter.h
+++ b/src/frontend/ir/ir_emitter.h
@@ -210,6 +210,8 @@ public:
     U128 VectorGreaterUnsigned(size_t esize, const U128& a, const U128& b);
     U128 VectorHalvingAddSigned(size_t esize, const U128& a, const U128& b);
     U128 VectorHalvingAddUnsigned(size_t esize, const U128& a, const U128& b);
+    U128 VectorHalvingSubSigned(size_t esize, const U128& a, const U128& b);
+    U128 VectorHalvingSubUnsigned(size_t esize, const U128& a, const U128& b);
     U128 VectorInterleaveLower(size_t esize, const U128& a, const U128& b);
     U128 VectorInterleaveUpper(size_t esize, const U128& a, const U128& b);
     U128 VectorLessEqualSigned(size_t esize, const U128& a, const U128& b);
diff --git a/src/frontend/ir/opcodes.inc b/src/frontend/ir/opcodes.inc
index ba2960d8..48901af0 100644
--- a/src/frontend/ir/opcodes.inc
+++ b/src/frontend/ir/opcodes.inc
@@ -261,6 +261,12 @@ OPCODE(VectorHalvingAddS32,                 T::U128,        T::U128,        T::U
 OPCODE(VectorHalvingAddU8,                  T::U128,        T::U128,        T::U128                         )
 OPCODE(VectorHalvingAddU16,                 T::U128,        T::U128,        T::U128                         )
 OPCODE(VectorHalvingAddU32,                 T::U128,        T::U128,        T::U128                         )
+OPCODE(VectorHalvingSubS8,                  T::U128,        T::U128,        T::U128                         )
+OPCODE(VectorHalvingSubS16,                 T::U128,        T::U128,        T::U128                         )
+OPCODE(VectorHalvingSubS32,                 T::U128,        T::U128,        T::U128                         )
+OPCODE(VectorHalvingSubU8,                  T::U128,        T::U128,        T::U128                         )
+OPCODE(VectorHalvingSubU16,                 T::U128,        T::U128,        T::U128                         )
+OPCODE(VectorHalvingSubU32,                 T::U128,        T::U128,        T::U128                         )
 OPCODE(VectorInterleaveLower8,              T::U128,        T::U128,        T::U128                         )
 OPCODE(VectorInterleaveLower16,             T::U128,        T::U128,        T::U128                         )
 OPCODE(VectorInterleaveLower32,             T::U128,        T::U128,        T::U128                         )