koffi 2.15.0 → 2.16.0-beta.1

package/src/koffi/src/abi_x64_win.cc

@@ -29,8 +29,6 @@ extern "C" napi_value CallSwitchStack(Napi::Function *func, size_t argc, napi_va
                                       uint8_t *old_sp, Span<uint8_t> *new_stack,
                                       napi_value (*call)(Napi::Function *func, size_t argc, napi_value *argv));
 
-#include "trampolines/prototypes.inc"
-
 bool AnalyseFunction(Napi::Env, InstanceData *, FunctionInfo *func)
 {
     func->ret.regular = IsRegularSize(func->ret.type->size, 8);
@@ -47,154 +45,212 @@ bool AnalyseFunction(Napi::Env, InstanceData *, FunctionInfo *func)
 
 bool CallData::Prepare(const FunctionInfo *func, const Napi::CallbackInfo &info)
 {
-    uint64_t *args_ptr = nullptr;
+    uint64_t *args_ptr = AllocStack<uint64_t>(func->args_size);
 
-    // Pass return value in register or through memory
-    if (!AllocStack(func->args_size, 16, &args_ptr)) [[unlikely]]
+    if (!args_ptr) [[unlikely]]
         return false;
     if (!func->ret.regular) {
         return_ptr = AllocHeap(func->ret.type->size, 16);
         *(uint8_t **)(args_ptr++) = return_ptr;
     }
 
+    Size i = -1;
+
+#if defined(__GNUC__) || defined(__clang__)
+    static const void *const DispatchTable[] = {
+        #define PRIMITIVE(Name) && Name,
+        #include "primitives.inc"
+    };
+
+    #define LOOP
+    #define CASE(Primitive) \
+        do { \
+            PrimitiveKind next = func->primitives[++i]; \
+            goto *DispatchTable[(int)next]; \
+        } while (false); \
+        Primitive:
+    #define OR(Primitive) \
+        Primitive:
+#else
+    #define LOOP \
+        while (++i < func->parameters.len) \
+            switch (func->primitives[i])
+    #define CASE(Primitive) \
+        break; \
+        case PrimitiveKind::Primitive:
+    #define OR(Primitive) \
+        case PrimitiveKind::Primitive:
+#endif
+
 #define PUSH_INTEGER(CType) \
         do { \
-            if (!value.IsNumber() && !value.IsBigInt()) [[unlikely]] { \
+            const ParameterInfo &param = func->parameters[i]; \
+            Napi::Value value = info[param.offset]; \
+             \
+            CType v; \
+            if (!TryNumber(value, &v)) [[unlikely]] { \
                 ThrowError<Napi::TypeError>(env, "Unexpected %1 value, expected number", GetValueType(instance, value)); \
                 return false; \
             } \
              \
-            CType v = GetNumber<CType>(value); \
             *(args_ptr++) = (uint64_t)v; \
         } while (false)
 #define PUSH_INTEGER_SWAP(CType) \
         do { \
-            if (!value.IsNumber() && !value.IsBigInt()) [[unlikely]] { \
+            const ParameterInfo &param = func->parameters[i]; \
+            Napi::Value value = info[param.offset]; \
+             \
+            CType v; \
+            if (!TryNumber(value, &v)) [[unlikely]] { \
                 ThrowError<Napi::TypeError>(env, "Unexpected %1 value, expected number", GetValueType(instance, value)); \
                 return false; \
             } \
              \
-            CType v = GetNumber<CType>(value); \
             *(args_ptr++) = (uint64_t)ReverseBytes(v); \
         } while (false)
 
-    // Push arguments
-    for (Size i = 0; i < func->parameters.len; i++) {
-        const ParameterInfo &param = func->parameters[i];
-        K_ASSERT(param.directions >= 1 && param.directions <= 3);
+    LOOP {
+        CASE(Void) { K_UNREACHABLE(); };
 
-        Napi::Value value = info[param.offset];
+        CASE(Bool) {
+            const ParameterInfo &param = func->parameters[i];
+            Napi::Value value = info[param.offset];
 
-        switch (param.type->primitive) {
-            case PrimitiveKind::Void: { K_UNREACHABLE(); } break;
+            bool b;
+            if (napi_get_value_bool(env, value, &b) != napi_ok) [[unlikely]] {
+                ThrowError<Napi::TypeError>(env, "Unexpected %1 value, expected boolean", GetValueType(instance, value));
+                return false;
+            }
 
-            case PrimitiveKind::Bool: {
-                if (!value.IsBoolean()) [[unlikely]] {
-                    ThrowError<Napi::TypeError>(env, "Unexpected %1 value, expected boolean", GetValueType(instance, value));
-                    return false;
-                }
+            *(bool *)(args_ptr++) = b;
+        };
+
+        CASE(Int8) { PUSH_INTEGER(int8_t); };
+        CASE(UInt8) { PUSH_INTEGER(uint8_t); };
+        CASE(Int16) { PUSH_INTEGER(int16_t); };
+        CASE(Int16S) { PUSH_INTEGER_SWAP(int16_t); };
+        CASE(UInt16) { PUSH_INTEGER(uint16_t); };
+        CASE(UInt16S) { PUSH_INTEGER_SWAP(uint16_t); };
+        CASE(Int32) { PUSH_INTEGER(int32_t); };
+        CASE(Int32S) { PUSH_INTEGER_SWAP(int32_t); };
+        CASE(UInt32) { PUSH_INTEGER(uint32_t); };
+        CASE(UInt32S) { PUSH_INTEGER_SWAP(uint32_t); };
+        CASE(Int64) { PUSH_INTEGER(int64_t); };
+        CASE(Int64S) { PUSH_INTEGER_SWAP(int64_t); };
+        CASE(UInt64) { PUSH_INTEGER(uint64_t); };
+        CASE(UInt64S) { PUSH_INTEGER_SWAP(uint64_t); };
+
+        CASE(String) {
+            const ParameterInfo &param = func->parameters[i];
+            Napi::Value value = info[param.offset];
 
-                bool b = value.As<Napi::Boolean>();
+            const char *str;
+            if (!PushString(value, param.directions, &str)) [[unlikely]]
+                return false;
 
-                *(bool *)(args_ptr++) = b;
-            } break;
-            case PrimitiveKind::Int8: { PUSH_INTEGER(int8_t); } break;
-            case PrimitiveKind::UInt8: { PUSH_INTEGER(uint8_t); } break;
-            case PrimitiveKind::Int16: { PUSH_INTEGER(int16_t); } break;
-            case PrimitiveKind::Int16S: { PUSH_INTEGER_SWAP(int16_t); } break;
-            case PrimitiveKind::UInt16: { PUSH_INTEGER(uint16_t); } break;
-            case PrimitiveKind::UInt16S: { PUSH_INTEGER_SWAP(uint16_t); } break;
-            case PrimitiveKind::Int32: { PUSH_INTEGER(int32_t); } break;
-            case PrimitiveKind::Int32S: { PUSH_INTEGER_SWAP(int32_t); } break;
-            case PrimitiveKind::UInt32: { PUSH_INTEGER(uint32_t); } break;
-            case PrimitiveKind::UInt32S: { PUSH_INTEGER_SWAP(uint32_t); } break;
-            case PrimitiveKind::Int64: { PUSH_INTEGER(int64_t); } break;
-            case PrimitiveKind::Int64S: { PUSH_INTEGER_SWAP(int64_t); } break;
-            case PrimitiveKind::UInt64: { PUSH_INTEGER(uint64_t); } break;
-            case PrimitiveKind::UInt64S: { PUSH_INTEGER_SWAP(uint64_t); } break;
-            case PrimitiveKind::String: {
-                const char *str;
-                if (!PushString(value, param.directions, &str)) [[unlikely]]
-                    return false;
+            *(const char **)(args_ptr++) = str;
+        };
+        CASE(String16) {
+            const ParameterInfo &param = func->parameters[i];
+            Napi::Value value = info[param.offset];
 
-                *(const char **)(args_ptr++) = str;
-            } break;
-            case PrimitiveKind::String16: {
-                const char16_t *str16;
-                if (!PushString16(value, param.directions, &str16)) [[unlikely]]
-                    return false;
+            const char16_t *str16;
+            if (!PushString16(value, param.directions, &str16)) [[unlikely]]
+                return false;
 
-                *(const char16_t **)(args_ptr++) = str16;
-            } break;
-            case PrimitiveKind::String32: {
-                const char32_t *str32;
-                if (!PushString32(value, param.directions, &str32)) [[unlikely]]
-                    return false;
+            *(const char16_t **)(args_ptr++) = str16;
+        };
+        CASE(String32) {
+            const ParameterInfo &param = func->parameters[i];
+            Napi::Value value = info[param.offset];
 
-                *(const char32_t **)(args_ptr++) = str32;
-            } break;
-            case PrimitiveKind::Pointer: {
-                void *ptr;
-                if (!PushPointer(value, param.type, param.directions, &ptr)) [[unlikely]]
-                    return false;
+            const char32_t *str32;
+            if (!PushString32(value, param.directions, &str32)) [[unlikely]]
+                return false;
 
-                *(void **)(args_ptr++) = ptr;
-            } break;
-            case PrimitiveKind::Record:
-            case PrimitiveKind::Union: {
-                if (!IsObject(value)) [[unlikely]] {
-                    ThrowError<Napi::TypeError>(env, "Unexpected %1 value, expected object", GetValueType(instance, value));
-                    return false;
-                }
+            *(const char32_t **)(args_ptr++) = str32;
+        };
 
-                uint8_t *ptr;
-                if (param.regular) {
-                    ptr = (uint8_t *)(args_ptr++);
-                } else {
-                    ptr = AllocHeap(param.type->size, 16);
-                    *(uint8_t **)(args_ptr++) = ptr;
-                }
+        CASE(Pointer) {
+            const ParameterInfo &param = func->parameters[i];
+            Napi::Value value = info[param.offset];
 
-                Napi::Object obj = value.As<Napi::Object>();
-                if (!PushObject(obj, param.type, ptr))
-                    return false;
-            } break;
-            case PrimitiveKind::Array: { K_UNREACHABLE(); } break;
-            case PrimitiveKind::Float32: {
-                if (!value.IsNumber() && !value.IsBigInt()) [[unlikely]] {
-                    ThrowError<Napi::TypeError>(env, "Unexpected %1 value, expected number", GetValueType(instance, value));
-                    return false;
-                }
+            void *ptr;
+            if (!PushPointer(value, param.type, param.directions, &ptr)) [[unlikely]]
+                return false;
 
-                float f = GetNumber<float>(value);
+            *(void **)(args_ptr++) = ptr;
+        };
 
-                memset((uint8_t *)args_ptr + 4, 0, 4);
-                *(float *)(args_ptr++) = f;
-            } break;
-            case PrimitiveKind::Float64: {
-                if (!value.IsNumber() && !value.IsBigInt()) [[unlikely]] {
-                    ThrowError<Napi::TypeError>(env, "Unexpected %1 value, expected number", GetValueType(instance, value));
-                    return false;
-                }
+        CASE(Record) OR(Union) {
+            const ParameterInfo &param = func->parameters[i];
+            Napi::Value value = info[param.offset];
 
-                double d = GetNumber<double>(value);
-                *(double *)(args_ptr++) = d;
-            } break;
-            case PrimitiveKind::Callback: {
-                void *ptr;
-                if (!PushCallback(value, param.type, &ptr)) [[unlikely]]
-                    return false;
+            if (!IsObject(value)) [[unlikely]] {
+                ThrowError<Napi::TypeError>(env, "Unexpected %1 value, expected object", GetValueType(instance, value));
+                return false;
+            }
 
-                *(void **)(args_ptr++) = ptr;
-            } break;
+            uint8_t *ptr;
+            if (param.regular) {
+                ptr = (uint8_t *)(args_ptr++);
+            } else {
+                ptr = AllocHeap(param.type->size, 16);
+                *(uint8_t **)(args_ptr++) = ptr;
+            }
 
-            case PrimitiveKind::Prototype: { K_UNREACHABLE(); } break;
-        }
+            Napi::Object obj = value.As<Napi::Object>();
+            if (!PushObject(obj, param.type, ptr))
+                return false;
+        };
+        CASE(Array) { K_UNREACHABLE(); };
+
+        CASE(Float32) {
+            const ParameterInfo &param = func->parameters[i];
+            Napi::Value value = info[param.offset];
+
+            float f;
+            if (!TryNumber(value, &f)) [[unlikely]] {
+                ThrowError<Napi::TypeError>(env, "Unexpected %1 value, expected number", GetValueType(instance, value));
+                return false;
+            }
+
+            memset((uint8_t *)args_ptr + 4, 0, 4);
+            *(float *)(args_ptr++) = f;
+        };
+        CASE(Float64) {
+            const ParameterInfo &param = func->parameters[i];
+            Napi::Value value = info[param.offset];
+
+            double d;
+            if (!TryNumber(value, &d)) [[unlikely]] {
+                ThrowError<Napi::TypeError>(env, "Unexpected %1 value, expected number", GetValueType(instance, value));
+                return false;
+            }
+
+            *(double *)(args_ptr++) = d;
+        };
+        CASE(Callback) {
+            const ParameterInfo &param = func->parameters[i];
+            Napi::Value value = info[param.offset];
+
+            void *ptr;
+            if (!PushCallback(value, param.type, &ptr)) [[unlikely]]
+                return false;
+
+            *(void **)(args_ptr++) = ptr;
+        };
+
+        CASE(Prototype) { /* End loop */ };
     }
 
 #undef PUSH_INTEGER_SWAP
 #undef PUSH_INTEGER
 
+#undef OR
+#undef CASE
+#undef LOOP
+
     new_sp = mem->stack.end();
 
     return true;
@@ -302,17 +358,8 @@ Napi::Value CallData::Complete(const FunctionInfo *func)
         case PrimitiveKind::String: return result.ptr ? Napi::String::New(env, (const char *)result.ptr) : env.Null();
         case PrimitiveKind::String16: return result.ptr ? Napi::String::New(env, (const char16_t *)result.ptr) : env.Null();
         case PrimitiveKind::String32: return result.ptr ? MakeStringFromUTF32(env, (const char32_t *)result.ptr) : env.Null();
-        case PrimitiveKind::Pointer:
-        case PrimitiveKind::Callback: {
-            if (result.ptr) {
-                Napi::External<void> external = Napi::External<void>::New(env, result.ptr);
-                SetValueTag(external, func->ret.type->ref.marker);
-
-                return external;
-            } else {
-                return env.Null();
-            }
-        } break;
+        case PrimitiveKind::Pointer: return result.ptr ? WrapPointer(env, func->ret.type->ref.type, result.ptr) : env.Null();
+        case PrimitiveKind::Callback: return result.ptr ? WrapCallback(env, func->ret.type->ref.type, result.ptr) : env.Null();
         case PrimitiveKind::Record:
         case PrimitiveKind::Union: {
             const uint8_t *ptr = return_ptr ? (const uint8_t *)return_ptr
@@ -520,19 +567,23 @@ void CallData::Relay(Size idx, uint8_t *own_sp, uint8_t *caller_sp, bool switch_
                 Napi::Value arg = str32 ? MakeStringFromUTF32(env, str32) : env.Null();
                 arguments.Append(arg);
             } break;
-            case PrimitiveKind::Pointer:
-            case PrimitiveKind::Callback: {
+            case PrimitiveKind::Pointer: {
                 void *ptr2 = *(void **)(j < 4 ? gpr_ptr + j : args_ptr);
                 args_ptr += (j >= 4);
 
-                if (ptr2) {
-                    Napi::External<void> external = Napi::External<void>::New(env, ptr2);
-                    SetValueTag(external, param.type->ref.marker);
+                Napi::Value p = ptr2 ? WrapPointer(env, param.type->ref.type, ptr2) : env.Null();
+                arguments.Append(p);
 
-                    arguments.Append(external);
-                } else {
-                    arguments.Append(env.Null());
+                if (param.type->dispose) {
+                    param.type->dispose(env, param.type, ptr2);
                 }
+            } break;
+            case PrimitiveKind::Callback: {
+                void *ptr2 = *(void **)(j < 4 ? gpr_ptr + j : args_ptr);
+                args_ptr += (j >= 4);
+
+                Napi::Value p = ptr2 ? WrapCallback(env, param.type->ref.type, ptr2) : env.Null();
+                arguments.Append(p);
 
                 if (param.type->dispose) {
                     param.type->dispose(env, param.type, ptr2);
@@ -588,34 +639,34 @@ void CallData::Relay(Size idx, uint8_t *own_sp, uint8_t *caller_sp, bool switch_
 
 #define RETURN_INTEGER(CType) \
         do { \
-            if (!value.IsNumber() && !value.IsBigInt()) [[unlikely]] { \
+            CType v; \
+            if (!TryNumber(value, &v)) [[unlikely]] { \
                 ThrowError<Napi::TypeError>(env, "Unexpected %1 value, expected number", GetValueType(instance, value)); \
                 return; \
             } \
              \
-            CType v = GetNumber<CType>(value); \
             out_reg->rax = (uint64_t)v; \
         } while (false)
 #define RETURN_INTEGER_SWAP(CType) \
         do { \
-            if (!value.IsNumber() && !value.IsBigInt()) [[unlikely]] { \
+            CType v; \
+            if (!TryNumber(value, &v)) [[unlikely]] { \
                 ThrowError<Napi::TypeError>(env, "Unexpected %1 value, expected number", GetValueType(instance, value)); \
                 return; \
             } \
              \
-            CType v = GetNumber<CType>(value); \
             out_reg->rax = (uint64_t)ReverseBytes(v); \
         } while (false)
 
     switch (type->primitive) {
         case PrimitiveKind::Void: {} break;
         case PrimitiveKind::Bool: {
-            if (!value.IsBoolean()) [[unlikely]] {
+            bool b;
+            if (napi_get_value_bool(env, value, &b) != napi_ok) [[unlikely]] {
                 ThrowError<Napi::TypeError>(env, "Unexpected %1 value, expected boolean", GetValueType(instance, value));
                 return;
             }
 
-            bool b = value.As<Napi::Boolean>();
             out_reg->rax = (uint64_t)b;
         } break;
         case PrimitiveKind::Int8: { RETURN_INTEGER(int8_t); } break;
@@ -654,24 +705,9 @@ void CallData::Relay(Size idx, uint8_t *own_sp, uint8_t *caller_sp, bool switch_
             out_reg->rax = (uint64_t)str32;
         } break;
         case PrimitiveKind::Pointer: {
-            uint8_t *ptr;
-
-            if (CheckValueTag(value, type->ref.marker)) {
-                ptr = value.As<Napi::External<uint8_t>>().Data();
-            } else if (IsObject(value) && (type->ref.type->primitive == PrimitiveKind::Record ||
-                                           type->ref.type->primitive == PrimitiveKind::Union)) {
-                Napi::Object obj = value.As<Napi::Object>();
-
-                ptr = AllocHeap(type->ref.type->size, 16);
-
-                if (!PushObject(obj, type->ref.type, ptr))
-                    return;
-            } else if (IsNullOrUndefined(value)) {
-                ptr = nullptr;
-            } else {
-                ThrowError<Napi::TypeError>(env, "Unexpected %1 value, expected %2", GetValueType(instance, value), type->name);
+            void *ptr;
+            if (!PushPointer(value, type, 1, &ptr)) [[unlikely]]
                 return;
-            }
 
             out_reg->rax = (uint64_t)ptr;
         } break;
@@ -694,42 +730,28 @@ void CallData::Relay(Size idx, uint8_t *own_sp, uint8_t *caller_sp, bool switch_
         } break;
         case PrimitiveKind::Array: { K_UNREACHABLE(); } break;
         case PrimitiveKind::Float32: {
-            if (!value.IsNumber() && !value.IsBigInt()) [[unlikely]] {
+            float f;
+            if (!TryNumber(value, &f)) [[unlikely]] {
                 ThrowError<Napi::TypeError>(env, "Unexpected %1 value, expected number", GetValueType(instance, value));
                 return;
             }
 
-            float f = GetNumber<float>(value);
-
             memset((uint8_t *)&out_reg->xmm0 + 4, 0, 4);
             memcpy(&out_reg->xmm0, &f, 4);
         } break;
         case PrimitiveKind::Float64: {
-            if (!value.IsNumber() && !value.IsBigInt()) [[unlikely]] {
+            double d;
+            if (!TryNumber(value, &d)) [[unlikely]] {
                 ThrowError<Napi::TypeError>(env, "Unexpected %1 value, expected number", GetValueType(instance, value));
                 return;
             }
 
-            double d = GetNumber<double>(value);
             out_reg->xmm0 = d;
         } break;
         case PrimitiveKind::Callback: {
             void *ptr;
-
-            if (value.IsFunction()) {
-                Napi::Function func2 = value.As<Napi::Function>();
-
-                ptr = ReserveTrampoline(type->ref.proto, func2);
-                if (!ptr) [[unlikely]]
-                    return;
-            } else if (CheckValueTag(value, type->ref.marker)) {
-                ptr = value.As<Napi::External<void>>().Data();
-            } else if (IsNullOrUndefined(value)) {
-                ptr = nullptr;
-            } else {
-                ThrowError<Napi::TypeError>(env, "Unexpected %1 value, expected %2", GetValueType(instance, value), type->name);
+            if (!PushCallback(value, type, &ptr)) [[unlikely]]
                 return;
-            }
 
             out_reg->rax = (uint64_t)ptr;
         } break;
@@ -743,12 +765,6 @@ void CallData::Relay(Size idx, uint8_t *own_sp, uint8_t *caller_sp, bool switch_
     err_guard.Disable();
 }
 
-void *GetTrampoline(int16_t idx, const FunctionInfo *proto)
-{
-    bool xmm = proto->forward_fp || IsFloat(proto->ret.type);
-    return Trampolines[idx][xmm];
-}
-
 }
 
 #endif

package/src/koffi/src/abi_x64_win_asm.S

@@ -0,0 +1,144 @@
+# SPDX-License-Identifier: MIT
+# SPDX-FileCopyrightText: 2025 Niels Martignène <niels.martignene@protonmail.com>
+
+#define SYMBOL(Symbol) Symbol
+
+# Forward
+# ----------------------------
+
+.global ForwardCallG
+.global ForwardCallF
+.global ForwardCallD
+.global ForwardCallXG
+.global ForwardCallXF
+.global ForwardCallXD
+
+# Copy function pointer to RAX, in order to save it through argument forwarding.
+# Also make a copy of the SP to CallData::old_sp because the callback system might need it.
+# Save RSP in RBX (non-volatile), and use carefully assembled stack provided by caller.
+.macro prologue
+    endbr64
+    movq %rcx, %rax
+    push %rbp
+    movq %rsp, %rbp
+    movq %rsp, (%r8)
+    movq %rdx, %rsp
+.endm
+
+# Call native function.
+# Once done, restore normal stack pointer and return.
+# The return value is passed untouched through RAX or XMM0.
+.macro epilogue
+    call *%rax
+    movq %rbp, %rsp
+    pop %rbp
+    ret
+.endm
+
+# Prepare integer argument registers from array passed by caller.
+.macro forward_gpr
+    movq 24(%rdx), %r9
+    movq 16(%rdx), %r8
+    movq 0(%rdx), %rcx
+    movq 8(%rdx), %rdx
+.endm
+
+# Prepare XMM argument registers from array passed by caller.
+.macro forward_xmm
+    movsd 24(%rdx), %xmm3
+    movsd 16(%rdx), %xmm2
+    movsd 8(%rdx), %xmm1
+    movsd 0(%rdx), %xmm0
+.endm
+
+ForwardCallG:
+    prologue
+    forward_gpr
+    epilogue
+
+ForwardCallF:
+    prologue
+    forward_gpr
+    epilogue
+
+ForwardCallD:
+    prologue
+    forward_gpr
+    epilogue
+
+ForwardCallXG:
+    prologue
+    forward_xmm
+    forward_gpr
+    epilogue
+
+ForwardCallXF:
+    prologue
+    forward_xmm
+    forward_gpr
+    epilogue
+
+ForwardCallXD:
+    prologue
+    forward_xmm
+    forward_gpr
+    epilogue
+
+# Callbacks
+# ----------------------------
+
+.global RelayCallback
+.global CallSwitchStack
+
+# First, make a copy of argument registers.
+# Then call the C function RelayCallback with the following arguments:
+# static trampoline ID, a pointer to the saved GPR array, a pointer to the stack
+# arguments of this call, and a pointer to a struct that will contain the result registers.
+# After the call, simply load these registers from the output struct.
+.macro trampoline id
+    endbr64
+    subq $120, %rsp
+    movq %rcx, 32(%rsp)
+    movq %rdx, 40(%rsp)
+    movq %r8, 48(%rsp)
+    movq %r9, 56(%rsp)
+    movsd %xmm0, 64(%rsp)
+    movsd %xmm1, 72(%rsp)
+    movsd %xmm2, 80(%rsp)
+    movsd %xmm3, 88(%rsp)
+    movq $\id, %rcx
+    leaq 32(%rsp), %rdx
+    leaq 160(%rsp), %r8
+    leaq 96(%rsp), %r9
+    call RelayCallback
+    movq 96(%rsp), %rax
+    movsd 104(%rsp), %xmm0
+    addq $120, %rsp
+    ret
+.endm
+
+# When a callback is relayed, Koffi will call into Node.js and V8 to execute Javascript.
+# The problem is that we're still running on the separate Koffi stack, and V8 will
+# probably misdetect this as a "stack overflow". We have to restore the old
+# stack pointer, call Node.js/V8 and go back to ours.
+# The first three parameters (rcx, rdx, r8) are passed through untouched.
+CallSwitchStack:
+    endbr64
+    push %rbp
+    movq %rsp, %rbp
+    movq 56(%rsp), %rax
+    movq %rsp, %r10
+    movq 48(%rsp), %r11
+    subq 0(%r11), %r10
+    andq $-16, %r10
+    movq %r10, 8(%r11)
+    leaq -32(%r9), %rsp
+    call *%rax
+    movq %rbp, %rsp
+    pop %rbp
+    ret
+
+# Trampolines
+# ----------------------------
+
+#include "trampolines/gnu.inc"

package/src/koffi/src/abi_x64_win_asm.asm

@@ -103,32 +103,12 @@ ForwardCallXD endp
 extern RelayCallback : PROC
 public CallSwitchStack
 
-; First, make a copy of the GPR argument registers (rcx, rdx, r8, r9).
+; First, make a copy of argument registers.
 ; Then call the C function RelayCallback with the following arguments:
 ; static trampoline ID, a pointer to the saved GPR array, a pointer to the stack
 ; arguments of this call, and a pointer to a struct that will contain the result registers.
 ; After the call, simply load these registers from the output struct.
 trampoline macro ID
-    endbr64
-    sub rsp, 120
-    .allocstack 120
-    .endprolog
-    mov qword ptr [rsp+32], rcx
-    mov qword ptr [rsp+40], rdx
-    mov qword ptr [rsp+48], r8
-    mov qword ptr [rsp+56], r9
-    mov rcx, ID
-    lea rdx, qword ptr [rsp+32]
-    lea r8, qword ptr [rsp+160]
-    lea r9, qword ptr [rsp+96]
-    call RelayCallback
-    mov rax, qword ptr [rsp+96]
-    add rsp, 120
-    ret
-endm
-
-; Same thing, but also forward the XMM argument registers and load the XMM result registers.
-trampoline_vec macro ID
     endbr64
     sub rsp, 120
     .allocstack 120