koffi 0.9.5 → 0.9.6

package/CMakeLists.txt

@@ -20,6 +20,7 @@ if(MSVC)
 endif()
 
 set(KOFFI_SRC
+    src/call_arm32.cc
     src/call_arm64.cc
     src/call_x64_sysv.cc
     src/call_x64_win.cc
@@ -41,6 +42,8 @@ elseif(CMAKE_SIZEOF_VOID_P EQUAL 4)
         list(APPEND KOFFI_SRC src/call_x86_fwd.asm)
     elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "i[3456]86|x86|AMD64")
         list(APPEND KOFFI_SRC src/call_x86_fwd.S)
+    elseif(CMAKE_SYSTEM_PROCESSOR MATCHES "armv[678]l")
+        list(APPEND KOFFI_SRC src/call_arm32_fwd.S)
     endif()
 endif()
 

package/README.md

@@ -8,7 +8,8 @@ The following platforms __are supported__ at the moment:
 * Windows x86_64
 * Linux x86
 * Linux x86_64
-* Linux ARM64 *(Raspberry Pi OS 64-bit)*
+* Linux ARM32+VFP Little Endian *(Raspberry Pi OS 32-bit)*
+* Linux ARM64 Little Endian *(Raspberry Pi OS 64-bit)*
 
 The following platforms are __not yet supported__ but will be soon:
 
@@ -17,9 +18,7 @@ The following platforms are __not yet supported__ but will be soon:
 * FreeBSD/NetBSD/OpenBSD x86_64
 * FreeBSD/NetBSD/OpenBSD ARM64
 
-Support for ARM32 is not planned, though I don't mind working on it if there is enough demand.
-
-This is still in development, more tests will come in the near future.
+This is still in development, bugs are to expected. More tests will come in the near future.
 
 # Installation
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "koffi",
-  "version": "0.9.5",
+  "version": "0.9.6",
   "description": "Fast and simple FFI (foreign function interface) for Node.js",
   "keywords": ["foreign", "function", "interface", "ffi", "binding", "c", "napi"],
   "repository": {

package/src/call_arm32.cc

@@ -0,0 +1,514 @@
+// This program is free software: you can redistribute it and/or modify
+// it under the terms of the GNU Affero General Public License as published by
+// the Free Software Foundation, either version 3 of the License, or
+// (at your option) any later version.
+//
+// This program is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+// GNU Affero General Public License for more details.
+//
+// You should have received a copy of the GNU Affero General Public License
+// along with this program. If not, see https://www.gnu.org/licenses/.
+
+#if defined(__arm__)
+
+#include "vendor/libcc/libcc.hh"
+#include "ffi.hh"
+#include "call.hh"
+#include "util.hh"
+
+#include <napi.h>
+
+namespace RG {
+
+struct HfaRet {
+    double d0;
+    double d1;
+    double d2;
+    double d3;
+};
+
+extern "C" uint64_t ForwardCallGG(const void *func, uint8_t *sp);
+extern "C" float ForwardCallF(const void *func, uint8_t *sp);
+extern "C" HfaRet ForwardCallDDDD(const void *func, uint8_t *sp);
+
+extern "C" uint64_t ForwardCallXGG(const void *func, uint8_t *sp);
+extern "C" float ForwardCallXF(const void *func, uint8_t *sp);
+extern "C" HfaRet ForwardCallXDDDD(const void *func, uint8_t *sp);
+
+static bool IsHFA(const TypeInfo *type)
+{
+    if (type->primitive != PrimitiveKind::Record)
+        return false;
+
+    if (type->members.len < 1 || type->members.len > 4)
+        return false;
+    if (type->members[0].type->primitive != PrimitiveKind::Float32 &&
+            type->members[0].type->primitive != PrimitiveKind::Float64)
+        return false;
+
+    for (Size i = 1; i < type->members.len; i++) {
+        if (type->members[i].type != type->members[0].type)
+            return false;
+    }
+
+    return true;
+}
+
+bool AnalyseFunction(FunctionInfo *func)
+{
+    if (IsHFA(func->ret.type)) {
+        func->ret.vec_count = func->ret.type->members.len *
+                              (func->ret.type->members[0].type->size / 4);
+    } else if (func->ret.type->size <= 4) {
+        func->ret.gpr_count = 1;
+    } else {
+        func->ret.use_memory = true;
+    }
+
+    int gpr_avail = 4 - func->ret.use_memory;
+    int vec_avail = 16;
+    bool started_stack = false;
+
+    for (ParameterInfo &param: func->parameters) {
+        switch (param.type->primitive) {
+            case PrimitiveKind::Void: { RG_UNREACHABLE(); } break;
+
+            case PrimitiveKind::Bool:
+            case PrimitiveKind::Int8:
+            case PrimitiveKind::UInt8:
+            case PrimitiveKind::Int16:
+            case PrimitiveKind::UInt16:
+            case PrimitiveKind::Int32:
+            case PrimitiveKind::UInt32:
+            case PrimitiveKind::String:
+            case PrimitiveKind::Pointer: {
+                if (gpr_avail) {
+                    param.gpr_count = 1;
+                    gpr_avail--;
+                } else {
+                    started_stack = true;
+                }
+            } break;
+
+            case PrimitiveKind::Int64:
+            case PrimitiveKind::UInt64: {
+                if (gpr_avail >= 2) {
+                    param.gpr_count = 2;
+                    gpr_avail -= 2;
+                } else {
+                    started_stack = true;
+                }
+            } break;
+
+            case PrimitiveKind::Float32:
+            case PrimitiveKind::Float64: {
+                Size need = param.type->size / 4;
+
+                if (need <= vec_avail) {
+                    param.vec_count = need;
+                    vec_avail -= need;
+                } else {
+                    started_stack = true;
+                }
+            } break;
+
+            case PrimitiveKind::Record: {
+                if (IsHFA(param.type)) {
+                    int vec_count = (int)(param.type->members.len *
+                                          param.type->members[0].type->size / 4);
+
+                    if (vec_count <= vec_avail) {
+                        param.vec_count = vec_count;
+                        vec_avail -= vec_count;
+                    } else {
+                        vec_avail = 0;
+                        started_stack = true;
+                    }
+                } else if (param.type->size) {
+                    int gpr_count = (param.type->size + 3) / 4;
+
+                    if (gpr_count <= gpr_avail) {
+                        param.gpr_count = gpr_count;
+                        gpr_avail -= gpr_count;
+                    } else {
+                        if (!started_stack) {
+                            param.gpr_count = gpr_avail;
+                            gpr_avail = 0;
+
+                            started_stack = true;
+                        }
+                    }
+                }
+            } break;
+        }
+    }
+
+    func->forward_fp = (vec_avail < 16);
+
+    return true;
+}
+
+static bool PushHFA(const Napi::Object &obj, const TypeInfo *type, uint8_t *dest)
+{
+    Napi::Env env = obj.Env();
+    InstanceData *instance = env.GetInstanceData<InstanceData>();
+
+    RG_ASSERT(obj.IsObject());
+    RG_ASSERT(type->primitive == PrimitiveKind::Record);
+    RG_ASSERT(AlignUp(dest, type->members[0].type->size) == dest);
+
+    for (const RecordMember &member: type->members) {
+        Napi::Value value = obj.Get(member.name);
+
+        if (member.type->primitive == PrimitiveKind::Float32) {
+            if (!value.IsNumber() && !value.IsBigInt()) {
+                ThrowError<Napi::TypeError>(env, "Unexpected value %1 for member '%2', expected number", GetValueType(instance, value), member.name);
+                return false;
+            }
+
+            *(float *)dest = CopyNodeNumber<float>(value);
+        } else if (member.type->primitive == PrimitiveKind::Float64) {
+            if (!value.IsNumber() && !value.IsBigInt()) {
+                ThrowError<Napi::TypeError>(env, "Unexpected value %1 for member '%2', expected number", GetValueType(instance, value), member.name);
+                return false;
+            }
+
+            *(double *)dest = CopyNodeNumber<double>(value);
+        } else {
+            RG_UNREACHABLE();
+        }
+
+        dest += type->members[0].type->size;
+    }
+
+    return true;
+}
+
+static Napi::Object PopHFA(napi_env env, const uint8_t *ptr, const TypeInfo *type)
+{
+    RG_ASSERT(type->primitive == PrimitiveKind::Record);
+
+    Napi::Object obj = Napi::Object::New(env);
+
+    for (const RecordMember &member: type->members) {
+        if (member.type->primitive == PrimitiveKind::Float32) {
+            float f = *(float *)ptr;
+            obj.Set(member.name, Napi::Number::New(env, (double)f));
+        } else if (member.type->primitive == PrimitiveKind::Float64) {
+            double d = *(double *)ptr;
+            obj.Set(member.name, Napi::Number::New(env, d));
+        } else {
+            RG_UNREACHABLE();
+        }
+
+        ptr += member.type->size;
+    }
+
+    return obj;
+}
+
+Napi::Value TranslateCall(const Napi::CallbackInfo &info)
+{
+    Napi::Env env = info.Env();
+    InstanceData *instance = env.GetInstanceData<InstanceData>();
+
+    FunctionInfo *func = (FunctionInfo *)info.Data();
+    LibraryData *lib = func->lib.get();
+
+    RG_DEFER { lib->tmp_alloc.ReleaseAll(); };
+
+    // Sanity checks
+    if (info.Length() < (uint32_t)func->parameters.len) {
+        ThrowError<Napi::TypeError>(env, "Expected %1 arguments, got %2", func->parameters.len, info.Length());
+        return env.Null();
+    }
+
+    // Stack pointer and register
+    uint8_t *top_ptr = lib->stack.end();
+    uint8_t *return_ptr = nullptr;
+    uint8_t *args_ptr = nullptr;
+    uint32_t *gpr_ptr = nullptr;
+    uint32_t *vec_ptr = nullptr;
+    uint8_t *sp_ptr = nullptr;
+
+    // Return through registers unless it's too big
+    if (!func->ret.use_memory) {
+        args_ptr = top_ptr - func->scratch_size;
+        gpr_ptr = (uint32_t *)args_ptr - 4;
+        vec_ptr = gpr_ptr - 16;
+        sp_ptr = (uint8_t *)vec_ptr;
+
+#ifdef RG_DEBUG
+        memset(sp_ptr, 0, top_ptr - sp_ptr);
+#endif
+    } else {
+        return_ptr = top_ptr - AlignLen(func->ret.type->size, 16);
+
+        args_ptr = return_ptr - func->scratch_size;
+        gpr_ptr = (uint32_t *)args_ptr - 4;
+        vec_ptr = gpr_ptr - 16;
+        sp_ptr = (uint8_t *)vec_ptr;
+
+#ifdef RG_DEBUG
+        memset(sp_ptr, 0, top_ptr - sp_ptr);
+#endif
+
+        *(gpr_ptr++) = (uint32_t)return_ptr;
+    }
+
+    RG_ASSERT(AlignUp(lib->stack.ptr, 16) == lib->stack.ptr);
+    RG_ASSERT(AlignUp(lib->stack.end(), 16) == lib->stack.end());
+    RG_ASSERT(AlignUp(sp_ptr, 16) == sp_ptr);
+
+    // Push arguments
+    for (Size i = 0; i < func->parameters.len; i++) {
+        const ParameterInfo &param = func->parameters[i];
+        Napi::Value value = info[i];
+
+        switch (param.type->primitive) {
+            case PrimitiveKind::Void: { RG_UNREACHABLE(); } break;
+
+            case PrimitiveKind::Bool: {
+                if (RG_UNLIKELY(!value.IsBoolean())) {
+                    ThrowError<Napi::TypeError>(env, "Unexpected %1 value for argument %2, expected boolean", GetValueType(instance, value), i + 1);
+                    return env.Null();
+                }
+
+                bool b = value.As<Napi::Boolean>();
+
+                if (RG_LIKELY(param.gpr_count)) {
+                    *(gpr_ptr++) = (uint32_t)b;
+                } else {
+                    *(args_ptr++) = (uint8_t)b;
+                }
+            } break;
+            case PrimitiveKind::Int8:
+            case PrimitiveKind::UInt8:
+            case PrimitiveKind::Int16:
+            case PrimitiveKind::UInt16:
+            case PrimitiveKind::Int32:
+            case PrimitiveKind::UInt32: {
+                if (RG_UNLIKELY(!value.IsNumber() && !value.IsBigInt())) {
+                    ThrowError<Napi::TypeError>(env, "Unexpected %1 value for argument %2, expected number", GetValueType(instance, value), i + 1);
+                    return env.Null();
+                }
+
+                int64_t v = CopyNodeNumber<int64_t>(value);
+
+                if (RG_LIKELY(param.gpr_count)) {
+                    *(gpr_ptr++) = (uint32_t)v;
+                } else {
+                    args_ptr = AlignUp(args_ptr, param.type->align);
+                    memcpy(args_ptr, &v, param.type->size); // Little Endian
+                    args_ptr += param.type->size;
+                }
+            } break;
+            case PrimitiveKind::Int64:
+            case PrimitiveKind::UInt64: {
+                if (RG_UNLIKELY(!value.IsNumber() && !value.IsBigInt())) {
+                    ThrowError<Napi::TypeError>(env, "Unexpected %1 value for argument %2, expected number", GetValueType(instance, value), i + 1);
+                    return env.Null();
+                }
+
+                int64_t v = CopyNodeNumber<int64_t>(value);
+
+                if (RG_LIKELY(param.gpr_count)) {
+                    *(uint64_t *)gpr_ptr = (uint64_t)v;
+                    gpr_ptr += 2;
+                } else {
+                    args_ptr = AlignUp(args_ptr, 8);
+                    memcpy(args_ptr, &v, param.type->size); // Little Endian
+                    args_ptr += 8;
+                }
+            } break;
+            case PrimitiveKind::Float32: {
+                if (RG_UNLIKELY(!value.IsNumber() && !value.IsBigInt())) {
+                    ThrowError<Napi::TypeError>(env, "Unexpected %1 value for argument %2, expected number", GetValueType(instance, value), i + 1);
+                    return env.Null();
+                }
+
+                float f = CopyNodeNumber<float>(value);
+
+                if (RG_LIKELY(param.vec_count)) {
+                    memcpy(vec_ptr++, &f, 4);
+                } else {
+                    args_ptr = AlignUp(args_ptr, 4);
+                    memcpy(args_ptr, &f, 4);
+                    args_ptr += 4;
+                }
+            } break;
+            case PrimitiveKind::Float64: {
+                if (RG_UNLIKELY(!value.IsNumber() && !value.IsBigInt())) {
+                    ThrowError<Napi::TypeError>(env, "Unexpected %1 value for argument %2, expected number", GetValueType(instance, value), i + 1);
+                    return env.Null();
+                }
+
+                double d = CopyNodeNumber<double>(value);
+
+                if (RG_LIKELY(param.vec_count)) {
+                    memcpy(vec_ptr, &d, 8);
+                    vec_ptr += 2;
+                } else {
+                    args_ptr = AlignUp(args_ptr, 8);
+                    memcpy(args_ptr, &d, 8);
+                    args_ptr += 8;
+                }
+            } break;
+            case PrimitiveKind::String: {
+                if (RG_UNLIKELY(!value.IsString())) {
+                    ThrowError<Napi::TypeError>(env, "Unexpected %1 value for argument %2, expected string", GetValueType(instance, value), i + 1);
+                    return env.Null();
+                }
+
+                const char *str = CopyNodeString(value, &lib->tmp_alloc);
+
+                if (RG_LIKELY(param.gpr_count)) {
+                    *(gpr_ptr++) = (uint64_t)str;
+                } else {
+                    args_ptr = AlignUp(args_ptr, 4);
+                    *(const char **)args_ptr = str;
+                    args_ptr += 4;
+                }
+            } break;
+            case PrimitiveKind::Pointer: {
+                if (RG_UNLIKELY(!CheckValueTag(instance, value, param.type))) {
+                    ThrowError<Napi::TypeError>(env, "Unexpected %1 value for argument %2, expected %3", GetValueType(instance, value), i + 1, param.type->name);
+                    return env.Null();
+                }
+
+                void *ptr = value.As<Napi::External<void>>();
+
+                if (RG_LIKELY(param.gpr_count)) {
+                    *(gpr_ptr++) = (uint64_t)ptr;
+                } else {
+                    args_ptr = AlignUp(args_ptr, 4);
+                    *(void **)args_ptr = ptr;
+                    args_ptr += 4;
+                }
+            } break;
+
+            case PrimitiveKind::Record: {
+                if (RG_UNLIKELY(!value.IsObject())) {
+                    ThrowError<Napi::TypeError>(env, "Unexpected %1 value for argument %2, expected object", GetValueType(instance, value), i + 1);
+                    return env.Null();
+                }
+
+                Napi::Object obj = value.As<Napi::Object>();
+
+                if (param.vec_count) {
+                    if (!PushHFA(obj, param.type, (uint8_t *)vec_ptr))
+                        return env.Null();
+                    vec_ptr += param.vec_count;
+                } else {
+                    if (param.gpr_count) {
+                        RG_ASSERT(param.type->align <= 8);
+
+                        if (!PushObject(obj, param.type, &lib->tmp_alloc, (uint8_t *)gpr_ptr))
+                            return env.Null();
+
+                        args_ptr += AlignLen(param.type->size - param.gpr_count * 4, 4);
+                        gpr_ptr += param.gpr_count;
+                    } else if (param.type->size) {
+                        int16_t align = (param.type->align <= 4) ? 4 : 8;
+
+                        args_ptr = AlignUp(args_ptr, align);
+                        if (!PushObject(obj, param.type, &lib->tmp_alloc, args_ptr))
+                            return env.Null();
+                        args_ptr += AlignLen(param.type->size, 4);
+                    }
+                }
+            } break;
+        }
+    }
+
+    // DumpStack(func, MakeSpan(sp_ptr, top_ptr - sp_ptr));
+
+#define PERFORM_CALL(Suffix) \
+        (func->forward_fp ? ForwardCallX ## Suffix(func->func, sp_ptr) \
+                          : ForwardCall ## Suffix(func->func, sp_ptr))
+
+    // Execute and convert return value
+    switch (func->ret.type->primitive) {
+        case PrimitiveKind::Float32: {
+            float f = PERFORM_CALL(F);
+
+            return Napi::Number::New(env, (double)f);
+        } break;
+
+        case PrimitiveKind::Float64: {
+            HfaRet ret = PERFORM_CALL(DDDD);
+
+            return Napi::Number::New(env, (double)ret.d0);
+        } break;
+
+        case PrimitiveKind::Record: {
+            if (func->ret.gpr_count) {
+                uint64_t ret = PERFORM_CALL(GG);
+                uint32_t r0 = (uint32_t)ret;
+
+                Napi::Object obj = PopObject(env, (const uint8_t *)&r0, func->ret.type);
+                return obj;
+            } else if (func->ret.vec_count) {
+                HfaRet ret = PERFORM_CALL(DDDD);
+
+                Napi::Object obj = PopHFA(env, (const uint8_t *)&ret, func->ret.type);
+                return obj;
+            } else if (func->ret.type->size) {
+                RG_ASSERT(return_ptr);
+
+                uint64_t ret = PERFORM_CALL(GG);
+                uint64_t r0 = (uint32_t)ret;
+                RG_ASSERT(r0 == (uint64_t)return_ptr);
+
+                Napi::Object obj = PopObject(env, return_ptr, func->ret.type);
+                return obj;
+            } else {
+                PERFORM_CALL(GG);
+
+                Napi::Object obj = Napi::Object::New(env);
+                return obj;
+            }
+        } break;
+
+        default: {
+            uint64_t ret = PERFORM_CALL(GG);
+            uint32_t r0 = (uint32_t)ret;
+
+            switch (func->ret.type->primitive) {
+                case PrimitiveKind::Void: return env.Null();
+                case PrimitiveKind::Bool: return Napi::Boolean::New(env, r0);
+                case PrimitiveKind::Int8: return Napi::Number::New(env, (double)r0);
+                case PrimitiveKind::UInt8: return Napi::Number::New(env, (double)r0);
+                case PrimitiveKind::Int16: return Napi::Number::New(env, (double)r0);
+                case PrimitiveKind::UInt16: return Napi::Number::New(env, (double)r0);
+                case PrimitiveKind::Int32: return Napi::Number::New(env, (double)r0);
+                case PrimitiveKind::UInt32: return Napi::Number::New(env, (double)r0);
+                case PrimitiveKind::Int64: return Napi::BigInt::New(env, (int64_t)ret);
+                case PrimitiveKind::UInt64: return Napi::BigInt::New(env, ret);
+                case PrimitiveKind::Float32: { RG_UNREACHABLE(); } break;
+                case PrimitiveKind::Float64: { RG_UNREACHABLE(); } break;
+                case PrimitiveKind::String: return Napi::String::New(env, (const char *)r0);
+                case PrimitiveKind::Pointer: {
+                    void *ptr = (void *)r0;
+
+                    Napi::External<void> external = Napi::External<void>::New(env, ptr);
+                    SetValueTag(instance, external, func->ret.type);
+
+                    return external;
+                } break;
+
+                case PrimitiveKind::Record: { RG_UNREACHABLE(); } break;
+            }
+        } break;
+    }
+
+#undef PERFORM_CALL
+
+    RG_UNREACHABLE();
+}
+
+}
+
+#endif

package/src/call_arm32_fwd.S

@@ -0,0 +1,108 @@
+// This program is free software: you can redistribute it and/or modify
+// it under the terms of the GNU Affero General Public License as published by
+// the Free Software Foundation, either version 3 of the License, or
+// (at your option) any later version.
+//
+// This program is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+// GNU Affero General Public License for more details.
+
+// You should have received a copy of the GNU Affero General Public License
+// along with this program. If not, see https://www.gnu.org/licenses/.
+
+.syntax unified
+
+// These three are the same, but they differ (in the C side) by their return type.
+// Unlike the three next functions, these ones don't forward XMM argument registers.
+.global ForwardCallGG
+.global ForwardCallF
+.global ForwardCallDDDD
+
+// The X variants are slightly slower, and are used when XMM arguments must be forwarded.
+.global ForwardCallXGG
+.global ForwardCallXF
+.global ForwardCallXDDDD
+
+// Copy function pointer to r12, in order to save it through argument forwarding.
+// Save RSP in fp (non-volatile), and use carefully assembled stack provided by caller.
+.macro prologue
+    .cfi_startproc
+    push {fp, lr}
+    .cfi_def_cfa sp, 8
+    .cfi_offset 11, 4
+    .cfi_offset 14, 8
+    mov fp, sp
+    .cfi_def_cfa fp, 8
+    mov r12, r0
+    mov sp, r1
+    add sp, sp, #80
+.endm
+
+// Call native function.
+// Once done, restore normal stack pointer and return.
+// The return value is passed untouched through r0, r1, and or FP registers
+.macro epilogue
+    blx r12
+    mov sp, fp
+    .cfi_def_cfa sp, 8
+    pop {fp, lr}
+    .cfi_def_cfa sp, 0
+    .cfi_restore 11
+    .cfi_restore 14
+    bx lr
+    .cfi_endproc
+.endm
+
+// Prepare general purpose argument registers from array passed by caller.
+.macro forward_int
+    ldr r3, [r1, 76]
+    ldr r2, [r1, 72]
+    ldr r0, [r1, 64]
+    ldr r1, [r1, 68]
+.endm
+
+// Prepare vector argument registers from array passed by caller.
+.macro forward_vec
+    vldr d7, [r1, 56]
+    vldr d6, [r1, 48]
+    vldr d5, [r1, 40]
+    vldr d4, [r1, 32]
+    vldr d3, [r1, 24]
+    vldr d2, [r1, 16]
+    vldr d1, [r1, 8]
+    vldr d0, [r1, 0]
+.endm
+
+ForwardCallGG:
+    prologue
+    forward_int
+    epilogue
+
+ForwardCallF:
+    prologue
+    forward_int
+    epilogue
+
+ForwardCallDDDD:
+    prologue
+    forward_int
+    epilogue
+
+ForwardCallXGG:
+    prologue
+    forward_vec
+    forward_int
+    epilogue
+
+ForwardCallXF:
+    prologue
+    forward_vec
+    forward_int
+    epilogue
+
+ForwardCallXDDDD:
+    prologue
+    forward_vec
+    forward_int
+    epilogue

package/src/call_arm64.cc

@@ -66,6 +66,8 @@ bool AnalyseFunction(FunctionInfo *func)
         func->ret.vec_count = func->ret.type->members.len;
     } else if (func->ret.type->size <= 16) {
         func->ret.gpr_count = (func->ret.type->size + 7) / 8;
+    } else {
+        func->ret.use_memory = true;
     }
 
     int gpr_avail = 8;
@@ -104,17 +106,28 @@ bool AnalyseFunction(FunctionInfo *func)
                 if (IsHFA(param.type)) {
                     int vec_count = (int)param.type->members.len;
 
-                    param.vec_count = std::min(vec_avail, vec_count);
-                    vec_avail -= vec_count;
+                    if (vec_count <= vec_avail) {
+                        param.vec_count = vec_count;
+                        vec_avail -= vec_count;
+                    } else {
+                        vec_avail = 0;
+                    }
                 } else if (param.type->size <= 16) {
                     int gpr_count = (param.type->size + 7) / 8;
 
-                    param.gpr_count = std::min(gpr_avail, gpr_count);
-                    gpr_avail -= gpr_count;
+                    if (gpr_count <= gpr_avail) {
+                        param.gpr_count = gpr_count;
+                        gpr_avail -= gpr_count;
+                    } else {
+                        gpr_avail = 0;
+                    }
                 } else if (gpr_avail) {
                     // Big types (more than 16 bytes) are replaced by a pointer
-                    param.gpr_count = 1;
-                    gpr_avail -= 1;
+                    if (gpr_avail) {
+                        param.gpr_count = 1;
+                        gpr_avail -= 1;
+                    }
+                    param.use_memory = true;
                 }
             } break;
         }
@@ -209,11 +222,11 @@ Napi::Value TranslateCall(const Napi::CallbackInfo &info)
     uint8_t *sp_ptr = nullptr;
 
     // Return through registers unless it's too big
-    if (!func->ret.type->size || func->ret.vec_count || func->ret.gpr_count) {
+    if (!func->ret.use_memory) {
         args_ptr = scratch_ptr - AlignLen(8 * func->parameters.len, 16);
         vec_ptr = (uint64_t *)args_ptr - 8;
         gpr_ptr = vec_ptr - 9;
-        sp_ptr = (uint8_t *)(gpr_ptr - 7);
+        sp_ptr = (uint8_t *)gpr_ptr;
 
 #ifdef RG_DEBUG
         memset(sp_ptr, 0, top_ptr - sp_ptr);
@@ -224,7 +237,7 @@ Napi::Value TranslateCall(const Napi::CallbackInfo &info)
         args_ptr = return_ptr - AlignLen(8 * func->parameters.len, 16);
         vec_ptr = (uint64_t *)args_ptr - 8;
         gpr_ptr = vec_ptr - 9;
-        sp_ptr = (uint8_t *)(gpr_ptr - 7);
+        sp_ptr = (uint8_t *)gpr_ptr;
 
 #ifdef RG_DEBUG
         memset(sp_ptr, 0, top_ptr - sp_ptr);
@@ -235,7 +248,7 @@ Napi::Value TranslateCall(const Napi::CallbackInfo &info)
 
     RG_ASSERT(AlignUp(lib->stack.ptr, 16) == lib->stack.ptr);
     RG_ASSERT(AlignUp(lib->stack.end(), 16) == lib->stack.end());
-    RG_ASSERT(AlignUp(sp_ptr, 16) == sp_ptr);
+    RG_ASSERT(AlignUp(sp_ptr - 8, 16) == sp_ptr);
 
     // Push arguments
     for (Size i = 0; i < func->parameters.len; i++) {
@@ -359,7 +372,7 @@ Napi::Value TranslateCall(const Napi::CallbackInfo &info)
                     if (!PushHFA(obj, param.type, (uint8_t *)vec_ptr))
                         return env.Null();
                     vec_ptr += param.vec_count;
-                } else if (param.type->size <= 16) {
+                } else if (!param.use_memory) {
                     if (param.gpr_count) {
                         RG_ASSERT(param.type->align <= 8);
 
@@ -367,7 +380,9 @@ Napi::Value TranslateCall(const Napi::CallbackInfo &info)
                             return env.Null();
                         gpr_ptr += param.gpr_count;
                     } else if (param.type->size) {
-                        args_ptr = AlignUp(args_ptr, param.type->align);
+                        int16_t align = (param.type->align <= 4) ? 4 : 8;
+
+                        args_ptr = AlignUp(args_ptr, align);
                         if (!PushObject(obj, param.type, &lib->tmp_alloc, args_ptr))
                             return env.Null();
                         args_ptr += AlignLen(param.type->size, 8);
@@ -429,7 +444,7 @@ Napi::Value TranslateCall(const Napi::CallbackInfo &info)
                 RG_ASSERT(return_ptr);
 
                 X0X1Ret ret = PERFORM_CALL(GG);
-                RG_ASSERT(ret.x0 = (uint64_t)return_ptr);
+                RG_ASSERT(ret.x0 == (uint64_t)return_ptr);
 
                 Napi::Object obj = PopObject(env, return_ptr, func->ret.type);
                 return obj;

package/src/call_arm64_fwd.S

@@ -11,8 +11,8 @@
 // You should have received a copy of the GNU Affero General Public License
 // along with this program. If not, see https://www.gnu.org/licenses/.
 
-// These five are the same, but they differ (in the C side) by their return type.
-// Unlike the five next functions, these ones don't forward XMM argument registers.
+// These three are the same, but they differ (in the C side) by their return type.
+// Unlike the three next functions, these ones don't forward XMM argument registers.
 .global ForwardCallGG
 .global ForwardCallF
 .global ForwardCallDDDD
@@ -23,20 +23,17 @@
 .global ForwardCallXDDDD
 
 // Copy function pointer to r9, in order to save it through argument forwarding.
-// Save RSP in r19 (non-volatile), and use carefully assembled stack provided by caller.
+// Save RSP in r29 (non-volatile), and use carefully assembled stack provided by caller.
 .macro prologue
     .cfi_startproc
     stp x29, x30, [sp, -16]!
     .cfi_def_cfa sp, 16
     .cfi_offset 29, 16
     .cfi_offset 30, 8
-    str x19, [sp, -16]!
-    .cfi_def_cfa sp, 32
+    mov x29, sp
+    .cfi_def_cfa x29, 16
     mov x9, x0
-    mov x19, sp
-    .cfi_def_cfa x19, 32
-    mov sp, x1
-    add sp, sp, #192
+    add sp, x1, #136
 .endm
 
 // Call native function.
@@ -44,9 +41,7 @@
 // The return value is passed untouched through r0, r1, v0 and/or v1.
 .macro epilogue
     blr x9
-    mov sp, x19
-    .cfi_def_cfa sp, 32
-    ldr x19, [sp], 16
+    mov sp, x29
     .cfi_def_cfa sp, 16
     ldp x29, x30, [sp], 16
     .cfi_def_cfa sp, 0
@@ -58,27 +53,27 @@
 
 // Prepare general purpose argument registers from array passed by caller.
 .macro forward_int
-    ldr x8, [x1, 120]
-    ldr x7, [x1, 112]
-    ldr x6, [x1, 104]
-    ldr x5, [x1, 96]
-    ldr x4, [x1, 88]
-    ldr x3, [x1, 80]
-    ldr x2, [x1, 72]
-    ldr x0, [x1, 56]
-    ldr x1, [x1, 64]
+    ldr x8, [x1, 64]
+    ldr x7, [x1, 56]
+    ldr x6, [x1, 48]
+    ldr x5, [x1, 40]
+    ldr x4, [x1, 32]
+    ldr x3, [x1, 24]
+    ldr x2, [x1, 16]
+    ldr x0, [x1, 0]
+    ldr x1, [x1, 8]
 .endm
 
 // Prepare vector argument registers from array passed by caller.
 .macro forward_vec
-    ldr d7, [x1, 184]
-    ldr d6, [x1, 176]
-    ldr d5, [x1, 168]
-    ldr d4, [x1, 160]
-    ldr d3, [x1, 152]
-    ldr d2, [x1, 144]
-    ldr d1, [x1, 136]
-    ldr d0, [x1, 128]
+    ldr d7, [x1, 128]
+    ldr d6, [x1, 120]
+    ldr d5, [x1, 112]
+    ldr d4, [x1, 104]
+    ldr d3, [x1, 96]
+    ldr d2, [x1, 88]
+    ldr d1, [x1, 80]
+    ldr d0, [x1, 72]
 .endm
 
 ForwardCallGG:

package/src/call_x64_sysv.cc

@@ -228,7 +228,7 @@ Napi::Value TranslateCall(const Napi::CallbackInfo &info)
 
     RG_ASSERT(AlignUp(lib->stack.ptr, 16) == lib->stack.ptr);
     RG_ASSERT(AlignUp(lib->stack.end(), 16) == lib->stack.end());
-    RG_ASSERT(AlignUp(args_ptr, 16) == args_ptr);
+    RG_ASSERT(AlignUp(sp_ptr, 16) == sp_ptr);
 
     // Push arguments
     for (Size i = 0; i < func->parameters.len; i++) {

package/src/call_x86.cc

@@ -90,7 +90,7 @@ Napi::Value TranslateCall(const Napi::CallbackInfo &info)
 
     RG_ASSERT(AlignUp(lib->stack.ptr, 16) == lib->stack.ptr);
     RG_ASSERT(AlignUp(lib->stack.end(), 16) == lib->stack.end());
-    RG_ASSERT(AlignUp(args_ptr, 16) == args_ptr);
+    RG_ASSERT(AlignUp(sp_ptr, 16) == sp_ptr);
 
     // Push arguments
     for (Size i = 0; i < func->parameters.len; i++) {

package/src/ffi.cc

@@ -298,7 +298,7 @@ static Napi::Value LoadSharedLibrary(const Napi::CallbackInfo &info)
         }
 
         // That's enough extra space for pretty much every ABI supported, with 16 bytes of
-        // bonus for each in case we need to put pointers and exta-align them.
+        // bonus for each parameter in case we need to put pointers and exta-align them.
         extra_size = std::max(extra_size, func->scratch_size +
                                           AlignLen(8 * (func->parameters.len + 1), 16));
 

package/src/ffi.hh

@@ -96,7 +96,8 @@ struct ParameterInfo {
     int8_t gpr_count;
     int8_t xmm_count;
     bool gpr_first;
-#elif defined(__aarch64__)
+#elif defined(__arm__) || defined(__aarch64__)
+    bool use_memory;
     int8_t gpr_count;
     int8_t vec_count;
 #elif defined(__i386__) || defined(_M_IX86)
@@ -118,7 +119,7 @@ struct FunctionInfo {
 
     // ABI-specific part
 
-#if defined(__aarch64__) || defined(__x86_64__) || defined(_WIN64)
+#if defined(__arm__) || defined(__aarch64__) || defined(__x86_64__) || defined(_WIN64)
     bool forward_fp;
 #endif
 };

package/src/util.cc

@@ -284,9 +284,10 @@ void DumpStack(const FunctionInfo *func, Span<const uint8_t> sp)
 
     PrintLn(stderr, "Stack (%1 bytes) at 0x%2:", sp.len, sp.ptr);
     for (const uint8_t *ptr = sp.begin(); ptr < sp.end();) {
-        Print(stderr, "  [0x%1 %2 %3]  ", FmtArg(ptr).Pad0(-16), FmtArg((ptr - sp.begin()) / 8).Pad(-4),
-                                                                 FmtArg(ptr - sp.begin()).Pad(-4));
-        for (int i = 0; ptr < sp.end() && i < 8; i++, ptr++) {
+        Print(stderr, "  [0x%1 %2 %3]  ", FmtArg(ptr).Pad0(-16),
+                                          FmtArg((ptr - sp.begin()) / sizeof(void *)).Pad(-4),
+                                          FmtArg(ptr - sp.begin()).Pad(-4));
+        for (int i = 0; ptr < sp.end() && i < sizeof(void *); i++, ptr++) {
             Print(stderr, " %1", FmtHex(*ptr).Pad0(-2));
         }
         PrintLn(stderr);

package/vendor/dragonbox/include/dragonbox/dragonbox.h

@@ -2055,7 +2055,9 @@ namespace jkj::dragonbox {
                 // Using an upper bound on xi, we might be able to optimize the division
                 // better than the compiler; we are computing xi / big_divisor here.
                 ret_value.significand =
-                    div::divide_by_pow10<kappa + 1, significand_bits + kappa + 2, kappa + 1>(xi);
+                    div::divide_by_pow10<kappa + 1, carrier_uint,
+                                         (carrier_uint(1) << (significand_bits + 1)) * big_divisor -
+                                             1>(xi);
                 auto r = std::uint32_t(xi - big_divisor * ret_value.significand);
 
                 if (r != 0) {
@@ -2130,7 +2132,9 @@ namespace jkj::dragonbox {
                 // Using an upper bound on zi, we might be able to optimize the division better than
                 // the compiler; we are computing zi / big_divisor here.
                 ret_value.significand =
-                    div::divide_by_pow10<kappa + 1, significand_bits + kappa + 2, kappa + 1>(zi);
+                    div::divide_by_pow10<kappa + 1, carrier_uint,
+                                         (carrier_uint(1) << (significand_bits + 1)) * big_divisor -
+                                             1>(zi);
                 auto const r = std::uint32_t(zi - big_divisor * ret_value.significand);
 
                 if (r > deltai) {